memra_engine/

lib.rs

//! memra engine: Stage-1 correctness-first forward-pass kernels + ops, on sm_120 via cudarc.

use cudarc::driver::sys::CUfunction_attribute_enum::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES;
use cudarc::driver::{
    CudaContext, CudaFunction, CudaModule, CudaSlice, CudaStream, DeviceSlice, LaunchConfig,
    PushKernelArg,
};
use cudarc::nvrtc::Ptx;
use std::sync::{Arc, Mutex};

const GDN_K2_DYNAMIC_SHARED_BYTES: u32 = 67_072;

/// The default dynamic-shared-memory launch bound the naive SDPA family lives under: past
/// `T_kv * 4 > 48KB` (T_kv > 12288) the smem kernel cannot launch — the measured
/// dspark/full-attn long-ctx crash class. `sdpa_naive` dispatches to the byte-identical
/// gmem-scores twin above this line.
const SDPA_NAIVE_SMEM_MAX: usize = 48 * 1024;

/// Guard on the gmem twin's `n_head * T * T_kv * 4`-byte scores workspace. The shapes that
/// legitimately hit the smem bound are tall-KV blocks (T <= draft block size), which land in
/// the tens of MB; 1 GiB refuses a square T==T_kv misuse before it silently eats the card.
const SDPA_NAIVE_GMEM_WS_MAX: usize = 1 << 30;

#[cfg(debug_assertions)]
pub(crate) fn debug_assert_tensor_stream_device<T>(
    tensor: &CudaSlice<T>,
    stream: &CudaStream,
    site: &str,
) {
    let tensor_dev = tensor.ordinal();
    let stream_dev = stream.context().ordinal();
    assert_eq!(
        tensor_dev, stream_dev,
        "PP cross-device tensor read at {site}: tensor on dev{tensor_dev}, stream on dev{stream_dev}"
    );
}

fn ensure_tensor_stream_device<T>(
    tensor: &impl DeviceSlice<T>,
    stream: &CudaStream,
    site: &str,
) -> Result<(), Box<dyn std::error::Error>> {
    let tensor_dev = tensor.stream().context().ordinal();
    let stream_dev = stream.context().ordinal();
    if tensor_dev != stream_dev {
        return Err(format!(
            "PP cross-device tensor access at {site}: tensor on dev{tensor_dev}, \
             stream on dev{stream_dev}"
        )
        .into());
    }
    Ok(())
}

pub use memra_gguf;
pub use memra_runtime;

pub mod forward;
pub mod hybrid;
pub mod hybrid_forward;
pub mod model;
pub mod sigrouter_contract;
pub mod vision;
pub mod vision_gemma;
pub mod vision_pre;
/// The dual cache lives in the shared `memra-kv` crate (Phase D extraction); this
/// re-export keeps every `crate::cache::` / `memra_engine::cache::` path unchanged.
pub mod cache {
    pub use memra_kv::*;
}
pub mod decode;
pub mod decode_batch;
pub mod dflash;
pub mod eagle;
pub mod gemma_spec;
pub mod graph_update;
/// MLA (multi-head latent attention) CPU f32 reference — GLM-5.2 bring-up lane increment 1.
/// Naive vs absorbed decode forms + NORM/NEOX rope permutation, unit-tested; the permanent
/// oracle for the MLA kernel family (`research/mla-bringup-20260801/DESIGN.md`). No CUDA deps.
pub mod mla;
pub mod moesd;
pub mod parallel;
pub mod plan_backend;
pub mod pp;
pub mod round_stream;
pub mod spec;
pub mod tp;
pub use memra_sampling as sampler;

/// In-house MoE router GEMV on the spec-verify small-t path (DEFAULT ON since 2026-07-10:
/// battery green on 35B p2/p3 K=1..8, acceptance bit-identical, +2-4% spec e2e — replaces
/// ~240 per-column cuBLAS gemv launches/round). MEMRA_ROUTER_KERNEL=0 is the rollback seam.
/// MoE grouped f16 GEMM door (experimental until gated), f16-mirror numeric class:
/// per-layer expert dequant to f16 + one grouped f16 GEMM over the CSR groups.
///   MEMRA_MOE_F16G=1  cublasGemmGroupedBatchedEx (round 46 arc 2). The grouped API issues
///                     through cublas-internal streams NOT ordered with ours — v1 pays a full
///                     stream sync per projection (round-47 ledgered defect).
///   MEMRA_MOE_F16G=2  single-kernel grouped GEMM on the engine stream (round 49): ordered by
///                     construction, zero syncs, f32 C with the act row-scale folded in.
/// DEFAULT (2026-08-01, round 49 promotion): mode 1 on the Hopper lane — with the 41/41
/// dequant coverage fix the q35 board-2048 prime measured 5490 (MMQ) / 8380 (mode 1,
/// +53%) / 7990 (mode 2) x3 interleaved on the H100, argmax MATCH — the last board loss
/// flips. The 5090 measured FLAT (858GB/s makes the dequant-workspace traffic cancel the
/// GEMM win) — but that verdict is for expert banks the int8-MMA MMQ arm can take
/// (IQ3_S/IQ4_XS/Q4_0). MEMRA_MOE_F16G=0 kills anywhere.
///
/// HOPPER RE-VERDICT (2026-08-02, lane/h100-flip-full): mode 2 with full direct coverage
/// (Q4_K/Q6_K/IQ4_XS/IQ3_S tile loaders, lane/iq-direct-loaders) + the deep tail
/// (lane/sk-tail-form) FLIPS past cublas mode 1 on the H100 — q35 board-2048 prime
/// 13163.6 (mode 2, cross=32) vs 8626.5 (mode 1) vs 8073.4 (round-51 sk form), +52.6%,
/// interleaved x5 zero overlap, argmax MATCH 30/30. The round-54 NO-FLIP (8547 vs 8112)
/// was coverage-priced at 5.2% direct; ~100% coverage kills the workspace pass and the
/// verdict inverts. Hopper naked default -> mode 2 (this arm); the gemma (gelu) site
/// stays env-explicit-only via moe_f16g_gemma_on (Err => closed, unaffected by this arm).
///
/// MODE-2 DEFAULT (sm_120a naked, 2026-08-02, lane/f16g-default-rearb): with the direct
/// tile loaders covering Q4_K/Q6_K/IQ4_XS/IQ3_S, the sk visitor beats the int8-MMA MMQ
/// tiles on the IQ-bank models too (q35 board-2048 +33.9%, KAT pp512 +46.7% / pp2048
/// +30.6% — research/iq-direct-loaders-20260802 §3-5, confirmed + full battery in
/// research/f16g-default-rearb-20260802/), so every f16g-admitted expert layer rides
/// mode 2 naked. Decode/verify stay on dp4a (t >= 16 floor). f16-mirror numeric class
/// for naked q35/KAT prefill+prime — new token-sha anchors stamped in the rearb lane.
///
/// AUTO-KQUANT (mode 3, 2026-08-02, lane/q4k-expert-prefill): the previous sm_120a
/// default, kept reachable via MEMRA_MOE_F16G=3. The mode-2 sk form is admitted ONLY for
/// layers the MMA MMQ arm rejects (k-quant expert projections — Q3_K/Q4_K/Q6_K), i.e.
/// exactly where the baseline is the per-pair moe_pairs_matvec_q8_em fallback with zero
/// token reuse (Ornith-35B Q4_K_M board-2048 1098.2 -> 3453.7, 3.14x,
/// research/q4k-expert-prefill-20260802/). Its "IQ banks keep their measured-faster MMQ
/// tiles" ruling was priced BEFORE the IQ direct loaders and is refuted on the 5090 —
/// the k-quant-only admission survives as the rollback seam, not the default.
/// The gemma (gelu) site stays env-explicit-only (moe_f16g_gemma_on).
pub fn moe_f16g_mode() -> u8 {
    static M: std::sync::OnceLock<u8> = std::sync::OnceLock::new();
    *M.get_or_init(|| match std::env::var("MEMRA_MOE_F16G").as_deref() {
        Ok("0") => 0,
        Ok("2") => 2,
        Ok("3") => 3,
        Ok(_) => 1,
        // Both arches independently re-arbitrated to mode 2 on 2026-08-02
        // (5090: lane/f16g-default-rearb; H100: lane/h100-flip-full) — unset = 2 everywhere.
        Err(_) => 2,
    })
}
/// Mode-2 sk kernel form policy (round 51, lane/sk-bm128): the single-kernel grouped GEMM runs
/// as a persistent problem-visitor over the real CSR tiles with two tile forms. Returns
/// (shape_sel, cross) for the FFI:
///   MEMRA_F16G_SK=0    -> (-1, _): the round-49 grid-scan kernel (rollback seam).
///   MEMRA_F16G_SK=32   -> all groups on the 32x64x32 2-stage form (cross = i32::MAX).
///   MEMRA_F16G_SK=128  -> all groups on the 128x64x64 3-stage form (cross = 1; groups fall
///                         back to 32x64 in-launcher when the device/in_f can't take it).
///   unset              -> hybrid split: groups with m_e >= MEMRA_F16G_SK_CROSS ride the 128
///                         form. Default cross = 64 (5090 sweep 2026-08-01, receipts
///                         research/sk-bm128-20260801/; H100 re-swept on the direct+tail
///                         form 2026-08-02, lane/h100-flip-full: {16,32,64} ->
///                         12868/13192/13225 — 64 wins there too, the pre-direct 32
///                         verdict was stale).
pub fn moe_f16g_sk_params() -> (i32, i32) {
    static P: std::sync::OnceLock<(i32, i32)> = std::sync::OnceLock::new();
    *P.get_or_init(|| match std::env::var("MEMRA_F16G_SK").as_deref() {
        Ok("0") => (-1, 0),
        Ok("32") => (0, i32::MAX),
        Ok("128") => (0, 1),
        _ => {
            let cross = std::env::var("MEMRA_F16G_SK_CROSS")
                .ok()
                .and_then(|v| v.parse().ok())
                .unwrap_or(64);
            (0, cross)
        }
    })
}
/// DIRECT-FROM-QUANT sk tile loaders (lane/kquant-tile-loaders, 2026-08-02; IQ classes added
/// by lane/iq-direct-loaders): Q4_K/Q6_K/IQ4_XS/IQ3_S expert projections on the mode-2/3 sk
/// visitor forms dequant their weight tiles in-register from the quant superblocks instead of
/// running the per-(layer,projection) dequant pass into an f16 workspace (41.8% of Ornith-35B
/// t=512 kernel time — the pp512 wall, research/q4k-expert-prefill-20260802 §5; the IQ classes
/// are 94.8% of q35's bank bytes — the h100-sk-direct coverage pricing). Bit-identical to the
/// workspace path by construction (kernel-check "f16g-kq-direct" gates it bitwise) — a
/// data-movement change, not a numeric-class change. Default ON; MEMRA_F16G_DIRECT=0 reverts
/// to the workspace path everywhere; MEMRA_F16G_DIRECT=kq keeps the k-quant loaders and
/// reverts only the IQ classes (the iq-direct-loaders A/B seam — the pre-lane shipped config).
pub fn moe_f16g_direct_on(qtype: i32) -> bool {
    static M: std::sync::OnceLock<u8> = std::sync::OnceLock::new();
    let m = *M.get_or_init(|| match std::env::var("MEMRA_F16G_DIRECT").as_deref() {
        Ok("0") => 0,
        Ok("kq") => 1,
        _ => 2,
    });
    match m {
        0 => false,
        1 => qtype == QT_Q4_K || qtype == QT_Q6_K,
        _ => true,
    }
}
/// DEEP-TAIL sk form (lane/sk-tail-form, 2026-08-02): groups below the visitor crossover ride
/// a 32x64x64 3-STAGE cp.async tile instead of the round-51 32x64x32 2-stage — the same 32-row
/// tile (zero extra padding), 2 k-blocks in flight instead of 1 and half the syncs per k. The
/// H100 ncu pricing (research/sk-bm128-20260801) put the 2-stage tail at 31% of the sk GEMM
/// stage under q35's routing skew. Bit-identical to every other sk form by construction
/// (kernel-check "f16g-sk" gates all tail arms maxdiff==0); exists in both the workspace-f16
/// and direct-from-quant variants. Default ON; MEMRA_F16G_TAIL=0 = rollback to the 2-stage
/// tail. in_f % 64 != 0 falls back in-launcher.
pub fn moe_f16g_tail_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| std::env::var("MEMRA_F16G_TAIL").as_deref() != Ok("0"))
}

/// Per-model door for the gemma-MoE (gelu) grouped path: round 49's Hopper default
/// REGRESSED g26 board-2048 prefill -8.3% interleaved x5 on-box (def median 10380,
/// wild 8.9k-11.7k spread; off 11317, ±0.13%) — the +6-15% probe verdict didn't
/// survive the board workload (stale-verdict law, round 50). The silu/qwen class
/// keeps the round-49 default (q35 +53% board-2048). Explicit MEMRA_MOE_F16G=1/2
/// still opens this door for A/B.
pub fn moe_f16g_gemma_on() -> bool {
    static M: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *M.get_or_init(|| !matches!(std::env::var("MEMRA_MOE_F16G").as_deref(), Ok("0") | Err(_)))
}

/// Fused act-epilogue (silu/gelu-mul + q8_1_mmq quantize in one launch) for the MoE prefill
/// MMA arms. Byte-identical to the two-pass path (kernel-check gated) — default ON.
/// MEMRA_MOE_FUSE_ACTQ=0 is the rollback/A-B seam.
pub fn moe_fuse_actq_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| std::env::var("MEMRA_MOE_FUSE_ACTQ").as_deref() != Ok("0"))
}

/// PREFILL router m-invariance (lane/concat-prime-exact, 2026-08-02). The batched cuBLASLt
/// router GEMM changes a row's logits when OTHER rows join the call (probed: first change at
/// m=65 on the Ornith-35B router, 3.9e-3 — while the MMQ/f16 trunk GEMMs are bit-identical
/// across m). Feeding a top-k discontinuity, that made a served request's expert selection a
/// function of its CO-ARRIVALS under cross-request prime batching. The in-house router GEMV
/// is m-invariant, so prefill uses it too and routing depends on a session's own tokens only.
/// DEFAULT ON: it is the serving isolation contract, and it is the same kernel decode and spec
/// verify already use (dispatch parity, one router kernel for every t).
/// MEMRA_ROUTER_PREFILL_EXACT=0 reverts to the batched GEMM.
pub fn router_prefill_exact_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| std::env::var("MEMRA_ROUTER_PREFILL_EXACT").as_deref() != Ok("0"))
}

pub fn router_kernel_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| {
        let on = std::env::var("MEMRA_ROUTER_KERNEL").as_deref() != Ok("0");
        if !on {
            eprintln!("[memra] router kernel OFF (rollback: per-column cuBLAS gemv)");
        }
        on
    })
}

/// FAST-ROUTER batch twin (lane/fast-router, 2026-08-02). The concat-prime exactness fix
/// (router_prefill_exact_on) routes prefill through router_gemv — m-invariant, but a
/// per-(expert,token) GEMV program with zero operand reuse, so q35 board-2048 prefill paid
/// -10% on the 5090. router_gemv_f32_w8_batch register-tiles (8x8 expert-x-token) the same
/// per-row FP chains (BIT-IDENTICAL per row — kernel-check sweeps m=1..2048 on real router
/// weights), so the t crossover below is pure perf, not a numeric config. Swept on-box
/// (research/fast-router-20260802/crossover-router*.jsonl): plain wins t<=4, batch +7-9%
/// at t=8, 1.9x at t=16 rising to 3.45x at t=2048 — MIN_T=8. Decode t=1 and spec verify
/// t<8 keep the plain w8 form. MEMRA_ROUTER_BATCH=0 forces plain at every t (rollback
/// seam, perf-only: bits are equal by the kernel-check gate).
/// Killed arms (same sweep, JSONL is the record): the 8x16 tile lost to 8x8 at every t
/// (128-accumulator register pressure beats the halved w-traffic), and the same-shape
/// sigmoid_dot_rows twin (out_f=1) measured 0.62-0.89x at every prefill t
/// (launch-latency-bound, ~7us/layer at m=2048) — both bit-identity-PASSED before dying.
pub const ROUTER_BATCH_MIN_T: usize = 8;
pub fn router_batch_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| std::env::var("MEMRA_ROUTER_BATCH").as_deref() != Ok("0"))
}
mod cpu_experts;
#[cfg(memra_cutlass)]
pub mod cutlass_ffi;
pub mod dsv4_ffi;
pub mod dsv4_gpu;
pub mod f16_ffi;
pub mod fp8_ffi;
pub mod mmq_ffi;
pub mod moe_cache;
pub mod prime_graph;
pub mod spill;
mod spill_pread;

// Fatbins are EMBEDDED (crates-release lane, 2026-08-04): build.rs still writes them to
// OUT_DIR, but the bytes ship inside the binary via include_bytes! and load through
// cuModuleLoadData. Distribution contract: a prebuilt or cargo-installed binary must be
// self-contained — the old baked OUT_DIR *paths* pointed at the builder's temp dir and
// broke every machine that wasn't the build machine. Same bytes, same module image;
// the runtime MEMRA_GEMM_FATBIN tune-seam override below is preserved.
const FATBIN: &[u8] = include_bytes!(env!("MEMRA_ENGINE_FATBIN"));
const HYBRID_FATBIN: &[u8] = include_bytes!(env!("MEMRA_HYBRID_FATBIN"));
const QMATVEC_FATBIN: &[u8] = include_bytes!(env!("MEMRA_QMATVEC_FATBIN"));
const FLASH_FATBIN: &[u8] = include_bytes!(env!("MEMRA_FLASH_FATBIN"));
const GEMM_FATBIN: &[u8] = include_bytes!(env!("MEMRA_GEMM_FATBIN"));
const ROUTER_FATBIN: &[u8] = include_bytes!(env!("MEMRA_ROUTER_FATBIN"));
/// spec_sample.cu: sampled-spec primitives (Philox Gumbel-max / softmax gather / residual sampler).
const SAMPLE_FATBIN: &[u8] = include_bytes!(env!("MEMRA_SAMPLE_FATBIN"));

/// TUNE SEAM (tools/sweep): a RUNTIME `MEMRA_GEMM_FATBIN=<path>` overrides the baked-in
/// qmatvec_gemm.cu fatbin path (build.rs bakes the same name at COMPILE time via
/// cargo:rustc-env — that constant is the default). Lets the sweep harness swap in a
/// `-D`-tuned fatbin per process with NO rust rebuild. Unset at runtime => the
/// compile-time default (zero behavior change).
fn gemm_fatbin_bytes() -> std::borrow::Cow<'static, [u8]> {
    assert!(
        !(portable_mma_gated() && std::env::var_os("MEMRA_GEMM_FATBIN").is_some()),
        "MEMRA_GEMM_FATBIN overrides are not allowed in the portable CUDA lane"
    );
    match std::env::var("MEMRA_GEMM_FATBIN") {
        Ok(path) => std::borrow::Cow::Owned(
            std::fs::read(&path).unwrap_or_else(|e| panic!("MEMRA_GEMM_FATBIN read {path}: {e}")),
        ),
        Err(_) => std::borrow::Cow::Borrowed(GEMM_FATBIN),
    }
}

/// Phase A (ARCHITECTURE-H100.md): sm_90a re-enables the portable-PTX tensor-core paths
/// (int8 mma.m16n8k32/k16.s8, bf16 m16n8k16, ldmatrix, cp.async — all sm_80-class, native
/// on Hopper) that the portable boot lane gates off. Dispatch guards that used to test
/// `cfg!(memra_portable_cuda)` test this instead; sm_89 keeps the pure-portable behavior.
/// The sm_120a/sm_100a-only MMA kinds (mxf4nvf4, kind::f8f6f4) are NOT covered — their
/// launchers stay fail-closed stubs on 90a and their dispatch arms stay arch-gated.
pub(crate) const fn portable_mma_gated() -> bool {
    cfg!(memra_portable_cuda) && !cfg!(memra_hopper_mma)
}

/// Refuse an env force that would reach a kernel THIS BUILD DOES NOT CONTAIN.
///
/// Doors of the shape `MEMRA_X=1 => true` are arch-blind: they were written so an operator could
/// force a promoted path on, and the default arm (`cfg!(memra_hopper_mma)` or similar) is the only
/// thing that consulted the arch. On a portable build the forced path then reaches
/// `Engine::func`, which resolves lazily and ends in `panic!("kernel {name} not in any fatbin")` —
/// a confusing crash naming a kernel the operator never heard of, several frames from the switch
/// they actually flipped.
///
/// Found 2026-08-23 by tools/fatbin-lookup-census.py, which listed 20 looked-up kernels absent
/// from the sm_89 fatbins. 18 of those turned out to be correctly unreachable (the GDN varlen
/// chain is gated through `gdn_mma_enabled`, which starts with `!portable_mma_gated()`); these
/// env doors were the two that were genuinely reachable, and only by explicit operator action.
///
/// Same shape and same message style as `gemm_fatbin_bytes`'s refusal above — one idiom for
/// "this switch cannot work on this build", so it fails at the switch instead of at the lookup.
#[track_caller]
pub(crate) fn refuse_portable_force(var: &str, needs: &str) {
    assert!(
        !portable_mma_gated(),
        "{var} forces a kernel path this build does not contain: it needs {needs}, and this is a \
         portable-CUDA build (sm_89). Unset {var} — the default path serves this arch."
    );
}

/// The GDN K4/K5 mma pair's UNSET-env default — ONE definition for the three read sites
/// (gdn_mma_enabled, the k123 pre-work, gdn_scan_chunked's dispatch). They read the env
/// per call ON PURPOSE (kernel-check toggles it to pin both configs), so the shared part
/// is this compile-time constant: ON for Hopper-MMA builds (the original 90a promotion)
/// and for sm_120a builds (lane/moeprime-nvfp4-direct, 2026-08-21 — measured on one RTX
/// PRO 6000 ornith15 pp14715 +6-8% and the local 5090 q38-27b +1-2%, both orders both
/// rigs). A site defaulting differently from its peers arms the mma pre-work while the
/// scan takes the scalar route — measured as a 0.8% LOSS, the drift this helper kills.
pub(crate) const fn gdn_mma_default_on() -> bool {
    cfg!(memra_hopper_mma) || konst_eq(env!("MEMRA_BUILT_CUDA_ARCH"), "120a")
}

/// const str-eq (std `==` on &str is not const-stable on this toolchain floor).
const fn konst_eq(a: &str, b: &str) -> bool {
    let (a, b) = (a.as_bytes(), b.as_bytes());
    if a.len() != b.len() {
        return false;
    }
    let mut i = 0;
    while i < a.len() {
        if a[i] != b[i] {
            return false;
        }
        i += 1;
    }
    true
}

/// The legacy quantized prefill GEMMs are tuned and validated for sm_120a; sm_90a re-admits
/// them through the Hopper-MMA lane (int8 m16n8k32.s8 is sm_80-class PTX).  Keep the policy
/// in a pure helper so the dispatch guard can be regression-tested without constructing an
/// Engine or allocating a GPU tensor.
const fn legacy_quant_gemm_allowed(portable_cuda: bool, hopper_mma: bool, no_gemm: bool) -> bool {
    (!portable_cuda || hopper_mma) && !no_gemm
}

// ---- KV-cache format selection (kvbytes lane, 2026-07-08; default OFF = daily config) ----
// `MEMRA_KV_K` = q8_0 (default, 34 B/32elem) | fp8 (raw e4m3, 32 B — the -6% K-bytes arm)
// `MEMRA_KV_V` = q5_1 (default, 24 B/32elem) | q4_0 (18 B, -25% V bytes) | fp8 (32 B, +33%)
// A non-default format is a NEW NUMERIC CONFIG: its own run-gen argmax baseline is legal,
// but the gate battery (kernel-check, run-spec self-consistency) must pass WITHIN it and
// the choice is explicit env, never silent. flash_attn.cu is compiled once per format pair
// (build.rs); the kernels keep their names — Engine::new just loads the matching fatbin.
const FLASH_FATBIN_VQ4: &[u8] = include_bytes!(env!("MEMRA_FLASH_FATBIN_VQ4"));
const FLASH_FATBIN_VF8: &[u8] = include_bytes!(env!("MEMRA_FLASH_FATBIN_VF8"));
const FLASH_FATBIN_KF8: &[u8] = include_bytes!(env!("MEMRA_FLASH_FATBIN_KF8"));
const FLASH_FATBIN_KF8VQ4: &[u8] = include_bytes!(env!("MEMRA_FLASH_FATBIN_KF8VQ4"));
const FLASH_FATBIN_KF8VF8: &[u8] = include_bytes!(env!("MEMRA_FLASH_FATBIN_KF8VF8"));

/// KV format policy moved to the shared `memra-kv` crate (Phase D); re-exported so the
/// fatbin router below and every existing `crate::kv_blk_bytes()` call site is unchanged.
pub use memra_kv::{kv_blk_bytes, kv_cache_formats};

/// The flash_attn fatbin matching the selected KV formats.
fn flash_fatbin_bytes() -> &'static [u8] {
    match kv_cache_formats() {
        ("q8_0", "q5_1") => FLASH_FATBIN,
        ("q8_0", "q4_0") => FLASH_FATBIN_VQ4,
        ("q8_0", "fp8") => FLASH_FATBIN_VF8,
        ("fp8", "q5_1") => FLASH_FATBIN_KF8,
        ("fp8", "q4_0") => FLASH_FATBIN_KF8VQ4,
        ("fp8", "fp8") => FLASH_FATBIN_KF8VF8,
        other => unreachable!("kv_cache_formats returned {other:?}"),
    }
}

/// TUNE SEAM (tools/sweep): kernel1 (Q8_0/Q4_K/Q5_K) launch-tile override,
/// `MEMRA_GEMM_K1_LAUNCH="BM,BN,NWARP"`. MUST match the `-D K1_BM/K1_BN/NWARP` the swept
/// fatbin was compiled with (the .cu tile and the host launch grid/block have to agree —
/// the hardcoded (128,128,8) in qmatvec_gemm/qmatvec_gemm_raw is the shipped default).
/// Kernel2 (Q6_K/NVFP4) launch is untouched. Unset or malformed => None => shipped
/// defaults (zero behavior change).
fn k1_launch_override() -> Option<(u32, u32, u32)> {
    static K1: std::sync::OnceLock<Option<(u32, u32, u32)>> = std::sync::OnceLock::new();
    *K1.get_or_init(|| {
        let v = std::env::var("MEMRA_GEMM_K1_LAUNCH").ok()?;
        let p: Vec<u32> = v.split(',').filter_map(|s| s.trim().parse().ok()).collect();
        match p.as_slice() {
            [bm, bn, w] => Some((*bm, *bn, *w)),
            _ => None,
        }
    })
}

/// H100 wgmma prefill-GEMM seam (task 8, ARCHITECTURE-H100.md): OPT-IN (MEMRA_WGMMA=1).
/// v0 verdict (2026-07-26, N=5 pp512 9B-Q8_0): wgmma 3845 tok/s vs MMQ 8692 — the
/// standalone harness's "688us MMQ ref" was a pp2048-shape figure, so v0 (unpipelined,
/// 64x64 tile, wait_group<0> every 32-K step) is ~3x SLOWER per launch at m=512 model
/// shapes. Default stays MMQ until the pipelined version beats it N=5 (repo law).
/// Correctness stays pinned regardless: kernel-check's wgmma case is cfg-gated, not env-gated.
pub(crate) fn wgmma_gemm_enabled() -> bool {
    static V: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *V.get_or_init(|| std::env::var("MEMRA_WGMMA").as_deref() == Ok("1"))
}

/// TUNE SEAM: keys per FA-decode split (`MEMRA_FA_SPLIT` forces a fixed size; default 64). Smaller
/// splits raise grid.y so grid = n_head_kv * n_splits fills the 82 SMs at short/mid ctx (vec path
/// launches only n_head_kv=8 CTAs per split). Swept clock-locked 2026-07-03 (graph tg128): 32 beat
/// 64 at ctx 128/512 (+0.5/+1.2%) and lost at 2048 (-3%) — BUT the adaptive 32/64 default BROKE the
/// MTP spec-decode exact-match gate (run-spec K=1/2 self-consistency FAIL with 32; PASS with 64):
/// the split count changes the combine's FP summation order, and the spec verify's batched forward
/// only argmax-matches single-step decode under the 64-split order on real prompts. Spec exactness
/// (the bigger lever) outranks a <=1.2% decode win -> default stays FIXED 64; sweeps use the env.
/// Takes t_kv so eager, _dc capture, and fa_geom_eager stay signature-compatible for future
/// adaptive retries (any retry MUST pass run-spec self-consistency first).
/// Minimum t_kv for the warp-per-token vec FA path (below it the scalar path's 4x-more-blocks
/// hides latency better — measured crossover, see `fa_decode`). Shared by fa_decode / fa_decode_dc /
/// fa_geom_eager / fa_decode_rows-eligibility (spec verify) so the kernel pick NEVER diverges
/// between eager decode and the verify (the spec-exactness law).
pub const FA_VEC_MIN_TKV: usize = 96;
/// Env-overridable crossover (MEMRA_FA_VEC_MIN, default FA_VEC_MIN_TKV). The 96 floor was
/// measured on the qwen geometry (nkv=2); gemma4 SWA layers run nkv=8 = 4x the vec grid,
/// which moves the crossover — sweep per model, adopt per the battery.
pub fn fa_vec_min_tkv() -> usize {
    static V: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
    *V.get_or_init(|| {
        std::env::var("MEMRA_FA_VEC_MIN")
            .ok()
            .and_then(|v| v.parse().ok())
            .unwrap_or_else(|| FA_VEC_MIN_DEFAULT.load(std::sync::atomic::Ordering::Relaxed))
    })
}

/// f16-P/V class (DEFAULT since 2026-07-23 stamp v4; MEMRA_FA_F16PV=0 = f32-class rollback):
/// llama-fa=1-style f16 P + f16 P@V accumulation on the hd512/SWA prefill stamps
/// (KQ/softmax/normalize stay f32). Laptop stamp: 12B 1.045x, 31B 0.979x vs llama.
///
/// SPEC-SERVING FLIP (2026-07-26, the wkv acceptance-law pattern): with MEMRA_DRAFT set the
/// default is OFF. f16 P/V shifts the PRIME's hidden states/KV in the sub-argmax logit
/// space the drafter feeds on — argmax gates stay MATCH while depth acceptance falls off a
/// cliff (26B d1736 0.883 -> 0.405, -40% e2e; f16pv-off alone restores 0.846/314 tok/s —
/// the perf-ci acceptance battery is the only gate that sees this class). Explicit
/// MEMRA_FA_F16PV always wins; plain serving keeps the f16 prefill win.
pub fn fa_f16pv_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| {
        std::env::var("MEMRA_FA_F16PV")
            .map(|v| v != "0")
            .unwrap_or_else(|_| std::env::var("MEMRA_DRAFT").is_err())
    })
}

/// hd512 head-pair arm (DEFAULT since stamp v4; MEMRA_FA512_HP=0 reverts to sp16): GQA
/// ncols2=2 — 2 heads per CTA share each staged K/V tile, Q register-resident. Engages
/// when n_head is even and the GQA group (n_head/n_head_kv) is even.
pub fn fa512_hp_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| std::env::var("MEMRA_FA512_HP").as_deref() != Ok("0"))
}

/// SWA head-pair arm (DEFAULT since stamp v4; MEMRA_FAW_HP=0 reverts to p1): llama-class
/// windowed geometry — 32 q-rows x 2 heads per CTA sharing staged K/V, f16 P@V
/// accumulation. Even n_head and even GQA group required (guarded per call).
pub fn faw_hp_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| std::env::var("MEMRA_FAW_HP").as_deref() != Ok("0"))
}

/// 4-warp sp16 experiment arm (MEMRA_FA512_W4=1, requires the f16pv door): GEMM0 split-K
/// 4-way + GEMM1 4x128 O-dims. Own partial-sum order — oracle-band gated. Returns warp
/// count (2 = base sp16). 8-warp arm measured NEGATIVE 2026-07-23 (jsonl) and removed.
pub fn fa512_wide_warps() -> usize {
    static N: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
    *N.get_or_init(|| match std::env::var("MEMRA_FA512_W4").as_deref() {
        Ok("1") => 4,
        _ => 2,
    })
}

/// hd-512 vec crossover floor (MEMRA_FA512_MIN, default 512) — shared by fa_decode dispatch
/// and the gemma global-layer rows/parity call sites.
pub fn fa512_min_tkv() -> usize {
    static FA512_MIN: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
    *FA512_MIN.get_or_init(|| {
        std::env::var("MEMRA_FA512_MIN")
            .ok()
            .and_then(|v| v.parse().ok())
            .unwrap_or(512)
    })
}
/// Per-model crossover default, set at model load BEFORE the first decode (per-model
/// numeric-config adoption law). qwen keeps the measured 96; gemma4 (nkv=8 SWA) measured
/// vec-always fastest: 119.9 (96) / 130.0 (48) / 133.2 (1) tok/s tg128-regime, 2026-07-10.
pub static FA_VEC_MIN_DEFAULT: std::sync::atomic::AtomicUsize =
    std::sync::atomic::AtomicUsize::new(FA_VEC_MIN_TKV);
/// Per-model windowed-split default (MEMRA_FA_SPW overrides): gemma MoE (26B, nkv=8) measured
/// 32 (grid-limited t=1 under the raw-e4m3 sV ceiling, 2026-07-12); dense gemma (31B)
/// measured 64 (37.13/37.12 vs 36.87/36.86 at 1.7k, N=2 — different attention geometry).
pub static FA_SPW_DEFAULT: std::sync::atomic::AtomicUsize = std::sync::atomic::AtomicUsize::new(32);
/// Per-model hd512 (gemma globals) split default (MEMRA_FA_SP512 overrides): 26B measured 16
/// (2026-07-11 N=2), dense 31B measured 32 (36.86/36.93 vs 36.73/36.73 at 1.7k, 2026-07-12).
/// fused t=1 q4_0 pair/triple row mapping: true = mr1 (one row/warp). Per-model default
/// (dense gemma wins +1.1% short / +0.6% depth on the 31B; MoE 26B REGRESSES −1.2% —
/// its shared-expert fused2 shapes lose to the finer grid). MEMRA_Q40_MR env still wins.
pub static FUSED_MR1_DEFAULT: std::sync::atomic::AtomicBool =
    std::sync::atomic::AtomicBool::new(false);
/// Per-model router-GEMV form (2026-07-31): the 8-warp twin is +8.8% on the H100 q35
/// decode step (router was 14.8% of it) with argmax + spec self-consistency green on
/// qwen-class MoE both rigs. The gemma-4 26B knife-edge block (2026-07-31, single
/// synthetic prompt) was RE-ARBITRATED 2026-08-01 on 6 real prompts — gate outcomes
/// identical to the lone-warp arm, +13% g26 decode — so gemma4 rides the default too
/// (research/g26-decode-20260801/). MEMRA_ROUTER_V2 env overrides either way.
pub static ROUTER_W8_DEFAULT: std::sync::atomic::AtomicBool =
    std::sync::atomic::AtomicBool::new(true);
pub static FA_SP512_DEFAULT: std::sync::atomic::AtomicUsize =
    std::sync::atomic::AtomicUsize::new(16);
/// Per-model rms_norm block size (per-model numeric-config law: the per-thread partial-sum
/// split changes with blockDim -> different FP order -> battery-arbitrated per model).
/// qwen keeps the shipped 256; gemma4 adopts 1024 (single-row 2816-col norms are one-block
/// latency-bound at 256 threads — 7us/launch measured).
pub static RMS_BLOCK_DEFAULT: std::sync::atomic::AtomicU32 = std::sync::atomic::AtomicU32::new(256);
/// gemma4 fa split ladder switch (set at model load; see fa_split_keys).
pub static FA_SP_GEMMA: std::sync::atomic::AtomicBool = std::sync::atomic::AtomicBool::new(false);
/// Per-model stream-k entry override for SPEC serving (-1 = unset → env/default;
/// 0 = force tiling; 1 = admit the deterministic form selector). The former timing
/// selector made identical boots choose different fold orders; `MEMRA_MMQ_SK_FORM` is the
/// explicit numerical-form seam. mmq_ffi reads this before the env.
pub static MMQ_SK_FORCE: std::sync::atomic::AtomicI8 = std::sync::atomic::AtomicI8::new(-1);
/// Per-model FP8-KV door — lives in memra-kv next to the format policy it drives
/// (re-export keeps `crate::KV_FP8_FORCE` setters in model.rs/hybrid.rs working).
pub use memra_kv::KV_FP8_FORCE;
/// bf16 matvec family block size (MEMRA_MMV_BLOCK, default 128, clamped to [64, 256] and a
/// multiple of 32 — the f32acc twin's shared reduce caps at 256). NUMERIC-CLASS knob: the
/// per-thread stride and reduction order change with the block, same acceptance class as
/// MEMRA_RMS_BLOCK (fresh-tape identity + battery at the pinned value).
pub(crate) fn mmv_block() -> u32 {
    static V: std::sync::OnceLock<u32> = std::sync::OnceLock::new();
    *V.get_or_init(|| {
        std::env::var("MEMRA_MMV_BLOCK")
            .ok()
            .and_then(|v| v.parse().ok())
            .filter(|&b: &u32| (64..=256).contains(&b) && b % 32 == 0)
            .unwrap_or(128)
    })
}

/// MEMRA_STEP_TP_W8=1: q8_0 mirror of the step TP attention projections for DECODE.
///
/// NUMERIC-CLASS door, same class and acceptance as `MEMRA_STEP_TP_QKV_FUSED` /
/// `MEMRA_BF16_MMV` / `MEMRA_SEL_GU_WPR`: the per-row arithmetic becomes an int8 dp4a dot
/// with per-32 scales instead of a bf16xf32 fma chain, so a bit-tape cannot apply and the
/// acceptance is the argmax gate plus the boot battery. Motivation is measured, not assumed
/// (`decode-kernel-census`, 2026-08-25): the fused qkv shape runs 23.0 us in bf16 at
/// 1.83 TB/s and 14.0 us in q8_0 at 1.60, and o_proj 24.2 -> 11.7 us — together
/// ~-1.0 ms of a 13.16 ms token. Default OFF.
pub(crate) fn step_tp_w8_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| std::env::var("MEMRA_STEP_TP_W8").as_deref() == Ok("1"))
}

/// MEMRA_TOPK_FAST=1: barrier-lean sigmoid top-k twin (warp-local top-k + one merge).
/// Selection and weight arithmetic identical to the round-robin kernel — a latency twin.
/// MEMRA_SIG_EXPF_DEV=1: device-libm expf sigmoid router (numeric-class door — the
/// host-glibc transcription is FP64-rate-bound on consumer Blackwell). New tape + battery.
pub(crate) fn sig_expf_dev_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| std::env::var("MEMRA_SIG_EXPF_DEV").as_deref() == Ok("1"))
}

pub(crate) fn topk_fast_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| std::env::var("MEMRA_TOPK_FAST").as_deref() == Ok("1"))
}

/// Select the sigmoid-router kernel without ever sending a shape wider than the fast
/// kernels' fixed eight-pick scratch. The generic and dexp kernels support the full host
/// contract; both `_fast` twins index `[warp][8]` storage and would write out of bounds for
/// `n_used > 8` (Hermes `0d220d8c9a3eb634`).
fn sigmoid_topk_kernel(sig_expf: bool, fast: bool, n_used: usize) -> &'static str {
    match (sig_expf, fast && n_used <= 8) {
        (true, true) => "moe_router_sigmoid_topk_f32_dexp_fast",
        (true, false) => "moe_router_sigmoid_topk_f32_dexp",
        (false, true) => "moe_router_sigmoid_topk_f32_fast",
        (false, false) => "moe_router_sigmoid_topk_f32",
    }
}

#[cfg(test)]
mod sigmoid_topk_dispatch_tests {
    #[test]
    fn fast_kernel_refuses_wide_topk_and_composes_with_dexp() {
        use super::sigmoid_topk_kernel;

        assert_eq!(
            sigmoid_topk_kernel(false, true, 8),
            "moe_router_sigmoid_topk_f32_fast"
        );
        assert_eq!(
            sigmoid_topk_kernel(true, true, 8),
            "moe_router_sigmoid_topk_f32_dexp_fast"
        );
        assert_eq!(
            sigmoid_topk_kernel(false, true, 9),
            "moe_router_sigmoid_topk_f32"
        );
        assert_eq!(
            sigmoid_topk_kernel(true, true, 9),
            "moe_router_sigmoid_topk_f32_dexp"
        );
    }
}

pub(crate) fn rms_block() -> u32 {
    static V: std::sync::OnceLock<u32> = std::sync::OnceLock::new();
    *V.get_or_init(|| {
        std::env::var("MEMRA_RMS_BLOCK")
            .ok()
            .and_then(|v| v.parse().ok())
            .unwrap_or_else(|| RMS_BLOCK_DEFAULT.load(std::sync::atomic::Ordering::Relaxed))
    })
}

pub(crate) fn fa_split_keys(t_kv: usize, n_head_kv: usize) -> usize {
    static S: std::sync::OnceLock<Option<usize>> = std::sync::OnceLock::new();
    if let Some(forced) = *S.get_or_init(|| {
        std::env::var("MEMRA_FA_SPLIT")
            .ok()
            .and_then(|v| v.parse().ok())
            .filter(|&s: &usize| s >= 8 && s % 8 == 0)
    }) {
        return forced;
    }
    // CTX-ADAPTIVE default (2026-07-05 40k sweep: sp32 24.5 vs sp128 26.0 tok/s = +5.8% — at
    // deep ctx the n_splits count explodes (40k/32 = 1265 splits x 8 kv-heads) and the combine
    // + partial-buffer cost dominates; at short ctx small splits fill the SMs). Exactness: split
    // size only changes the PARTITION of keys; the rows/combine order per split is fixed and the
    // gate battery (kernel-check + run-spec K=1..8) arbitrates every default change.
    //
    // SM-AWARE SHORT-CTX RUNG (2026-07-06 rtx6000): the 32-key rung was tuned on the 82-SM 5090.
    // On 188 SMs the vec grid (n_head_kv x n_splits CTAs) starves at short ctx — the 35B has
    // n_head_kv=2, so ctx128/split32 = 8 CTAs on 188 SMs. Measured on rtx6000 (N=1 sweep + N=3
    // interleaved confirm): 35B ctx128 sp16 179 vs sp32 161 (+11%), ctx512 178 vs 158, ctx2048
    // flat, ctx>=4096 sp64 edges sp16 by ~3%; 27B ctx128 70.9 vs 66.3 (+7%); 9B 177 vs 163
    // (+9%). Rigs <=100 SMs keep the validated 5090 ladder EXACTLY (default unchanged there —
    // rig-divergence law: this branch is measured on 188 SMs only).
    // gemma4 all-16 ladder probe REVERTED (2026-07-10): +1.3 plain at d1736 (157.5 vs 156.2)
    // but depth VERIFY collapsed (spec 203.5 -> 169 — the windowed rows' per-row combine over
    // 64 splits). The mixed default (swa nkv=8 -> 32, globals nkv=2 -> 8-ladder) stays; a
    // caller-split policy would break row-vs-decode split parity. FA_SP_GEMMA kept as a seam.
    if FA_SP_GEMMA.load(std::sync::atomic::Ordering::Relaxed)
        && std::env::var("MEMRA_FA_SP16").as_deref() == Ok("1")
    {
        return if t_kv <= 8192 {
            16
        } else if t_kv <= 16384 {
            64
        } else {
            128
        };
    }
    let big_rig = fa_sm_count() >= 128;
    if big_rig {
        let _ = n_head_kv;
        if t_kv <= 2048 {
            // MEMRA_FA_SP_SHORT=N: the SHORT rung only (the SWA layers' capped t_kv lands
            // here on step37: 33 of 45 layers at t_kv=512). At 16 the tile loop runs
            // HALF-EMPTY (FA_DEC_TILE=32 -> nt=16 per split), so the V staging pass moves a
            // half tile per iteration and the combine carries 2x the partials; 32 makes each
            // split exactly one full tile. A global MEMRA_FA_SPLIT cannot isolate this — it
            // moves the deep-ctx rung too, where more splits measured worse.
            // NUMERIC-CLASS door (key partition -> different per-split partials/combine):
            // new tape + battery, exactly like every other split-ladder change.
            static SHORT: std::sync::OnceLock<Option<usize>> = std::sync::OnceLock::new();
            if let Some(sp) = *SHORT.get_or_init(|| {
                std::env::var("MEMRA_FA_SP_SHORT")
                    .ok()
                    .and_then(|v| v.parse().ok())
                    .filter(|&s: &usize| s >= 8 && s % 8 == 0)
            }) {
                return sp;
            }
            16
        } else if t_kv <= 16384 {
            64
        } else {
            128
        }
    } else if n_head_kv <= 4 {
        // KV-HEAD-AWARE RUNG (2026-07-08, 5090): the 8192->32 rung was validated on kv=8 models
        // (27B/9B: 8 heads x n_splits fills 82 SMs). The 35B has n_head_kv=2 — at ctx512/sp32
        // the vec grid is 2 x 20 = 40 CTAs on 82 SMs (half idle). Measured (35B, run-gen 128tok
        // N=1 sweep + N=3 confirm): sp8 162.1 / sp16 161.3 / sp32 159.4 at short ctx.
        // DEPTH TAPER (same day, the deep-ctx lesson re-learned on this rung): sp8 at d6257 =
        // 782 splits -> combine + partial-buffer cost dominates (141.2 tok/s); the d6257 sweep
        // says sp64 = 153.0 (sp16/32 147, sp96 147.6, sp128 141). Few-kv-head models need the
        // taper EARLIER than kv=8 (per-split grid 4x thinner, same per-split combine cost).
        // Crossover hunt: sp8 vs sp64 = 156.7/155.9 at d3072, 151.7/155.6 at d4096 -> boundary 3072.
        // RUNG RE-SWEPT UNDER THE DEEP KERNEL (2026-08-02, lane/ladder-3072 — the stale-verdict
        // law: the 3072 boundary was calibrated on the conflicted v4 core; the deep rewrite cut
        // vec cost ~1.2-1.4x while combine scales with n_splits, so sp8's combine bill
        // dominates far earlier). Kernel receipts (quiet-rig nsys, deep vec + combine us):
        // d1024 sp8 17.1 vs sp64 10.6; d2048 31.0 vs 12.2; d3072 44.0 vs 18.3. e2e run-gen
        // tg128 N=3 interleaved (KAT + q35, research/ladder-3072-20260802/): sp8 loses at
        // EVERY depth >= 1024 (KAT d2048 182.6 vs 188.0 = -2.9%, d3072 175.9 vs 186.4 =
        // -5.6%; q35 d4096 169.2 vs 182.6 = -7.4%); d512 flat (+-0.2%, inside noise). sp32
        // ties sp64 within noise in the mid band and loses at d4096 -> no extra rung.
        // Boundary 3072 -> 512: sp8 keeps only the short-ctx band it was validated on
        // (ctx128-512); sp64 takes over where the deep kernel made combine the bill.
        if t_kv <= 512 {
            8
        } else if t_kv <= 16384 {
            64
        } else {
            128
        }
    } else {
        if t_kv <= 8192 {
            32
        } else if t_kv <= 16384 {
            64
        } else {
            128
        }
    }
}

/// SM count of device 0, cached (used by fa_split_keys' rig-size rung; primary-context query,
/// same attribute Engine::batched_variant reads).
pub(crate) fn fa_sm_count() -> i32 {
    static N: std::sync::OnceLock<i32> = std::sync::OnceLock::new();
    *N.get_or_init(|| {
        cudarc::driver::result::init().ok();
        cudarc::driver::result::device::get(0)
            .and_then(|d| unsafe { cudarc::driver::result::device::get_attribute(
                d, cudarc::driver::sys::CUdevice_attribute_enum::CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT) })
            .unwrap_or(82)
    })
}

/// FA-prefill kernel-name suffix for a head_dim (the template-stamped twins in flash_attn.cu):
/// 256 = the original names (qwen35 class, dispatch unchanged), 128 = `_hd128` (MiniMax-M3).
/// Any other dim errors — callers gate to sdpa_naive before dispatching FA.
fn fa_hd_suffix(head_dim: usize) -> Result<&'static str, Box<dyn std::error::Error>> {
    match head_dim {
        256 => Ok(""),
        128 => Ok("_hd128"),
        d => Err(format!(
            "fa_prefill: no kernel stamped for head_dim={d} (only 256/128); \
                          callers must gate to sdpa_naive"
        )
        .into()),
    }
}

/// Quant type codes matching qmatvec.cu QType enum.
pub const QT_Q8_0: i32 = 0;
pub const QT_Q4_K: i32 = 1;
pub const QT_Q6_K: i32 = 2;
pub const QT_Q5_K: i32 = 3;
pub const QT_Q3_K: i32 = 4;
pub const QT_IQ4_XS: i32 = 5;
pub const QT_IQ3_S: i32 = 6;
pub const QT_NVFP4: i32 = 7;
/// Checkpoint-native FP8-E4M3 (MEMRA_ST_E4M3, lane e4m3dec): raw safetensors e4m3 weight bytes
/// [out_f, in_f] row-major (row_bytes == in_f), per-tensor f32 weight_scale in GpuTensor `scale`
/// (fused at the mmvq write / post-matmul scale_inplace). Decode = qmatvec_e4m3_mmvq (+ _b2/_b4/_b8
/// batched twins); prefill (m>=16) = the cuBLASLt FP8 GEMM on the SAME resident bytes (fp8_ffi.rs)
/// — ONE weight copy total, no Q8_0 re-encode duplicate.
pub const QT_F8_E4M3: i32 = 10;
/// Device-side tag for the A6 SPLIT-PLANE repacked NVFP4 layout (Stage-A generic kernel only;
/// GpuTensor keeps qtype=QT_NVFP4 + an `rp` flag — this tag never lives in a GpuTensor).
pub const QT_NVFP4_RP: i32 = 9;
/// Unquantized f32 weight (safetensors MoE Path A: experts dequantized to f32 host-resident).
pub const QT_F32: i32 = 8;
pub const QT_BF16: i32 = 11;
pub const QT_Q4_0: i32 = 12; // gemma-4 QAT GGUF weight format (18B/32: fp16 d + nibbles)
/// GGUF Q2_K. Appended after the existing Q4_0 code so kernel ABI values do not move.
/// Mixed-expert artifacts use the generic f32-dequant staged kernel until a target-rig-gated
/// dp4a/MMQ implementation exists.
pub const QT_Q2_K: i32 = 13;
/// Checkpoint-native FP8-E4M3 with a BLOCK-128 weight-scale GRID (lane/fp8-blk128-decode,
/// 2026-08-05) — the Qwen-official FP8 / DeepSeek-V3 scale class. Same raw e4m3 bytes as
/// `QT_F8_E4M3` ([out_f, in_f] row-major, row_bytes == in_f), but the dequant scale is
/// `GpuTensor::Quant.blk` (`Fp8BlockScales`, [ceil(out_f/128), ceil(in_f/128)] f32) and the
/// scalar `scale` field is 1.0 by the layout contract.
///
/// WHY A DISTINCT CODE rather than `QT_F8_E4M3` + a `blk` flag: every existing QT_F8_E4M3
/// consumer (qmatvec_e4m3_mmvq and its batched/fused twins, e4m3_fused_params,
/// matmul_pre_dual_noscale's F8 arm, try_fp8_gemm) threads exactly ONE scalar weight scale. Under
/// a shared code, any consumer that was not taught the grid would still MATCH and would dequant
/// every tile at scale 1.0 — a silent numeric corruption. Under a distinct code every untaught
/// consumer refuses loudly instead (`mmvq_supports`/`gemm_supports`/`mmq_supports` return false;
/// the mmvq name match panics), so a missed dispatch site is a crash or a refusal receipt, never
/// wrong numbers. Decode = `qmatvec_e4m3_blk_mmvq`; prefill (m>=16) = the per-block FP8 MMQ tile
/// on the SAME resident bytes+grid (fp8_ffi::try_fp8_blk_mmq) — ONE weight copy total.
pub const QT_F8_E4M3_BLK: i32 = 14;

/// Engine device context: CUDA context, stream, loaded kernel modules, cuBLASLt (via runtime::Gpu).
pub struct Engine {
    pub gpu: memra_runtime::Gpu,
    module: Arc<CudaModule>,
    hybrid: Arc<CudaModule>,
    qmatvec: Arc<CudaModule>,
    flash: Arc<CudaModule>,
    /// FP8-GLOBALS module (2026-07-11): the kf8vf8 fatbin loaded ALONGSIDE the default —
    /// gemma GLOBAL layers (hd512) append + attend in e4m3 (dequant-latency arc, HANDOVER).
    /// Lazy: loaded on first global-format use; None until then.
    flash_g: std::sync::OnceLock<Arc<CudaModule>>,
    gemm: Arc<CudaModule>,
    router: Arc<CudaModule>,
    /// Sampled-spec kernels (research/sampled-spec-impl-map.md piece A).
    sample: Arc<CudaModule>,
    /// EDGE-1 §B: one shared SLRU expert-residency cache, lazily built on first MoE dispatch under
    /// MEMRA_MOE_CACHE. `Mutex` makes it multi-agent safe (§E.2); the lock covers only lookup/admit/
    /// memcpy-issue (µs), NOT the GEMM, so streams still overlap. `None` => cache disabled.
    moe_cache: Mutex<Option<crate::moe_cache::MoeSlotCache>>,
    /// MEMRA_STEP_TP_W8, hybrid half: q8_0 mirrors of bf16 GEMV weights that do NOT live in a
    /// TP resident bank (the LM head, the shared expert, the dense-FFN layers), keyed by the
    /// bf16 slab's device pointer and built on first decode use. The mirror is 1.0625 B/w
    /// against bf16's 2, and the raw slab stays resident, so prefill keeps its arithmetic.
    w8_mirrors: Mutex<std::collections::HashMap<u64, CudaSlice<u8>>>,
    /// Per-`in_f` q8_1 activation scratch for those mirrors (allocating per call would cost
    /// more than the door saves).
    w8_act: Mutex<std::collections::HashMap<usize, (CudaSlice<i8>, CudaSlice<f32>)>>,
    /// Exact retained expert-block lengths collected after model load. Mixed-layout models use
    /// this inventory to preallocate fixed-address size classes instead of sizing every slot to
    /// the single largest block. The cache still owns every address for its full lifetime.
    moe_cache_layout: Mutex<Option<Vec<usize>>>,
    /// CAPTURE-RETAIN mode (graph arc, 2026-07-12): while a graph capture (and its allocator
    /// warmups) runs, every Engine allocation is ALSO kept alive here — a captured graph's
    /// transient buffers must never return to the pool, or later allocations (e.g. the spec
    /// verify between replays) reuse their addresses and the replay reads/writes live memory
    /// (the draft-graph corruption root cause). Fast-path cost when off: one relaxed atomic.
    capture_keep_on: std::sync::atomic::AtomicBool,
    /// VERIFY-EXACT scope (dflash lane, 2026-07-13): when set, matmul/matmul_pre skip the
    /// m>=16 prefill-GEMM branches so a t>=16 batched VERIFY rides the decode-exact b-tier
    /// class (the parity law). The t=16 dflash verify tripped the GEMM threshold — 770us/
    /// matmul (54% of the round) AND a different FP order than decode (issue-10 landmine).
    verify_exact: std::sync::atomic::AtomicBool,
    capture_keep: Mutex<Vec<Box<dyn std::any::Any + Send>>>,
    /// EDGE-1 §C.2: dedicated H2D copy stream for async prefetch (event-synced to the compute stream).
    pub copy_stream: Arc<CudaStream>,
    /// Resident CUTLASS NVFP4 prefill scratch (workspace + a_packed + sfa_linear + sfa_sw + y + alpha),
    /// allocated ONCE and grown to the largest prefill GEMM shape, then reused per-call. Removes the
    /// 6 fresh allocations + alpha htod that `cutlass_fp4_gemm` did every prefill matmul (~200/prefill).
    /// Safe as a single shared buffer because all GPU compute serializes on the one `gpu.stream` worker
    /// thread (the server runs one GPU worker; no concurrent CUTLASS GEMMs share this scratch). `None`
    /// until the first CUTLASS FP4 GEMM. Mutex guards lazy build/grow only (matches `moe_cache`).
    #[cfg(memra_cutlass)]
    cutlass_scratch: Mutex<Option<crate::cutlass_ffi::CutlassScratch>>,
    /// FP8-ACT PREFILL scratch (MEMRA_PP_FP8): quantized-activation buffer + scale block + cuBLASLt
    /// workspace, allocated once and grown to the largest prefill m*k (see fp8_ffi.rs). `None`
    /// until the first FP8 prefill GEMM; Mutex guards lazy build/grow only (matches cutlass_scratch).
    fp8_scratch: Mutex<Option<crate::fp8_ffi::Fp8Scratch>>,
    /// f16-P/V door: pooled V re-encode buffer (bf16->f16) for the hd512 _pre path. Lazy-grow;
    /// per-call cudaMalloc was a laptop-regression suspect (VRAM pressure, 31B nkv=4 = 4x bytes).
    fa_vf16_scratch: Mutex<Option<CudaSlice<u8>>>,
    /// Pooled fa-decode split partials (part_o, part_m, part_l): per-call zeros() was 3
    /// alloc+memset pairs per fa launch (~144 mem nodes per decode token — the graph door's
    /// residual launch tax) — lazy-grow, memset-prefix per use, stream-ordered reuse.
    fa_part_pool: Mutex<Option<(CudaSlice<f32>, CudaSlice<f32>, CudaSlice<f32>)>>,
    /// Retired fa-part pool generations (#68): old buffers whose addresses captured graphs may
    /// have baked — kept alive for the Engine's lifetime instead of returning to the async pool
    /// (see the RETIRE-ON-GROW comment at the realloc sites). Doubling growth bounds the total.
    fa_part_retired: Mutex<Vec<(CudaSlice<f32>, CudaSlice<f32>, CudaSlice<f32>)>>,
    /// name -> resolved CudaFunction (capture-safe lookups; see `func`).
    fn_cache: Mutex<std::collections::HashMap<String, CudaFunction>>,
    f16_scratch: Mutex<Option<crate::f16_ffi::F16Scratch>>,
    /// RANK1 LEVER (parallel argmax): resident pass-1 partials scratch (part_v[NB] f32, part_i[NB] i32),
    /// allocated ONCE on first parallel-argmax call and reused. Stable pointers so the 2-pass argmax
    /// is CUDA-graph-capturable (the buffer is referenced by both captured passes; lazy-allocated
    /// before capture under the generate_graph tracking-off window so it carries no events).
    argmax_partials: Mutex<Option<(CudaSlice<f32>, CudaSlice<i32>)>>,
    /// ARC B (chunk-prime dequant-once): resident bf16 K/V workspace for `fa_prefill_view_ws`
    /// ((K bytes, V bytes) u8 buffers holding [t_kv, kv_dim] bf16). Grown lazily to the largest
    /// (t_kv, kv_dim) seen, REUSED across layers/chunks/calls (contents rewritten per launch —
    /// safe because all compute serializes on the one gpu.stream). ~82MB at 40k ctx on the 27B.
    prime_deqw_ws: Mutex<Option<(CudaSlice<u8>, CudaSlice<u8>)>>,
    /// LAUNCH-STRUCTURE STAGE 1: persistent PINNED (cacheable, flags=0) host staging buffer for the
    /// fused-router sel/w readback — one async DtoH pair + ONE sync instead of two synced dtohs.
    /// Grown lazily; reused every MoE layer (single-threaded decode serializes on the sync).
    router_stage: Mutex<Option<PinnedStage>>,
}

/// FAVENDOR lane env gate (2026-07-08): MEMRA_FA_V2=1 dispatches the llama-fattn-vec-mechanism
/// decode kernels (fa_decode_vec_q_v2 / fa_decode_vec_q_rows_v2 / fa_decode_vec_q_v2_dc):
/// tile-batched online softmax (one alpha rescale per 32-key tile instead of per key) + wide-load
/// block dequant in the staging phase. NOTE rev2: llama's register streaming (no smem) was ALSO
/// tried and measured 2x WORSE at depth in our gqa-warps frame — the smem KV-tile broadcast stays
/// (see the kernel comment). NEW NUMERIC CONFIG (tile-level softmax regrouping changes FP order vs
/// the per-key twins) — own argmax baseline; eager decode, the spec-verify rows path AND the
/// graph _dc path switch TOGETHER (the spec-exactness law). Default OFF. Read per call (not
/// OnceLock) so the gate battery can A/B within one process, matching the MEMRA_NO_FA_VEC pattern.
fn fa_v2_on() -> bool {
    // DEFAULT ON since 2026-07-08 (MEMRA_FA_V2=0 reverts): tile-batched online softmax, e2e
    // measured across every model x depth — 35B 168.7->173.4 (d512) / 153.1->158.5 (d6257),
    // 9B 131.2->132.7 / 108.4->124.5 (+15% — the engine-wide depth-slope fix), 27B 47.2->47.7 /
    // 42.2->44.9. One-time numeric-config change; kernel-check + argmax + spec self-consistency
    // + graph bit-identity green on all three models.
    std::env::var("MEMRA_FA_V2")
        .map(|v| v != "0")
        .unwrap_or(true)
}

/// FA v3 gate (default ON since 2026-07-09; MEMRA_FA_V3=0 reverts to v2 — research/fa/fa_v3_design.md):
/// HYBRID decode twins (fa_decode_vec_q_v3 / _rows_v3 / _v3_dc): llama's int8-dp4a K.Q with
/// register-quantized Q (no K dequant, no K smem) + OUR CTA-shared staged bf16 V tile + OUR
/// split partition/combine. NEW NUMERIC CONFIG (int8 Q quantization changes the K.Q accumulation
/// vs the bf16-roundtrip FMA chain) — own argmax baseline; eager decode, the spec-verify rows
/// path AND the graph _dc path switch TOGETHER (the spec-exactness law). Read per call so the
/// gate battery can A/B within one process (the MEMRA_FA_V2 pattern).
pub(crate) fn fa_v3_on() -> bool {
    // DEFAULT ON since 2026-07-09 (MEMRA_FA_V3=0 reverts to v2): dp4a-K hybrid FA decode —
    // fa kernel -21-23% at depth (micro), 35B spec p3 +5% (190->200, the last spec cell),
    // d6257 +1.7%. Own numeric config; full battery green on 35B+9B incl graph bit-identity.
    std::env::var("MEMRA_FA_V3")
        .map(|v| v != "0")
        .unwrap_or(true)
}

/// The v3 dp4a K path reads RAW q8_0 bytes (34B blocks) and stages q5_1 V verbatim — it is only
/// correct on the DEFAULT KV formats — and needs dpl % 4 == 0 consecutive quants per lane
/// (head_dim % 128 == 0; both daily models are hd256). All three dispatch sites share this
/// predicate so the twins can never diverge.
fn fa_v4_mode() -> &'static str {
    static M: std::sync::OnceLock<String> = std::sync::OnceLock::new();
    M.get_or_init(|| std::env::var("MEMRA_FA_V4").unwrap_or_default())
}
fn fa_v4_on() -> bool {
    fa_v4_mode() != "0"
} // DEFAULT ON 2026-07-10 (MEMRA_FA_V4=0 rollback)
/// t_kv-conditional v4 pick (gemma depth lesson 2026-07-10: v4's key-per-lane pipeline starves
/// at the 1024-window with short splits — MEMRA_FA_V4=0 measured depth plain 158.0 vs 156.7).
/// Threshold MEMRA_FA_V4_MAX (default usize::MAX = unchanged behavior; gemma sets 1024 at load
/// via FA_V4_MAX_DEFAULT). Applied at EVERY dispatch site (eager, rows, rows_w, dc) so verify
/// stays kernel-family-identical to decode at the same t_kv.
/// Per-model deep-ctx smem floor default (MEMRA_FA_SMEM_TKV env overrides): gemma pushes it
/// above the 1024 window so the windowed decode + verify rows share the REGISTER family.
pub static FA_SMEM_TKV_DEFAULT: std::sync::atomic::AtomicUsize =
    std::sync::atomic::AtomicUsize::new(1024);
pub static FA_V4_MAX_DEFAULT: std::sync::atomic::AtomicUsize =
    std::sync::atomic::AtomicUsize::new(usize::MAX);
pub fn fa_v4_at_pub(t_kv: usize) -> bool {
    fa_v4_at(t_kv)
}
fn fa_v4_at(t_kv: usize) -> bool {
    static M: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
    let mx = *M.get_or_init(|| {
        std::env::var("MEMRA_FA_V4_MAX")
            .ok()
            .and_then(|v| v.parse().ok())
            .unwrap_or_else(|| FA_V4_MAX_DEFAULT.load(std::sync::atomic::Ordering::Relaxed))
    });
    fa_v4_on() && t_kv < mx
}
/// FA-DEEP gate (2026-08-02, lane fa-decode-deep): deep-ctx v4 twins
/// (fa_decode_vec_q_v4_deep / _deep_dc) — the depth-decode lane's priced fix. Unlike
/// v2/v3/v4 this is NOT a numeric config: the deep twins run the v4 program VERBATIM
/// (same split partition, same softmax/accumulation order, same partials/combine) and only
/// move the smem physical layout (bank de-conflict row pads) + the load schedule (next-tile
/// L2 prefetch) — kernel-check pins bitdiff==0 vs the v4 twins across depths, so eager /
/// rows-verify / graph / seqs stay mutually bit-identical wherever the threshold falls.
/// Engages at t_kv >= MEMRA_FA_DEEP_MIN. The swept floor is 0 = ALWAYS ON where v4 ran
/// (fa-deep-bench fine grid 96..6144, 2026-08-02: deep flat-or-better at EVERY depth,
/// 1.01-1.26x, no losing cell — so there is no engagement boundary and no new
/// capture-recapture edge; the env stays as a sweep/diagnostic seam only).
/// MEMRA_FA_DEEP=0 is the rollback seam. Read per call so the battery + bench can A/B
/// within one process (the v2/v3 pattern).
pub const FA_DEEP_MIN_DEFAULT: usize = 0;
fn fa_deep_at(t_kv: usize) -> bool {
    if std::env::var("MEMRA_FA_DEEP").as_deref() == Ok("0") {
        return false;
    }
    let min = std::env::var("MEMRA_FA_DEEP_MIN")
        .ok()
        .and_then(|v| v.parse().ok())
        .unwrap_or(FA_DEEP_MIN_DEFAULT);
    t_kv >= min
}
/// Public twin (kernel-check builds the deep-vs-v4 bit pin; bench sweeps the floor).
pub fn fa_deep_at_pub(t_kv: usize) -> bool {
    fa_deep_at(t_kv)
}

fn fa_v3_active(head_dim: usize) -> bool {
    // v3's dp4a-K walk reads raw q8_0 K bytes — no e4m3 arm; the fp8-KV arm (MEMRA_KV_FP8)
    // must fall back like any non-default KV format (the rows_dc stream path asserts on it).
    fa_v3_on()
        && head_dim % 128 == 0
        && kv_cache_formats() == ("q8_0", "q5_1")
        && !Engine::kv_fp8_on()
}

/// BATCHED-TICK increment 2 (2026-08-01): true iff a row at this t_kv would take the v4
/// eager arm in `fa_decode_kvmod`'s dispatch — the exact precondition for the z-batched
/// `fa_decode_vec_q_seqs_v4` twin to reproduce its per-seq program bit-identically.
/// Mirrors the kvmod predicates: vec on + above the vec floor + hd256 + inside the v4
/// window + the PRODUCTION v4 body (the noB3/stage phase probes are wrong-output) + the
/// default flash module (no fp8-KV g-module). Callers must ALSO group rows on one
/// `fa_split_keys` rung (the rows-twins' straddle law) before batching.
pub fn fa_seqs_eligible(t_kv: usize, head_dim: usize) -> bool {
    std::env::var("MEMRA_NO_FA_VEC").is_err()
        && t_kv >= fa_vec_min_tkv()
        && head_dim == 256
        && fa_v4_at(t_kv)
        && !matches!(fa_v4_mode(), "noB3" | "stage")
        && !Engine::kv_fp8_on()
}
/// Public twin of the crate-private split ladder (kernel-check builds the seqs-vs-loop pin).
pub fn fa_split_keys_pub(t_kv: usize, n_head_kv: usize) -> usize {
    fa_split_keys(t_kv, n_head_kv)
}

/// A raw pinned (page-locked, CACHEABLE — flags=0, not write-combined) host allocation for
/// DtoH staging. cudarc's `alloc_pinned` uses CU_MEMHOSTALLOC_WRITECOMBINED, which is right for
/// HtoD streams but pathologically slow for host READS — the router readback is host-read-heavy,
/// so we allocate through `result::malloc_host` with flags=0 directly.
struct PinnedStage {
    ptr: *mut u8,
    cap: usize,
}
unsafe impl Send for PinnedStage {}
impl PinnedStage {
    fn new(cap: usize) -> Result<Self, Box<dyn std::error::Error>> {
        let ptr = unsafe { cudarc::driver::result::malloc_host(cap, 0)? } as *mut u8;
        Ok(PinnedStage { ptr, cap })
    }
}
impl Drop for PinnedStage {
    fn drop(&mut self) {
        let _ = unsafe { cudarc::driver::result::free_host(self.ptr as _) };
    }
}

/// Number of pass-1 blocks for the parallel argmax (fan-out across SMs to saturate HBM). 256 blocks
/// x 256 threads = 65536 threads covering the 248K-vocab scan in ~4 strided loads/thread.
pub const ARGMAX_NB: usize = 256;

/// crate-visible alias for the batched FA3 shim entry (hybrid_forward's batch arm).
pub(crate) use memra_fa3_vl as fa3_vl_raw;

unsafe extern "C" {
    /// FA3 v10 shim (cu/fa3_prefill.cu): TMA-swizzled wgmma FA, fresh causal hd256.
    fn memra_fa3_prefill(
        q16: *const core::ffi::c_void,
        k16: *const core::ffi::c_void,
        v16: *const core::ffi::c_void,
        o: *mut f32,
        t: i32,
        h: i32,
        hkv: i32,
        d: i32,
        scale: f32,
        stream: *mut core::ffi::c_void,
    ) -> i32;
    /// batched varlen twin: host arrays of device pointers per seq (B <= 8).
    pub(crate) fn memra_fa3_vl(
        q16s: *const *const core::ffi::c_void,
        k16s: *const *const core::ffi::c_void,
        v16s: *const *const core::ffi::c_void,
        os: *const *mut f32,
        ts: *const i32,
        b: i32,
        h: i32,
        hkv: i32,
        d: i32,
        scale: f32,
        stream: *mut core::ffi::c_void,
    ) -> i32;
}

/// STAGE-2 GROUPED DECODE: 8 expert weight-block device pointers passed BY VALUE as one kernel
/// param (matches the CUDA `wptr8_t` struct: 8x 64-bit pointers, `#[repr(C)]` => identical
/// layout). The pointers are SLRU cache-slot base addresses — fixed for the engine's lifetime
/// (slots are never re-allocated), so passing raw values is stable across the launch.
#[repr(C)]
#[derive(Clone, Copy)]
pub struct WPtr8(pub [u64; 8]);
unsafe impl cudarc::driver::DeviceRepr for WPtr8 {}

/// task #18 varlen GDN: per-seq args for gdn_chunk_{state,output}_mma_vl — one launch
/// runs all B<=8 sequences' K4/K5 (CUDA `gdnseq_t`/`gdnvl_t`, layout-identical repr(C)).
/// Raw addresses are valid for the launch: every referenced buffer outlives the call and
/// all work is on the single compute stream (same discipline as the f16 GEMM FFI).
#[repr(C)]
#[derive(Clone, Copy, Default)]
pub struct GdnSeqVl {
    pub kb16: u64,
    pub gcum: u64,
    pub beta: u64,
    pub u: u64,
    pub wb16: u64,
    pub y: u64,
    pub ssnap: u64,
    pub state_in: u64,
    pub state_out: u64,
    pub q: u64,
    pub p: u64,
    pub o: u64,
    pub k: u64,
    pub v: u64,
    pub g: u64,
    pub a: u64,
    pub w: u64,
    pub t: i32,
    pub nc: i32,
}
unsafe impl cudarc::driver::DeviceRepr for GdnSeqVl {}
#[repr(C)]
#[derive(Clone, Copy)]
pub struct GdnVl8(pub [GdnSeqVl; 8]);
unsafe impl cudarc::driver::DeviceRepr for GdnVl8 {}

/// task #22: per-seq wgmma-fused extras (CUDA `gdnw_t`/`gdnwvl_t`) — qb16 mirror +
/// pre-masked Pb16, riding NEXT TO GdnSeqVl so the base struct stays untouched.
#[repr(C)]
#[derive(Clone, Copy, Default)]
pub struct GdnWVl {
    pub qb16: u64,
    pub pb16: u64,
}
unsafe impl cudarc::driver::DeviceRepr for GdnWVl {}
#[repr(C)]
#[derive(Clone, Copy)]
pub struct GdnWVl8(pub [GdnWVl; 8]);
unsafe impl cudarc::driver::DeviceRepr for GdnWVl8 {}

/// task #18 increment 3: per-seq PREP/TAIL args (CUDA `gdnprep_t`/`gdnprepvl_t`).
#[repr(C)]
#[derive(Clone, Copy, Default)]
pub struct GdnPrepVl {
    pub qkv: u64,
    pub conv_state: u64,
    pub conv_out: u64,
    pub q_g: u64,
    pub k_g: u64,
    pub v_g: u64,
    pub q_l2: u64,
    pub k_l2: u64,
    pub beta_raw: u64,
    pub alpha: u64,
    pub beta: u64,
    pub g_log: u64,
    pub o: u64,
    pub z: u64,
    pub gn: u64,
    pub gn16: u64,
    pub kb16: u64,
    pub qb16: u64,
    pub t: i32,
    pub pad: i32,
}
unsafe impl cudarc::driver::DeviceRepr for GdnPrepVl {}
#[repr(C)]
#[derive(Clone, Copy)]
pub struct GdnPrepVl8(pub [GdnPrepVl; 8]);
unsafe impl cudarc::driver::DeviceRepr for GdnPrepVl8 {}

/// task #18 (attn side): per-seq varlen FA args (CUDA `faseq_t`/`favl_t`).
#[repr(C)]
#[derive(Clone, Copy, Default)]
pub struct FaSeqVl {
    pub q: u64,
    pub k16: u64,
    pub v16: u64,
    pub o: u64,
    pub kf: u64,
    pub vf: u64,
    pub t: i32,
    pub pad: i32,
}
unsafe impl cudarc::driver::DeviceRepr for FaSeqVl {}
#[repr(C)]
#[derive(Clone, Copy)]
pub struct FaVl8(pub [FaSeqVl; 8]);
unsafe impl cudarc::driver::DeviceRepr for FaVl8 {}

/// task #18 (attn pre-FA): per-seq split/norm/rope/append args (CUDA `attnpre_t`).
#[repr(C)]
#[derive(Clone, Copy, Default)]
pub struct AttnPreVl {
    pub qf: u64,
    pub kf: u64,
    pub vf: u64,
    pub q: u64,
    pub gate: u64,
    pub qn: u64,
    pub kn: u64,
    pub kc: u64,
    pub vc: u64,
    pub t: i32,
    pub pad: i32,
}
unsafe impl cudarc::driver::DeviceRepr for AttnPreVl {}
#[repr(C)]
#[derive(Clone, Copy)]
pub struct AttnPreVl8(pub [AttnPreVl; 8]);
unsafe impl cudarc::driver::DeviceRepr for AttnPreVl8 {}

/// task #18 increment 2: one sequence's FULL chunk-buffer set (alloc-only; the
/// varlen K1-K5 chain fills them).
pub struct GdnChunkBufs {
    pub gcum: CudaSlice<f32>,
    pub a: CudaSlice<f32>,
    pub p: CudaSlice<f32>,
    pub u: CudaSlice<f32>,
    pub w: CudaSlice<f32>,
    pub kb16: CudaSlice<u8>,
    pub wb16: CudaSlice<u8>,
    pub y16: CudaSlice<u8>,
    pub ssnap16: CudaSlice<u8>,
    pub qb16: CudaSlice<u8>,
    pub pb16: CudaSlice<u8>,
    pub o: CudaSlice<f32>,
    pub t: usize,
    pub nc: usize,
}

/// STAGE-2 GROUPED DECODE: the 8 routed-expert weights by value (CUDA `f32x8_t`).
#[repr(C)]
#[derive(Clone, Copy)]
pub struct F32x8(pub [f32; 8]);
unsafe impl cudarc::driver::DeviceRepr for F32x8 {}

/// Harness timing contract: wall nanos of the LAST generate/generate_spec prompt prime on this
/// process. Bench binaries read it right after the call to print gen-only throughput without the
/// prime-subtraction hack (which amplifies prime jitter into the gen number at long prompts).
pub static PRIME_NANOS: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(0);

/// RAII guard from `Engine::exact_scope`: restores the pre-scope `verify_exact` value on
/// drop, so error propagation (`?`) can never leave the engine latched in the
/// decode-exact matmul program (hermes finding, fixed 2026-08-23). Holds the flag, not
/// the Engine, so the restoration contract is unit-testable without a GPU.
#[must_use = "dropping immediately ends the exact scope"]
pub struct ExactScope<'a> {
    flag: &'a std::sync::atomic::AtomicBool,
    prev: bool,
}

impl<'a> ExactScope<'a> {
    pub(crate) fn set(flag: &'a std::sync::atomic::AtomicBool, on: bool) -> Self {
        let prev = flag.load(std::sync::atomic::Ordering::Relaxed);
        flag.store(on, std::sync::atomic::Ordering::Relaxed);
        ExactScope { flag, prev }
    }
}

impl Drop for ExactScope<'_> {
    fn drop(&mut self) {
        self.flag
            .store(self.prev, std::sync::atomic::Ordering::Relaxed);
    }
}

#[cfg(test)]
mod exact_scope_tests {
    use std::sync::atomic::{AtomicBool, Ordering};

    #[test]
    fn error_path_restores_verify_exact() {
        // TOOTH (hermes finding, fixed 2026-08-23): dspark_spec_session_burst called
        // set_verify_exact(true)/(false) manually with `?`s in between — any error left
        // the engine latched in the decode-exact matmul program for every later request.
        // The RAII scope must restore across an error propagation.
        let flag = AtomicBool::new(false);
        let failing = |flag: &AtomicBool| -> Result<(), &'static str> {
            let _scope = super::ExactScope::set(flag, true);
            assert!(flag.load(Ordering::Relaxed), "scope arms the flag");
            Err("draft forward failed")? // the `?` exit the manual pair leaked on
        };
        assert!(failing(&flag).is_err());
        assert!(
            !flag.load(Ordering::Relaxed),
            "error propagation must restore the pre-scope value"
        );
        // Nested/previous-value contract: a scope entered while already ON restores ON.
        let flag = AtomicBool::new(true);
        {
            let _scope = super::ExactScope::set(&flag, true);
        }
        assert!(flag.load(Ordering::Relaxed));
        // Early drop ends the scope exactly where the manual `false` used to sit.
        let flag = AtomicBool::new(false);
        let scope = super::ExactScope::set(&flag, true);
        drop(scope);
        assert!(!flag.load(Ordering::Relaxed));
    }
}

impl Engine {
    pub fn new(ordinal: usize) -> Result<Self, Box<dyn std::error::Error>> {
        let gpu = memra_runtime::Gpu::new(ordinal)?;
        // ARCH GUARD (unified dual-arch engine): the fatbins carry single-arch SASS, so a
        // binary/device mismatch otherwise dies at first module load with an opaque CUDA
        // error. Fail early with the rebuild hint instead. MEMRA_ARCH_CHECK=0 skips.
        if std::env::var("MEMRA_ARCH_CHECK").as_deref() != Ok("0") {
            use cudarc::driver::sys::CUdevice_attribute_enum as A;
            let (maj, min) = cudarc::driver::result::device::get(ordinal as i32)
                .and_then(|d| unsafe {
                    Ok((
                        cudarc::driver::result::device::get_attribute(
                            d,
                            A::CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR,
                        )?,
                        cudarc::driver::result::device::get_attribute(
                            d,
                            A::CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR,
                        )?,
                    ))
                })
                .unwrap_or((0, 0));
            let built = env!("MEMRA_BUILT_CUDA_ARCH");
            let ok = matches!(
                (built, maj, min),
                ("120a", 12, 0) | ("120a", 12, 1) | ("100a", 10, 0) | ("90a", 9, 0) | ("89", 8, 9)
            );
            if !ok {
                return Err(format!(
                    "memra was built for sm_{built} but device {ordinal} reports compute \
                     capability {maj}.{min}. Rebuild on this machine (MEMRA_CUDA_ARCH \
                     auto-detects the GPU) or set MEMRA_ARCH_CHECK=0 to bypass."
                )
                .into());
            }
        }
        // Default async-pool RELEASE_THRESHOLD is 0: freed blocks return to the OS at every
        // sync, so cuMemAllocAsync NODES inside captured graphs re-map memory on EVERY
        // cuGraphLaunch (measured 226us/launch on the gemma graph door, 2026-07-23 osrt).
        // Pinning the threshold keeps the pool cached -> alloc nodes become pointer bumps.
        unsafe {
            use cudarc::driver::sys;
            let dev: sys::CUdevice = ordinal as sys::CUdevice;
            let mut pool: sys::CUmemoryPool = std::ptr::null_mut();
            if sys::cuDeviceGetDefaultMemPool(&mut pool, dev) == sys::CUresult::CUDA_SUCCESS {
                let mut thresh: u64 = u64::MAX;
                let _ = sys::cuMemPoolSetAttribute(
                    pool,
                    sys::CUmemPool_attribute_enum::CU_MEMPOOL_ATTR_RELEASE_THRESHOLD,
                    &mut thresh as *mut u64 as *mut core::ffi::c_void,
                );
            }
        }
        let module = gpu.ctx.load_module(Ptx::from_binary(FATBIN.to_vec()))?;
        let hybrid = gpu
            .ctx
            .load_module(Ptx::from_binary(HYBRID_FATBIN.to_vec()))?;
        let qmatvec = gpu
            .ctx
            .load_module(Ptx::from_binary(QMATVEC_FATBIN.to_vec()))?;
        let flash = gpu
            .ctx
            .load_module(Ptx::from_binary(flash_fatbin_bytes().to_vec()))?;
        let gemm = gpu
            .ctx
            .load_module(Ptx::from_binary(gemm_fatbin_bytes().into_owned()))?;
        let router = gpu
            .ctx
            .load_module(Ptx::from_binary(ROUTER_FATBIN.to_vec()))?;
        let sample = gpu
            .ctx
            .load_module(Ptx::from_binary(SAMPLE_FATBIN.to_vec()))?;
        let copy_stream = gpu.ctx.new_stream()?;
        // DECODE EVENT-TRACKING ELISION — DEFAULT ON (2026-07-05; MEMRA_EVT=1 = escape hatch).
        // cudarc is in multi-stream mode (main stream +
        // copy_stream are both created streams), so with tracking on EVERY launch arg records a
        // read/write CudaEvent and inserts cuStreamWaitEvent on prior events. On the 35B MoE decode
        // that is ~19k cuStreamWaitEvent + ~9k cuEventRecord + ~6k event create/destroy per token
        // (~7 ms/tok host time, measured nsys 2026-07-04 rtx6000), and +4.6% measured on 27B decode —
        // protecting NOTHING: every hot-path kernel/memcpy runs on the ONE gpu.stream.
        // CROSS-STREAM HAZARD AUDIT: MoeSlotCache in-memory prefetch uses copy_stream. Every
        // overwrite explicitly records the prior compute point and makes copy_stream wait; every
        // consumer explicitly waits for the copy completion event. The opt-in positioned-read
        // proof stays on gpu.stream and retains an explicit event solely to guard pinned-source
        // reuse. Graph-capture sites use only gpu.stream, so these handoffs never rely on cudarc's
        // implicit event tracking.
        // SAFETY: single-stream ordering is total; the runtime mem-pool is configured with
        // internal-dependency reuse (memra-runtime), so alloc reuse is stream-ordered too.
        if std::env::var("MEMRA_EVT")
            .map(|v| v == "1")
            .unwrap_or(false)
        {
            // escape hatch: keep cudarc's implicit cross-stream event tracking.
        } else {
            unsafe {
                gpu.ctx.disable_event_tracking();
            }
        }
        Ok(Self {
            gpu,
            module,
            hybrid,
            qmatvec,
            flash,
            flash_g: std::sync::OnceLock::new(),
            gemm,
            router,
            sample,
            moe_cache: Mutex::new(None),
            w8_mirrors: Mutex::new(std::collections::HashMap::new()),
            w8_act: Mutex::new(std::collections::HashMap::new()),
            moe_cache_layout: Mutex::new(None),
            copy_stream,
            capture_keep_on: std::sync::atomic::AtomicBool::new(false),
            verify_exact: std::sync::atomic::AtomicBool::new(false),
            capture_keep: Mutex::new(Vec::new()),
            argmax_partials: Mutex::new(None),
            prime_deqw_ws: Mutex::new(None),
            router_stage: Mutex::new(None),
            fp8_scratch: Mutex::new(None),
            fa_vf16_scratch: Mutex::new(None),
            fa_part_pool: Mutex::new(None),
            fa_part_retired: Mutex::new(Vec::new()),
            fn_cache: Mutex::new(Default::default()),
            f16_scratch: Mutex::new(None),
            #[cfg(memra_cutlass)]
            cutlass_scratch: Mutex::new(None),
        })
    }

    pub fn ctx(&self) -> &Arc<CudaContext> {
        &self.gpu.ctx
    }

    /// Bytes the async pool holds MAPPED but NOT LIVE (reserved - used), i.e. freed blocks
    /// parked in the pool because `Engine::new` pins RELEASE_THRESHOLD to u64::MAX above.
    ///
    /// Why this is a public engine surface: `mem_get_info`'s `free` DOES NOT SEE these bytes —
    /// they are mapped to this process, so `free` counts them as gone, yet the very next
    /// `alloc_u8` is satisfied from them without touching `free` at all. Any admission or
    /// budget decision that reads `free` alone therefore under-counts real headroom by exactly
    /// this amount. Effective allocatable headroom is `free + pool_cached_bytes()`.
    ///
    /// MEASURED SIZE (c=64 serve burst, 9B NVFP4 + draft, 24GB card, 2026-08-06): 34-89 MB
    /// during the burst — SMALL. The admission gate adds it because a term that can only ever
    /// under-count headroom does not belong in a gate that queues real work, but the honest
    /// reading of this number is that pool caching is NOT where a long-running server's VRAM
    /// hides on this path: reserved ~= used throughout, so the memory the driver reports as
    /// gone is genuinely LIVE (see `pool_reserved_used` for the diagnostic pair).
    ///
    /// Returns 0 if the pool cannot be queried (never a false-positive headroom claim).
    pub fn pool_cached_bytes(&self) -> usize {
        let (reserved, used) = self.pool_reserved_used();
        reserved.saturating_sub(used)
    }

    /// Bytes the driver's per-device CUDA GRAPH memory pool currently holds RESERVED
    /// (`cuDeviceGetGraphMemAttribute` RESERVED_MEM_CURRENT) — the backing store of every
    /// captured alloc node, which on this engine means the dspark verify-graph pool
    /// (decode/step graphs bake pre-allocated buffers and own no alloc nodes). This memory
    /// is DISTINCT from the async pool above: `mem_get_info`'s `free` already excludes it,
    /// it is never released back (the vgraph pool has no eviction by design), and it GROWS
    /// as new (segment, vt)/(vt, rung, hi) keys capture — the growth is what
    /// `dspark_vg_admission_debt` charges at admission. Returns 0 if the attribute cannot
    /// be queried (never a false headroom claim, matching `pool_cached_bytes`).
    pub fn device_graph_mem_reserved(&self) -> usize {
        use cudarc::driver::sys as cus;
        let Ok(dev) = cudarc::driver::result::device::get(self.gpu.ctx.ordinal() as i32) else {
            return 0;
        };
        let mut bytes: u64 = 0;
        let rc = unsafe {
            cus::cuDeviceGetGraphMemAttribute(
                dev,
                cus::CUgraphMem_attribute::CU_GRAPH_MEM_ATTR_RESERVED_MEM_CURRENT,
                &mut bytes as *mut u64 as *mut std::ffi::c_void,
            )
        };
        if rc == cus::cudaError_enum::CUDA_SUCCESS {
            bytes as usize
        } else {
            0
        }
    }

    /// Raw async-pool occupancy: (RESERVED_MEM_CURRENT, USED_MEM_CURRENT) in bytes. Reserved is
    /// what the pool has mapped from the driver; used is what is live inside it. Exposed for
    /// admission/VRAM diagnostics — the pair distinguishes "memory is parked in the pool and
    /// `free` cannot see it" (reserved >> used) from "memory is genuinely held live by some
    /// owner" (reserved ~= used), which are opposite bugs with opposite fixes.
    /// (0, 0) if the pool cannot be queried.
    /// Release every CACHED (freed-but-retained) block of the default async mempool
    /// back to the driver (deploy-headroom lane, 2026-08-27). The boot-time
    /// RELEASE_THRESHOLD=u64::MAX pin keeps freed blocks cached for graph-launch speed,
    /// which is right for steady serving and wrong at a blue/green overlap: a green
    /// PROCESS cannot use blue's cached pool. cuMemPoolTrimTo(0) frees only unused
    /// blocks — live allocations are untouched; later allocs re-map once. Returns the
    /// bytes released (reserved delta), 0 if the pool cannot be queried.
    pub fn pool_trim_to_zero(&self) -> usize {
        use cudarc::driver::sys;
        let (before, _) = self.pool_reserved_used();
        unsafe {
            let mut pool: sys::CUmemoryPool = std::ptr::null_mut();
            if sys::cuDeviceGetDefaultMemPool(&mut pool, self.gpu.ctx.ordinal() as sys::CUdevice)
                != sys::CUresult::CUDA_SUCCESS
            {
                return 0;
            }
            let _ = sys::cuMemPoolTrimTo(pool, 0);
        }
        let (after, _) = self.pool_reserved_used();
        before.saturating_sub(after)
    }

    pub fn pool_reserved_used(&self) -> (usize, usize) {
        use cudarc::driver::sys;
        unsafe {
            let mut pool: sys::CUmemoryPool = std::ptr::null_mut();
            if sys::cuDeviceGetDefaultMemPool(&mut pool, self.gpu.ctx.ordinal() as sys::CUdevice)
                != sys::CUresult::CUDA_SUCCESS
            {
                return (0, 0);
            }
            let (mut reserved, mut used) = (0u64, 0u64);
            if sys::cuMemPoolGetAttribute(
                pool,
                sys::CUmemPool_attribute_enum::CU_MEMPOOL_ATTR_RESERVED_MEM_CURRENT,
                &mut reserved as *mut u64 as *mut core::ffi::c_void,
            ) != sys::CUresult::CUDA_SUCCESS
            {
                return (0, 0);
            }
            if sys::cuMemPoolGetAttribute(
                pool,
                sys::CUmemPool_attribute_enum::CU_MEMPOOL_ATTR_USED_MEM_CURRENT,
                &mut used as *mut u64 as *mut core::ffi::c_void,
            ) != sys::CUresult::CUDA_SUCCESS
            {
                return (0, 0);
            }
            (reserved as usize, used as usize)
        }
    }

    /// Ambient stream (by value since M1-PP2 increment 2): the thread's pp2 stage stream
    /// when a stage scope is active, else the main compute stream — see `Gpu::stream`.
    pub fn stream(&self) -> Arc<CudaStream> {
        self.gpu.stream()
    }
    /// FP8-GLOBALS switch (MEMRA_GEMMA_GKV, default ON): gemma global (hd512) layers keep
    /// their KV in e4m3 — the dequant-latency arc (HANDOVER). Windowed layers stay q8_0/q5_1.
    pub fn gkv_on() -> bool {
        memra_kv::gkv_on()
    }

    /// FP8-WINDOWED switch (MEMRA_GEMMA_WKV — measured 2026-07-12 in a validity-gated
    /// window: 1.7k 174.1-174.4 vs 168.6-169.4 default (+3%), 4.9k 158.7-160.4; vs llama
    /// same-window 159.5-160.2 / 140.6 = 1.09x / 1.13x): gemma windowed (hd256 SWA)
    /// layers hold e4m3 KV and ride the format-aware v4 lane from the kf8vf8 module.
    /// SERVING-MODE DEFAULT (2026-07-12, the 31B spec unlock): fp8-windowed KV GUTS the
    /// MTP drafter's acceptance — its single swa attention reads the windowed cache and
    /// e4m3 noise flips its argmaxes (31B short accept .758 -> 1.000 with q8/q5, spec 88
    /// -> 122.7 vs llama-mtp 112; depth .59 -> .78; 26B depth .57 -> .89). So the default
    /// keys on serving intent: SPEC serving (MEMRA_DRAFT set) -> OFF, plain -> ON (its
    /// depth-plain +3% stands). Explicit MEMRA_GEMMA_WKV always wins. GKV (globals) stays
    /// ON for both — no acceptance cost measured.
    pub fn wkv_on() -> bool {
        memra_kv::wkv_on()
    }

    /// QWEN FP8-KV switch (MEMRA_KV_FP8 explicit; else the per-model KV_FP8_FORCE door set
    /// at model load; else OFF). Non-gemma full-attn layers hold e4m3 K/V via the kf8vf8
    /// module. Per-model verdict 2026-07-12: 9B +0.7-4% scaling with depth, 27B flat,
    /// 35B −2% (fp8 format-gates its v3 dp4a lane) — so the 9B class defaults ON
    /// (adopted 2026-07-28 with the deferred acceptance battery), others stay OFF.
    pub fn kv_fp8_on() -> bool {
        memra_kv::kv_fp8_on()
    }

    /// fa kernel routed by head_dim: hd512 (gemma globals) resolves from the kf8vf8 module
    /// when the fp8-globals arm is on; everything else from the default flash module.
    fn fa_func(&self, name: &str, head_dim: usize) -> CudaFunction {
        if head_dim == 512 && Self::gkv_on() {
            self.func_g(name)
        } else {
            self.func(name)
        }
    }

    /// Kernel from the FP8-GLOBALS (kf8vf8) flash module — gemma global-layer arm only.
    /// Format-AGNOSTIC kernels (e.g. fa_decode_combine_f32) are not compiled into the
    /// per-format fatbins; fall back to the base modules for those.
    fn func_g(&self, name: &str) -> CudaFunction {
        let m = self.flash_g.get_or_init(|| {
            self.gpu
                .ctx
                .load_module(cudarc::nvrtc::Ptx::from_binary(
                    FLASH_FATBIN_KF8VF8.to_vec(),
                ))
                .expect("load kf8vf8 flash fatbin (fp8-globals arm)")
        });
        let key = format!("g:{name}");
        if let Some(f) = self.fn_cache.lock().unwrap().get(&key) {
            return f.clone();
        }
        let f = match m.load_function(name) {
            Ok(f) => f,
            Err(_) => self.func(name),
        };
        self.fn_cache.lock().unwrap().insert(key, f.clone());
        f
    }

    fn func(&self, name: &str) -> CudaFunction {
        // Resolution cache: cuModuleGetFunction fails inside a CUDA-graph capture region,
        // so capture-time lookups MUST be host-memory hits (warmups populate the cache).
        if let Some(f) = self.fn_cache.lock().unwrap().get(name) {
            return f.clone();
        }
        let f = self
            .module
            .load_function(name)
            .or_else(|_| self.hybrid.load_function(name))
            .or_else(|_| self.qmatvec.load_function(name))
            .or_else(|_| self.flash.load_function(name))
            .or_else(|_| self.gemm.load_function(name))
            .or_else(|_| self.router.load_function(name))
            .or_else(|_| self.sample.load_function(name))
            .unwrap_or_else(|_| panic!("kernel {name} not in any fatbin"));
        self.fn_cache
            .lock()
            .unwrap()
            .insert(name.to_string(), f.clone());
        f
    }

    /// Scatter trimmed draft logits into full-vocab space: dst = -inf everywhere, then
    /// dst[d2t[i]] = src[i]. Two launches (fill, scatter) — no grid-wide sync needed.
    pub fn scatter_trim_logits(
        &self,
        src: &CudaSlice<f32>,
        d2t: &CudaSlice<u32>,
        dst: &mut CudaSlice<f32>,
        d_vocab: usize,
        n_vocab: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f1 = self.func("scatter_trim_logits_f32");
        let f2 = self.func("scatter_trim_logits_pass2_f32");
        let (dv, nv) = (d_vocab as i32, n_vocab as i32);
        let cfg1 = LaunchConfig {
            grid_dim: (256, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b1 = self.gpu.stream();
        let mut b1 = __s_b1.launch_builder(&f1);
        b1.arg(src).arg(d2t).arg(&mut *dst).arg(&dv).arg(&nv);
        unsafe {
            b1.launch(cfg1)?;
        }
        let cfg2 = LaunchConfig {
            grid_dim: (d_vocab.div_ceil(256) as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&f2);
        b2.arg(src).arg(d2t).arg(&mut *dst).arg(&dv);
        unsafe {
            b2.launch(cfg2)?;
        }
        Ok(())
    }

    // ---- FILTERED-SPEC (feat/filtered-spec): top-k/p/min-p transforms applied symmetrically
    // to p and q — rejection sampling stays distribution-exact for the filtered target. ----

    /// Per-row filtered-softmax stats: out[r] = (threshold_e, renorm_mass_e, row_max) for the
    /// filter (top_k, top_p, min_p) at `temp`. Rows index into x with row_stride f32s.
    #[allow(clippy::too_many_arguments)]
    pub fn filter_stats(
        &self,
        x: &CudaSlice<f32>,
        row_stride: usize,
        rows: &CudaSlice<i32>,
        out_th: &mut CudaSlice<f32>,
        out_z: &mut CudaSlice<f32>,
        out_max: &mut CudaSlice<f32>,
        n: usize,
        nrow: usize,
        temp: f32,
        top_k: i32,
        top_p: f32,
        min_p: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        // A top-K-selection form of this kernel (3 vocab passes vs the search's ~51) was
        // implemented and REFUTED on 2026-08-21 (lane/moebatch-q35moe): the 248k-vocab row is
        // L2-resident, so the extra passes are near-free while the per-thread selection list
        // spills to local memory — B=8 tick 12.8/11.2 ms (cap 64/32) vs 10.4 ms for this
        // kernel, and serve c8 agg ~648 vs ~666. The receipts row is the record; the real
        // filtered-sampling win was batching the per-row launches (decode_batch.rs).
        //
        // COOPERATIVE MULTI-BLOCK FORM (lane/samplat, 2026-08-21): the surviving inefficiency
        // was WIDTH, not passes — one block per row left ~94% of the device idle for ~620us
        // per B=8 serve tick (5.9% of the tick, box4 nsys receipt). filter_stats_coop_f32
        // splits each row across 16 blocks with grid-synced bisection totals — same algorithm,
        // slice-partial f32 sums (accepted device-sampling class; sample-check arbitrates).
        // MEMRA_FILTER_COOP=0 is the rollback seam to the single-block form.
        //
        // DETERMINISTIC KEYING (hermes finding, fixed 2026-08-23): the old admission
        // `16*nrow <= sm_count` fell back to the single-block program PER CALL when a tick
        // carried too many rows — and the two programs are NOT bit-identical (measured
        // ~1e-7 rel on the renorm mass: different f32 partial-sum order), so a request's
        // sampling threshold arithmetic depended on how many rows shared its serve tick.
        // Coop is now THE program on every coop-capable device: rows are CHUNKED to the
        // co-residency cap (sm_count/16 rows per cooperative launch) and each row's
        // arithmetic uses only its own 16 slices + its own ws region, so the per-row bits
        // are independent of batch width by construction — the kernel-check
        // FILTER-COOP-CHUNK cell pins exactly that. The single-block program remains only
        // behind the deployment-keyed seams: MEMRA_FILTER_COOP=0, or a device with
        // sm_count < 16 (fixed per device class, never per call).
        static COOP_ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let coop_on =
            *COOP_ON.get_or_init(|| std::env::var("MEMRA_FILTER_COOP").as_deref() != Ok("0"));
        if coop_on && self.sm_count() >= 16 {
            let cap = self.sm_count() as usize / 16;
            let mut done = 0usize;
            while done < nrow {
                let chunk = cap.min(nrow - done);
                self.filter_stats_coop_chunk(
                    x, row_stride, rows, done, out_th, out_z, out_max, n, chunk, temp, top_k,
                    top_p, min_p,
                )?;
                done += chunk;
            }
            return Ok(());
        }
        self.filter_stats_plain_program(
            x, row_stride, rows, out_th, out_z, out_max, n, nrow, temp, top_k, top_p, min_p,
        )
    }

    /// One cooperative `filter_stats` launch over rows `row0..row0+chunk` (pub so the
    /// kernel-check FILTER-COOP-CHUNK cell can pin batch-width independence directly).
    /// The kernel indexes `rows`/outputs by blockIdx.y, so the chunk is expressed as
    /// sub-views at `row0` — per-row arithmetic is untouched by the offset.
    #[allow(clippy::too_many_arguments)]
    pub fn filter_stats_coop_chunk(
        &self,
        x: &CudaSlice<f32>,
        row_stride: usize,
        rows: &CudaSlice<i32>,
        row0: usize,
        out_th: &mut CudaSlice<f32>,
        out_z: &mut CudaSlice<f32>,
        out_max: &mut CudaSlice<f32>,
        n: usize,
        chunk: usize,
        temp: f32,
        top_k: i32,
        top_p: f32,
        min_p: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let (ni, nr, rs) = (n as i32, chunk as i32, row_stride as i64);
        let f = self.func("filter_stats_coop_f32");
        let mut ws = self.alloc_uninit::<f32>(chunk * (2 * 16 + 2))?;
        let cfg = LaunchConfig {
            grid_dim: (16, chunk as u32, 1),
            block_dim: (512, 1, 1),
            shared_mem_bytes: 0,
        };
        let rows_v = rows.slice(row0..row0 + chunk);
        let mut th_v = out_th.slice_mut(row0..row0 + chunk);
        let mut z_v = out_z.slice_mut(row0..row0 + chunk);
        let mut mx_v = out_max.slice_mut(row0..row0 + chunk);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x)
            .arg(&rs)
            .arg(&rows_v)
            .arg(&mut th_v)
            .arg(&mut z_v)
            .arg(&mut mx_v)
            .arg(&mut ws)
            .arg(&ni)
            .arg(&nr)
            .arg(&temp)
            .arg(&top_k)
            .arg(&top_p)
            .arg(&min_p);
        unsafe {
            b.launch_cooperative(cfg)?;
        }
        Ok(())
    }

    /// The single-block-per-row `filter_stats` program (the pre-coop form; the
    /// MEMRA_FILTER_COOP=0 rollback and the occupancy fallback). Gate-callable twin of
    /// `filter_stats_coop_program`.
    #[allow(clippy::too_many_arguments)]
    pub fn filter_stats_plain_program(
        &self,
        x: &CudaSlice<f32>,
        row_stride: usize,
        rows: &CudaSlice<i32>,
        out_th: &mut CudaSlice<f32>,
        out_z: &mut CudaSlice<f32>,
        out_max: &mut CudaSlice<f32>,
        n: usize,
        nrow: usize,
        temp: f32,
        top_k: i32,
        top_p: f32,
        min_p: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let (ni, nr, rs) = (n as i32, nrow as i32, row_stride as i64);
        let f = self.func("filter_stats_f32");
        let cfg = LaunchConfig {
            grid_dim: (nrow as u32, 1, 1),
            block_dim: (1024, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x)
            .arg(&rs)
            .arg(rows)
            .arg(&mut *out_th)
            .arg(&mut *out_z)
            .arg(&mut *out_max)
            .arg(&ni)
            .arg(&nr)
            .arg(&temp)
            .arg(&top_k)
            .arg(&top_p)
            .arg(&min_p);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// out[pair] = filtered-softmax prob of ids[pair] in row rows[pair] (th/z per PAIR).
    #[allow(clippy::too_many_arguments)]
    pub fn softmax_gather_filtered(
        &self,
        x: &CudaSlice<f32>,
        row_stride: usize,
        ids: &CudaSlice<u32>,
        rows: &CudaSlice<i32>,
        th: &CudaSlice<f32>,
        z: &CudaSlice<f32>,
        out: &mut CudaSlice<f32>,
        n: usize,
        npair: usize,
        temp: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("softmax_gather_filtered_f32");
        let (ni, np, rs) = (n as i32, npair as i32, row_stride as i64);
        let cfg = LaunchConfig {
            grid_dim: (npair as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x)
            .arg(&rs)
            .arg(ids)
            .arg(rows)
            .arg(th)
            .arg(z)
            .arg(&mut *out)
            .arg(&ni)
            .arg(&np)
            .arg(&temp);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Filtered residual sample: token ~ norm(max(0, fp - fq)) with fp/fq the filtered softmaxes.
    #[allow(clippy::too_many_arguments)]
    pub fn residual_sample_filtered(
        &self,
        p: &CudaSlice<f32>,
        q: Option<&CudaSlice<f32>>,
        n: usize,
        temp: f32,
        seed: u64,
        stream_pos: u32,
        p_stats: (f32, f32, f32),
        q_stats: (f32, f32, f32),
        out_tok: &mut CudaSlice<u32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("residual_sample_filtered_f32");
        let (ni, slo, shi) = (n as i32, (seed & 0xFFFF_FFFF) as u32, (seed >> 32) as u32);
        let has_q: i32 = q.is_some() as i32;
        let qbuf = q.unwrap_or(p);
        let (pm, pth, pz) = p_stats;
        let (qm, qth, qz) = q_stats;
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (1024, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(p)
            .arg(qbuf)
            .arg(&has_q)
            .arg(&ni)
            .arg(&temp)
            .arg(&slo)
            .arg(&shi)
            .arg(&stream_pos)
            .arg(&pm)
            .arg(&pth)
            .arg(&pz)
            .arg(&qm)
            .arg(&qth)
            .arg(&qz)
            .arg(&mut *out_tok);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Sparse-q filtered residual sample: token ~ norm(max(0, fp - q)) where fp is the
    /// FILTERED softmax of `p` (stats from `filter_stats`) and q is a PROBABILITY vector
    /// supported on `cand_ids` (<=32 ids — the DFlash2 selector's candidate-set proposal;
    /// lane/dspark-sampled-admission-20260820). Same event semantics/Philox tag as
    /// `residual_sample_filtered` — one uniform per (seed, stream_pos).
    #[allow(clippy::too_many_arguments)]
    pub fn residual_sample_sparse_q(
        &self,
        p: &CudaSlice<f32>,
        cand_ids: &CudaSlice<u32>,
        q_probs: &CudaSlice<f32>,
        n_cand: usize,
        n: usize,
        temp: f32,
        seed: u64,
        stream_pos: u32,
        p_stats: (f32, f32, f32),
        out_tok: &mut CudaSlice<u32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        assert!(
            n_cand >= 1 && n_cand <= 32,
            "residual_sample_sparse_q supports 1..=32 candidates, got {n_cand}"
        );
        let f = self.func("residual_sample_sparse_q_f32");
        let (ni, nc) = (n as i32, n_cand as i32);
        let (slo, shi) = ((seed & 0xFFFF_FFFF) as u32, (seed >> 32) as u32);
        let (pm, pth, pz) = p_stats;
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (1024, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(p)
            .arg(cand_ids)
            .arg(q_probs)
            .arg(&nc)
            .arg(&ni)
            .arg(&temp)
            .arg(&slo)
            .arg(&shi)
            .arg(&stream_pos)
            .arg(&pm)
            .arg(&pth)
            .arg(&pz)
            .arg(&mut *out_tok);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Gumbel-max draw from the FILTERED distribution (masked perturb; argmax after).
    #[allow(clippy::too_many_arguments)]
    pub fn gumbel_perturb_filtered(
        &self,
        x: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        n: usize,
        seed: u64,
        stream_pos: u32,
        temp: f32,
        row_max: f32,
        th: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gumbel_perturb_filtered_f32");
        let (ni, slo, shi) = (n as i32, (seed & 0xFFFF_FFFF) as u32, (seed >> 32) as u32);
        let cfg = LaunchConfig {
            grid_dim: (n.div_ceil(256) as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x)
            .arg(&mut *y)
            .arg(&ni)
            .arg(&slo)
            .arg(&shi)
            .arg(&stream_pos)
            .arg(&temp)
            .arg(&row_max)
            .arg(&th);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Keskar penalties applied IN PLACE to a logits buffer: history token ids get
    /// rep-divided/multiplied + freq*count + presence subtracted. Symmetric p/q usage keeps
    /// filtered rejection sampling exact for the penalized target.
    #[allow(clippy::too_many_arguments)]
    pub fn penalize_logits(
        &self,
        x: &mut CudaSlice<f32>,
        hist: &CudaSlice<u32>,
        n_hist: usize,
        rep: f32,
        freq: f32,
        present: f32,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if n_hist == 0 {
            return Ok(());
        }
        let f = self.func("penalize_logits_f32");
        let (nh, ni) = (n_hist as i32, n as i32);
        let cfg = LaunchConfig {
            grid_dim: (n_hist.div_ceil(128) as u32, 1, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&mut *x)
            .arg(hist)
            .arg(&nh)
            .arg(&rep)
            .arg(&freq)
            .arg(&present)
            .arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Rows variant: penalize `nrow` contiguous rows of length n in one launch.
    #[allow(clippy::too_many_arguments)]
    pub fn penalize_logits_rows(
        &self,
        x: &mut CudaSlice<f32>,
        hist: &CudaSlice<u32>,
        n_hist: usize,
        rep: f32,
        freq: f32,
        present: f32,
        n: usize,
        nrow: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if n_hist == 0 || nrow == 0 {
            return Ok(());
        }
        let f = self.func("penalize_logits_rows_f32");
        let (nh, ni, nr) = (n_hist as i32, n as i32, nrow as i32);
        let cfg = LaunchConfig {
            grid_dim: (n_hist.div_ceil(128) as u32, nrow as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&mut *x)
            .arg(hist)
            .arg(&nh)
            .arg(&rep)
            .arg(&freq)
            .arg(&present)
            .arg(&ni)
            .arg(&nr);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Heterogeneous serving-batch penalties over host-maintained sparse window counts.
    /// `offsets[r]..offsets[r+1]` indexes the unique positive-count `(id,count)` entries for logits row
    /// `rows[r]`; each row may carry independent repetition/frequency/presence coefficients.
    /// One thread owns one distinct logit, so the kernel needs neither atomics nor the
    /// history-squared dedup scan used by the speculative raw-history oracle.
    #[allow(clippy::too_many_arguments)]
    pub fn penalize_logits_sparse_rows(
        &self,
        x: &mut CudaSlice<f32>,
        ids: &[u32],
        counts: &[u32],
        offsets: &[i32],
        rows: &[i32],
        reps: &[f32],
        freqs: &[f32],
        presents: &[f32],
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let nrow = rows.len();
        if nrow == 0 {
            return Ok(());
        }
        let _ni = i32::try_from(n).map_err(|_| "sparse penalty logits width must fit CUDA i32")?;
        let _nr = i32::try_from(nrow).map_err(|_| "sparse penalty row count must fit CUDA i32")?;
        let entry_count =
            i32::try_from(ids.len()).map_err(|_| "sparse penalty entry count must fit CUDA i32")?;
        if ids.len() != counts.len()
            || offsets.len() != nrow + 1
            || reps.len() != nrow
            || freqs.len() != nrow
            || presents.len() != nrow
            || offsets.first().copied() != Some(0)
            || offsets.last().copied() != Some(entry_count)
        {
            return Err("sparse penalty row metadata shape mismatch".into());
        }
        if counts.contains(&0) {
            return Err("sparse penalty counts must be positive".into());
        }
        let mut max_len = 0usize;
        for pair in offsets.windows(2) {
            if pair[0] < 0 || pair[1] < pair[0] {
                return Err("sparse penalty offsets must be monotonic".into());
            }
            max_len = max_len.max((pair[1] - pair[0]) as usize);
        }
        if max_len == 0 {
            return Ok(());
        }

        let mut seen = std::collections::HashSet::with_capacity(ids.len());
        for (r, &row) in rows.iter().enumerate() {
            if row < 0 || (row as usize + 1).saturating_mul(n) > x.len() {
                return Err("sparse penalty row index exceeds logits shape".into());
            }
            let begin = offsets[r] as usize;
            let end = offsets[r + 1] as usize;
            for &id in &ids[begin..end] {
                if id as usize >= n {
                    return Err("sparse penalty token id exceeds logits row".into());
                }
                if !seen.insert((row, id)) {
                    return Err("sparse penalty entries must be unique per logits row".into());
                }
            }
        }

        // SAFETY: the checks above establish every invariant of the launch-only helper.
        unsafe {
            self.penalize_logits_sparse_rows_unchecked(
                x, ids, counts, offsets, rows, reps, freqs, presents, n,
            )
        }
    }

    /// Launch-only form for the serving hot path, whose `HashMap`-backed producer already
    /// guarantees unique ids and whose rows are enumerated from the live batch.
    ///
    /// # Safety
    ///
    /// Shapes must match the safe wrapper, offsets must be monotonic and in bounds, every row
    /// must index `x`, and each `(row,id)` pair must occur at most once. Token ids outside the
    /// logits row are safe no-ops because the kernel bounds-checks them before computing `x`.
    #[allow(clippy::too_many_arguments)]
    pub(crate) unsafe fn penalize_logits_sparse_rows_unchecked(
        &self,
        x: &mut CudaSlice<f32>,
        ids: &[u32],
        counts: &[u32],
        offsets: &[i32],
        rows: &[i32],
        reps: &[f32],
        freqs: &[f32],
        presents: &[f32],
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let nrow = rows.len();
        if nrow == 0 {
            return Ok(());
        }
        let max_len = offsets
            .windows(2)
            .map(|pair| (pair[1] - pair[0]) as usize)
            .max()
            .unwrap_or(0);
        if max_len == 0 {
            return Ok(());
        }
        let ids_d = self.htod_u32_v(ids)?;
        let counts_d = self.htod_u32_v(counts)?;
        let offsets_d = self.htod_i32(offsets)?;
        let rows_d = self.htod_i32(rows)?;
        let reps_d = self.htod(reps)?;
        let freqs_d = self.htod(freqs)?;
        let presents_d = self.htod(presents)?;
        let f = self.func("penalize_logits_sparse_rows_f32");
        let ni = i32::try_from(n).map_err(|_| "sparse penalty logits width must fit CUDA i32")?;
        let nr = i32::try_from(nrow).map_err(|_| "sparse penalty row count must fit CUDA i32")?;
        let cfg = LaunchConfig {
            grid_dim: (max_len.div_ceil(128) as u32, nrow as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&mut *x)
            .arg(&ids_d)
            .arg(&counts_d)
            .arg(&offsets_d)
            .arg(&rows_d)
            .arg(&reps_d)
            .arg(&freqs_d)
            .arg(&presents_d)
            .arg(&ni)
            .arg(&nr);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// ROW-INCREMENTAL penalties (dspark penalized-sampled admission): row r of `x`
    /// penalizes over the last `min(win, n_hist0 + r)` entries of `hist[..n_hist0 + r]`,
    /// where `hist` = [session window (n_hist0) ++ per-row drafted tokens (nrow-1)]. This
    /// is the within-round evolving penalty state block drafting needs: verify row r's
    /// target is penalized by every token committed before it INCLUDING same-round
    /// accepts — `penalize_logits_rows` (one shared window) is the frozen-window
    /// approximation this exists to replace on the dspark route.
    #[allow(clippy::too_many_arguments)]
    pub fn penalize_logits_rows_inc(
        &self,
        x: &mut CudaSlice<f32>,
        hist: &CudaSlice<u32>,
        n_hist0: usize,
        rep: f32,
        freq: f32,
        present: f32,
        n: usize,
        nrow: usize,
        win: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if nrow == 0 || win == 0 || (n_hist0 == 0 && nrow == 1) {
            return Ok(());
        }
        debug_assert!(
            hist.len() >= n_hist0 + nrow - 1,
            "rows-inc hist must carry n_hist0 + nrow - 1 ids"
        );
        let f = self.func("penalize_logits_rows_inc_f32");
        let max_len = win.min(n_hist0 + nrow - 1).max(1);
        let (nh, ni, nr, wi) = (n_hist0 as i32, n as i32, nrow as i32, win as i32);
        let cfg = LaunchConfig {
            grid_dim: (max_len.div_ceil(128) as u32, nrow as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&mut *x)
            .arg(hist)
            .arg(&nh)
            .arg(&rep)
            .arg(&freq)
            .arg(&present)
            .arg(&ni)
            .arg(&nr)
            .arg(&wi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// WEIGHT PREFETCH (SOTA item 3, 2026-07-13, DEFAULT ON): during a bandwidth-idle
    /// window (the fa launch reads KV, not weights) prefetch the NEXT matvec's
    /// decode-plane bytes into L2 so it reads L2-warm. Value-free scheduling op — same
    /// class as prefetch_l2 (numerics untouched by construction). Wired only where it
    /// measured positive: the E4B dc attn arm (+0.65%). 26B (flat — MoE ffn dominates),
    /// 31B (−0.2% — decode at the DRAM wall) and the ffn gate/up cascade (−1% — 29MB/layer
    /// floods the fill path) all probed and NOT wired. MEMRA_WPF=0 rollback seam.
    pub fn wpf_level() -> u32 {
        static ON: std::sync::OnceLock<u32> = std::sync::OnceLock::new();
        *ON.get_or_init(|| {
            std::env::var("MEMRA_WPF")
                .ok()
                .and_then(|v| v.parse().ok())
                .unwrap_or(1)
        })
    }

    /// PDL launch arm (SOTA item 2, 2026-07-13, DEFAULT ON): the six MEMRA_PDL_ENTRY glue
    /// kernels launch through cuLaunchKernelEx with PROGRAMMATIC_STREAM_SERIALIZATION — the
    /// grid launches while the predecessor drains (~120ns/kernel back, pdl_probe), the
    /// kernels' entry grid-dep sync restores read order (SASS-audited: ACQBULK precedes
    /// every LDG in all six). Valid windows: E4B +1.0-1.2% (128 AND 384-tok gens);
    /// 26B/31B/qwen flat no-harm. Battery: kernel-check GREEN, run-gen tokens IDENTICAL x3
    /// gemma, spec 64/64 E4B K=1/4/8 + 26B/31B K=4 + qwen PASS. Works eager AND under
    /// capture (capture encodes native programmatic edges — the post-capture edge-REWRITE
    /// arm died: engine graphs hold cuMemAllocAsync alloc nodes, edge edits on those return
    /// CUDA_ERROR_NOT_SUPPORTED). MEMRA_PDL=0 rollback seam.
    /// See the `verify_exact` field. Scoped by the dflash round around its t=16 verify.
    /// Prefer `exact_scope` — the RAII form — anywhere a `?` can exit the scope: a manual
    /// true/false pair leaves the flag LATCHED engine-wide when an error propagates
    /// between the two calls (hermes finding on dspark_spec_session_burst, fixed
    /// 2026-08-23), and every later request then runs the exact-GEMM program.
    pub fn set_verify_exact(&self, on: bool) {
        self.verify_exact
            .store(on, std::sync::atomic::Ordering::Relaxed);
    }
    pub(crate) fn verify_exact_on(&self) -> bool {
        self.verify_exact.load(std::sync::atomic::Ordering::Relaxed)
    }

    /// RAII scope over `verify_exact`: sets the flag to `on` now and restores the
    /// PREVIOUS value on drop — unwind, early `return`, and every `?` exit included.
    /// This is the required form for any scope an error can leave (see
    /// `set_verify_exact`); dropping the guard early (`drop(scope)`) ends the scope
    /// exactly where the manual `set_verify_exact(false)` used to sit.
    pub fn exact_scope(&self, on: bool) -> ExactScope<'_> {
        ExactScope::set(&self.verify_exact, on)
    }

    /// m=1 norm+rope+append fold seam (2026-07-23): MEMRA_QKV_APPEND=0 reverts to the
    /// fused-norm-rope + standalone-append pair (the exact-oracle bisect arm).
    pub fn qkv_append_on() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| {
            std::env::var("MEMRA_QKV_APPEND")
                .map(|v| v != "0")
                .unwrap_or(true)
        })
    }

    /// PDL wave-B1a seam: the four dense-glue kernels (rms_norm_f32, add_rms_norm_f32,
    /// add_scale_rms_norm_q8_1, quantize_q8_1). MEMRA_PDL_WB=0 reverts alone.
    pub fn pdl_wb_on() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| {
            std::env::var("MEMRA_PDL_WB")
                .map(|v| v != "0")
                .unwrap_or(true)
        })
    }

    /// Trunk-kernels norm ILP seam (lane/dspark-trunk-kernels-20260820): the T-row verify
    /// norms (rms_norm_f32 / add_rms_norm_f32 at grid=T, block=256) are serial-latency
    /// chains — 20 strided scalar load->fma rounds measured 11.8-12.2us/inst (nsys-B verify
    /// scope: 130 inst/rd = 1.51 ms/rd). The `_v2` twins unroll the element loop 4-deep
    /// (independent loads in flight; SAME per-thread element order into ONE accumulator,
    /// reduce VERBATIM) — BIT-IDENTICAL per row at every (ncols, blockDim).
    /// MEMRA_NORM_ILP=0 reverts to the v1 kernels alone.
    pub fn norm_ilp_on() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| {
            std::env::var("MEMRA_NORM_ILP")
                .map(|v| v != "0")
                .unwrap_or(true)
        })
    }

    /// Trunk-kernels FFN dual seam (lane/dspark-trunk-kernels-20260820): the qwen35
    /// t-parallel verify FFN pair rides the PROVEN dual gate+up doors
    /// (`matmul_decode_exact_dual_pre` + `silu_mul_scaled_q8_1`, the q27 verify shape —
    /// bit-identical per (tensor,token,row), kernel-check-pinned, MEMRA_SPEC_DUAL_T
    /// receipts) instead of two singles + silu_mul + a standalone quantize. The doors
    /// existed but the qwen35 body never called them (nsys-B verify scope: gate+up singles
    /// = 107 launches/rd at grid 4352). MEMRA_TK_FFN_DUAL=0 reverts to the singles chain.
    pub fn tk_ffn_dual_on() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| {
            std::env::var("MEMRA_TK_FFN_DUAL")
                .map(|v| v != "0")
                .unwrap_or(true)
        })
    }

    /// PDL wave-A seam: the mmvq matvec PDL launches only (the six glue kernels keep
    /// their own MEMRA_PDL master seam). MEMRA_PDL_MMVQ=0 reverts wave-A alone — the
    /// per-model no-harm bisect knob.
    pub fn pdl_mmvq_on() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| {
            std::env::var("MEMRA_PDL_MMVQ")
                .map(|v| v != "0")
                .unwrap_or(true)
        })
    }

    pub fn pdl_on() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| std::env::var("MEMRA_PDL").map(|v| v != "0").unwrap_or(true))
    }

    /// PDL wave-B seam (gap-diagnosis arc, GAP-DIAGNOSIS.md verdict 8): the gemma
    /// NVFP4mix decode chain's hot kernels — nvfp4 fused2/mr2 and the q8_0 `_rp`
    /// singles — join the wave-A launch class. Scheduling-only (the entry macro waits
    /// on the producer before any read), bit-identical by construction.
    /// MEMRA_PDL_NVFP4=0 reverts wave-B alone.
    pub fn pdl_nvfp4q8_on() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| {
            std::env::var("MEMRA_PDL_NVFP4")
                .map(|v| v != "0")
                .unwrap_or(true)
        })
    }

    /// Raw CUfunction for a PDL-attributed launch: the SAME kernels.fatbin loaded once more
    /// through the raw driver API (cudarc hides its CUfunction handles; a duplicate module
    /// of tiny glue kernels is free). Resolved lazily per name, cached process-wide.
    /// Fused t=1 q4_0 mr policy: env MEMRA_Q40_MR wins (1/2); else the per-model
    /// FUSED_MR1_DEFAULT (dense gemma = mr1, MoE = mr2 — see the static's doc).
    fn q40_mr1_on() -> bool {
        static Q40MR: std::sync::OnceLock<Option<u32>> = std::sync::OnceLock::new();
        match *Q40MR.get_or_init(|| {
            std::env::var("MEMRA_Q40_MR")
                .ok()
                .and_then(|v| v.parse().ok())
        }) {
            Some(v) => v == 1,
            None => crate::FUSED_MR1_DEFAULT.load(std::sync::atomic::Ordering::Relaxed),
        }
    }

    /// PDL wave-B2: flash-module PDL functions. `g` selects the kf8vf8 flavor — the
    /// caller MUST pass the SAME flavor its builder launch would resolve (fa_func/func_g
    /// mirror); the flavors differ semantically (KV byte formats), a wrong-module launch
    /// writes wrong bytes silently.
    fn pdl_func_flash(
        &self,
        g: bool,
        name: &'static str,
    ) -> Result<cudarc::driver::sys::CUfunction, Box<dyn std::error::Error>> {
        use cudarc::driver::sys as cu;
        // PER-CONTEXT caches (M1-PP2 cross-device fix, 8x box 2026-08-02): CUmodule and
        // CUfunction handles are CONTEXT-scoped, and a remote-stage Engine
        // (MEMRA_PP_DEVICES=a,b) lives in the other device's primary context. The old
        // process-wide OnceLock cache handed stage 1 the dev-a handles, so every stage-1
        // launch_pdl* died CUDA_ERROR_INVALID_HANDLE. Key module + function caches by
        // this engine's CUcontext; single-context runs behave exactly as before.
        static MODS: std::sync::Mutex<Option<std::collections::HashMap<(usize, bool), usize>>> =
            std::sync::Mutex::new(None);
        static FNS: std::sync::Mutex<
            Option<std::collections::HashMap<(usize, bool, &'static str), usize>>,
        > = std::sync::Mutex::new(None);
        let ctx_key = self.ctx().cu_ctx() as usize;
        if let Some(&f) = FNS
            .lock()
            .unwrap()
            .get_or_insert_with(Default::default)
            .get(&(ctx_key, g, name))
        {
            return Ok(f as cu::CUfunction);
        }
        let module = {
            let mut mods = MODS.lock().unwrap();
            let map = mods.get_or_insert_with(Default::default);
            match map.get(&(ctx_key, g)) {
                Some(&m) => m,
                None => {
                    let m = self.pdl_load_module_in_ctx(if g {
                        FLASH_FATBIN_KF8VF8
                    } else {
                        FLASH_FATBIN
                    })?;
                    map.insert((ctx_key, g), m);
                    m
                }
            }
        };
        let cname = std::ffi::CString::new(name)?;
        let mut f: cu::CUfunction = std::ptr::null_mut();
        let r = unsafe { cu::cuModuleGetFunction(&mut f, module as cu::CUmodule, cname.as_ptr()) };
        if r != cu::CUresult::CUDA_SUCCESS {
            return Err(format!("pdl_func_flash {name} (g={g}): {r:?}").into());
        }
        FNS.lock()
            .unwrap()
            .get_or_insert_with(Default::default)
            .insert((ctx_key, g, name), f as usize);
        Ok(f)
    }

    /// Load a fatbin as a raw CUmodule IN THIS ENGINE'S CONTEXT. `cuModuleLoadData` binds
    /// the module to the thread's CURRENT context — a remote-stage engine must not
    /// inherit the primary's (the INVALID_HANDLE class above). Restores the caller's
    /// current context before returning.
    fn pdl_load_module_in_ctx(&self, bytes: &[u8]) -> Result<usize, Box<dyn std::error::Error>> {
        use cudarc::driver::sys as cu;
        let mut prev: cu::CUcontext = std::ptr::null_mut();
        unsafe {
            cu::cuCtxGetCurrent(&mut prev).result()?;
        }
        self.ctx().bind_to_thread()?;
        let mut m: cu::CUmodule = std::ptr::null_mut();
        let r = unsafe { cu::cuModuleLoadData(&mut m, bytes.as_ptr() as *const std::ffi::c_void) };
        let restore = if prev.is_null() {
            cu::CUresult::CUDA_SUCCESS
        } else {
            unsafe { cu::cuCtxSetCurrent(prev) }
        };
        if r != cu::CUresult::CUDA_SUCCESS {
            return Err(format!("pdl module load: {r:?}").into());
        }
        if restore != cu::CUresult::CUDA_SUCCESS {
            return Err(format!("pdl module load: ctx restore {restore:?}").into());
        }
        Ok(m as usize)
    }

    /// Raw CUfunction for prebuilt-args dispatch experiments (M4 probe): same duplicate
    /// raw-module loading as the PDL path, WITHOUT the PDL launch attribute.
    pub fn raw_kernel_function(
        &self,
        name: &'static str,
    ) -> Result<cudarc::driver::sys::CUfunction, Box<dyn std::error::Error>> {
        self.pdl_func(name)
    }

    fn pdl_func(
        &self,
        name: &'static str,
    ) -> Result<cudarc::driver::sys::CUfunction, Box<dyn std::error::Error>> {
        use cudarc::driver::sys as cu;
        // PER-CONTEXT caches — same M1-PP2 cross-device fix as pdl_func_flash (handles
        // are context-scoped; key everything by this engine's CUcontext).
        static MODULES: std::sync::Mutex<Option<std::collections::HashMap<usize, usize>>> =
            std::sync::Mutex::new(None);
        // PDL wave-A: the mmvq kernels live in the qmatvec fatbin, not kernels.cu — second
        // duplicate module, loaded lazily on the first kernels-module miss.
        static QMODULES: std::sync::Mutex<Option<std::collections::HashMap<usize, usize>>> =
            std::sync::Mutex::new(None);
        static FNS: std::sync::Mutex<
            Option<std::collections::HashMap<(usize, &'static str), usize>>,
        > = std::sync::Mutex::new(None);
        let ctx_key = self.ctx().cu_ctx() as usize;
        if let Some(&f) = FNS
            .lock()
            .unwrap()
            .get_or_insert_with(Default::default)
            .get(&(ctx_key, name))
        {
            return Ok(f as cu::CUfunction);
        }
        let module = {
            let mut mods = MODULES.lock().unwrap();
            let map = mods.get_or_insert_with(Default::default);
            match map.get(&ctx_key) {
                Some(&m) => m,
                None => {
                    let m = self.pdl_load_module_in_ctx(FATBIN)?;
                    map.insert(ctx_key, m);
                    m
                }
            }
        };
        let cname = std::ffi::CString::new(name)?;
        let mut f: cu::CUfunction = std::ptr::null_mut();
        let mut r =
            unsafe { cu::cuModuleGetFunction(&mut f, module as cu::CUmodule, cname.as_ptr()) };
        if r == cu::CUresult::CUDA_ERROR_NOT_FOUND {
            let qmodule = {
                let mut mods = QMODULES.lock().unwrap();
                let map = mods.get_or_insert_with(Default::default);
                match map.get(&ctx_key) {
                    Some(&m) => m,
                    None => {
                        let m = self.pdl_load_module_in_ctx(QMATVEC_FATBIN)?;
                        map.insert(ctx_key, m);
                        m
                    }
                }
            };
            r = unsafe { cu::cuModuleGetFunction(&mut f, qmodule as cu::CUmodule, cname.as_ptr()) };
        }
        if r != cu::CUresult::CUDA_SUCCESS {
            return Err(format!("pdl_func {name}: {r:?}").into());
        }
        FNS.lock()
            .unwrap()
            .get_or_insert_with(Default::default)
            .insert((ctx_key, name), f as usize);
        Ok(f)
    }

    /// cuLaunchKernelEx with CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_STREAM_SERIALIZATION on the
    /// compute stream. ONLY legal for kernels whose entry carries MEMRA_PDL_ENTRY.
    ///
    /// # Safety
    /// `params` must match the kernel's exact parameter list (order, types, count) —
    /// a mismatch corrupts the launch silently.
    /// Flash-module twin of `launch_pdl` — `g` picks the kf8vf8 flavor (must mirror the
    /// builder path's fa_func/func_g choice exactly).
    ///
    /// # Safety
    /// Same contract as `launch_pdl`.
    unsafe fn launch_pdl_flash(
        &self,
        g: bool,
        name: &'static str,
        grid: (u32, u32, u32),
        block: (u32, u32, u32),
        smem: u32,
        params: &mut [*mut std::ffi::c_void],
    ) -> Result<(), Box<dyn std::error::Error>> {
        use cudarc::driver::sys as cu;
        let f = self.pdl_func_flash(g, name)?;
        if smem > 0 {
            // mirror the builder path's opt-in ceiling (idempotent host-side set).
            let r =
                unsafe {
                    cu::cuFuncSetAttribute(f,
                cu::CUfunction_attribute_enum::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                smem as i32)
                };
            if r != cu::CUresult::CUDA_SUCCESS {
                return Err(format!("pdl smem attr {name}: {r:?}").into());
            }
        }
        let mut attr = cu::CUlaunchAttribute {
            id: cu::CUlaunchAttributeID::CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_STREAM_SERIALIZATION,
            pad: [0; 4],
            value: cu::CUlaunchAttributeValue {
                programmaticStreamSerializationAllowed: 1,
            },
        };
        let cfg = cu::CUlaunchConfig {
            gridDimX: grid.0,
            gridDimY: grid.1,
            gridDimZ: grid.2,
            blockDimX: block.0,
            blockDimY: block.1,
            blockDimZ: block.2,
            sharedMemBytes: smem,
            hStream: self.gpu.stream().cu_stream(),
            attrs: &mut attr,
            numAttrs: 1,
        };
        let r = unsafe { cu::cuLaunchKernelEx(&cfg, f, params.as_mut_ptr(), std::ptr::null_mut()) };
        if r != cu::CUresult::CUDA_SUCCESS {
            return Err(format!("launch_pdl_flash {name}: {r:?}").into());
        }
        Ok(())
    }

    unsafe fn launch_pdl(
        &self,
        name: &'static str,
        grid: (u32, u32, u32),
        block: (u32, u32, u32),
        params: &mut [*mut std::ffi::c_void],
    ) -> Result<(), Box<dyn std::error::Error>> {
        use cudarc::driver::sys as cu;
        let f = self.pdl_func(name)?;
        let mut attr = cu::CUlaunchAttribute {
            id: cu::CUlaunchAttributeID::CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_STREAM_SERIALIZATION,
            pad: [0; 4],
            value: cu::CUlaunchAttributeValue {
                programmaticStreamSerializationAllowed: 1,
            },
        };
        let cfg = cu::CUlaunchConfig {
            gridDimX: grid.0,
            gridDimY: grid.1,
            gridDimZ: grid.2,
            blockDimX: block.0,
            blockDimY: block.1,
            blockDimZ: block.2,
            sharedMemBytes: 0,
            hStream: self.gpu.stream().cu_stream(),
            attrs: &mut attr,
            numAttrs: 1,
        };
        let r = unsafe { cu::cuLaunchKernelEx(&cfg, f, params.as_mut_ptr(), std::ptr::null_mut()) };
        if r != cu::CUresult::CUDA_SUCCESS {
            return Err(format!("launch_pdl {name}: {r:?}").into());
        }
        Ok(())
    }

    /// L2-prefetch a quant weight's DECODE plane (the rp4 split-plane mirror when present —
    /// that is what the m<=8 dispatch reads — else the raw block bytes). No-op on float arms.
    pub fn prefetch_weight_l2(
        &self,
        w: &crate::model::GpuTensor,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if let crate::model::GpuTensor::Quant { bytes, rp4, .. } = w {
            let p = rp4.as_ref().unwrap_or(bytes);
            self.prefetch_l2(p, p.len())?;
        }
        Ok(())
    }

    /// DSpark markov chain ops (dflash lane): gather one bf16 row of a [V, rank] table
    /// by the DEVICE token id at tok[idx] into f32.
    pub fn gather_row_bf16(
        &self,
        table: &CudaSlice<u8>,
        tok: &CudaSlice<u32>,
        idx: usize,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gather_row_bf16_f32");
        let cfg = LaunchConfig {
            grid_dim: (ncols.div_ceil(256) as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, ix) = (ncols as i32, idx as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(table).arg(tok).arg(&ix).arg(dst).arg(&nc);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// DFlash2 grouped dynamic causal conv (dflash lane, DFLASH2-EVAL-20260820.md):
    /// out[p,c] = sum_{o<ksize, o<=p} (base[half][o][c] + dyn[p][half][o][group(c)])
    /// * x[p-o][c]. `dyn_` is the kernel_projection GEMM output [rows, 2*ksize*groups];
    /// `base` is base_kernel [2, ksize, hidden] flattened; `half` picks prepare(0) /
    /// finish(1).
    #[allow(clippy::too_many_arguments)]
    pub fn dflash2_dynconv(
        &self,
        x: &CudaSlice<f32>,
        dyn_: &CudaSlice<f32>,
        base: &CudaSlice<f32>,
        out: &mut CudaSlice<f32>,
        rows: usize,
        hidden: usize,
        group_size: usize,
        ksize: usize,
        half: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        assert_eq!(hidden % group_size, 0, "hidden % group_size != 0");
        let f = self.func("dflash2_dynconv_f32");
        let n = rows * hidden;
        let cfg = LaunchConfig {
            grid_dim: (n.div_ceil(256) as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (ri, hi, gi, ki, hf) = (
            rows as i32,
            hidden as i32,
            group_size as i32,
            ksize as i32,
            half as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x)
            .arg(dyn_)
            .arg(base)
            .arg(out)
            .arg(&ri)
            .arg(&hi)
            .arg(&gi)
            .arg(&ki)
            .arg(&hf);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Per-row top-k (k <= 32) over a [n_rows, n_cols] logits matrix (DFlash2
    /// candidate selector). Returns (values [n_rows, k], column indices [n_rows, k]),
    /// value-descending, ties to the lower index.
    pub fn topk_rows(
        &self,
        logits: &CudaSlice<f32>,
        n_rows: usize,
        n_cols: usize,
        k: usize,
    ) -> Result<(CudaSlice<f32>, CudaSlice<u32>), Box<dyn std::error::Error>> {
        assert!(k <= 32 && k >= 1, "topk_rows supports 1..=32, got {k}");
        assert!(k <= n_cols, "topk_rows: k {k} > n_cols {n_cols}");
        let f = self.func("topk_rows_f32");
        let nth = 256usize;
        let mut vals = self.uninit(n_rows * k)?;
        let mut idxs = self.gpu.stream().alloc_zeros::<u32>(n_rows * k)?;
        let cfg = LaunchConfig {
            grid_dim: (n_rows as u32, 1, 1),
            block_dim: (nth as u32, 1, 1),
            shared_mem_bytes: (nth * k * 8) as u32,
        };
        let (nr, nc, ki) = (n_rows as i32, n_cols as i32, k as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(logits)
            .arg(&nr)
            .arg(&nc)
            .arg(&ki)
            .arg(&mut vals)
            .arg(&mut idxs);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((vals, idxs))
    }

    /// logits[row_off .. row_off+n] += bias[0..n] (in place, one row).
    pub fn add_row_inplace(
        &self,
        logits: &mut CudaSlice<f32>,
        bias: &CudaSlice<f32>,
        n: usize,
        row_off: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("add_row_inplace_f32");
        let cfg = LaunchConfig {
            grid_dim: (n.div_ceil(256) as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (ni, off) = (n as i32, row_off as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(logits).arg(bias).arg(&ni).arg(&off);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// L2 prefetch of a device byte range (latency-hiding arc; value-free scheduling op).
    pub fn prefetch_l2(
        &self,
        p: &CudaSlice<u8>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("prefetch_l2_bytes");
        let lines = n.div_ceil(128);
        let ni = n as i64;
        let cfg = LaunchConfig {
            grid_dim: (lines.div_ceil(256) as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(p).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// MoE router GEMV (MEMRA_ROUTER_KERNEL): deterministic warp-per-(expert,token) f32 dot.
    /// Different FP order than the cuBLAS path it replaces — battery-gated numeric config.
    pub fn router_gemv(
        &self,
        w: &CudaSlice<f32>,
        x: &CudaSlice<f32>,
        n_embd: usize,
        n_experts: usize,
        t: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        // float4 v2 probed 2026-07-14: +0.25% but flips near-tie routing (new FP order,
        // stream differs) — too small to justify a numeric config change; deleted.
        // w8 twin (2026-07-31): on the 132-SM H100 the lone-warp form is 14.8% of the q35
        // decode step (latency-bound) — the calculus flipped. MEMRA_ROUTER_V2=0 reverts to
        // the warp form (rollback seam; new FP order, battery-arbitrated per model).
        let w8 = match std::env::var("MEMRA_ROUTER_V2").as_deref() {
            Ok("0") => false,
            Ok(_) => true,
            Err(_) => ROUTER_W8_DEFAULT.load(std::sync::atomic::Ordering::Relaxed),
        };
        // FAST-ROUTER batch twin (lane/fast-router, 2026-08-02): at prefill m the per-(e,tok)
        // w8 form re-streams both operand rows per output (GEMV program at GEMM shape — the
        // concat-prime exactness fix paid -10% q35 board-2048 prefill through it). The batch
        // twin (8x8 expert-x-token register tile) is BIT-IDENTICAL per row (same k order,
        // same tree, same fold — kernel-check sweeps m=1..2048 on real router weights), so
        // the crossover is pure perf, not a numeric config. MIN_T from the on-box sweep
        // (research/fast-router-20260802/crossover-router*.jsonl); decode t=1 and small-t
        // spec verify keep the plain w8 form. MEMRA_ROUTER_BATCH=0: rollback seam
        // (perf-only, bits equal).
        let batch = w8 && t >= ROUTER_BATCH_MIN_T && router_batch_on();
        self.router_gemv_form(w, x, n_embd, n_experts, t, w8, batch)
    }

    /// Form-explicit router GEMV launch (kernel-check bit-identity gate + crossover bench
    /// force both forms; `batch` requires `w8`).
    pub fn router_gemv_form(
        &self,
        w: &CudaSlice<f32>,
        x: &CudaSlice<f32>,
        n_embd: usize,
        n_experts: usize,
        t: usize,
        w8: bool,
        batch: bool,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        debug_assert!(!batch || w8, "batch twin exists for the w8 form only");
        let mut y = self.alloc_uninit::<f32>(t * n_experts)?;
        let f = if batch {
            self.func("router_gemv_f32_w8_batch")
        } else if w8 {
            self.func("router_gemv_f32_w8")
        } else {
            self.func("router_gemv_f32")
        };
        let (ne, nx, ti) = (n_embd as i32, n_experts as i32, t as i32);
        let cfg = if batch {
            LaunchConfig {
                grid_dim: (n_experts.div_ceil(8) as u32, t.div_ceil(8) as u32, 1),
                block_dim: (32, 8, 1),
                shared_mem_bytes: 0,
            }
        } else {
            LaunchConfig {
                grid_dim: (n_experts as u32, t as u32, 1),
                block_dim: (32, if w8 { 8 } else { 1 }, 1),
                shared_mem_bytes: 0,
            }
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(w).arg(x).arg(&mut y).arg(&ne).arg(&nx).arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(y)
    }

    /// `router_gemv` (decode form selection) writing into a caller-owned [t*n_experts]
    /// buffer — token-graph alloc-free.
    pub fn router_gemv_into(
        &self,
        w: &CudaSlice<f32>,
        x: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        n_embd: usize,
        n_experts: usize,
        t: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if y.len() < t * n_experts {
            return Err("router_gemv_into output too small".into());
        }
        let w8 = match std::env::var("MEMRA_ROUTER_V2").as_deref() {
            Ok("0") => false,
            Ok(_) => true,
            Err(_) => ROUTER_W8_DEFAULT.load(std::sync::atomic::Ordering::Relaxed),
        };
        let f = if w8 {
            self.func("router_gemv_f32_w8")
        } else {
            self.func("router_gemv_f32")
        };
        let (ne, nx, ti) = (n_embd as i32, n_experts as i32, t as i32);
        let cfg = LaunchConfig {
            grid_dim: (n_experts as u32, t as u32, 1),
            block_dim: (32, if w8 { 8 } else { 1 }, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(w).arg(x).arg(&mut *y).arg(&ne).arg(&nx).arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// f32 row permute: dst[idx[i], :] = src[i, :] (grouped-GEMM CSR -> pair-id reorder).
    pub fn rows_permute(
        &self,
        src: &CudaSlice<f32>,
        idx: &CudaSlice<i32>,
        nrows: usize,
        ncols: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let mut dst = self.alloc_uninit::<f32>(nrows * ncols)?;
        let f = self.func("rows_permute_f32");
        let (nc, nr) = (ncols as i32, nrows as i32);
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(src).arg(idx).arg(&mut dst).arg(&nc).arg(&nr);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(dst)
    }

    /// shexp gate fused dot: g[tok] = sigmoid(dot(x[tok,:], w)) — replaces the per-layer
    /// cuBLASLt m=1 GEMM + separate sigmoid launch on the qwen35moe decode path (the
    /// splitKreduce x40/step dig, 2026-07-31). One fold order for every t, so the t=1
    /// decode chain and the small-t spec-verify chain match per row by construction.
    pub fn sigmoid_dot_rows(
        &self,
        x: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        n_embd: usize,
        t: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        // MEMRA_SHEXP_DOT=0: rollback seam to the cuBLASLt linear + sigmoid pair (numeric
        // config; same class as MEMRA_ROUTER_V2).
        static OFF: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        if *OFF.get_or_init(|| std::env::var("MEMRA_SHEXP_DOT").as_deref() == Ok("0")) {
            let gs = self.linear(x, w, t, n_embd, 1)?;
            let mut g = self.uninit(t)?;
            self.sigmoid(&gs, &mut g, t)?;
            return Ok(g);
        }
        // FAST-ROUTER lane note (2026-08-02): a register-tiled 8-token batch twin of this
        // kernel was built, proven bit-identical, and measured SLOWER at every prefill t on
        // the 5090 (0.62-0.89x — launch-latency-bound op, ~7us/layer at m=2048;
        // research/fast-router-20260802/crossover-router.jsonl). Dispatch arm killed per
        // flags doctrine; this per-token form serves every t.
        let mut g = self.alloc_uninit::<f32>(t)?;
        let f = self.func("sigmoid_dot_rows_f32");
        let (ne, ti) = (n_embd as i32, t as i32);
        let cfg = LaunchConfig {
            grid_dim: (t as u32, 1, 1),
            block_dim: (32, 8, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(w).arg(&mut g).arg(&ne).arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(g)
    }

    /// `sigmoid_dot_rows` writing into a caller-owned [t] buffer (token-graph alloc-free).
    pub fn sigmoid_dot_rows_into(
        &self,
        x: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        g: &mut CudaSlice<f32>,
        n_embd: usize,
        t: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if g.len() < t {
            return Err("sigmoid_dot_rows_into output too small".into());
        }
        let f = self.func("sigmoid_dot_rows_f32");
        let (ne, ti) = (n_embd as i32, t as i32);
        let cfg = LaunchConfig {
            grid_dim: (t as u32, 1, 1),
            block_dim: (32, 8, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(w).arg(&mut *g).arg(&ne).arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// ROUND-STREAM stream rollback: all counters <- pos_start + base + n_acc.
    pub fn spec_rollback_stream(
        &self,
        len_ptrs: &CudaSlice<u64>,
        pos_start: &CudaSlice<i32>,
        acc: &CudaSlice<u32>,
        base: usize,
        n_rows: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("spec_rollback_stream");
        let (b, nr) = (base as i32, n_rows as i32);
        let cfg = LaunchConfig {
            grid_dim: (n_rows.div_ceil(64) as u32, 1, 1),
            block_dim: (64, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_bl = self.gpu.stream();
        let mut bl = __s_bl.launch_builder(&f);
        bl.arg(len_ptrs).arg(pos_start).arg(acc).arg(&b).arg(&nr);
        unsafe {
            bl.launch(cfg)?;
        }
        Ok(())
    }

    /// PLAIN-DECODE GRAPH ring store: ring[(pos_start - base) % cap] = vam[0].
    pub fn plain_tok_ring(
        &self,
        vam: &CudaSlice<u32>,
        pos_start: &CudaSlice<i32>,
        base: usize,
        ring: &mut CudaSlice<u32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("plain_tok_ring");
        let (b, cap) = (base as i32, ring.len() as i32);
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_bl = self.gpu.stream();
        let mut bl = __s_bl.launch_builder(&f);
        bl.arg(vam).arg(pos_start).arg(&b).arg(&mut *ring).arg(&cap);
        unsafe {
            bl.launch(cfg)?;
        }
        Ok(())
    }

    /// ROUND-STREAM stage (c) 4 epilogue: ring commit + tiny counter copies.
    pub fn spec_ring_commit(
        &self,
        vtok: &CudaSlice<u32>,
        acc: &CudaSlice<u32>,
        brk: &CudaSlice<u32>,
        ring: &mut CudaSlice<u32>,
        pend: &mut CudaSlice<u32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("spec_ring_commit");
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(vtok).arg(acc).arg(brk).arg(ring).arg(pend);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }
    pub fn i32_copy_add(
        &self,
        src: &CudaSlice<i32>,
        dst: &mut CudaSlice<i32>,
        delta: i32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("i32_copy_add");
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(src).arg(dst).arg(&delta);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }
    pub fn u32_copy(
        &self,
        src: &CudaSlice<u32>,
        dst: &mut CudaSlice<u32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("u32_copy");
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(src).arg(dst);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// ROUND-GRAPH adaptive depth: brk[0] <- clamp(acc[0] + 1, floor, cap) — the host
    /// adaptive policy as a captured device op (policy-identical: the accept walk depth
    /// caps acceptance exactly like drafting fewer tokens).
    pub fn spec_adapt_k(
        &self,
        acc: &CudaSlice<u32>,
        brk: &mut CudaSlice<u32>,
        floor: usize,
        cap: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("spec_adapt_k");
        let (fl, cp) = (floor as i32, cap as i32);
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(acc).arg(brk).arg(&fl).arg(&cp);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// ROUND-STREAM stage (c) 3: accept walk fully device-driven (brk + assembled vtok).
    pub fn spec_accept_greedy_dc(
        &self,
        preds: &CudaSlice<u32>,
        vtok: &CudaSlice<u32>,
        last_pred: &CudaSlice<u32>,
        brk: &CudaSlice<u32>,
        out: &mut CudaSlice<u32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("spec_accept_greedy_dc");
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(preds).arg(vtok).arg(last_pred).arg(brk).arg(out);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// ROUND-STREAM stage (c) 2: verify-chain device-pos entries.
    pub fn pos_iota(
        &self,
        pos0: &CudaSlice<i32>,
        out: &mut CudaSlice<i32>,
        t: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("pos_iota_i32");
        let ti = t as i32;
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (t.max(1) as u32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(pos0).arg(out).arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }
    #[allow(clippy::too_many_arguments)]
    pub fn append_kv_quantized_rows_dc(
        &self,
        k_rows: &CudaSlice<f32>,
        v_rows: &CudaSlice<f32>,
        kc: &mut CudaSlice<u8>,
        vc: &mut CudaSlice<u8>,
        t0_dev: &CudaSlice<i32>,
        t: usize,
        kv_dim_k: usize,
        kv_dim_v: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        g: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = if g {
            self.func_g("append_quantize_kv_q8_0_q5_1_rows_dc")
        } else {
            self.func("append_quantize_kv_q8_0_q5_1_rows_dc")
        };
        let nblk = (kv_dim_k.max(kv_dim_v) / 32) as u32;
        let cfg = LaunchConfig {
            grid_dim: (nblk, t as u32, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (kdk, kdv) = (kv_dim_k as i32, kv_dim_v as i32);
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(k_rows)
            .arg(v_rows)
            .arg(kc)
            .arg(vc)
            .arg(t0_dev)
            .arg(&kdk)
            .arg(&kdv)
            .arg(&ktb)
            .arg(&vtb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// t=1 dc append with a FUSED len_d increment (wave 5c) — one launch replaces
    /// append_rows_dc + inc_seqlen. Single block (read-before-inc ordering).
    #[allow(clippy::too_many_arguments)]
    pub fn append_kv_quantized_row_dc_inc(
        &self,
        k_row: &CudaSlice<f32>,
        v_row: &CudaSlice<f32>,
        kc: &mut CudaSlice<u8>,
        vc: &mut CudaSlice<u8>,
        t0_dev: &mut CudaSlice<i32>,
        kv_dim_k: usize,
        kv_dim_v: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        g: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = if g {
            self.func_g("append_quantize_kv_q8_0_q5_1_dc_inc")
        } else {
            self.func("append_quantize_kv_q8_0_q5_1_dc_inc")
        };
        let nthreads = ((kv_dim_k.max(kv_dim_v) / 32) * 32).min(1024) as u32;
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (nthreads, 1, 1),
            shared_mem_bytes: 0,
        };
        let (kdk, kdv) = (kv_dim_k as i32, kv_dim_v as i32);
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(k_row)
            .arg(v_row)
            .arg(kc)
            .arg(vc)
            .arg(t0_dev)
            .arg(&kdk)
            .arg(&kdv)
            .arg(&ktb)
            .arg(&vtb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// ROUND-STREAM: draft-chain pack + in-graph d2t remap (see kernels.cu headers).
    pub fn pack_tok_p(
        &self,
        tok: &CudaSlice<u32>,
        p: &CudaSlice<f32>,
        out: &mut CudaSlice<u32>,
        slot: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("pack_tok_p");
        let sl = slot as i32;
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(tok).arg(p).arg(out).arg(&sl);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }
    pub fn tok_map_u32(
        &self,
        tok: &mut CudaSlice<u32>,
        map: &CudaSlice<u32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("tok_map_u32");
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(tok).arg(map);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// ROUND-STREAM stage (c) 1: device verify-token assembly + p-min break derivation.
    #[allow(clippy::too_many_arguments)]
    pub fn spec_assemble_verify(
        &self,
        tokp: &CudaSlice<u32>,
        pend: &CudaSlice<u32>,
        d2t: Option<&CudaSlice<u32>>,
        vtok: &mut CudaSlice<u32>,
        brk: &mut CudaSlice<u32>,
        p_min: f32,
        k: usize,
        pmin0: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("spec_assemble_verify");
        let (ki, pm) = (k as i32, if pmin0 { 1i32 } else { 0i32 });
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        match d2t {
            Some(m) => {
                b.arg(tokp)
                    .arg(pend)
                    .arg(m)
                    .arg(vtok)
                    .arg(brk)
                    .arg(&p_min)
                    .arg(&ki)
                    .arg(&pm);
                unsafe {
                    b.launch(cfg)?;
                }
            }
            None => {
                let null: u64 = 0;
                b.arg(tokp)
                    .arg(pend)
                    .arg(&null)
                    .arg(vtok)
                    .arg(brk)
                    .arg(&p_min)
                    .arg(&ki)
                    .arg(&pm);
                unsafe {
                    b.launch(cfg)?;
                }
            }
        }
        Ok(())
    }

    /// ROUND-STREAM stage (b) 3b: recur-restore twins with device-j (see hybrid.cu headers).
    #[allow(clippy::too_many_arguments)]
    pub fn ssm_conv_ring_rebuild_dc(
        &self,
        qkv_tm: &CudaSlice<f32>,
        ring_old: &CudaSlice<f32>,
        conv_state: &mut CudaSlice<f32>,
        conv_dim: usize,
        acc: &CudaSlice<u32>,
        base: usize,
        t_v: usize,
        d_conv: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("ssm_conv_ring_rebuild_f32_dc");
        let n = conv_dim * (d_conv - 1);
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let (cd, b0, tv, dc) = (conv_dim as i32, base as i32, t_v as i32, d_conv as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(qkv_tm)
            .arg(ring_old)
            .arg(conv_state)
            .arg(&cd)
            .arg(acc)
            .arg(&b0)
            .arg(&tv)
            .arg(&dc);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }
    #[allow(clippy::too_many_arguments)]
    pub fn gdn_scan_s128_dc(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        g: &CudaSlice<f32>,
        beta: &CudaSlice<f32>,
        state_in: &CudaSlice<f32>,
        state_out: &mut CudaSlice<f32>,
        o: &mut CudaSlice<f32>,
        n_head: usize,
        acc: &CudaSlice<u32>,
        base: usize,
        t_v: usize,
        scale: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gdn_scan_s128_dc");
        const S_V: u32 = 128;
        const WARP: u32 = 32;
        const COLS_PER_BLOCK: u32 = 4;
        let cfg = LaunchConfig {
            grid_dim: (n_head as u32, 1, S_V / COLS_PER_BLOCK),
            block_dim: (WARP, COLS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (h, b0, tv) = (n_head as i32, base as i32, t_v as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(v)
            .arg(g)
            .arg(beta)
            .arg(state_in)
            .arg(state_out)
            .arg(o)
            .arg(&h)
            .arg(acc)
            .arg(&b0)
            .arg(&tv)
            .arg(&scale);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// ROUND-STREAM stage (b) 3a: device per-layer KV-len rollback (see spec_rollback_kv).
    pub fn spec_rollback_kv(
        &self,
        len_ptrs: &CudaSlice<u64>,
        saved: &CudaSlice<i32>,
        acc: &CudaSlice<u32>,
        base: usize,
        n_layer: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("spec_rollback_kv");
        let (b, nl) = (base as i32, n_layer as i32);
        let cfg = LaunchConfig {
            grid_dim: (n_layer.div_ceil(64) as u32, 1, 1),
            block_dim: (64, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_bl = self.gpu.stream();
        let mut bl = __s_bl.launch_builder(&f);
        bl.arg(len_ptrs).arg(saved).arg(acc).arg(&b).arg(&nl);
        unsafe {
            bl.launch(cfg)?;
        }
        Ok(())
    }

    /// OPTIPIPE increment 1: derive the K=1 successor-valid bit on device.
    pub fn spec_fork_valid(
        &self,
        acc: &CudaSlice<u32>,
        optimistic_pending: u32,
        valid: &mut CudaSlice<u32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("spec_fork_valid");
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (1, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_bl = self.gpu.stream();
        let mut bl = __s_bl.launch_builder(&f);
        bl.arg(acc).arg(&optimistic_pending).arg(valid);
        unsafe {
            bl.launch(cfg)?;
        }
        Ok(())
    }

    /// OPTIPIPE increment 1: leave stage-local KV lengths on hit, restore them on miss.
    pub fn spec_fork_reconcile_kv(
        &self,
        len_ptrs: &CudaSlice<u64>,
        saved: &CudaSlice<i32>,
        acc: &CudaSlice<u32>,
        valid: &CudaSlice<u32>,
        base: usize,
        n_layer: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("spec_fork_reconcile_kv");
        let (b, nl) = (base as i32, n_layer as i32);
        let cfg = LaunchConfig {
            grid_dim: (n_layer.div_ceil(64) as u32, 1, 1),
            block_dim: (64, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_bl = self.gpu.stream();
        let mut bl = __s_bl.launch_builder(&f);
        bl.arg(len_ptrs)
            .arg(saved)
            .arg(acc)
            .arg(valid)
            .arg(&b)
            .arg(&nl);
        unsafe {
            bl.launch(cfg)?;
        }
        Ok(())
    }

    /// OPTIPIPE increment 1: conditionally restore one stage-owned recurrent-state buffer.
    pub fn spec_fork_restore_f32(
        &self,
        snapshot: &CudaSlice<f32>,
        state: &mut CudaSlice<f32>,
        valid: &CudaSlice<u32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        assert_eq!(
            snapshot.len(),
            state.len(),
            "fork recurrent snapshot shape mismatch"
        );
        let f = self.func("spec_fork_restore_f32");
        let n = state.len() as i32;
        let blocks = state.len().div_ceil(256).min(65535).max(1) as u32;
        let cfg = LaunchConfig {
            grid_dim: (blocks, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_bl = self.gpu.stream();
        let mut bl = __s_bl.launch_builder(&f);
        bl.arg(snapshot).arg(state).arg(valid).arg(&n);
        unsafe {
            bl.launch(cfg)?;
        }
        Ok(())
    }

    /// ROUND-STREAM stage (b): device next-round seed gather (see spec_seed_gather header).
    /// Caller D2Ds h_seed into fill_prev after (both slots carry the same value in every arm).
    pub fn spec_seed_gather(
        &self,
        vx: &CudaSlice<f32>,
        fill_prev: &CudaSlice<f32>,
        acc: &CudaSlice<u32>,
        h_seed: &mut CudaSlice<f32>,
        base: usize,
        n_embd: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("spec_seed_gather");
        let (b, ne) = (base as i32, n_embd as i32);
        let cfg = LaunchConfig {
            grid_dim: (n_embd.div_ceil(256) as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_bl = self.gpu.stream();
        let mut bl = __s_bl.launch_builder(&f);
        bl.arg(vx)
            .arg(fill_prev)
            .arg(acc)
            .arg(h_seed)
            .arg(&b)
            .arg(&ne);
        unsafe {
            bl.launch(cfg)?;
        }
        Ok(())
    }

    /// ROUND-STREAM stage (a): device greedy accept walk (see spec_accept_greedy header).
    pub fn spec_accept_greedy(
        &self,
        preds: &CudaSlice<u32>,
        draft: &CudaSlice<u32>,
        last_pred: u32,
        base: usize,
        k_round: usize,
        out: &mut CudaSlice<u32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("spec_accept_greedy");
        let (b, k) = (base as i32, k_round as i32);
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_bl = self.gpu.stream();
        let mut bl = __s_bl.launch_builder(&f);
        bl.arg(preds)
            .arg(draft)
            .arg(&last_pred)
            .arg(&b)
            .arg(&k)
            .arg(out);
        unsafe {
            bl.launch(cfg)?;
        }
        Ok(())
    }

    // ================= SAMPLED-SPEC PRIMITIVES (spec_sample.cu, piece A) =================
    // Counter-based randomness: every call takes (seed, stream_pos) — the caller owns the
    // event counter (one per sampled token). temp <= 0 arms are exact greedy limits.

    /// y = x/temp + Gumbel(Philox(seed, stream_pos)) over n logits (then run device argmax on y
    /// = one categorical sample at temperature `temp`). temp<=0: y = x (pure copy).
    pub fn gumbel_perturb(
        &self,
        x: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        n: usize,
        seed: u64,
        stream_pos: u32,
        temp: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gumbel_perturb_f32");
        let (ni, slo, shi) = (n as i32, (seed & 0xFFFF_FFFF) as u32, (seed >> 32) as u32);
        let cfg = LaunchConfig {
            grid_dim: (n.div_ceil(256) as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x)
            .arg(&mut *y)
            .arg(&ni)
            .arg(&slo)
            .arg(&shi)
            .arg(&stream_pos)
            .arg(&temp);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// GRAMMAR TOKEN MASK (constrained decoding, lane/constrained-full): ban every vocab id
    /// whose bit is unset in the packed llguidance bitset, IN PLACE on row `col` of a stacked
    /// [B, n_vocab] logits buffer. `mask` = the SimpleVob u32 words H2D'd verbatim
    /// (~n_vocab/8 bytes/step — trivial on PCIe); ids >= 32*mask_words (padded lm_head tail)
    /// are banned too, the device twin of constrained::apply_mask. Banned value -FLT_MAX ==
    /// the argmax/gumbel kernels' init sentinel, so a fully-banned tail can never win and
    /// ordering matches the host -inf mask bit-for-bit for every finite logit.
    pub fn mask_logits_col(
        &self,
        logits: &mut CudaSlice<f32>,
        mask: &CudaSlice<u32>,
        col: usize,
        n: usize,
        mask_words: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("mask_logits_f32");
        let (ci, ni, mw) = (col as i32, n as i32, mask_words as i32);
        let cfg = LaunchConfig {
            grid_dim: (n.div_ceil(256).min(1024) as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&mut *logits).arg(mask).arg(&ci).arg(&ni).arg(&mw);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Column-`col` twin of `gumbel_perturb` over stacked logits [B, n_vocab] (the batched
    /// serving tick's device sampler): y = x[col]/temp + gumbel(seed, stream_pos, lane).
    /// SAME kernel/Philox mapping as `gumbel_perturb` — bit-identical perturbation for the
    /// same (seed, stream_pos, temp) regardless of which batch column the row sits in
    /// (the lane index is the in-row position; `col` only moves the input pointer). That
    /// pointer-invariance IS the serving isolation contract for sampled rows.
    pub fn gumbel_perturb_col(
        &self,
        x: &CudaSlice<f32>,
        col: usize,
        y: &mut CudaSlice<f32>,
        n: usize,
        seed: u64,
        stream_pos: u32,
        temp: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gumbel_perturb_f32");
        let (ni, slo, shi) = (n as i32, (seed & 0xFFFF_FFFF) as u32, (seed >> 32) as u32);
        let col_view = x.slice(col * n..(col + 1) * n);
        let cfg = LaunchConfig {
            grid_dim: (n.div_ceil(256) as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&col_view)
            .arg(&mut *y)
            .arg(&ni)
            .arg(&slo)
            .arg(&shi)
            .arg(&stream_pos)
            .arg(&temp);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Filtered twin of `gumbel_perturb_col`: the per-row (row_max, th) floor comes from
    /// DEVICE buffers (`filter_stats` output slots at `stat_idx`) — one filtered draw from
    /// the top-k/top-p/min-p-truncated softmax with no stat D2H and no row copy. Same
    /// Philox mapping as every gumbel kernel (pointer-invariant across batch columns —
    /// the serving isolation contract for sampled rows).
    #[allow(clippy::too_many_arguments)]
    pub fn gumbel_perturb_filtered_col(
        &self,
        x: &CudaSlice<f32>,
        col: usize,
        y: &mut CudaSlice<f32>,
        n: usize,
        seed: u64,
        stream_pos: u32,
        temp: f32,
        stat_max: &CudaSlice<f32>,
        stat_th: &CudaSlice<f32>,
        stat_idx: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gumbel_perturb_filtered_col_f32");
        let (ni, slo, shi) = (n as i32, (seed & 0xFFFF_FFFF) as u32, (seed >> 32) as u32);
        let (ci, si) = (col as i32, stat_idx as i32);
        let cfg = LaunchConfig {
            grid_dim: (n.div_ceil(256) as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x)
            .arg(&ci)
            .arg(&mut *y)
            .arg(&ni)
            .arg(&slo)
            .arg(&shi)
            .arg(&stream_pos)
            .arg(&temp)
            .arg(stat_max)
            .arg(stat_th)
            .arg(&si);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// In-graph sampling-event counter bump (spec_sample.cu kernel 5): ctr[0] += 1. The sampled
    /// graph-draft chain replays with FIXED kernel args, so the Philox event counter must be
    /// DEVICE data — the host seeds it once per round; every replay bumps it before the perturb
    /// reads it (counter is data, not state — graph-replay-safe).
    pub fn sctr_inc(&self, ctr: &mut CudaSlice<u32>) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("memra_sctr_inc");
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (1, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&mut *ctr);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Graph-capturable `gumbel_perturb`: the sampling-event counter comes from DEVICE memory
    /// (`ctr[0]`) instead of a host scalar. Identical math to `gumbel_perturb` at
    /// stream_pos == ctr[0] (same Philox call, same lane mapping) — the eager and graph sampled
    /// chains produce bit-identical perturbations for the same (seed, counter, temp).
    pub fn gumbel_perturb_ctr(
        &self,
        x: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        n: usize,
        seed: u64,
        ctr: &CudaSlice<u32>,
        temp: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gumbel_perturb_ctr_f32");
        let (ni, slo, shi) = (n as i32, (seed & 0xFFFF_FFFF) as u32, (seed >> 32) as u32);
        let cfg = LaunchConfig {
            grid_dim: (n.div_ceil(256) as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x)
            .arg(&mut *y)
            .arg(&ni)
            .arg(&slo)
            .arg(&shi)
            .arg(ctr)
            .arg(&temp);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// out[pair] = softmax_temp(x[rows[pair]])[ids[pair]] for npair (row, id) pairs; rows index
    /// into x with `row_stride` f32s per row. temp<=0: out = 1.0 iff id is the row argmax
    /// (smallest-index tie-break — matches the argmax-gate contract).
    pub fn softmax_gather(
        &self,
        x: &CudaSlice<f32>,
        row_stride: usize,
        ids: &CudaSlice<u32>,
        rows: &CudaSlice<i32>,
        out: &mut CudaSlice<f32>,
        n: usize,
        npair: usize,
        temp: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("softmax_gather_f32");
        let (ni, rs) = (n as i32, row_stride as i64);
        let np = npair as i32;
        let cfg = LaunchConfig {
            grid_dim: (npair as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x)
            .arg(&rs)
            .arg(ids)
            .arg(rows)
            .arg(&mut *out)
            .arg(&ni)
            .arg(&np)
            .arg(&temp);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Sample token from norm(max(0, softmax_temp(p) - softmax_temp(q))) (q = None -> plain
    /// categorical from softmax_temp(p)). Row stats (max, sumexp at temp) must be precomputed
    /// (softmax_gather's pass-1 values; see spec.rs caller). Deterministic fixed-order CDF walk.
    pub fn residual_sample(
        &self,
        p: &CudaSlice<f32>,
        q: Option<&CudaSlice<f32>>,
        n: usize,
        temp: f32,
        seed: u64,
        stream_pos: u32,
        out_tok: &mut CudaSlice<u32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("residual_sample_f32");
        let (ni, slo, shi) = (n as i32, (seed & 0xFFFF_FFFF) as u32, (seed >> 32) as u32);
        let nth = 1024u32;
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (nth, 1, 1),
            shared_mem_bytes: 0,
        };
        let has_q: i32 = q.is_some() as i32;
        let qbuf = q.unwrap_or(p); // dummy when absent; kernel gates on has_q
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(p)
            .arg(qbuf)
            .arg(&has_q)
            .arg(&ni)
            .arg(&temp)
            .arg(&slo)
            .arg(&shi)
            .arg(&stream_pos)
            .arg(&mut *out_tok);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Access the shared MoE residency cache (EDGE-1 §B), building it on first use under
    /// MEMRA_MOE_CACHE. The closure runs while the lock is held — keep it to lookup/admit/issue, not
    /// the GEMM. `max_block_bytes` sizes the slots (largest of gate/up/down). Returns the closure's
    /// result. If MEMRA_MOE_CACHE is unset this is never called (the caller checks the env first).
    pub fn with_moe_cache<R>(
        &self,
        max_block_bytes: usize,
        f: impl FnOnce(
            &mut crate::moe_cache::MoeSlotCache,
            &Engine,
        ) -> Result<R, Box<dyn std::error::Error>>,
    ) -> Result<R, Box<dyn std::error::Error>> {
        let mut guard = self.moe_cache.lock().unwrap();
        if guard.is_none() {
            *guard = Some(crate::moe_cache::MoeSlotCache::new(self, max_block_bytes)?);
        }
        let cache = guard.as_mut().unwrap();
        f(cache, self)
    }

    /// Freeze the already-built MoE residency set. This never constructs a cache: callers use it
    /// only after a real prefill has populated the machine-specific CPU/GPU working set.
    pub fn freeze_moe_cache(&self) {
        if let Some(cache) = self.moe_cache.lock().unwrap().as_mut() {
            cache.freeze();
        }
    }

    /// The current residency set as (layer, proj, ex) triples, or None if no cache was built.
    /// Never constructs a cache.
    pub fn export_moe_residency(&self) -> Option<Vec<(u16, u8, u16)>> {
        self.moe_cache
            .lock()
            .unwrap()
            .as_ref()
            .map(crate::moe_cache::MoeSlotCache::export_residency)
    }

    pub(crate) fn moe_cache_frozen(&self) -> bool {
        self.moe_cache
            .lock()
            .unwrap()
            .as_ref()
            .is_some_and(crate::moe_cache::MoeSlotCache::is_frozen)
    }

    /// A frozen heterogeneous CPU/GPU expert split cannot use Hy3's ordinary batched prefill
    /// efficiently: T>=PRIME_MIN_T bypasses the CPU backend and transiently rereads every missing
    /// expert through the GPU spill path. Replay the short prompt through decode after freezing,
    /// while leaving the profiling warmup's established batched behavior untouched.
    /// (`pub`: run-gen's #46 batched-prime gate skips itself when generation will take the
    /// tokenwise arm anyway.)
    pub fn frozen_cpu_experts_prefer_tokenwise_prime(&self) -> bool {
        crate::cpu_experts::configured()
            && self.moe_cache_frozen()
            && std::env::var("MEMRA_CPU_EXPERT_BATCHED_PRIME").as_deref() != Ok("1")
    }

    /// Install the loaded model's exact retained expert-block inventory before lazy cache build.
    pub(crate) fn configure_moe_cache_layout(&self, block_bytes: Vec<usize>) {
        assert!(
            self.moe_cache.lock().unwrap().is_none(),
            "MoE cache layout configured after cache construction"
        );
        *self.moe_cache_layout.lock().unwrap() = Some(block_bytes);
    }

    pub(crate) fn moe_cache_layout(&self) -> Option<Vec<usize>> {
        self.moe_cache_layout.lock().unwrap().clone()
    }

    /// True if the MoE residency cache is enabled (MEMRA_MOE_CACHE set).
    pub fn moe_cache_enabled() -> bool {
        std::env::var("MEMRA_MOE_CACHE").as_deref() != Ok("0")
    }

    /// Snapshot the MoE cache counters (hits, misses, staged_bytes, n_slots) for the §D.4 PCIe gate.
    /// Returns None if the cache was never built (disabled or no MoE forward ran).
    pub fn moe_cache_stats(&self) -> Option<(u64, u64, u64, usize)> {
        let guard = self.moe_cache.lock().unwrap();
        guard
            .as_ref()
            .map(|c| (c.hits, c.misses, c.staged_bytes, c.n_slots()))
    }

    /// Experimental CPU expert backend counters: completed layer calls, experts served, and the
    /// sum of backend wall nanoseconds. The timer includes explicit disk->RAM fills on cache misses;
    /// callers compare a before/after snapshot around a decode window.
    pub fn cpu_expert_stats(
        &self,
    ) -> Option<(u64, u64, u64, u64, u64, u64, u64, u64, u64, u64, u64)> {
        crate::cpu_experts::configured().then(crate::cpu_experts::stats)
    }

    /// Caller-blocked nanoseconds at CPU expert joins. Compare before/after snapshots to measure
    /// the backend tail that resident-GPU expert work did not hide.
    pub fn cpu_expert_predictor_stats(&self) -> (u64, u64) {
        crate::cpu_experts::predictor_stats()
    }

    pub fn cpu_expert_exposed_wait_ns(&self) -> Option<u64> {
        crate::cpu_experts::configured().then(crate::cpu_experts::exposed_wait_ns)
    }

    /// CPU-routed expert selections grouped by how many of their three projections were already
    /// resident in HBM. This makes otherwise-stranded partial residency visible to tuning runs.
    pub fn cpu_expert_gpu_residency_stats(&self) -> Option<(u64, u64, u64)> {
        crate::cpu_experts::configured().then(crate::cpu_experts::incomplete_gpu_residency_stats)
    }

    /// Positioned-read proof-backend counters:
    /// `(reads, bytes, read_errors, short_reads, mmap_fallbacks, buffer_waits, ring_full)`.
    pub fn moe_pread_stats(&self) -> Option<(u64, u64, u64, u64, u64, u64, u64)> {
        let guard = self.moe_cache.lock().unwrap();
        guard
            .as_ref()
            .and_then(|cache| cache.pread_stats())
            .map(|stats| {
                (
                    stats.reads,
                    stats.bytes,
                    stats.read_errors,
                    stats.short_reads,
                    stats.fallbacks,
                    stats.buffer_waits,
                    stats.ring_full,
                )
            })
    }

    /// Spill configuration values that warned and substituted their documented defaults.
    pub fn spill_config_fallbacks(&self) -> u64 {
        crate::spill_pread::config_fallbacks()
    }

    /// Reset the MoE cache perf counters (to separate warmup from steady-state windows).
    pub fn moe_cache_reset_counters(&self) {
        if let Some(c) = self.moe_cache.lock().unwrap().as_mut() {
            c.reset_counters();
        }
    }

    pub fn htod_bytes(&self, v: &[u8]) -> Result<CudaSlice<u8>, Box<dyn std::error::Error>> {
        Ok(self.gpu.stream().clone_htod(v)?)
    }

    /// `htod_bytes` with a mapped (uninit) tail pad: the wide-load expert dots read up to 6B
    /// past the final q4_0 block through their aligned window — the bytes never reach a
    /// result (funnelshift discards them) but must be mapped memory.
    pub fn htod_bytes_padded(
        &self,
        v: &[u8],
        pad: usize,
    ) -> Result<CudaSlice<u8>, Box<dyn std::error::Error>> {
        let mut d = self.alloc_u8_uninit(v.len() + pad)?;
        {
            let mut view = d.slice_mut(0..v.len());
            self.gpu.stream().memcpy_htod(v, &mut view)?;
        }
        Ok(d)
    }

    /// Device-to-device copy of `src` into `dst[off..off+len]` (f32). For in-place KV append.
    pub fn copy_into(
        &self,
        dst: &mut CudaSlice<f32>,
        off: usize,
        src: &CudaSlice<f32>,
        len: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let mut view = dst.slice_mut(off..off + len);
        self.gpu
            .stream()
            .memcpy_dtod(&src.slice(0..len), &mut view)?;
        Ok(())
    }

    /// D2D copy with an offset on BOTH sides. `copy_into` always reads the source from 0,
    /// which cannot express "copy the TAIL of this buffer" — the shape a sliding-window draft
    /// KV export needs (lane/dspark-draft-plane-20260827).
    pub fn copy_range_into(
        &self,
        dst: &mut CudaSlice<f32>,
        dst_off: usize,
        src: &CudaSlice<f32>,
        src_off: usize,
        len: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let mut view = dst.slice_mut(dst_off..dst_off + len);
        self.gpu
            .stream()
            .memcpy_dtod(&src.slice(src_off..src_off + len), &mut view)?;
        Ok(())
    }

    /// View a sub-range of a device buffer (for attending over [0..len) of a KV cache).
    /// u8 twin of copy_into (D2D byte-range copy at an offset).
    pub fn copy_u8_into(
        &self,
        dst: &mut CudaSlice<u8>,
        off: usize,
        src: &CudaSlice<u8>,
        len: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let mut view = dst.slice_mut(off..off + len);
        self.gpu
            .stream()
            .memcpy_dtod(&src.slice(0..len), &mut view)?;
        Ok(())
    }

    /// D2D byte-range copy with explicit source and destination offsets.
    pub fn copy_u8_range_into(
        &self,
        dst: &mut CudaSlice<u8>,
        dst_off: usize,
        src: &CudaSlice<u8>,
        src_off: usize,
        len: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let mut dst_view = dst.slice_mut(dst_off..dst_off + len);
        self.gpu
            .stream()
            .memcpy_dtod(&src.slice(src_off..src_off + len), &mut dst_view)?;
        Ok(())
    }

    /// Resolve an absolute append slot to the Step35 SWA layer's physical rows. At wrap, copy
    /// only the aligned live prefix through temporary device storage and rebase it at row zero,
    /// keeping the audited attention range contiguous without changing its absolute start.
    pub fn prepare_kv_append(
        &self,
        kv: &mut crate::cache::KvLayer,
        retain_from: usize,
        append_rows: usize,
    ) -> Result<usize, Box<dyn std::error::Error>> {
        let Some(plan) = kv
            .ring
            .as_ref()
            .map(|ring| ring.append_plan(kv.len, retain_from, append_rows))
            .transpose()?
        else {
            return Ok(kv.len);
        };
        match plan {
            crate::cache::KvRingAppend::Contiguous { write_row } => Ok(write_row),
            crate::cache::KvRingAppend::Rebase {
                src_row,
                keep_rows,
                new_base,
                write_row,
            } => {
                if keep_rows > 0 {
                    let k_len = keep_rows * kv.k_tok_bytes;
                    let v_len = keep_rows * kv.v_tok_bytes;
                    let mut k_tmp = self.alloc_u8_uninit(k_len)?;
                    let mut v_tmp = self.alloc_u8_uninit(v_len)?;
                    self.copy_u8_range_into(&mut k_tmp, 0, &kv.k, src_row * kv.k_tok_bytes, k_len)?;
                    self.copy_u8_range_into(&mut v_tmp, 0, &kv.v, src_row * kv.v_tok_bytes, v_len)?;
                    self.copy_u8_into(&mut kv.k, 0, &k_tmp, k_len)?;
                    self.copy_u8_into(&mut kv.v, 0, &v_tmp, v_len)?;
                }
                kv.ring.as_mut().unwrap().apply_rebase(new_base);
                Ok(write_row)
            }
        }
    }

    /// H2D write of `src` into `dst[off..off+src.len()]` (u8). In-place row updates for the
    /// adaptive trim head: no realloc, so captured graphs keep their baked addresses.
    pub fn htod_u8_into(
        &self,
        dst: &mut CudaSlice<u8>,
        off: usize,
        src: &[u8],
    ) -> Result<(), Box<dyn std::error::Error>> {
        let mut view = dst.slice_mut(off..off + src.len());
        self.gpu.stream().memcpy_htod(src, &mut view)?;
        Ok(())
    }

    pub fn view<'a>(&self, b: &'a CudaSlice<f32>, len: usize) -> cudarc::driver::CudaView<'a, f32> {
        b.slice(0..len)
    }

    /// View the first `len` BYTES of a u8 device buffer (quantized KV cache: [0..t_kv*tok_bytes)).
    /// Byte-range view (gemma4 R6 window offset into the quantized KV stream).
    pub fn view_u8_range<'a>(
        &self,
        b: &'a CudaSlice<u8>,
        start: usize,
        end: usize,
    ) -> cudarc::driver::CudaView<'a, u8> {
        b.slice(start..end)
    }
    pub fn view_u8<'a>(
        &self,
        b: &'a CudaSlice<u8>,
        len: usize,
    ) -> cudarc::driver::CudaView<'a, u8> {
        b.slice(0..len)
    }

    /// Append-quantize ONE token's post-RoPE K (q8_0) and V (q5_1) into the resident byte caches at
    /// token index `t` (KVQUANT-PLAN §C). One CTA (one warp) per 32-element block; the kernel writes
    /// the f16 scale(s) + packed quants for K and V. k_row/v_row are f32 [kv_dim_k]/[kv_dim_v].
    pub fn append_kv_quantized(
        &self,
        k_row: &CudaSlice<f32>,
        v_row: &CudaSlice<f32>,
        kc: &mut CudaSlice<u8>,
        vc: &mut CudaSlice<u8>,
        t: usize,
        kv_dim_k: usize,
        kv_dim_v: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        g: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = if g {
            self.func_g("append_quantize_kv_q8_0_q5_1")
        } else {
            self.func("append_quantize_kv_q8_0_q5_1")
        };
        let nblk = (kv_dim_k.max(kv_dim_v) / 32) as u32;
        let cfg = LaunchConfig {
            grid_dim: (nblk, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (ti, kdk, kdv) = (t as i32, kv_dim_k as i32, kv_dim_v as i32);
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(k_row)
            .arg(v_row)
            .arg(kc)
            .arg(vc)
            .arg(&ti)
            .arg(&kdk)
            .arg(&kdv)
            .arg(&ktb)
            .arg(&vtb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Device-counter variant of `append_kv_quantized` (CUDA-GRAPH-PLAN Phase 2): the write slot
    /// `t` is read from `t_dev[0]` (a resident device i32[1]) instead of a host int arg, so the
    /// launch args are FIXED across decode steps (graph-capturable). Identical quant math.
    pub fn append_kv_quantized_dc(
        &self,
        k_row: &CudaSlice<f32>,
        v_row: &CudaSlice<f32>,
        kc: &mut CudaSlice<u8>,
        vc: &mut CudaSlice<u8>,
        t_dev: &CudaSlice<i32>,
        kv_dim_k: usize,
        kv_dim_v: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        g: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let nblk = (kv_dim_k.max(kv_dim_v) / 32) as u32;
        let (kdk, kdv) = (kv_dim_k as i32, kv_dim_v as i32);
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        // PDL wave-B2: flash-module flavor mirrors the builder path's g flag exactly.
        if Self::pdl_on() && Self::pdl_wb_on() {
            use cudarc::driver::{DevicePtr, DevicePtrMut};
            let s = &self.gpu.stream();
            let (pk, _g0) = k_row.device_ptr(s);
            let (pv, _g1) = v_row.device_ptr(s);
            let (pkc, _g2) = kc.device_ptr_mut(s);
            let (pvc, _g3) = vc.device_ptr_mut(s);
            let (pt, _g4) = t_dev.device_ptr(s);
            let mut ps = [
                &pk as *const _ as *mut std::ffi::c_void,
                &pv as *const _ as *mut _,
                &pkc as *const _ as *mut _,
                &pvc as *const _ as *mut _,
                &pt as *const _ as *mut _,
                &kdk as *const _ as *mut _,
                &kdv as *const _ as *mut _,
                &ktb as *const _ as *mut _,
                &vtb as *const _ as *mut _,
            ];
            unsafe {
                self.launch_pdl_flash(
                    g,
                    "append_quantize_kv_q8_0_q5_1_dc",
                    (nblk, 1, 1),
                    (32, 1, 1),
                    0,
                    &mut ps,
                )?;
            }
            return Ok(());
        }
        let f = if g {
            self.func_g("append_quantize_kv_q8_0_q5_1_dc")
        } else {
            self.func("append_quantize_kv_q8_0_q5_1_dc")
        };
        let cfg = LaunchConfig {
            grid_dim: (nblk, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(k_row)
            .arg(v_row)
            .arg(kc)
            .arg(vc)
            .arg(t_dev)
            .arg(&kdk)
            .arg(&kdv)
            .arg(&ktb)
            .arg(&vtb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Append-quantize T token rows in one shot (BATCHED PROMPT PRIME). k_rows/v_rows are
    /// token-major [T, kv_dim] post-RoPE f32; rows land at cache slots t0..t0+T. Default = the
    /// batched `_rows` kernel: one (nblk, T) launch whose per-(block,token) warp program is the
    /// per-token append kernel verbatim -> every written row is BIT-IDENTICAL to T sequential
    /// `append_kv_quantized_view` calls (kernel_check pins the bytes). MEMRA_PRIME_APPEND_LOOP=1
    /// forces the T-launch per-row loop (the A/B seam that measured the launch overhead).
    #[allow(clippy::too_many_arguments)]
    pub fn append_kv_quantized_rows(
        &self,
        k_rows: &CudaSlice<f32>,
        v_rows: &CudaSlice<f32>,
        kc: &mut CudaSlice<u8>,
        vc: &mut CudaSlice<u8>,
        t0: usize,
        t: usize,
        kv_dim_k: usize,
        kv_dim_v: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        g: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if std::env::var("MEMRA_PRIME_APPEND_LOOP").is_ok() {
            for i in 0..t {
                let k_row = k_rows.slice(i * kv_dim_k..(i + 1) * kv_dim_k);
                let v_row = v_rows.slice(i * kv_dim_v..(i + 1) * kv_dim_v);
                self.append_kv_quantized_view(
                    &k_row,
                    &v_row,
                    kc,
                    vc,
                    t0 + i,
                    kv_dim_k,
                    kv_dim_v,
                    k_tok_bytes,
                    v_tok_bytes,
                    g,
                )?;
            }
            return Ok(());
        }
        let f = if g {
            self.func_g("append_quantize_kv_q8_0_q5_1_rows")
        } else {
            self.func("append_quantize_kv_q8_0_q5_1_rows")
        };
        let nblk = (kv_dim_k.max(kv_dim_v) / 32) as u32;
        let cfg = LaunchConfig {
            grid_dim: (nblk, t as u32, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (t0i, kdk, kdv) = (t0 as i32, kv_dim_k as i32, kv_dim_v as i32);
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(k_rows)
            .arg(v_rows)
            .arg(kc)
            .arg(vc)
            .arg(&t0i)
            .arg(&kdk)
            .arg(&kdv)
            .arg(&ktb)
            .arg(&vtb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Increment a device i32[1] counter in place (p[0] += 1) via the resident `inc_i32` kernel.
    /// Used to advance the device-resident seqlen/pos counters inside the decode-dc path (and,
    /// later, inside a captured graph) without a host round-trip.
    pub fn inc_seqlen(&self, p: &mut CudaSlice<i32>) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("inc_i32");
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (1, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(p);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Like `append_kv_quantized` but k_row/v_row are CudaViews (one token's row sliced out of a
    /// token-major [T, kv_dim] activation buffer — the MTP verify path appends T tokens).
    pub fn append_kv_quantized_view(
        &self,
        k_row: &cudarc::driver::CudaView<f32>,
        v_row: &cudarc::driver::CudaView<f32>,
        kc: &mut CudaSlice<u8>,
        vc: &mut CudaSlice<u8>,
        t: usize,
        kv_dim_k: usize,
        kv_dim_v: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        g: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let stream = self.gpu.stream();
        ensure_tensor_stream_device(k_row, &stream, "append_kv_quantized_view.k_row")?;
        ensure_tensor_stream_device(v_row, &stream, "append_kv_quantized_view.v_row")?;
        ensure_tensor_stream_device(kc, &stream, "append_kv_quantized_view.k_cache")?;
        ensure_tensor_stream_device(vc, &stream, "append_kv_quantized_view.v_cache")?;
        let f = if g {
            self.func_g("append_quantize_kv_q8_0_q5_1")
        } else {
            self.func("append_quantize_kv_q8_0_q5_1")
        };
        let nblk = (kv_dim_k.max(kv_dim_v) / 32) as u32;
        let cfg = LaunchConfig {
            grid_dim: (nblk, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (ti, kdk, kdv) = (t as i32, kv_dim_k as i32, kv_dim_v as i32);
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let mut b = stream.launch_builder(&f);
        b.arg(k_row)
            .arg(v_row)
            .arg(kc)
            .arg(vc)
            .arg(&ti)
            .arg(&kdk)
            .arg(&kdv)
            .arg(&ktb)
            .arg(&vtb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Device-to-device copy of a CudaView `src` into `dst[off..off+len]` (f32). Like `copy_into`
    /// but the source is a sub-view (e.g. one column of a token-major activation buffer).
    pub fn copy_view_into(
        &self,
        dst: &mut CudaSlice<f32>,
        off: usize,
        src: &cudarc::driver::CudaView<f32>,
        len: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let mut view = dst.slice_mut(off..off + len);
        self.gpu
            .stream()
            .memcpy_dtod(&src.slice(0..len), &mut view)?;
        Ok(())
    }

    /// Real device-to-device COPY of `src` into a freshly allocated buffer (NOT an Arc clone).
    /// Used for cache snapshots (MTP-PLAN §D.4): `CudaSlice::clone()` only bumps a refcount and
    /// would alias the live buffer; this allocs new device memory and memcpy_dtod's the contents.
    pub fn clone_dtod(
        &self,
        src: &CudaSlice<f32>,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let mut dst = self.gpu.stream().alloc_zeros::<f32>(src.len())?;
        self.gpu.stream().memcpy_dtod(src, &mut dst)?;
        Ok(dst)
    }

    /// D2D row extraction: copy a view (e.g. one row of a [B, n] batch buffer) into `dst`.
    /// Stream-ordered, async — decode_batch's per-sequence row plumbing.
    pub fn dtod_copy_view(
        &self,
        src: &cudarc::driver::CudaView<f32>,
        dst: &mut CudaSlice<f32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        self.gpu.stream().memcpy_dtod(src, dst)?;
        Ok(())
    }

    /// D2D i8 twin of `dtod_copy_view` (q8_1 activation rows).
    pub fn dtod_copy_view_i8(
        &self,
        src: &cudarc::driver::CudaView<i8>,
        dst: &mut CudaSlice<i8>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        self.gpu.stream().memcpy_dtod(src, dst)?;
        Ok(())
    }

    /// D2D row placement: copy `src` into `dst[offset .. offset+src.len()]`.
    pub fn dtod_copy_into(
        &self,
        src: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        offset: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let n = src.len();
        let mut dv = dst.slice_mut(offset..offset + n);
        self.gpu.stream().memcpy_dtod(src, &mut dv)?;
        Ok(())
    }

    /// Batched uniform-size D2D copy (engine-bundle slice 1, DSF-ROUNDCOST-20260820 §1.1):
    /// `n` disjoint regions of `words` f32 each; `table` = [src_0..src_{n-1}, dst_0..dst_{n-1}]
    /// raw device pointers. ONE kernel launch replaces `n` memcpy_dtod dispatches — the dspark
    /// round's snap/commit copy dribble (~0.9 ms/round of dispatch serialization measured).
    /// Bytes and stream order are identical to the memcpy sequence it replaces.
    pub fn copy_batch_uniform_f32(
        &self,
        table: &CudaSlice<u64>,
        n: usize,
        words: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if n == 0 || words == 0 {
            return Ok(());
        }
        debug_assert!(
            table.len() >= 2 * n,
            "pointer table must hold n srcs + n dsts"
        );
        let f = self.func("copy_batch_uniform_f32");
        // Enough blocks to stream a multi-MB region, few enough that (chunks x n) stays a
        // sane grid: 512K-word ssm regions get 48 grid-striding blocks each.
        let chunks = (words / 4).max(1).div_ceil(256).min(48) as u32;
        let (ni, wi) = (n as i32, words as i32);
        let cfg = LaunchConfig {
            grid_dim: (chunks, n as u32, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s = self.gpu.stream();
        let mut b = __s.launch_builder(&f);
        b.arg(table).arg(&ni).arg(&wi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// H2D refresh of an EXISTING u64 pointer table IN PLACE (stable pointer — the batched
    /// state-copy tables are refreshed per round because the GDN ssm handles ping-pong).
    pub fn htod_u64_into(
        &self,
        v: &[u64],
        dst: &mut CudaSlice<u64>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let mut view = dst.slice_mut(0..v.len());
        self.gpu.stream().memcpy_htod(v, &mut view)?;
        Ok(())
    }

    /// Indirect-source copy (engine-bundle slice 3): the src ADDRESS is loaded from a
    /// device pointer-table entry at run time, so a captured graph follows the gdn
    /// ping-pong through the same table its scan kernels read — a baked memcpy node
    /// would keep the capture-time physical buffer. `dst_off`/`words` in f32 elements.
    pub fn copy_indirect_src_f32(
        &self,
        src_entry: &cudarc::driver::CudaView<u64>,
        dst: &mut CudaSlice<f32>,
        dst_off: usize,
        words: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("copy_indirect_src_f32");
        let chunks = (words / 4).max(1).div_ceil(256).min(48) as u32;
        let wi = words as i32;
        let cfg = LaunchConfig {
            grid_dim: (chunks, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let mut dv = dst.slice_mut(dst_off..dst_off + words);
        let __s = self.gpu.stream();
        let mut b = __s.launch_builder(&f);
        b.arg(src_entry).arg(&mut dv).arg(&wi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Uninitialized i8 device buffer (decode_batch q8_1 row scratch).
    pub fn uninit_i8(&self, n: usize) -> Result<CudaSlice<i8>, Box<dyn std::error::Error>> {
        self.alloc_uninit::<i8>(n)
    }

    /// Resident-quantized linear (Stage-A: f32 dequant-in-kernel). y[m,out]=x[m,in]@W[out,in]^T.
    pub fn qmatvec(
        &self,
        w: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        qtype: i32,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let f = self.func("qmatvec_f32");
        let mut y = self.alloc_uninit::<f32>(m * out_f)?; // full-overwrite output: skip memset
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, m as u32, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi, qt, rb) =
            (in_f as i32, out_f as i32, m as i32, qtype, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(w)
            .arg(x)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&qt)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(y)
    }

    /// Allocate a reusable u8 GPU scratch buffer (for staged expert weights).
    pub fn alloc_u8(&self, n: usize) -> Result<CudaSlice<u8>, Box<dyn std::error::Error>> {
        let s = self.gpu.stream().alloc_zeros::<u8>(n)?;
        self.keep_if_capturing(&s);
        Ok(s)
    }

    /// Uninitialized u8 scratch — skips alloc_zeros' memset. ONLY for staging buffers whose read
    /// range is fully overwritten by a stage_expert H2D before any kernel reads it (LAUNCH-STRUCTURE
    /// STAGE 2: the per-layer MoE scratch trio was 3 dead ~1MB memsets per layer per decode token).
    pub fn alloc_u8_uninit(&self, n: usize) -> Result<CudaSlice<u8>, Box<dyn std::error::Error>> {
        let s = unsafe { self.gpu.stream().alloc::<u8>(n)? };
        self.keep_if_capturing(&s);
        Ok(s)
    }

    /// Zero a SUB-RANGE of an f32 buffer (CudaViewMut) — the row-sized memset the moe_out
    /// memset-elision uses for tokens that fall off the gdec fast path (LAUNCH-STRUCTURE STAGE 2).
    pub fn memset_zeros_view(
        &self,
        dst: &mut cudarc::driver::CudaViewMut<f32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        self.gpu.stream().memset_zeros(dst)?;
        Ok(())
    }

    /// EDGE-1 staging: copy `host_bytes` (a sub-slice of a HostExps buffer) into `scratch`
    /// at byte offset `off` (async H2D on the default stream). Length is host_bytes.len().
    /// The qmatvec_view that reads `scratch[off..]` is enqueued on the SAME stream after this,
    /// so ordering is guaranteed without an explicit sync (Stage-1; Stage-2 prefetch on a 2nd
    /// stream would require an event).
    pub fn stage_expert(
        &self,
        host_bytes: &[u8],
        scratch: &mut CudaSlice<u8>,
        off: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let mut dst = scratch.slice_mut(off..off + host_bytes.len()); // CudaViewMut<u8>
        self.gpu.stream().memcpy_htod(host_bytes, &mut dst)?; // accepts &[u8] HostSlice src
        Ok(())
    }

    /// EDGE-1 §A: fused MoE router. `logits` is the router output [t, n_expert] (device, f32, the
    /// `gate_inp @ z` result). Returns (sel_idx [t, n_used] i32, sel_w [t, n_used] f32): the top-k
    /// expert ids (DESC by prob, ascending-index tiebreak) and renormalized weights. Replaces the
    /// host dtoh + softmax-256 + stable DESC top-8 sort + renorm (hybrid_forward.rs ~281-298).
    /// One CTA per token row, 256 threads (one per expert).
    pub fn moe_router_topk(
        &self,
        logits: &CudaSlice<f32>,
        t: usize,
        n_expert: usize,
        n_used: usize,
    ) -> Result<(CudaSlice<i32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        let f = self.func("moe_router_topk_f32");
        let mut sel_idx = self.alloc_uninit::<i32>(t * n_used)?; // kernel fully overwrites
        let mut sel_w = self.alloc_uninit::<f32>(t * n_used)?; // kernel fully overwrites
        let cfg = LaunchConfig {
            grid_dim: (t as u32, 1, 1),
            block_dim: (n_expert as u32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (ne, nu) = (n_expert as i32, n_used as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(logits)
            .arg(&mut sel_idx)
            .arg(&mut sel_w)
            .arg(&ne)
            .arg(&nu);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((sel_idx, sel_w))
    }

    /// gemma4 twin: per-expert output scale folded into the topk renorm write (replaces the
    /// separate moe_w_exscale launch; value chain identical: (w/ws) * s[sel]).
    pub fn moe_router_topk_scaled(
        &self,
        logits: &CudaSlice<f32>,
        t: usize,
        n_expert: usize,
        n_used: usize,
        ex_scale: &CudaSlice<f32>,
    ) -> Result<(CudaSlice<i32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        // barrier-lean v2 twin (per-warp top-k + one-warp merge) FALSIFIED 2026-07-14:
        // bit-identical streams but −1.4% (26B plain N=3 interleaved) — at t=1 the grid is
        // ONE block, so the 6.6us is launch/dependency overhead, not the barrier chain;
        // fewer barriers bought nothing and the merge structure cost. jsonl is the record.
        let f = self.func("moe_router_topk_scaled_f32");
        let mut sel_idx = self.alloc_uninit::<i32>(t * n_used)?;
        let mut sel_w = self.alloc_uninit::<f32>(t * n_used)?;
        let cfg = LaunchConfig {
            grid_dim: (t as u32, 1, 1),
            block_dim: (n_expert as u32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (ne, nu) = (n_expert as i32, n_used as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(logits)
            .arg(&mut sel_idx)
            .arg(&mut sel_w)
            .arg(&ne)
            .arg(&nu)
            .arg(ex_scale);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((sel_idx, sel_w))
    }

    /// LAUNCH-STRUCTURE STAGE 1 (2026-07-05): fused router + SINGLE-SYNC host readback. The old
    /// MEMRA_FUSED_ROUTER path lost 2% at t=1 because it paid TWO full stream syncs (dtoh_i32 then
    /// dtoh, each = clone_dtoh + synchronize) + two alloc_zeros memsets per MoE layer, where the
    /// host route pays ONE sync on the 1KB logits dtoh. This variant: uninit outputs (kernel fully
    /// overwrites), both DtoH copies issued ASYNC into a persistent PINNED host staging buffer
    /// (flags=0 — cacheable, NOT cudarc's WRITECOMBINED default, so the host-side reads of sel/w
    /// stay cached), then ONE synchronize. Numerics identical to `moe_router_topk` (same kernel).
    pub fn moe_router_topk_host(
        &self,
        logits: &CudaSlice<f32>,
        t: usize,
        n_expert: usize,
        n_used: usize,
    ) -> Result<(Vec<u32>, Vec<f32>), Box<dyn std::error::Error>> {
        let f = self.func("moe_router_topk_f32");
        let n = t * n_used;
        let mut sel_idx = self.alloc_uninit::<i32>(n)?;
        let mut sel_w = self.alloc_uninit::<f32>(n)?;
        let cfg = LaunchConfig {
            grid_dim: (t as u32, 1, 1),
            block_dim: (n_expert as u32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (ne, nu) = (n_expert as i32, n_used as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(logits)
            .arg(&mut sel_idx)
            .arg(&mut sel_w)
            .arg(&ne)
            .arg(&nu);
        unsafe {
            b.launch(cfg)?;
        }
        // single-sync readback: sel (i32) at offset 0, w (f32) at offset n*4 of the pinned stage.
        let bytes = n * 8;
        let mut guard = self.router_stage.lock().unwrap();
        if guard.as_ref().map(|p| p.cap < bytes).unwrap_or(true) {
            *guard = Some(PinnedStage::new(bytes.max(4096))?);
        }
        let stage = guard.as_mut().unwrap();
        let (si, sw) = unsafe {
            (
                std::slice::from_raw_parts_mut(stage.ptr as *mut i32, n),
                std::slice::from_raw_parts_mut(stage.ptr.add(n * 4) as *mut f32, n),
            )
        };
        self.gpu.stream().memcpy_dtoh(&sel_idx, si)?; // async (pinned dst)
        self.gpu.stream().memcpy_dtoh(&sel_w, sw)?; // async (pinned dst)
        self.gpu.stream().synchronize()?; // ONE sync for both
        Ok((si.iter().map(|&i| i as u32).collect(), sw.to_vec()))
    }

    /// Device sigmoid router for Step-3.7 / DeepSeek-V3-class MoEs. `correction_bias` is added
    /// only to the top-k key; returned weights use the un-biased sigmoid score. `active` masks
    /// original expert ids before top-k. Exact key ties choose the smaller original id.
    #[allow(clippy::too_many_arguments)]
    pub fn moe_router_sigmoid_topk(
        &self,
        logits: &CudaSlice<f32>,
        t: usize,
        n_expert: usize,
        n_used: usize,
        active_count: usize,
        correction_bias: &CudaSlice<f32>,
        active: &CudaSlice<u8>,
        scaling_factor: f32,
        route_norm: bool,
    ) -> Result<(CudaSlice<i32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        crate::sigrouter_contract::validate_active_count(n_used, active_count)?;
        if n_expert == 0 || n_expert > 1024 || n_used == 0 || n_used > n_expert {
            return Err(format!(
                "sigmoid router shape unsupported: n_expert={n_expert}, n_used={n_used}",
            )
            .into());
        }
        if logits.len() < t * n_expert
            || correction_bias.len() != n_expert
            || active.len() != n_expert
        {
            return Err(format!(
                "sigmoid router buffer mismatch: logits={} bias={} active={} expected logits>={} row={}",
                logits.len(), correction_bias.len(), active.len(), t * n_expert, n_expert,
            ).into());
        }
        let f = self.func(crate::sigmoid_topk_kernel(
            crate::sig_expf_dev_on(),
            crate::topk_fast_on(),
            n_used,
        ));
        let mut sel_idx = self.alloc_uninit::<i32>(t * n_used)?;
        let mut sel_w = self.alloc_uninit::<f32>(t * n_used)?;
        let threads = n_expert.div_ceil(32) * 32;
        let cfg = LaunchConfig {
            grid_dim: (t as u32, 1, 1),
            block_dim: (threads as u32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (ne, nu, rn) = (n_expert as i32, n_used as i32, i32::from(route_norm));
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(logits)
            .arg(correction_bias)
            .arg(active)
            .arg(&mut sel_idx)
            .arg(&mut sel_w)
            .arg(&ne)
            .arg(&nu)
            .arg(&scaling_factor)
            .arg(&rn);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((sel_idx, sel_w))
    }

    /// `moe_router_sigmoid_topk` writing into caller-owned buffers (alloc-free: child graphs
    /// cannot contain mem nodes, so the token-graph e-sections pre-own every output).
    #[allow(clippy::too_many_arguments)]
    /// Ring a doorbell flag at a RAW device address (see `memra_ring_flag`): one store of
    /// `value`, fenced. Used by a peer rank to signal join readiness into root memory, where
    /// the model engine can wait on it with a same-device stream memop.
    pub fn ring_flag_raw(&self, ptr: u64, value: u32) -> Result<(), Box<dyn std::error::Error>> {
        if ptr == 0 {
            return Err("ring_flag_raw: unarmed flag".into());
        }
        let f = self.func("memra_ring_flag");
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&ptr).arg(&value);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// One-launch mirror of a routed selection (`sel` int32 + `route_w` f32) — see
    /// `moe_sel_w_mirror`. Replaces the two tiny D2D copies the rank pull used to issue.
    pub fn moe_sel_w_mirror(
        &self,
        sel_src: &CudaSlice<i32>,
        w_src: &CudaSlice<f32>,
        sel_dst: &mut CudaSlice<i32>,
        w_dst: &mut CudaSlice<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if n == 0
            || n > 32
            || sel_src.len() < n
            || w_src.len() < n
            || sel_dst.len() < n
            || w_dst.len() < n
        {
            return Err(format!("moe_sel_w_mirror geometry n={n}").into());
        }
        let f = self.func("moe_sel_w_mirror");
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let ni = n as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(sel_src).arg(w_src).arg(sel_dst).arg(w_dst).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn moe_router_sigmoid_topk_into(
        &self,
        logits: &CudaSlice<f32>,
        t: usize,
        n_expert: usize,
        n_used: usize,
        active_count: usize,
        correction_bias: &CudaSlice<f32>,
        active: &CudaSlice<u8>,
        scaling_factor: f32,
        route_norm: bool,
        sel_idx: &mut CudaSlice<i32>,
        sel_w: &mut CudaSlice<f32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        crate::sigrouter_contract::validate_active_count(n_used, active_count)?;
        if n_expert == 0
            || n_expert > 1024
            || n_used == 0
            || n_used > 32 // the kernels' shared pick cache (s_pick_w) is sized 32
            || n_used > n_expert
            || logits.len() < t * n_expert
            || correction_bias.len() != n_expert
            || active.len() != n_expert
            || sel_idx.len() < t * n_used
            || sel_w.len() < t * n_used
        {
            return Err("sigmoid router _into geometry mismatch".into());
        }
        let f = self.func(crate::sigmoid_topk_kernel(
            crate::sig_expf_dev_on(),
            crate::topk_fast_on(),
            n_used,
        ));
        let threads = n_expert.div_ceil(32) * 32;
        let cfg = LaunchConfig {
            grid_dim: (t as u32, 1, 1),
            block_dim: (threads as u32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (ne, nu, rn) = (n_expert as i32, n_used as i32, i32::from(route_norm));
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(logits)
            .arg(correction_bias)
            .arg(active)
            .arg(&mut *sel_idx)
            .arg(&mut *sel_w)
            .arg(&ne)
            .arg(&nu)
            .arg(&scaling_factor)
            .arg(&rn);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Single-sync pinned readback twin of `moe_router_sigmoid_topk`. This preserves the existing
    /// grouped/staged dispatch contract while replacing the full-logit DtoH plus host sigmoid/sort.
    #[allow(clippy::too_many_arguments)]
    pub fn moe_router_sigmoid_topk_host(
        &self,
        logits: &CudaSlice<f32>,
        t: usize,
        n_expert: usize,
        n_used: usize,
        active_count: usize,
        correction_bias: &CudaSlice<f32>,
        active: &CudaSlice<u8>,
        scaling_factor: f32,
        route_norm: bool,
    ) -> Result<(Vec<u32>, Vec<f32>), Box<dyn std::error::Error>> {
        let (sel_idx, sel_w) = self.moe_router_sigmoid_topk(
            logits,
            t,
            n_expert,
            n_used,
            active_count,
            correction_bias,
            active,
            scaling_factor,
            route_norm,
        )?;
        let n = t * n_used;
        let bytes = n * 8;
        let mut guard = self.router_stage.lock().unwrap();
        if guard.as_ref().map(|p| p.cap < bytes).unwrap_or(true) {
            *guard = Some(PinnedStage::new(bytes.max(4096))?);
        }
        let stage = guard.as_mut().unwrap();
        let (si, sw) = unsafe {
            (
                std::slice::from_raw_parts_mut(stage.ptr as *mut i32, n),
                std::slice::from_raw_parts_mut(stage.ptr.add(n * 4) as *mut f32, n),
            )
        };
        self.gpu.stream().memcpy_dtoh(&sel_idx, si)?;
        self.gpu.stream().memcpy_dtoh(&sel_w, sw)?;
        self.gpu.stream().synchronize()?;
        Ok((si.iter().map(|&i| i as u32).collect(), sw.to_vec()))
    }

    /// EDGE-1 §C.2: async H2D of `host_bytes` into `scratch[off..]` on the COPY stream, returning a
    /// recorded event the compute stream can `wait` on before the dependent GEMM. Used for in-token
    /// expert prefetch (pipeline by one). `host_bytes` should be pinned for a true DMA (§C.1).
    pub fn stage_expert_async(
        &self,
        host_bytes: &[u8],
        scratch: &mut CudaSlice<u8>,
        off: usize,
    ) -> Result<cudarc::driver::CudaEvent, Box<dyn std::error::Error>> {
        let mut dst = scratch.slice_mut(off..off + host_bytes.len());
        self.copy_stream.memcpy_htod(host_bytes, &mut dst)?;
        Ok(self.copy_stream.record_event(None)?)
    }

    /// Make the compute stream wait for an async copy event (the consumer side of `stage_expert_async`).
    pub fn compute_wait(
        &self,
        ev: &cudarc::driver::CudaEvent,
    ) -> Result<(), Box<dyn std::error::Error>> {
        self.gpu.stream().wait(ev)?;
        Ok(())
    }

    /// qmatvec over a byte sub-range of a (resident/scratch) CudaSlice<u8> holding ONE expert
    /// matrix. x is a CudaView<f32> (a sliced row of z, or a sliced activation). Reuses the
    /// validated qmatvec_f32 dequant path (NOT a fast path — the correctness gate). The
    /// CudaView base+offset pointer is honored by the launch arg.
    pub fn qmatvec_view(
        &self,
        w: &CudaSlice<u8>,
        range: std::ops::Range<usize>,
        x: &cudarc::driver::CudaView<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        qtype: i32,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let f = self.func("qmatvec_f32");
        let wv = w.slice(range); // CudaView<u8>, offset honored
        let mut y = self.alloc_uninit::<f32>(m * out_f)?; // full-overwrite output: skip memset
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, m as u32, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi, qt, rb) =
            (in_f as i32, out_f as i32, m as i32, qtype, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&wv)
            .arg(x)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&qt)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(y)
    }

    /// STAGE-2 GROUPED DECODE (2026-07-04): one MoE layer's gate+up+SiLU for all `n_used` routed
    /// experts of ONE token in ONE launch (replaces 8x qmatvec(gate) + 8x qmatvec(up) + 8x
    /// silu_mul = 24 launches). `gp`/`up` are the 8 expert weight-block device pointers (SLRU
    /// cache slots — fixed-address, stable for the launch). Returns act [n_used, n_ff].
    /// BIT-IDENTICAL to the sequential chain: each dot reproduces qmatvec_f32's exact 256-thread
    /// reduction; the SiLU epilogue is silu_mul_f32's exact expression (see kernel header).
    #[allow(clippy::too_many_arguments)]
    /// dp4a q8 twins (MoE expert dp4a arc, 2026-07-06): same contract as the _f32 versions but
    /// consume a PRE-QUANTIZED q8_1 activation. FP-order differs from _f32 (int dot + warp tree)
    /// — the argmax/stream-identity battery arbitrates; MEMRA_MOE_Q8=0 restores f32.
    pub fn moe_gate_up_silu8_q8(
        &self,
        gp: WPtr8,
        up: WPtr8,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        in_f: usize,
        n_ff: usize,
        n_used: usize,
        qt_g: i32,
        qt_u: i32,
        rb_g: usize,
        rb_u: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let f = self.func("moe_gate_up_silu8_q8");
        let mut act = self.alloc_uninit::<f32>(n_used * n_ff)?;
        let cfg = LaunchConfig {
            grid_dim: (n_ff as u32, n_used as u32, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, nff, rbg, rbu) = (in_f as i32, n_ff as i32, rb_g as i64, rb_u as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&gp)
            .arg(&up)
            .arg(aq)
            .arg(ad)
            .arg(&mut act)
            .arg(&inf)
            .arg(&nff)
            .arg(&qt_g)
            .arg(&qt_u)
            .arg(&rbg)
            .arg(&rbu);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(act)
    }

    #[allow(clippy::too_many_arguments)]
    pub fn moe_down8_fma_q8(
        &self,
        dp: WPtr8,
        w: F32x8,
        aq2: &CudaSlice<i8>,
        ad2: &CudaSlice<f32>,
        dst: &mut cudarc::driver::CudaViewMut<f32>,
        in_f: usize,
        out_f: usize,
        n_used: usize,
        qt: i32,
        rb: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("moe_down8_fma_q8");
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, nu, rbi) = (in_f as i32, out_f as i32, n_used as i32, rb as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&dp)
            .arg(&w)
            .arg(aq2)
            .arg(ad2)
            .arg(dst)
            .arg(&inf)
            .arg(&outf)
            .arg(&nu)
            .arg(&qt)
            .arg(&rbi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// q8 sequential expert matvec (staged path twin of qmatvec_view for IQ3_S/IQ4_XS).
    pub fn qmatvec_expert_q8(
        &self,
        w: &CudaSlice<u8>,
        range: std::ops::Range<usize>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        qtype: i32,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let f = self.func("qmatvec_expert_q8");
        let wv = w.slice(range);
        let mut y = self.alloc_uninit::<f32>(m * out_f)?;
        const ROWS: u32 = 4; // MEMRA_MMVQ_ROWS
        let cfg = LaunchConfig {
            grid_dim: ((out_f as u32 + ROWS - 1) / ROWS, m as u32, 1),
            block_dim: (32, ROWS, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi, rbi) = (in_f as i32, out_f as i32, m as i32, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&wv)
            .arg(aq)
            .arg(ad)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&qtype)
            .arg(&rbi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(y)
    }

    pub fn moe_gate_up_silu8(
        &self,
        gp: WPtr8,
        up: WPtr8,
        x: &cudarc::driver::CudaView<f32>,
        in_f: usize,
        n_ff: usize,
        n_used: usize,
        qt_g: i32,
        qt_u: i32,
        rb_g: usize,
        rb_u: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let f = self.func("moe_gate_up_silu8_f32");
        let mut act = self.alloc_uninit::<f32>(n_used * n_ff)?; // fully overwritten
        let cfg = LaunchConfig {
            grid_dim: (n_ff as u32, n_used as u32, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, nff, rbg, rbu) = (in_f as i32, n_ff as i32, rb_g as i64, rb_u as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&gp)
            .arg(&up)
            .arg(x)
            .arg(&mut act)
            .arg(&inf)
            .arg(&nff)
            .arg(&qt_g)
            .arg(&qt_u)
            .arg(&rbg)
            .arg(&rbu);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(act)
    }

    /// STAGE-2 GROUPED DECODE: one MoE layer's down-proj + weighted accumulation for all `n_used`
    /// routed experts in ONE launch (replaces 8x qmatvec(down) + 8x axpy = 16 launches), writing
    /// the token's moe_out row DIRECTLY (`dst` is the zeroed row; the in-kernel slot-ordered
    /// __fmaf_rn chain starting at 0.0f reproduces the sequential axpy_f32 accumulation into the
    /// zeroed row bit-for-bit — the A2 byte-identity scheme at m=1).
    #[allow(clippy::too_many_arguments)]
    pub fn moe_down8_fma_into(
        &self,
        dp: WPtr8,
        w: F32x8,
        act: &CudaSlice<f32>,
        dst: &mut cudarc::driver::CudaViewMut<f32>,
        in_f: usize,
        out_f: usize,
        n_used: usize,
        qt: i32,
        rb: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("moe_down8_fma_f32");
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, nu, rbv) = (in_f as i32, out_f as i32, n_used as i32, rb as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&dp)
            .arg(&w)
            .arg(act)
            .arg(dst)
            .arg(&inf)
            .arg(&outf)
            .arg(&nu)
            .arg(&qt)
            .arg(&rbv);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// LAUNCH-STRUCTURE STAGE 3: device-dispatch twin of `moe_gate_up_silu8` for FULLY-RESIDENT
    /// layers. The expert ids come from the router kernel's DEVICE `sel` output (no DtoH) and the
    /// weight pointers from the per-layer device table `[3, n_expert]` of slot base addresses.
    /// BIT-IDENTICAL math (same grid/block/reduction; only the pointer/id source differs).
    #[allow(clippy::too_many_arguments)]
    /// dp4a q8 twin of the _dev pair (resident-experts arc).
    ///
    /// GEOMETRY VARIANTS (multirow/occupancy arc 2026-07-05): all outputs are BIT-IDENTICAL to
    /// the base one-warp-per-(row,slot) kernel (same expert_dot_g g-order + warp tree per row;
    /// down's FMA chain stays slot-ordered serial). Seams:
    ///   MEMRA_MOE_DEVQ8_GU   = 0(base) | 1 | 2 | 4 -> _r{1,2,4} multirow twin (RPW rows/warp)
    ///                       | s2 (gate/up warp split) | s2z (s2 + WPB rows packed per block)
    ///                       | gs4 (gate/up x low/high-group 4-warp split, nsb==64 only)
    ///                       | u64 (nsb==64 unrolled ILP twin, geometry unchanged)
    ///   MEMRA_MOE_DEVQ8_WPB  = warps per block for _r twins / z-rows for s2z (default 4)
    ///   MEMRA_MOE_DEVQ8_DOWN = auto(default: w8h2 when in_f==512 & n_used<=8 — measured +3.8%
    ///                       decode on 35B/rtx6000) | 0 (base one-warp serial-slot) | 1 | 2 | 4 ->
    ///                       _w8r{1,2,4} slot-parallel twin | h2 (half-warp dual-row, nsb==16
    ///                       only) | w8h2 (h2 x slot-parallel)
    #[allow(clippy::too_many_arguments)]
    /// MoE PREFILL pair-batch matvec: one launch covers all (token,expert) pairs for one proj.
    #[allow(clippy::too_many_arguments)]
    pub fn moe_pairs_matvec_q8(
        &self,
        table: &CudaSlice<u64>,
        proj: i32,
        pair_tok: &CudaSlice<i32>,
        pair_ex: &CudaSlice<i32>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        in_f: usize,
        out_f: usize,
        n_expert: usize,
        n_pairs: usize,
        qtype: i32,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let f = self.func("moe_pairs_matvec_q8");
        let mut y = self.alloc_uninit::<f32>(n_pairs * out_f)?;
        const ROWS: u32 = 4;
        let cfg = LaunchConfig {
            grid_dim: ((out_f as u32 + ROWS - 1) / ROWS, n_pairs as u32, 1),
            block_dim: (32, ROWS, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, ne, np, rbi) = (
            in_f as i32,
            out_f as i32,
            n_expert as i32,
            n_pairs as i32,
            row_bytes as i64,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(table)
            .arg(&proj)
            .arg(pair_tok)
            .arg(pair_ex)
            .arg(aq)
            .arg(ad)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&ne)
            .arg(&np)
            .arg(&qtype)
            .arg(&rbi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(y)
    }

    /// Expert-major pair matvec (weight-reuse across each expert's token group).
    #[allow(clippy::too_many_arguments)]
    pub fn moe_pairs_matvec_q8_em(
        &self,
        table: &CudaSlice<u64>,
        proj: i32,
        ex_ids: &CudaSlice<i32>,
        ex_off: &CudaSlice<i32>,
        ex_pairs: &CudaSlice<i32>,
        pair_tok: &CudaSlice<i32>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        in_f: usize,
        out_f: usize,
        n_expert: usize,
        n_active: usize,
        n_pairs: usize,
        qtype: i32,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let f = self.func("moe_pairs_matvec_q8_em");
        let mut y = self.alloc_uninit::<f32>(n_pairs * out_f)?;
        const ROWS: u32 = 4;
        let cfg = LaunchConfig {
            grid_dim: ((out_f as u32 + ROWS - 1) / ROWS, n_active as u32, 1),
            block_dim: (32, ROWS, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, ne, na, rbi) = (
            in_f as i32,
            out_f as i32,
            n_expert as i32,
            n_active as i32,
            row_bytes as i64,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(table)
            .arg(&proj)
            .arg(ex_ids)
            .arg(ex_off)
            .arg(ex_pairs)
            .arg(pair_tok)
            .arg(aq)
            .arg(ad)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&ne)
            .arg(&na)
            .arg(&qtype)
            .arg(&rbi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(y)
    }

    // Decode-once expert-major MMQ (rung 3). Same CSR inputs/geometry as _em; kernel dequants each
    // weight group once per (row,group) then dp4a's across the expert's token group.
    #[allow(clippy::too_many_arguments)]
    pub fn moe_pairs_matvec_q8_dec(
        &self,
        table: &CudaSlice<u64>,
        proj: i32,
        ex_ids: &CudaSlice<i32>,
        ex_off: &CudaSlice<i32>,
        ex_pairs: &CudaSlice<i32>,
        pair_tok: &CudaSlice<i32>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        in_f: usize,
        out_f: usize,
        n_expert: usize,
        n_active: usize,
        n_pairs: usize,
        qtype: i32,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let f = self.func("moe_pairs_matvec_q8_dec");
        let mut y = self.alloc_uninit::<f32>(n_pairs * out_f)?;
        const ROWS: u32 = 4;
        let cfg = LaunchConfig {
            grid_dim: ((out_f as u32 + ROWS - 1) / ROWS, n_active as u32, 1),
            block_dim: (32, ROWS, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, ne, na, rbi) = (
            in_f as i32,
            out_f as i32,
            n_expert as i32,
            n_active as i32,
            row_bytes as i64,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(table)
            .arg(&proj)
            .arg(ex_ids)
            .arg(ex_off)
            .arg(ex_pairs)
            .arg(pair_tok)
            .arg(aq)
            .arg(ad)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&ne)
            .arg(&na)
            .arg(&qtype)
            .arg(&rbi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(y)
    }

    pub fn moe_pairs_gelu_mul(
        &self,
        gate: &CudaSlice<f32>,
        up: &CudaSlice<f32>,
        n: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let f = self.func("moe_pairs_gelu_mul");
        let mut act = self.alloc_uninit::<f32>(n)?;
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let nl = n as i64;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate).arg(up).arg(&mut act).arg(&nl);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(act)
    }

    pub fn moe_pairs_silu_mul(
        &self,
        gate: &CudaSlice<f32>,
        up: &CudaSlice<f32>,
        n: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let f = self.func("moe_pairs_silu_mul");
        let mut act = self.alloc_uninit::<f32>(n)?;
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let nl = n as i64;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate).arg(up).arg(&mut act).arg(&nl);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(act)
    }

    #[allow(clippy::too_many_arguments)]
    pub fn moe_pairs_scatter(
        &self,
        y_down: &CudaSlice<f32>,
        pair_w: &CudaSlice<f32>,
        tok_pair_off: &CudaSlice<i32>,
        tok_pair_ids: &CudaSlice<i32>,
        moe_out: &mut CudaSlice<f32>,
        t: usize,
        n_embd: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("moe_pairs_scatter");
        let cfg = LaunchConfig {
            grid_dim: (((n_embd + 255) / 256) as u32, t as u32, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let ne = n_embd as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(y_down)
            .arg(pair_w)
            .arg(tok_pair_off)
            .arg(tok_pair_ids)
            .arg(moe_out)
            .arg(&ne);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// gemma4 GELU twin of moe_gate_up_silu8_dev_q8 (base geometry — slot-packed j8/j8r2
    /// twins probed 2026-08-01 g26 decode dig: bit-identical rows, -2.5%/-2.9% whole-model
    /// decode x3 interleaved -> refuted and killed; research/g26-decode-20260801/receipts.md).
    #[allow(clippy::too_many_arguments)]
    pub fn moe_gate_up_gelu8_dev_q8(
        &self,
        table: &CudaSlice<u64>,
        sel: &cudarc::driver::CudaView<i32>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        in_f: usize,
        n_ff: usize,
        n_used: usize,
        n_expert: usize,
        qt_g: i32,
        qt_u: i32,
        rb_g: usize,
        rb_u: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let mut act = self.alloc_uninit::<f32>(n_used * n_ff)?;
        let (inf, nff, ne, rbg, rbu) = (
            in_f as i32,
            n_ff as i32,
            n_expert as i32,
            rb_g as i64,
            rb_u as i64,
        );
        let f = self.func("moe_gate_up_gelu8_dev_q8");
        let cfg = LaunchConfig {
            grid_dim: (n_ff as u32, n_used as u32, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(table)
            .arg(sel)
            .arg(aq)
            .arg(ad)
            .arg(&mut act)
            .arg(&inf)
            .arg(&nff)
            .arg(&ne)
            .arg(&qt_g)
            .arg(&qt_u)
            .arg(&rbg)
            .arg(&rbu);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(act)
    }

    /// gemma4 GELU rows twin (verify): one launch over (n_ff, n_used, t).
    #[allow(clippy::too_many_arguments)]
    pub fn moe_gate_up_gelu8_dev_q8_rows(
        &self,
        table: &CudaSlice<u64>,
        sel: &CudaSlice<i32>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        t: usize,
        in_f: usize,
        n_ff: usize,
        n_used: usize,
        n_expert: usize,
        qt_g: i32,
        qt_u: i32,
        rb_g: usize,
        rb_u: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let mut act = self.alloc_uninit::<f32>(t * n_used * n_ff)?;
        let (inf, nff, ne, rbg, rbu, nu) = (
            in_f as i32,
            n_ff as i32,
            n_expert as i32,
            rb_g as i64,
            rb_u as i64,
            n_used as i32,
        );
        let f = self.func("moe_gate_up_gelu8_dev_q8_rows");
        let cfg = LaunchConfig {
            grid_dim: (n_ff as u32, n_used as u32, t as u32),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(table)
            .arg(sel)
            .arg(aq)
            .arg(ad)
            .arg(&mut act)
            .arg(&inf)
            .arg(&nff)
            .arg(&ne)
            .arg(&qt_g)
            .arg(&qt_u)
            .arg(&rbg)
            .arg(&rbu)
            .arg(&nu);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(act)
    }

    /// gemma4 GELU CSR twin (verify dedup: owner block serves every pair of its expert).
    #[allow(clippy::too_many_arguments)]
    pub fn moe_gate_up_gelu8_dev_q8_csr(
        &self,
        table: &CudaSlice<u64>,
        sel: &CudaSlice<i32>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        n_pairs: usize,
        in_f: usize,
        n_ff: usize,
        n_used: usize,
        n_expert: usize,
        qt_g: i32,
        qt_u: i32,
        rb_g: usize,
        rb_u: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let mut act = self.alloc_uninit::<f32>(n_pairs * n_ff)?;
        let (inf, nff, ne, rbg, rbu, nu, npi) = (
            in_f as i32,
            n_ff as i32,
            n_expert as i32,
            rb_g as i64,
            rb_u as i64,
            n_used as i32,
            n_pairs as i32,
        );
        let f = self.func("moe_gate_up_gelu8_dev_q8_csr");
        let cfg = LaunchConfig {
            grid_dim: (n_ff as u32, n_pairs as u32, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(table)
            .arg(sel)
            .arg(aq)
            .arg(ad)
            .arg(&mut act)
            .arg(&inf)
            .arg(&nff)
            .arg(&ne)
            .arg(&qt_g)
            .arg(&qt_u)
            .arg(&rbg)
            .arg(&rbu)
            .arg(&nu)
            .arg(&npi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(act)
    }

    /// gemma4 generic down rows twin (verify): one launch over (out_f, 1, t).
    #[allow(clippy::too_many_arguments)]
    pub fn moe_down8_fma_dev_q8_rows_g(
        &self,
        table: &CudaSlice<u64>,
        sel: &CudaSlice<i32>,
        w: &CudaSlice<f32>,
        aq2: &CudaSlice<i8>,
        ad2: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        t: usize,
        in_f: usize,
        out_f: usize,
        n_used: usize,
        n_expert: usize,
        qt: i32,
        rb: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let (inf, outf, nu, ne, rbi) = (
            in_f as i32,
            out_f as i32,
            n_used as i32,
            n_expert as i32,
            rb as i64,
        );
        // Exact Step-3.7 B=1 shape: expose the eight independent slot dots as
        // eight warps, then replay the original slot-ordered FMA chain. Every
        // other shape retains the generic one-warp rows kernel.
        let step_b1_w8 = t == 1 && in_f == 1280 && out_f == 4096 && n_used == 8 && qt == QT_IQ4_XS;
        let f = self.func(if step_b1_w8 {
            "moe_down8_fma_dev_q8_rows_w8"
        } else {
            "moe_down8_fma_dev_q8_rows_g"
        });
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, 1, t as u32),
            block_dim: (32, if step_b1_w8 { 8 } else { 1 }, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(table)
            .arg(sel)
            .arg(w)
            .arg(aq2)
            .arg(ad2)
            .arg(dst)
            .arg(&inf)
            .arg(&outf)
            .arg(&nu)
            .arg(&ne)
            .arg(&qt)
            .arg(&rbi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// rp_q4 microprobe (2026-07-10 verify-trunk lever): b4 GGUF-block layout vs the Q4_0
    /// split-plane twin on the wq-class shape. Returns (blk_us, rp_us) after asserting bitwise
    /// identity. Bench-only surface (rp_q4_probe bin); no production dispatch reads this.
    pub fn rp_probe_q4(&self, m: usize) -> Result<(f64, f64), Box<dyn std::error::Error>> {
        let (out_f, in_f) = (2048usize, 2816usize);
        let nblk = in_f / 32;
        let mut seed = 0x9E3779B97F4A7C15u64;
        let mut rng = move || {
            seed = seed
                .wrapping_mul(6364136223846793005)
                .wrapping_add(1442695040888963407);
            (seed >> 33) as u8
        };
        let mut w = vec![0u8; out_f * nblk * 18];
        for b in w.iter_mut() {
            *b = rng();
        }
        for r in 0..out_f {
            for g in 0..nblk {
                let off = (r * nblk + g) * 18;
                w[off] = 0x00;
                w[off + 1] = 0x2C; // sane half d
            }
        }
        let qplane = out_f * nblk * 16;
        let mut wrp = vec![0u8; w.len()];
        for r in 0..out_f {
            for g in 0..nblk {
                let src = &w[(r * nblk + g) * 18..(r * nblk + g) * 18 + 18];
                wrp[qplane + (r * nblk + g) * 2..qplane + (r * nblk + g) * 2 + 2]
                    .copy_from_slice(&src[0..2]);
                wrp[(r * nblk + g) * 16..(r * nblk + g) * 16 + 16].copy_from_slice(&src[2..18]);
            }
        }
        let w_d = self.htod_bytes(&w)?;
        let wrp_d = self.htod_bytes(&wrp)?;
        let mut aq = vec![0i8; m * in_f];
        for v in aq.iter_mut() {
            *v = rng() as i8;
        }
        let aq_d = self.htod_i8(&aq)?;
        let ad_d = self.htod(&vec![0.03125f32; m * nblk])?;
        let mut y0 = self.alloc_uninit::<f32>(m * out_f)?;
        let mut y1 = self.alloc_uninit::<f32>(m * out_f)?;
        const RPB: u32 = 4;
        let cfg = LaunchConfig {
            grid_dim: ((out_f as u32).div_ceil(RPB), 1, 1),
            block_dim: (32, RPB, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi) = (in_f as i32, out_f as i32, m as i32);
        let (rb, qp) = ((nblk * 18) as i64, qplane as i64);
        let fb = self.func("qmatvec_q4_0_mmvq_b4");
        let fr = self.func("qmatvec_q4_0_mmvq_b4_rp");
        {
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&fb);
            b.arg(&w_d)
                .arg(&aq_d)
                .arg(&ad_d)
                .arg(&mut y0)
                .arg(&inf)
                .arg(&outf)
                .arg(&mi)
                .arg(&rb);
            unsafe {
                b.launch(cfg)?;
            }
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&fr);
            b.arg(&wrp_d)
                .arg(&aq_d)
                .arg(&ad_d)
                .arg(&mut y1)
                .arg(&inf)
                .arg(&outf)
                .arg(&mi)
                .arg(&qp);
            unsafe {
                b.launch(cfg)?;
            }
        }
        self.gpu.stream().synchronize()?;
        let (h0, h1) = (self.dtoh(&y0)?, self.dtoh(&y1)?);
        let nd = h0
            .iter()
            .zip(&h1)
            .filter(|(a, b)| a.to_bits() != b.to_bits())
            .count();
        if nd != 0 {
            return Err(format!("rp twin not bitwise: {nd}/{} diffs", h0.len()).into());
        }
        let mut time = |rp: bool| -> Result<f64, Box<dyn std::error::Error>> {
            self.gpu.stream().synchronize()?;
            let t0 = std::time::Instant::now();
            for _ in 0..500 {
                if rp {
                    let __s_b = self.gpu.stream();
                    let mut b = __s_b.launch_builder(&fr);
                    b.arg(&wrp_d)
                        .arg(&aq_d)
                        .arg(&ad_d)
                        .arg(&mut y1)
                        .arg(&inf)
                        .arg(&outf)
                        .arg(&mi)
                        .arg(&qp);
                    unsafe {
                        b.launch(cfg)?;
                    }
                } else {
                    let __s_b = self.gpu.stream();
                    let mut b = __s_b.launch_builder(&fb);
                    b.arg(&w_d)
                        .arg(&aq_d)
                        .arg(&ad_d)
                        .arg(&mut y0)
                        .arg(&inf)
                        .arg(&outf)
                        .arg(&mi)
                        .arg(&rb);
                    unsafe {
                        b.launch(cfg)?;
                    }
                }
            }
            self.gpu.stream().synchronize()?;
            Ok(t0.elapsed().as_secs_f64() * 1e6 / 500.0)
        };
        let _ = time(false)?;
        let _ = time(true)?; // warm
        Ok((time(false)?, time(true)?))
    }

    /// Build the Q4_0 split-plane decode mirror for a 2D Quant tensor (device-side permutation,
    /// q4_0_split_rp_build). Raw bytes stay resident (prefill/gemm/Stage-A); the m<=8 decode
    /// dispatch prefers the mirror (_rp twins). No-op unless (Q4_0, 2D, mirror absent).
    /// VRAM cost == the tensor's weight size. MEMRA_Q4RP=0 disables at the call sites.
    pub fn build_q4_rp4(
        &self,
        t: &mut crate::model::GpuTensor,
    ) -> Result<(), Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        let GpuTensor::Quant {
            bytes,
            qtype,
            row_bytes,
            ne,
            rp4,
            ..
        } = t
        else {
            return Ok(());
        };
        if *qtype != QT_Q4_0 || rp4.is_some() || ne.len() != 2 {
            return Ok(());
        }
        let (in_f, out_f) = (ne[0] as usize, ne[1] as usize);
        if in_f % 32 != 0 || *row_bytes != (in_f / 32) * 18 {
            return Ok(());
        }
        let nblk = in_f / 32;
        let mut dst = self.alloc_uninit::<u8>(out_f * nblk * 18)?;
        let f = self.func("q4_0_split_rp_build");
        let n = (out_f * nblk) as i32;
        let cfg = LaunchConfig {
            grid_dim: (((out_f * nblk) as u32).div_ceil(256), 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (of, nb) = (out_f as i32, nblk as i32);
        let _ = n;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&*bytes).arg(&mut dst).arg(&of).arg(&nb);
        unsafe {
            b.launch(cfg)?;
        }
        *rp4 = Some(dst);
        Ok(())
    }

    /// Q8_0 twin of `build_q4_rp4` (H100 coalescing fix, 2026-07-26 ncu: GGUF 34B-stride
    /// weight loads hold Max Bandwidth at 41-46%; the split mirror makes them aligned 16B
    /// ldcs). Raw bytes stay resident (prefill GEMM/MMQ/fused m=1 launches read GGUF layout);
    /// the mmvq/batched decode arms prefer the mirror via `rp4`. Bit-identical outputs.
    pub fn build_q8_rp4(
        &self,
        t: &mut crate::model::GpuTensor,
    ) -> Result<(), Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        let GpuTensor::Quant {
            bytes,
            qtype,
            row_bytes,
            ne,
            rp4,
            ..
        } = t
        else {
            return Ok(());
        };
        if *qtype != QT_Q8_0 || rp4.is_some() || ne.len() != 2 {
            return Ok(());
        }
        let (in_f, out_f) = (ne[0] as usize, ne[1] as usize);
        if in_f % 32 != 0 || *row_bytes != (in_f / 32) * 34 {
            return Ok(());
        }
        *rp4 = Some(self.build_q8_rp4_raw(bytes, in_f, out_f)?);
        Ok(())
    }

    /// Raw rp-mirror build for gates/benches: split GGUF Q8_0 bytes into the qplane+dplane
    /// mirror without a GpuTensor (same kernel the loader path above uses).
    pub fn build_q8_rp4_raw(
        &self,
        bytes: &CudaSlice<u8>,
        in_f: usize,
        out_f: usize,
    ) -> Result<CudaSlice<u8>, Box<dyn std::error::Error>> {
        assert!(in_f % 32 == 0);
        let nblk = in_f / 32;
        let mut dst = self.alloc_uninit::<u8>(out_f * nblk * 34)?;
        let f = self.func("q8_0_split_rp_build");
        let cfg = LaunchConfig {
            grid_dim: (((out_f * nblk) as u32).div_ceil(256), 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (of, nb) = (out_f as i32, nblk as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&*bytes).arg(&mut dst).arg(&of).arg(&nb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(dst)
    }

    /// K-quant twins of `build_q8_rp4` (H100 K-quant coalescing fix, 2026-08-01 ncu on the
    /// q27 Q4_K_M decode: q4_K mmvq DRAM 41-54% with 65% excessive sectors, q6_K 40% with
    /// 78% — the 144B/210B superblock strides land every 4B weight load off-sector). The
    /// mirror re-packs each tensor into planes (q4_K: qs ++ 16B meta; q6_K: ql ++ qh ++
    /// scales ++ d — same total bytes) so every quant fetch is an aligned 16B ldcs. Raw
    /// bytes stay resident (prefill GEMM/dequant/Stage-A read GGUF layout); the mmvq/batched
    /// decode arms prefer the mirror via `rp4`. Bit-identical outputs.
    pub fn build_q4k_rp4(
        &self,
        t: &mut crate::model::GpuTensor,
    ) -> Result<(), Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        let GpuTensor::Quant {
            bytes,
            qtype,
            row_bytes,
            ne,
            rp4,
            ..
        } = t
        else {
            return Ok(());
        };
        if *qtype != QT_Q4_K || rp4.is_some() || ne.len() != 2 {
            return Ok(());
        }
        let (in_f, out_f) = (ne[0] as usize, ne[1] as usize);
        if in_f % 256 != 0 || *row_bytes != (in_f / 256) * 144 {
            return Ok(());
        }
        *rp4 = Some(self.build_kq_rp4_raw(bytes, in_f, out_f, QT_Q4_K)?);
        Ok(())
    }

    pub fn build_q6k_rp4(
        &self,
        t: &mut crate::model::GpuTensor,
    ) -> Result<(), Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        let GpuTensor::Quant {
            bytes,
            qtype,
            row_bytes,
            ne,
            rp4,
            ..
        } = t
        else {
            return Ok(());
        };
        if *qtype != QT_Q6_K || rp4.is_some() || ne.len() != 2 {
            return Ok(());
        }
        let (in_f, out_f) = (ne[0] as usize, ne[1] as usize);
        if in_f % 256 != 0 || *row_bytes != (in_f / 256) * 210 {
            return Ok(());
        }
        *rp4 = Some(self.build_kq_rp4_raw(bytes, in_f, out_f, QT_Q6_K)?);
        Ok(())
    }

    /// Raw K-quant rp-mirror build for gates/benches (same kernels the loader path uses).
    pub fn build_kq_rp4_raw(
        &self,
        bytes: &CudaSlice<u8>,
        in_f: usize,
        out_f: usize,
        qtype: i32,
    ) -> Result<CudaSlice<u8>, Box<dyn std::error::Error>> {
        assert!(in_f % 256 == 0);
        let nsbk = in_f / 256;
        let (sb_bytes, kname) = match qtype {
            QT_Q4_K => (144usize, "q4_K_split_rp_build"),
            QT_Q6_K => (210usize, "q6_K_split_rp_build"),
            _ => return Err(format!("build_kq_rp4_raw: qtype {qtype} has no rp mirror").into()),
        };
        let mut dst = self.alloc_uninit::<u8>(out_f * nsbk * sb_bytes)?;
        let f = self.func(kname);
        let cfg = LaunchConfig {
            grid_dim: (((out_f * nsbk) as u32).div_ceil(256), 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (of, nb) = (out_f as i32, nsbk as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&*bytes).arg(&mut dst).arg(&of).arg(&nb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(dst)
    }

    /// MEMRA_KQRP seam: the K-quant (q4_K/q6_K) split-plane decode mirrors at model load.
    /// Default follows the Q8RP convention — ON on the Hopper lane (80GB pays the mirror
    /// VRAM), OFF elsewhere (a 24GB card cannot hold model + mirror + KV for the big trunks).
    pub fn kqrp_enabled() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| match std::env::var("MEMRA_KQRP").as_deref() {
            Ok("0") => false,
            Ok(_) => true,
            Err(_) => cfg!(memra_hopper_mma),
        })
    }

    /// IN-PLACE split-plane swap (the 31B dense arc): build the split layout and REPLACE the
    /// GGUF bytes (zero extra steady-state VRAM — the transient peak is one tensor's size).
    /// The tensor's `rp` flag then routes every consumer (mmvq/batched `_rp` twins, the
    /// `qmatvec_gemm_q4_0_rp` prefill kernel). Callers gate on the fast path being active —
    /// the Stage-A f32 oracle (`MEMRA_FAST=0`) reads GGUF layout and must never see a swap.
    pub fn build_q4_rp_swap(
        &self,
        t: &mut crate::model::GpuTensor,
    ) -> Result<bool, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        // QTYPE GUARD IN THE SWAP ITSELF (gemma4 NVFP4mix prefill-NaN, 2026-08-17):
        // `rp4` is a SHARED field — the Q8RP walk parks Q8_0 MIRRORS there, and this
        // fn used to `take()` whatever it found. On a Q8_0-carrying gemma4-dense trunk
        // the swap hijacked those mirrors: `bytes` became split-plane in place, the
        // m<=16 `_rp` dispatch read them correctly (masking the corruption from every
        // decode pin), and every GGUF-layout prefill consumer (MMQ + GEMM) read the
        // fp16 d-plane as weights -> layer-0 NaN, <pad>-spam serving. Only a tensor
        // this fn's OWN builder serves may ever be swapped; everything else refuses
        // here, regardless of walk ordering.
        if !matches!(t, GpuTensor::Quant { qtype, .. } if *qtype == QT_Q4_0) {
            return Ok(false);
        }
        self.build_q4_rp4(t)?;
        self.gpu.stream().synchronize()?; // build kernel reads the GGUF bytes — drain BEFORE dropping them
        let GpuTensor::Quant { bytes, rp4, rp, .. } = t else {
            return Ok(false);
        };
        match rp4.take() {
            Some(split) => {
                *bytes = split; // the GGUF-layout buffer drops here
                *rp = true;
                Ok(true)
            }
            None => Ok(false),
        }
    }

    /// MEMRA_Q4RP seam (default ON): the Q4_0 split-plane decode mirror at model load.
    pub fn q4rp_enabled() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| {
            std::env::var("MEMRA_Q4RP")
                .map(|v| v != "0")
                .unwrap_or(true)
        })
    }

    /// gemma4-E4B: dense [t][row_elems] gather of layer il's rows from the strided prologue
    /// buffer ([t][n_layer][n_epl]; off = il*n_epl, stride = n_layer*n_epl).
    pub fn copy_rows_strided(
        &self,
        src: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        row_elems: usize,
        n_rows: usize,
        src_stride: usize,
        src_off: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("copy_rows_strided_f32");
        let cfg = LaunchConfig {
            grid_dim: (((row_elems as u32 + 255) / 256).max(1), n_rows as u32, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (re, nr) = (row_elems as i32, n_rows as i32);
        let (st, off) = (src_stride as i64, src_off as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(src)
            .arg(&mut *dst)
            .arg(&re)
            .arg(&nr)
            .arg(&st)
            .arg(&off);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Place dense `[row][row_elems]` source rows into one column range of a strided destination.
    ///
    /// This is a byte-preserving layout operation. It exists so multi-GPU collectives can move
    /// one dense shard per rank and reconstruct the canonical token-major matrix without issuing
    /// one peer copy per token.
    pub fn place_rows_strided(
        &self,
        src: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        row_elems: usize,
        n_rows: usize,
        dst_stride: usize,
        dst_off: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if row_elems == 0 || n_rows == 0 {
            return Err("strided row placement requires nonzero rows and row width".into());
        }
        let src_len = n_rows
            .checked_mul(row_elems)
            .ok_or("strided row placement source size overflow")?;
        let dst_len = n_rows
            .checked_sub(1)
            .and_then(|rows| rows.checked_mul(dst_stride))
            .and_then(|base| base.checked_add(dst_off))
            .and_then(|base| base.checked_add(row_elems))
            .ok_or("strided row placement destination size overflow")?;
        let row_end = dst_off
            .checked_add(row_elems)
            .ok_or("strided row placement row size overflow")?;
        if src.len() < src_len || dst.len() < dst_len || row_end > dst_stride {
            return Err(format!(
                "strided row placement geometry mismatch: src={} need_src={src_len} \
                 dst={} need_dst={dst_len} row_elems={row_elems} rows={n_rows} \
                 dst_stride={dst_stride} dst_off={dst_off}",
                src.len(),
                dst.len(),
            )
            .into());
        }
        if row_elems > i32::MAX as usize || n_rows > i32::MAX as usize {
            return Err("strided row placement exceeds CUDA kernel geometry".into());
        }
        let f = self.func("place_rows_strided_f32");
        let cfg = LaunchConfig {
            grid_dim: (((row_elems as u32 + 255) / 256).max(1), n_rows as u32, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (re, nr) = (row_elems as i32, n_rows as i32);
        let (st, off) = (dst_stride as i64, dst_off as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(src)
            .arg(&mut *dst)
            .arg(&re)
            .arg(&nr)
            .arg(&st)
            .arg(&off);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Async device u32 store (value rides the kernel ARG — no host-memory transfer/sync).
    pub fn u32_set_k(
        &self,
        dst: &mut CudaSlice<u32>,
        v: u32,
        idx: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("u32_set_k");
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (1, 1, 1),
            shared_mem_bytes: 0,
        };
        let ii = idx as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(dst).arg(&v).arg(&ii);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// counter += v (device-slot append advance; the +1 twin is `inc_seqlen`).
    pub fn i32_add_k(
        &self,
        d: &mut CudaSlice<i32>,
        v: i32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("i32_add_k");
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(d).arg(&v);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// pos rows from a device counter: dst[i] = ctr[0] + i (verify-stream rope positions).
    pub fn i32_iota_from(
        &self,
        ctr: &CudaSlice<i32>,
        dst: &mut CudaSlice<i32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("i32_iota_from");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let ni = n as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(ctr).arg(dst).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// In-place trim-id translate: buf[idx] = map[buf[idx]] (FR-Spec d2t, async single-slot).
    pub fn u32_map_k(
        &self,
        buf: &mut CudaSlice<u32>,
        map: &CudaSlice<u32>,
        idx: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("u32_map_k");
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (1, 1, 1),
            shared_mem_bytes: 0,
        };
        let ii = idx as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(buf).arg(map).arg(&ii);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Pack a[off..off+n1] ++ b[0..n2] into one buffer (single dtoh follows).
    #[allow(clippy::too_many_arguments)]
    pub fn u32_pack2(
        &self,
        a: &CudaSlice<u32>,
        off_a: usize,
        n1: usize,
        b_in: &CudaSlice<u32>,
        n2: usize,
        out: &mut CudaSlice<u32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("u32_pack2");
        let cfg = LaunchConfig::for_num_elems((n1 + n2) as u32);
        let (oa, i1, i2) = (off_a as i32, n1 as i32, n2 as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(a).arg(&oa).arg(&i1).arg(b_in).arg(&i2).arg(out);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// gemma4 R3 device fold: w[i] *= s[sel[i]] over the router's [n] (sel, w) pair.
    pub fn moe_w_exscale(
        &self,
        w: &mut CudaSlice<f32>,
        sel: &CudaSlice<i32>,
        s: &CudaSlice<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("moe_w_exscale");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let ni = n as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(w).arg(sel).arg(s).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Down-projection macro fold: w[i] *= macros[2*n_expert + sel[i]] on the device router
    /// weights (one launch per MoE layer, only for macro-carrying artifacts — see MoeWeights).
    pub fn moe_w_scale_by_expert(
        &self,
        w: &mut CudaSlice<f32>,
        sel: &CudaSlice<i32>,
        macros: &CudaSlice<f32>,
        n_expert: usize,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("moe_w_scale_by_expert");
        let cfg = LaunchConfig {
            grid_dim: (n.div_ceil(64) as u32, 1, 1),
            block_dim: (64, 1, 1),
            shared_mem_bytes: 0,
        };
        let (ne, nn) = (n_expert as i32, n as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(w).arg(sel).arg(macros).arg(&ne).arg(&nn);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn moe_gate_up_silu8_dev_q8(
        &self,
        table: &CudaSlice<u64>,
        sel: &cudarc::driver::CudaView<i32>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        in_f: usize,
        n_ff: usize,
        n_used: usize,
        n_expert: usize,
        qt_g: i32,
        qt_u: i32,
        rb_g: usize,
        rb_u: usize,
        macros: &CudaSlice<f32>,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        static GU: std::sync::OnceLock<(String, u32)> = std::sync::OnceLock::new();
        let (mode, wpb) = GU.get_or_init(|| {
            let mode = std::env::var("MEMRA_MOE_DEVQ8_GU").unwrap_or_default();
            let wpb = std::env::var("MEMRA_MOE_DEVQ8_WPB")
                .ok()
                .and_then(|v| v.parse().ok())
                .unwrap_or(4u32)
                .clamp(1, 16);
            (mode, wpb)
        });
        let (mode, wpb) = (mode.as_str(), *wpb);
        let mut act = self.alloc_uninit::<f32>(n_used * n_ff)?;
        let (inf, nff, ne, rbg, rbu) = (
            in_f as i32,
            n_ff as i32,
            n_expert as i32,
            rb_g as i64,
            rb_u as i64,
        );
        let (f, cfg) = match mode {
            "1" | "2" | "4" => {
                let rpw: u32 = mode.parse().unwrap();
                let f = self.func(match rpw {
                    1 => "moe_gate_up_silu8_dev_q8_r1",
                    2 => "moe_gate_up_silu8_dev_q8_r2",
                    _ => "moe_gate_up_silu8_dev_q8_r4",
                });
                let rows_per_block = (rpw * wpb) as usize;
                let gx = n_ff.div_ceil(rows_per_block) as u32;
                (
                    f,
                    LaunchConfig {
                        grid_dim: (gx, n_used as u32, 1),
                        block_dim: (32, wpb, 1),
                        shared_mem_bytes: 0,
                    },
                )
            }
            "j8" if n_used <= 32 => (
                self.func("moe_gate_up_silu8_dev_q8_j8"),
                LaunchConfig {
                    grid_dim: (n_ff as u32, 1, 1),
                    block_dim: (32, n_used as u32, 1),
                    shared_mem_bytes: 0,
                },
            ),
            // SMEM-GRID twins (IQ3_S 2KB grid copied to shared, static smem — bit-identical dots)
            "vsm2" => {
                let f = self.func("moe_gate_up_silu8_dev_q8_vsm2");
                let sh = (rb_g + rb_u) as u32;
                use cudarc::driver::sys::CUfunction_attribute_enum as A;
                f.set_attribute(
                    A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                    sh as i32,
                )?;
                (
                    f,
                    LaunchConfig {
                        grid_dim: (n_ff as u32, n_used as u32, 1),
                        block_dim: (32, 1, 1),
                        shared_mem_bytes: sh,
                    },
                )
            }
            "vsm" => {
                let f = self.func("moe_gate_up_silu8_dev_q8_vsm");
                let sh = (rb_g + rb_u) as u32;
                use cudarc::driver::sys::CUfunction_attribute_enum as A;
                f.set_attribute(
                    A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                    sh as i32,
                )?;
                (
                    f,
                    LaunchConfig {
                        grid_dim: (n_ff as u32, n_used as u32, 1),
                        block_dim: (32, 1, 1),
                        shared_mem_bytes: sh,
                    },
                )
            }
            "sg" => (
                self.func("moe_gate_up_silu8_dev_q8_sg"),
                LaunchConfig {
                    grid_dim: (n_ff as u32, n_used as u32, 1),
                    block_dim: (32, 1, 1),
                    shared_mem_bytes: 0,
                },
            ),
            "j8sg" if n_used <= 32 => (
                self.func("moe_gate_up_silu8_dev_q8_j8sg"),
                LaunchConfig {
                    grid_dim: (n_ff as u32, 1, 1),
                    block_dim: (32, n_used as u32, 1),
                    shared_mem_bytes: 0,
                },
            ),
            "u64" if in_f == 2048 => (
                self.func("moe_gate_up_silu8_dev_q8_u64"),
                LaunchConfig {
                    grid_dim: (n_ff as u32, n_used as u32, 1),
                    block_dim: (32, 1, 1),
                    shared_mem_bytes: 0,
                },
            ),
            "gs4" if in_f == 2048 => (
                self.func("moe_gate_up_silu8_dev_q8_gs4"),
                LaunchConfig {
                    grid_dim: (n_ff as u32, n_used as u32, 1),
                    block_dim: (32, 4, 1),
                    shared_mem_bytes: 0,
                },
            ),
            // _v twin (down8 lane 2026-07-08): wide-load IQ4_XS dot, base geometry, bit-identical.
            "v" | "" => (
                self.func("moe_gate_up_silu8_dev_q8_v"),
                LaunchConfig {
                    grid_dim: (n_ff as u32, n_used as u32, 1),
                    block_dim: (32, 1, 1),
                    shared_mem_bytes: 0,
                },
            ),
            "s2" => (
                self.func("moe_gate_up_silu8_dev_q8_s2"),
                LaunchConfig {
                    grid_dim: (n_ff as u32, n_used as u32, 1),
                    block_dim: (32, 2, 1),
                    shared_mem_bytes: 0,
                },
            ),
            "s2z" => {
                let rz = wpb.min(16); // s2z smem tile is [16][2]
                (
                    self.func("moe_gate_up_silu8_dev_q8_s2z"),
                    LaunchConfig {
                        grid_dim: (n_ff.div_ceil(rz as usize) as u32, n_used as u32, 1),
                        block_dim: (32, 2, rz),
                        shared_mem_bytes: 0,
                    },
                )
            }
            _ => (
                self.func("moe_gate_up_silu8_dev_q8"),
                LaunchConfig {
                    grid_dim: (n_ff as u32, n_used as u32, 1),
                    block_dim: (32, 1, 1),
                    shared_mem_bytes: 0,
                },
            ),
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(table)
            .arg(sel)
            .arg(aq)
            .arg(ad)
            .arg(&mut act)
            .arg(&inf)
            .arg(&nff)
            .arg(&ne)
            .arg(&qt_g)
            .arg(&qt_u)
            .arg(&rbg)
            .arg(&rbu)
            .arg(macros);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(act)
    }

    #[allow(clippy::too_many_arguments)]
    pub fn moe_down8_fma_dev_q8(
        &self,
        table: &CudaSlice<u64>,
        sel: &cudarc::driver::CudaView<i32>,
        w: &cudarc::driver::CudaView<f32>,
        aq2: &CudaSlice<i8>,
        ad2: &CudaSlice<f32>,
        dst: &mut cudarc::driver::CudaViewMut<f32>,
        in_f: usize,
        out_f: usize,
        n_used: usize,
        n_expert: usize,
        qt: i32,
        rb: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        static DOWN: std::sync::OnceLock<String> = std::sync::OnceLock::new();
        let mode = DOWN.get_or_init(|| std::env::var("MEMRA_MOE_DEVQ8_DOWN").unwrap_or_default());
        let (inf, outf, nu, ne, rbi) = (
            in_f as i32,
            out_f as i32,
            n_used as i32,
            n_expert as i32,
            rb as i64,
        );
        // the w8 twins' smem tile is [RPW][8] — n_used must fit the 8-slot tile;
        // the h2 twins are nsb==16 (in_f==512) shape-gated.
        let (f, cfg) = match mode.as_str() {
            m @ ("1" | "2" | "4") if n_used <= 8 => {
                let rpw: usize = m.parse().unwrap();
                let f = self.func(match rpw {
                    1 => "moe_down8_fma_dev_q8_w8r1",
                    2 => "moe_down8_fma_dev_q8_w8r2",
                    _ => "moe_down8_fma_dev_q8_w8r4",
                });
                (
                    f,
                    LaunchConfig {
                        grid_dim: (out_f.div_ceil(rpw) as u32, 1, 1),
                        block_dim: (32, n_used as u32, 1),
                        shared_mem_bytes: 0,
                    },
                )
            }
            "h2" if in_f == 512 => (
                self.func("moe_down8_fma_dev_q8_h2"),
                LaunchConfig {
                    grid_dim: (out_f.div_ceil(2) as u32, 1, 1),
                    block_dim: (32, 1, 1),
                    shared_mem_bytes: 0,
                },
            ),
            // "" = AUTO gemma shape (in_f==704): w8r2 measured +1 tok/s vs base (sweep
            // 1/2/4 -> 133.6/134.2/133.6, 2026-07-10); slot-ordered chain preserved.
            "" if in_f == 704 && n_used <= 8 => (
                self.func("moe_down8_fma_dev_q8_w8r2"),
                LaunchConfig {
                    grid_dim: (out_f.div_ceil(2) as u32, 1, 1),
                    block_dim: (32, n_used as u32, 1),
                    shared_mem_bytes: 0,
                },
            ),
            // "" = AUTO: the measured winner for the 35B expert shape (arc 2026-07-05, +3.8%);
            // any shape the h2 kernels can't take (nsb!=16 / n_used>8) falls to base via `_`.
            // _v twins (down8 lane 2026-07-08): wide-load IQ4_XS dot, bit-identical outputs.
            "w8h2v" | "" if in_f == 512 && n_used <= 8 => (
                self.func("moe_down8_fma_dev_q8_w8h2v"),
                LaunchConfig {
                    grid_dim: (out_f.div_ceil(2) as u32, 1, 1),
                    block_dim: (32, n_used as u32, 1),
                    shared_mem_bytes: 0,
                },
            ),
            "w8h2r2v" if in_f == 512 && n_used <= 8 => (
                self.func("moe_down8_fma_dev_q8_w8h2r2v"),
                LaunchConfig {
                    grid_dim: (out_f.div_ceil(4) as u32, 1, 1),
                    block_dim: (32, n_used as u32, 1),
                    shared_mem_bytes: 0,
                },
            ),
            "w8h2r2" if in_f == 512 && n_used <= 8 => (
                self.func("moe_down8_fma_dev_q8_w8h2r2"),
                LaunchConfig {
                    grid_dim: (out_f.div_ceil(4) as u32, 1, 1),
                    block_dim: (32, n_used as u32, 1),
                    shared_mem_bytes: 0,
                },
            ),
            "w8h2" if in_f == 512 && n_used <= 8 => (
                self.func("moe_down8_fma_dev_q8_w8h2"),
                LaunchConfig {
                    grid_dim: (out_f.div_ceil(2) as u32, 1, 1),
                    block_dim: (32, n_used as u32, 1),
                    shared_mem_bytes: 0,
                },
            ),
            _ => (
                self.func("moe_down8_fma_dev_q8"),
                LaunchConfig {
                    grid_dim: (out_f as u32, 1, 1),
                    block_dim: (32, 1, 1),
                    shared_mem_bytes: 0,
                },
            ),
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(table)
            .arg(sel)
            .arg(w)
            .arg(aq2)
            .arg(ad2)
            .arg(dst)
            .arg(&inf)
            .arg(&outf)
            .arg(&nu)
            .arg(&ne)
            .arg(&qt)
            .arg(&rbi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// SMALL-M VERIFY rows twin (MEMRA_SPEC_M2, lane/spec-m2): ONE launch covers all `t` tokens
    /// of the spec verify's MoE dev gate/up (grid.z = token) — the _v geometry per token, with
    /// tok-offset sel/aq/ad/act pointers matching the serial loop's slices. BIT-IDENTICAL per
    /// token (see the kernel header). aq/ad are the BATCHED z-quantize ([t, in_f] rows —
    /// quantize_q8_1's per-32-block program is row-independent, so batched rows == the serial
    /// loop's per-token quantize_q8_1_view bytes). Returns act [t, n_used, n_ff].
    #[allow(clippy::too_many_arguments)]
    pub fn moe_gate_up_silu8_dev_q8_rows(
        &self,
        table: &CudaSlice<u64>,
        sel: &CudaSlice<i32>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        t: usize,
        in_f: usize,
        n_ff: usize,
        n_used: usize,
        n_expert: usize,
        qt_g: i32,
        qt_u: i32,
        rb_g: usize,
        rb_u: usize,
        macros: &CudaSlice<f32>,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let f = self.func("moe_gate_up_silu8_dev_q8_v_rows");
        let mut act = self.alloc_uninit::<f32>(t * n_used * n_ff)?;
        let cfg = LaunchConfig {
            grid_dim: (n_ff as u32, n_used as u32, t as u32),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, nff, ne, nu, rbg, rbu) = (
            in_f as i32,
            n_ff as i32,
            n_expert as i32,
            n_used as i32,
            rb_g as i64,
            rb_u as i64,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(table)
            .arg(sel)
            .arg(aq)
            .arg(ad)
            .arg(&mut act)
            .arg(&inf)
            .arg(&nff)
            .arg(&ne)
            .arg(&qt_g)
            .arg(&qt_u)
            .arg(&rbg)
            .arg(&rbu)
            .arg(&nu)
            .arg(macros);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(act)
    }

    /// SMALL-M VERIFY rows twin of the down proj: w8h2v geometry per token on a grid.z token
    /// axis. Caller gates the w8h2v shape contract (in_f == 512, n_used <= 8) — same gate as
    /// the AUTO dispatch in `moe_down8_fma_dev_q8`. aq2/ad2 = batched act quantize
    /// ([t*n_used, in_f] rows). dst rows are FULLY overwritten per token.
    #[allow(clippy::too_many_arguments)]
    pub fn moe_down8_fma_dev_q8_rows(
        &self,
        table: &CudaSlice<u64>,
        sel: &CudaSlice<i32>,
        w: &CudaSlice<f32>,
        aq2: &CudaSlice<i8>,
        ad2: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        t: usize,
        in_f: usize,
        out_f: usize,
        n_used: usize,
        n_expert: usize,
        qt: i32,
        rb: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        assert!(
            in_f == 512 && n_used <= 8,
            "down rows twin is w8h2v shape-gated"
        );
        let f = self.func("moe_down8_fma_dev_q8_w8h2v_rows");
        let cfg = LaunchConfig {
            grid_dim: (out_f.div_ceil(2) as u32, 1, t as u32),
            block_dim: (32, n_used as u32, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, nu, ne, rbi) = (
            in_f as i32,
            out_f as i32,
            n_used as i32,
            n_expert as i32,
            rb as i64,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(table)
            .arg(sel)
            .arg(w)
            .arg(aq2)
            .arg(ad2)
            .arg(dst)
            .arg(&inf)
            .arg(&outf)
            .arg(&nu)
            .arg(&ne)
            .arg(&qt)
            .arg(&rbi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// CSR gate/up v3 (owner-scan dedup, no build kernel): qtypes {IQ4_XS, IQ3_S} (caller
    /// gates), grid.y = pair index; the first pair of each expert serves all its pairs.
    /// Bit-identical to moe_gate_up_silu8_dev_q8_v_rows (explicit-intrinsic accumulate).
    #[allow(clippy::too_many_arguments)]
    pub fn moe_gate_up_silu8_dev_q8_csr(
        &self,
        table: &CudaSlice<u64>,
        sel: &CudaSlice<i32>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        n_pairs: usize,
        in_f: usize,
        n_ff: usize,
        n_used: usize,
        n_expert: usize,
        qt_g: i32,
        qt_u: i32,
        rb_g: usize,
        rb_u: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        // NVFP4 experts take the NVFP4-specialized owner-scan twin (lane/moebatch-q35moe);
        // host gate guarantees qt_g == qt_u within a supported class.
        let f = if qt_g == crate::QT_NVFP4 {
            self.func("moe_gate_up_silu8_dev_q8_csr_nvfp4")
        } else {
            self.func("moe_gate_up_silu8_dev_q8_csr_iq4")
        };
        let mut act = self.alloc_uninit::<f32>(n_pairs * n_ff)?;
        let cfg = LaunchConfig {
            grid_dim: (n_ff as u32, n_pairs as u32, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, nff, ne, nu, npi, rbg, rbu) = (
            in_f as i32,
            n_ff as i32,
            n_expert as i32,
            n_used as i32,
            n_pairs as i32,
            rb_g as i64,
            rb_u as i64,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(table)
            .arg(sel)
            .arg(aq)
            .arg(ad)
            .arg(&mut act)
            .arg(&inf)
            .arg(&nff)
            .arg(&ne)
            .arg(&qt_g)
            .arg(&qt_u)
            .arg(&rbg)
            .arg(&rbu)
            .arg(&nu)
            .arg(&npi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(act)
    }

    /// TEST SEAM (down8 lane 2026-07-08): launch a down dev_q8 variant BY NAME with its
    /// canonical geometry, bypassing the env-cached dispatch so moe-devq8-check can byte-
    /// compare variants in one process. Variants: "base", "w8h2", "w8h2r2", "w8h2v", "w8h2r2v".
    #[allow(clippy::too_many_arguments)]
    pub fn moe_down8_fma_dev_q8_variant(
        &self,
        variant: &str,
        table: &CudaSlice<u64>,
        sel: &cudarc::driver::CudaView<i32>,
        w: &cudarc::driver::CudaView<f32>,
        aq2: &CudaSlice<i8>,
        ad2: &CudaSlice<f32>,
        dst: &mut cudarc::driver::CudaViewMut<f32>,
        in_f: usize,
        out_f: usize,
        n_used: usize,
        n_expert: usize,
        qt: i32,
        rb: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let (inf, outf, nu, ne, rbi) = (
            in_f as i32,
            out_f as i32,
            n_used as i32,
            n_expert as i32,
            rb as i64,
        );
        let (f, cfg) = match variant {
            "w8h2" | "w8h2v" => (
                self.func(if variant == "w8h2" {
                    "moe_down8_fma_dev_q8_w8h2"
                } else {
                    "moe_down8_fma_dev_q8_w8h2v"
                }),
                LaunchConfig {
                    grid_dim: (out_f.div_ceil(2) as u32, 1, 1),
                    block_dim: (32, n_used as u32, 1),
                    shared_mem_bytes: 0,
                },
            ),
            "w8h2r2" | "w8h2r2v" => (
                self.func(if variant == "w8h2r2" {
                    "moe_down8_fma_dev_q8_w8h2r2"
                } else {
                    "moe_down8_fma_dev_q8_w8h2r2v"
                }),
                LaunchConfig {
                    grid_dim: (out_f.div_ceil(4) as u32, 1, 1),
                    block_dim: (32, n_used as u32, 1),
                    shared_mem_bytes: 0,
                },
            ),
            _ => (
                self.func("moe_down8_fma_dev_q8"),
                LaunchConfig {
                    grid_dim: (out_f as u32, 1, 1),
                    block_dim: (32, 1, 1),
                    shared_mem_bytes: 0,
                },
            ),
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(table)
            .arg(sel)
            .arg(w)
            .arg(aq2)
            .arg(ad2)
            .arg(dst)
            .arg(&inf)
            .arg(&outf)
            .arg(&nu)
            .arg(&ne)
            .arg(&qt)
            .arg(&rbi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// TEST SEAM (down8 lane): gate_up twin of the above. Variants: "base", "v".
    #[allow(clippy::too_many_arguments)]
    pub fn moe_gate_up_silu8_dev_q8_variant(
        &self,
        variant: &str,
        table: &CudaSlice<u64>,
        sel: &cudarc::driver::CudaView<i32>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        in_f: usize,
        n_ff: usize,
        n_used: usize,
        n_expert: usize,
        qt_g: i32,
        qt_u: i32,
        rb_g: usize,
        rb_u: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let mut act = self.alloc_uninit::<f32>(n_used * n_ff)?;
        let (inf, nff, ne, rbg, rbu) = (
            in_f as i32,
            n_ff as i32,
            n_expert as i32,
            rb_g as i64,
            rb_u as i64,
        );
        let f = self.func(if variant == "v" {
            "moe_gate_up_silu8_dev_q8_v"
        } else {
            "moe_gate_up_silu8_dev_q8"
        });
        let cfg = LaunchConfig {
            grid_dim: (n_ff as u32, n_used as u32, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(table)
            .arg(sel)
            .arg(aq)
            .arg(ad)
            .arg(&mut act)
            .arg(&inf)
            .arg(&nff)
            .arg(&ne)
            .arg(&qt_g)
            .arg(&qt_u)
            .arg(&rbg)
            .arg(&rbu);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(act)
    }

    pub fn moe_gate_up_silu8_dev(
        &self,
        table: &CudaSlice<u64>,
        sel: &cudarc::driver::CudaView<i32>,
        x: &cudarc::driver::CudaView<f32>,
        in_f: usize,
        n_ff: usize,
        n_used: usize,
        n_expert: usize,
        qt_g: i32,
        qt_u: i32,
        rb_g: usize,
        rb_u: usize,
        macros: &CudaSlice<f32>,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let f = self.func("moe_gate_up_silu8_dev");
        let mut act = self.alloc_uninit::<f32>(n_used * n_ff)?; // fully overwritten
        let cfg = LaunchConfig {
            grid_dim: (n_ff as u32, n_used as u32, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, nff, ne, rbg, rbu) = (
            in_f as i32,
            n_ff as i32,
            n_expert as i32,
            rb_g as i64,
            rb_u as i64,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(table)
            .arg(sel)
            .arg(x)
            .arg(&mut act)
            .arg(&inf)
            .arg(&nff)
            .arg(&ne)
            .arg(&qt_g)
            .arg(&qt_u)
            .arg(&rbg)
            .arg(&rbu)
            .arg(macros);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(act)
    }

    /// LAUNCH-STRUCTURE STAGE 3: device-dispatch twin of `moe_down8_fma_into` — expert ids AND
    /// renormalized weights read from the router kernel's device output. BIT-IDENTICAL chain.
    #[allow(clippy::too_many_arguments)]
    pub fn moe_down8_fma_dev(
        &self,
        table: &CudaSlice<u64>,
        sel: &cudarc::driver::CudaView<i32>,
        w: &cudarc::driver::CudaView<f32>,
        act: &CudaSlice<f32>,
        dst: &mut cudarc::driver::CudaViewMut<f32>,
        in_f: usize,
        out_f: usize,
        n_used: usize,
        n_expert: usize,
        qt: i32,
        rb: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("moe_down8_fma_dev");
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, nu, ne, rbv) = (
            in_f as i32,
            out_f as i32,
            n_used as i32,
            n_expert as i32,
            rb as i64,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(table)
            .arg(sel)
            .arg(w)
            .arg(act)
            .arg(dst)
            .arg(&inf)
            .arg(&outf)
            .arg(&nu)
            .arg(&ne)
            .arg(&qt)
            .arg(&rbv);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// dst[i] += alpha * src[i], i in 0..n. dst is a CudaViewMut (a row of moe_out).
    pub fn axpy_into(
        &self,
        src: &CudaSlice<f32>,
        alpha: f32,
        dst: &mut cudarc::driver::CudaViewMut<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("axpy_f32");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let (a, ni) = (alpha, n as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(src).arg(dst).arg(&a).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Host-oracle twin of `axpy_into` with separate RN multiply and add.
    pub fn axpy_host_into(
        &self,
        src: &cudarc::driver::CudaView<'_, f32>,
        alpha: f32,
        dst: &mut cudarc::driver::CudaViewMut<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("axpy_host_f32");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let (a, ni) = (alpha, n as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(src).arg(dst).arg(&a).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// dst[r*ncols + c] += src[r*ncols + c] * scale[r]. Per-row scalar accumulate (shared expert).
    pub fn add_scaled_rows(
        &self,
        src: &CudaSlice<f32>,
        scale: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("add_scaled_rows_f32");
        let cfg = LaunchConfig::for_num_elems((ncols * nrows) as u32);
        let (nc, nr) = (ncols as i32, nrows as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(src).arg(scale).arg(dst).arg(&nc).arg(&nr);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    // ======== A2 GROUPED MoE PREFILL KERNELS ========

    /// Gather m_e rows from src[T, ncols] into dst[m_e, ncols] using index array idx[m_e].
    pub fn gather_rows(
        &self,
        src: &CudaSlice<f32>,
        idx: &CudaSlice<i32>,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
        m_e: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gather_rows_f32");
        let cfg = LaunchConfig::for_num_elems((m_e * ncols) as u32);
        let (nc, me) = (ncols as i32, m_e as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(src).arg(idx).arg(dst).arg(&nc).arg(&me);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Scatter expert outputs into per-token slots: dst[tok_idx[r], slot_idx[r], :] = src[r, :] * weight[r].
    /// dst is [T, n_used, ncols], zero-initialized. Each (expert, token) pair maps to a unique slot.
    /// Scatter expert outputs into per-token slots (raw copy, no weight multiply).
    /// Weight stored into wbuf[tok*n_used + slot] for FMA in reduce step.
    pub fn scatter_slot(
        &self,
        src: &CudaSlice<f32>,
        tok_idx: &CudaSlice<i32>,
        slot_idx: &CudaSlice<i32>,
        weight: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        wbuf: &mut CudaSlice<f32>,
        ncols: usize,
        n_used: usize,
        m_e: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("scatter_add_slot_f32");
        let cfg = LaunchConfig::for_num_elems((m_e * ncols) as u32);
        let (nc, nu, me) = (ncols as i32, n_used as i32, m_e as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(src)
            .arg(tok_idx)
            .arg(slot_idx)
            .arg(weight)
            .arg(dst)
            .arg(wbuf)
            .arg(&nc)
            .arg(&nu)
            .arg(&me);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Reduce n_used slots per token: dst[t, col] = sum_s slots[t, s, col].
    /// Reduce n_used slots per token: dst[t, col] = sum_s FMA(wbuf[t,s], slots[t,s,col], acc).
    /// Uses FMA for bit-identity with the sequential axpy path.
    pub fn reduce_slots(
        &self,
        slots: &CudaSlice<f32>,
        wbuf: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
        n_used: usize,
        t: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("reduce_slots_f32");
        let cfg = LaunchConfig::for_num_elems((t * ncols) as u32);
        let (nc, nu, ti) = (ncols as i32, n_used as i32, t as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(slots).arg(wbuf).arg(dst).arg(&nc).arg(&nu).arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Canonical slot-order reduction with separately rounded multiply and add.
    ///
    /// This is the one-launch twin of repeated `axpy_host_into` calls. It preserves the official
    /// Step host-oracle arithmetic while allowing owner outputs to remain device-resident.
    pub fn reduce_slots_host(
        &self,
        slots: &CudaSlice<f32>,
        wbuf: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
        n_used: usize,
        t: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("reduce_slots_host_f32");
        let cfg = LaunchConfig::for_num_elems((t * ncols) as u32);
        let (nc, nu, ti) = (ncols as i32, n_used as i32, t as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(slots).arg(wbuf).arg(dst).arg(&nc).arg(&nu).arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Stage-B: quantize activation [m,in] f32 -> q8_1 (int8 qs + per-block f32 scale).
    /// Quantize an activation [m, in_f] to q8_1 (int8 qs + per-32 f32 scale). Public so the
    /// forward can quantize a SHARED activation ONCE and feed it to several matmuls (gate+up
    /// share `z`; q/k/v and wqkv/gate/beta/alpha share `h`) — quantize_q8_1 was 13.5% of decode
    /// GPU time, ~half of it redundant re-quantization of the same row.
    /// quantize_q8_1 over a CudaView (a sliced z-row) — same kernel, offset-honoring arg.
    pub fn quantize_q8_1_view(
        &self,
        x: &cudarc::driver::CudaView<f32>,
        m: usize,
        in_f: usize,
    ) -> Result<(CudaSlice<i8>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        let f = self.func("quantize_q8_1");
        let nblk = in_f / 32;
        let mut q = self.alloc_uninit::<i8>(m * in_f)?;
        let mut d = self.alloc_uninit::<f32>(m * nblk)?;
        let cfg = LaunchConfig::for_num_elems((m * in_f) as u32);
        let (inf, mi) = (in_f as i32, m as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(&mut q).arg(&mut d).arg(&inf).arg(&mi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((q, d))
    }

    pub fn quantize_q8_1(
        &self,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
    ) -> Result<(CudaSlice<i8>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        let nblk = in_f / 32;
        let mut q = self.alloc_uninit::<i8>(m * in_f)?; // full-overwrite output: skip memset
        let mut d = self.alloc_uninit::<f32>(m * nblk)?; // full-overwrite output: skip memset
        // WARP-PER-BLOCK kernel: one warp per 32-block -> m*in_f threads total.
        let cfg = LaunchConfig::for_num_elems((m * in_f) as u32);
        let (inf, mi) = (in_f as i32, m as i32);
        if Self::pdl_on() && Self::pdl_wb_on() {
            {
                use cudarc::driver::{DevicePtr, DevicePtrMut};
                let s = &self.gpu.stream();
                let (px, _g0) = x.device_ptr(s);
                let (pq, _g1) = q.device_ptr_mut(s);
                let (pd, _g2) = d.device_ptr_mut(s);
                let mut ps = [
                    &px as *const _ as *mut std::ffi::c_void,
                    &pq as *const _ as *mut _,
                    &pd as *const _ as *mut _,
                    &inf as *const _ as *mut _,
                    &mi as *const _ as *mut _,
                ];
                unsafe {
                    self.launch_pdl("quantize_q8_1", cfg.grid_dim, cfg.block_dim, &mut ps)?;
                }
            }
            return Ok((q, d));
        }
        let f = self.func("quantize_q8_1");
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(&mut q).arg(&mut d).arg(&inf).arg(&mi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((q, d))
    }

    /// Stage-C FP4: quantize activation [m,in] f32 -> e2m1 nibbles (aq4: u32 [m, in/8]) + per-16
    /// UE4M3 scale (ad4: u8 [m, in/16]), the layout the mxf4nvf4 block-scale GEMM B-operand wants.
    /// in_f must be a multiple of 64 (one NVFP4 K-block). One thread per (token, 16-block).
    pub fn quantize_fp4_act(
        &self,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
    ) -> Result<(CudaSlice<u32>, CudaSlice<u8>), Box<dyn std::error::Error>> {
        let f = self.func("quantize_fp4_act");
        let nb16 = in_f / 16;
        let mut aq4 = self.alloc_uninit::<u32>(m * (in_f / 8))?; // full-overwrite output: skip memset
        let mut ad4 = self.alloc_uninit::<u8>(m * nb16)?; // full-overwrite output: skip memset
        let cfg = LaunchConfig::for_num_elems((m * nb16) as u32);
        let (inf, mi) = (in_f as i32, m as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(&mut aq4).arg(&mut ad4).arg(&inf).arg(&mi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((aq4, ad4))
    }

    /// Stage-C FP4 GEMM (NVFP4 weights): native mxf4nvf4 block-scale tensor-core matmul. Feeds raw
    /// e2m1 weight nibbles + raw UE4M3 micro-scales directly to mma.sync.m16n8k64 (762 TFLOP/s peak,
    /// 3.5x int8). Activation `x` is quantized to FP4 e2m1 here. NVFP4 per-tensor macro-scale applied
    /// post (scale==1.0 -> no-op). `bytes` = raw NVFP4 weight rows. Used by the MEMRA_FP4 prefill path.
    pub fn qmatvec_gemm_nvfp4_fp4(
        &self,
        bytes: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
        scale: f32,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        assert!(
            in_f % 64 == 0,
            "FP4 GEMM requires in_f % 64 == 0, got {in_f}"
        );
        let (aq4, ad4) = self.quantize_fp4_act(x, m, in_f)?;
        let mut y = self.fp4_gemm_launch(bytes, &aq4, &ad4, m, in_f, out_f, row_bytes)?;
        if scale != 1.0 {
            self.scale_inplace(&mut y, scale, m * out_f)?;
        }
        Ok(y)
    }

    /// Shared mxf4 GEMM launch (pre-quantized FP4 activation aq4/ad4). Same CTA tile as the int8 GEMM
    /// (BM=64 rows x BN=128 tokens, 4 warps). No macro-scale applied here.
    fn fp4_gemm_launch(
        &self,
        bytes: &CudaSlice<u8>,
        aq4: &CudaSlice<u32>,
        ad4: &CudaSlice<u8>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let f = self.func("qmatvec_gemm_nvfp4_fp4");
        let mut y = self.alloc_uninit::<f32>(m * out_f)?; // full-overwrite output: skip memset
        const BM: u32 = 64;
        const BN: u32 = 256;
        let cfg = LaunchConfig {
            grid_dim: ((out_f as u32 + BM - 1) / BM, (m as u32 + BN - 1) / BN, 1),
            block_dim: (32, 4, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi, rb) = (in_f as i32, out_f as i32, m as i32, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(bytes)
            .arg(aq4)
            .arg(ad4)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(y)
    }

    /// Test entry (kernel_check): run the FP4 GEMM from raw bytes; NO macro-scale (caller compares bare).
    pub fn qmatvec_gemm_nvfp4_fp4_raw(
        &self,
        bytes: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        assert!(
            in_f % 64 == 0,
            "FP4 GEMM requires in_f % 64 == 0, got {in_f}"
        );
        let (aq4, ad4) = self.quantize_fp4_act(x, m, in_f)?;
        self.fp4_gemm_launch(bytes, &aq4, &ad4, m, in_f, out_f, row_bytes)
    }

    /// Stage-B: Q8_0 weight x q8_1 activation int8 dp4a matmul. y[m,out]=x@W^T.
    pub fn qmatvec_q8_0_fast(
        &self,
        w: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
        let f = self.func("qmatvec_q8_0_dp4a");
        let mut y = self.alloc_uninit::<f32>(m * out_f)?; // full-overwrite output: skip memset
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, m as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi, rb) = (in_f as i32, out_f as i32, m as i32, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(w)
            .arg(&aq)
            .arg(&ad)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(y)
    }

    /// Stage-B: Q4_K weight x q8_1 activation int8 dp4a (decode). Min-offset via q8_1 sum term.
    #[allow(non_snake_case)] // Keep GGUF qtype spelling visible at the public kernel boundary.
    pub fn qmatvec_q4_K_fast(
        &self,
        w: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
        let f = self.func("qmatvec_q4_K_dp4a");
        let mut y = self.alloc_uninit::<f32>(m * out_f)?; // full-overwrite output: skip memset
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, m as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi, rb) = (in_f as i32, out_f as i32, m as i32, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(w)
            .arg(&aq)
            .arg(&ad)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(y)
    }

    /// Stage-B: Q6_K weight x q8_1 activation int8 dp4a (decode, symmetric).
    #[allow(non_snake_case)] // Keep GGUF qtype spelling visible at the public kernel boundary.
    pub fn qmatvec_q6_K_fast(
        &self,
        w: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
        let f = self.func("qmatvec_q6_K_dp4a");
        let mut y = self.alloc_uninit::<f32>(m * out_f)?; // full-overwrite output: skip memset
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, m as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi, rb) = (in_f as i32, out_f as i32, m as i32, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(w)
            .arg(&aq)
            .arg(&ad)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(y)
    }

    /// Stage-B: Q5_K weight x q8_1 activation int8 dp4a (decode). Min-offset via q8_1 sum term.
    #[allow(non_snake_case)] // Keep GGUF qtype spelling visible at the public kernel boundary.
    pub fn qmatvec_q5_K_fast(
        &self,
        w: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        self.qmatvec_dp4a_named(
            "qmatvec_q5_K_dp4a",
            &w.slice(0..w.len()),
            x,
            m,
            in_f,
            out_f,
            row_bytes,
        )
    }
    /// Stage-B: Q3_K weight x q8_1 activation int8 dp4a (decode, symmetric).
    #[allow(non_snake_case)] // Keep GGUF qtype spelling visible at the public kernel boundary.
    pub fn qmatvec_q3_K_fast(
        &self,
        w: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        self.qmatvec_dp4a_named(
            "qmatvec_q3_K_dp4a",
            &w.slice(0..w.len()),
            x,
            m,
            in_f,
            out_f,
            row_bytes,
        )
    }
    /// A6 split-plane twin of `qmatvec_nvfp4_fast` (weights repacked; used by the rp gates).
    pub fn qmatvec_nvfp4_fast_rp(
        &self,
        w: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        assert!(
            in_f % 64 == 0,
            "NVFP4 dp4a requires in_f % 64 == 0, got {in_f}"
        );
        self.qmatvec_dp4a_named(
            "qmatvec_nvfp4_dp4a_rp",
            &w.slice(0..w.len()),
            x,
            m,
            in_f,
            out_f,
            row_bytes,
        )
    }
    /// Stage-B: NVFP4 weight x q8_1 activation int8 dp4a (decode, symmetric, codebook lookup).
    pub fn qmatvec_nvfp4_fast(
        &self,
        w: &cudarc::driver::CudaView<'_, u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        // B1: the NVFP4 dp4a kernel maps two 32-elem q8_1 blocks onto one 64-elem block_nvfp4
        // (sblk = g >> 1). in_f must be a multiple of 64 or the last block reads a partial superblock.
        assert!(
            in_f % 64 == 0,
            "NVFP4 dp4a requires in_f % 64 == 0, got {in_f}"
        );
        self.qmatvec_dp4a_named("qmatvec_nvfp4_dp4a", w, x, m, in_f, out_f, row_bytes)
    }
    /// v2-layout twin of `qmatvec_nvfp4_fast` for the slot-major expert banks
    /// (MEMRA_NVFP4_BANK_V2) — bit-identical per row, coalesced reads.
    pub fn qmatvec_nvfp4_fast_v2(
        &self,
        w: &cudarc::driver::CudaView<'_, u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        assert!(
            in_f % 64 == 0,
            "NVFP4 dp4a requires in_f % 64 == 0, got {in_f}"
        );
        self.qmatvec_dp4a_named("qmatvec_nvfp4_dp4a_v2", w, x, m, in_f, out_f, row_bytes)
    }
    /// Stage-B (optional perf): IQ4_XS codebook int8 dp4a.
    #[allow(non_snake_case)] // Keep GGUF qtype spelling visible at the public kernel boundary.
    pub fn qmatvec_iq4_XS_fast(
        &self,
        w: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        self.qmatvec_dp4a_named(
            "qmatvec_iq4_XS_dp4a",
            &w.slice(0..w.len()),
            x,
            m,
            in_f,
            out_f,
            row_bytes,
        )
    }

    /// Shared dp4a launcher: quantize_q8_1 then call the named kernel (grid (out,m), block 64).
    fn qmatvec_dp4a_named(
        &self,
        name: &str,
        w: &cudarc::driver::CudaView<'_, u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
        let f = self.func(name);
        let mut y = self.alloc_uninit::<f32>(m * out_f)?; // full-overwrite output: skip memset
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, m as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi, rb) = (in_f as i32, out_f as i32, m as i32, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(w)
            .arg(&aq)
            .arg(&ad)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(y)
    }

    /// NVFP4 dp4a matvec over PRE-QUANTIZED q8_1 activations, writing a caller-provided output.
    /// Same kernel and math as `qmatvec_nvfp4_fast` (which quantizes internally and allocates
    /// its output); this entry exists so a routed-expert program can quantize one activation
    /// ONCE and reuse it across every expert's gate/up, feed `silu_mul_scaled_q8_1`'s q8_1
    /// straight into down, and keep persistent output workspaces — zero per-expert allocations.
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_nvfp4_fast_prequant_into(
        &self,
        w: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        assert!(
            in_f % 64 == 0,
            "NVFP4 dp4a requires in_f % 64 == 0, got {in_f}"
        );
        if y.len() < m * out_f {
            return Err(format!(
                "NVFP4 prequant output {} is shorter than {m}x{out_f}",
                y.len()
            )
            .into());
        }
        let f = self.func("qmatvec_nvfp4_dp4a");
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, m as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi, rb) = (in_f as i32, out_f as i32, m as i32, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(w)
            .arg(aq)
            .arg(ad)
            .arg(y)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Fused QKV F32 matvec (one launch for all three rank-local projections; see the kernel
    /// doc for the numeric-class note). Requires `in_f % 4 == 0`.
    #[allow(clippy::too_many_arguments)]
    pub fn matvec_f32_qkv_into(
        &self,
        wq: &CudaSlice<f32>,
        wk: &CudaSlice<f32>,
        wv: &CudaSlice<f32>,
        wg: &CudaSlice<f32>,
        x: &CudaSlice<f32>,
        yq: &mut CudaSlice<f32>,
        yk: &mut CudaSlice<f32>,
        yv: &mut CudaSlice<f32>,
        yg: &mut CudaSlice<f32>,
        in_f: usize,
        out_q: usize,
        out_kv: usize,
        out_g: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if in_f % 4 != 0
            || wq.len() != out_q * in_f
            || wk.len() != out_kv * in_f
            || wv.len() != out_kv * in_f
            || wg.len() < out_g * in_f
            || x.len() < in_f
            || yq.len() < out_q
            || yk.len() < out_kv
            || yv.len() < out_kv
            || (out_g > 0 && yg.len() < out_g)
        {
            return Err(format!(
                "fused QKV geometry in={in_f} out_q={out_q} out_kv={out_kv} out_g={out_g} \
                 wq={} wk={} wv={} wg={}",
                wq.len(),
                wk.len(),
                wv.len(),
                wg.len()
            )
            .into());
        }
        let f = self.func("matvec_f32_qkv");
        let cfg = LaunchConfig {
            grid_dim: ((out_q + 2 * out_kv + out_g) as u32, 1, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, oq, okv, og) = (in_f as i32, out_q as i32, out_kv as i32, out_g as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(wq)
            .arg(wk)
            .arg(wv)
            .arg(wg)
            .arg(x)
            .arg(yq)
            .arg(yk)
            .arg(yv)
            .arg(yg)
            .arg(&inf)
            .arg(&oq)
            .arg(&okv)
            .arg(&og);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Selected-experts batched twin of `qmatvec_nvfp4_fast_prequant_into`: one launch covers
    /// every selected expert, weights indexed `sel[t] * expert_stride` into a contiguous
    /// per-rank bank, activations advancing `act_row_stride`/`ad_row_stride` elements per
    /// selection (0 for a shared input). Per (expert, row) bit-identical to the per-expert
    /// kernel — the batching only removes host launch latency.
    #[allow(clippy::too_many_arguments)]
    /// FUSION #2a: gate+up sweeps in one launch (v2 banks only; identical geometry both
    /// banks, caller-guarded). Per-row bit-identical to two qmatvec_nvfp4_sel_into calls.
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_nvfp4_sel_gu_into(
        &self,
        gate_bank: &CudaSlice<u8>,
        up_bank: &CudaSlice<u8>,
        sel: &CudaSlice<i32>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        yg: &mut CudaSlice<f32>,
        yu: &mut CudaSlice<f32>,
        n_sel: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
        expert_stride: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        assert!(in_f % 64 == 0, "NVFP4 dp4a requires in_f % 64 == 0");
        if yg.len() < n_sel * out_f || yu.len() < n_sel * out_f || sel.len() < n_sel {
            return Err("NVFP4 gu sel geometry".into());
        }
        // MEMRA_SEL_GU_RPW=2|4: multirow twin (activation group read once, reused across
        // RPW rows' gate+up dots) — bit-identical per row, one block per RPW rows.
        static RPW: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
        let rpw = *RPW.get_or_init(|| {
            std::env::var("MEMRA_SEL_GU_RPW")
                .ok()
                .and_then(|v| v.parse().ok())
                .filter(|r| *r == 2 || *r == 4)
                .unwrap_or(1)
        });
        let rpw = if out_f % rpw == 0 { rpw } else { 1 };
        // MEMRA_SEL_GU_WPR=1: warp-per-row (NUMERIC-CLASS — per-row reduction order changes;
        // acceptance is the argmax gate + battery, the QKV_FUSED/BF16_MMV class).
        static WPR: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let wpr = *WPR.get_or_init(|| std::env::var("MEMRA_SEL_GU_WPR").as_deref() == Ok("1"));
        let f = self.func(match (wpr, rpw) {
            (true, _) => "qmatvec_nvfp4_dp4a_sel_v2_gu_wpr",
            (_, 4) => "qmatvec_nvfp4_dp4a_sel_v2_gu_r4",
            (_, 2) => "qmatvec_nvfp4_dp4a_sel_v2_gu_r2",
            _ => "qmatvec_nvfp4_dp4a_sel_v2_gu",
        });
        let cfg = LaunchConfig {
            grid_dim: if wpr {
                (((2 * out_f) as u32).div_ceil(4), n_sel as u32, 1)
            } else if rpw == 1 {
                ((2 * out_f) as u32, n_sel as u32, 1)
            } else {
                ((out_f / rpw) as u32, n_sel as u32, 1)
            },
            block_dim: if wpr { (32, 4, 1) } else { (128, 1, 1) },
            shared_mem_bytes: 0,
        };
        let (inf, outf, ns) = (in_f as i32, out_f as i32, n_sel as i32);
        let (rb, es) = (row_bytes as i64, expert_stride as i64);
        let (ars, adrs) = (0i64, 0i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate_bank)
            .arg(up_bank)
            .arg(sel)
            .arg(aq)
            .arg(ad)
            .arg(yg)
            .arg(yu)
            .arg(&inf)
            .arg(&outf)
            .arg(&ns)
            .arg(&rb)
            .arg(&es)
            .arg(&ars)
            .arg(&adrs);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// MEMRA_SEL_DOWN8=1: the DOWN sweep and the route-weight combine in ONE launch
    /// (`qmatvec_nvfp4_dp4a_sel_v2_down8`, the q8 `down8 w8` occupancy arm ported to the
    /// NVFP4 banks). Block = (32, n_sel): one warp per slot instead of one warp per
    /// (row, slot), and the n_sel x out_f partial buffer disappears. Bit-identical to
    /// `qmatvec_nvfp4_sel_into` + `axpy_rows_seq_md_into` — same dot program, same reduce
    /// tree, same slot-ordered chain. Requires the v2 banks and nsb <= 32 (the fit-block
    /// class the reduce identity is argued at).
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_nvfp4_sel_down8_into(
        &self,
        bank: &CudaSlice<u8>,
        sel: &CudaSlice<i32>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        route_w: &CudaSlice<f32>,
        md: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        n_sel: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
        expert_stride: usize,
        act_row_stride: usize,
        ad_row_stride: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if in_f % 64 != 0
            || n_sel == 0
            || n_sel > 8
            || (in_f >> 5) > 32
            || dst.len() < out_f
            || sel.len() < n_sel
            || route_w.len() < n_sel
        {
            return Err(format!(
                "NVFP4 sel down8 geometry in_f={in_f} out_f={out_f} n_sel={n_sel} dst={}",
                dst.len()
            )
            .into());
        }
        if !crate::tp::nvfp4_bank_v2_on() {
            return Err("NVFP4 sel down8 requires the v2 banks (MEMRA_NVFP4_BANK_V2=1)".into());
        }
        let f = self.func("qmatvec_nvfp4_dp4a_sel_v2_down8");
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, 1, 1),
            block_dim: (32, n_sel as u32, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, ns) = (in_f as i32, out_f as i32, n_sel as i32);
        let (rb, es) = (row_bytes as i64, expert_stride as i64);
        let (ars, adrs) = (act_row_stride as i64, ad_row_stride as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(bank)
            .arg(sel)
            .arg(aq)
            .arg(ad)
            .arg(route_w)
            .arg(md)
            .arg(dst)
            .arg(&inf)
            .arg(&outf)
            .arg(&ns)
            .arg(&rb)
            .arg(&es)
            .arg(&ars)
            .arg(&adrs);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// T-ROW twin of the down8 fusion (spec verify + batched serving MoE): one block per
    /// (output row, token row) = the exact t=1 down8 program per token — bit-identical
    /// per row to its own down8/axpy pair at any t.
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_nvfp4_sel_down8_rows_into(
        &self,
        bank: &CudaSlice<u8>,
        sel: &CudaSlice<i32>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        route_w: &CudaSlice<f32>,
        md: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        t: usize,
        n_sel_col: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
        expert_stride: usize,
        act_row_stride: usize,
        ad_row_stride: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let n_sel = t * n_sel_col;
        if in_f % 64 != 0
            || n_sel_col == 0
            || n_sel_col > 8
            || t == 0
            || t > 64
            || (in_f >> 5) > 32
            || dst.len() < t * out_f
            || sel.len() < n_sel
            || route_w.len() < n_sel
        {
            return Err(format!(
                "NVFP4 sel down8 rows geometry in_f={in_f} out_f={out_f} t={t} dst={}",
                dst.len()
            )
            .into());
        }
        if !crate::tp::nvfp4_bank_v2_on() {
            return Err(
                "NVFP4 sel down8 rows requires the v2 banks (MEMRA_NVFP4_BANK_V2=1)".into(),
            );
        }
        let f = self.func("qmatvec_nvfp4_dp4a_sel_v2_down8_rows");
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, t as u32, 1),
            block_dim: (32, n_sel_col as u32, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, nsc) = (in_f as i32, out_f as i32, n_sel_col as i32);
        let (rb, es) = (row_bytes as i64, expert_stride as i64);
        let (ars, adrs) = (act_row_stride as i64, ad_row_stride as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(bank)
            .arg(sel)
            .arg(aq)
            .arg(ad)
            .arg(route_w)
            .arg(md)
            .arg(dst)
            .arg(&inf)
            .arg(&outf)
            .arg(&nsc)
            .arg(&rb)
            .arg(&es)
            .arg(&ars)
            .arg(&adrs);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// EP2 owner-guarded gate+up sweep: full-width rows, pairs whose expert this rank
    /// does not own exit immediately. Per-pair dot == the _sel_v2 gu body.
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_nvfp4_sel_gu_ep_into(
        &self,
        gate_bank: &CudaSlice<u8>,
        up_bank: &CudaSlice<u8>,
        sel: &CudaSlice<i32>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        yg: &mut CudaSlice<f32>,
        yu: &mut CudaSlice<f32>,
        n_sel: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
        expert_stride: usize,
        owner: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        assert!(in_f % 64 == 0, "NVFP4 dp4a requires in_f % 64 == 0");
        if yg.len() < n_sel * out_f || yu.len() < n_sel * out_f || sel.len() < n_sel {
            return Err("NVFP4 gu ep geometry".into());
        }
        let f = self.func("qmatvec_nvfp4_dp4a_sel_v2_gu_ep");
        let cfg = LaunchConfig {
            grid_dim: ((2 * out_f) as u32, n_sel as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, ns, own) = (in_f as i32, out_f as i32, n_sel as i32, owner as i32);
        let (rb, es) = (row_bytes as i64, expert_stride as i64);
        let (ars, adrs) = (0i64, 0i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate_bank)
            .arg(up_bank)
            .arg(sel)
            .arg(aq)
            .arg(ad)
            .arg(yg)
            .arg(yu)
            .arg(&inf)
            .arg(&outf)
            .arg(&ns)
            .arg(&rb)
            .arg(&es)
            .arg(&ars)
            .arg(&adrs)
            .arg(&own);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// EP2 owner-guarded SwiGLU (q8_1 emission), clamped or plain by `limit`.
    #[allow(clippy::too_many_arguments)]
    pub fn silu_mul_scaled_q8_1_sel_ep_into(
        &self,
        gate: &CudaSlice<f32>,
        up: &CudaSlice<f32>,
        gmac: &CudaSlice<f32>,
        umac: &CudaSlice<f32>,
        sel: &CudaSlice<i32>,
        limit: Option<f32>,
        out_q: &mut CudaSlice<i8>,
        out_d: &mut CudaSlice<f32>,
        n_per: usize,
        n_sel: usize,
        owner: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if n_per % 32 != 0 || out_q.len() < n_sel * n_per || out_d.len() < n_sel * n_per / 32 {
            return Err("NVFP4 silu ep geometry".into());
        }
        let f = self.func("silu_mul_scaled_q8_1_sel_ep");
        let warps = n_sel * n_per / 32;
        let cfg = LaunchConfig {
            grid_dim: ((warps as u32).div_ceil(4), 1, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (np, ns, own) = (n_per as i32, n_sel as i32, owner as i32);
        let (lim, has) = match limit {
            Some(l) => (l, 1i32),
            None => (0.0f32, 0i32),
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate)
            .arg(up)
            .arg(gmac)
            .arg(umac)
            .arg(sel)
            .arg(&lim)
            .arg(&has)
            .arg(out_q)
            .arg(out_d)
            .arg(&np)
            .arg(&ns)
            .arg(&own);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// EP2 owner-guarded down + owned-slot combine in one launch (block `(32, n_sel)`).
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_nvfp4_sel_down8_ep_into(
        &self,
        bank: &CudaSlice<u8>,
        sel: &CudaSlice<i32>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        route_w: &CudaSlice<f32>,
        md: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        n_sel: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
        expert_stride: usize,
        act_row_stride: usize,
        ad_row_stride: usize,
        owner: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if in_f % 64 != 0 || n_sel == 0 || n_sel > 8 || (in_f >> 5) > 64 || dst.len() < out_f {
            return Err("NVFP4 down8 ep geometry".into());
        }
        let f = self.func("qmatvec_nvfp4_dp4a_sel_v2_down8_ep");
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, 1, 1),
            block_dim: (32, n_sel as u32, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, ns, own) = (in_f as i32, out_f as i32, n_sel as i32, owner as i32);
        let (rb, es) = (row_bytes as i64, expert_stride as i64);
        let (ars, adrs) = (act_row_stride as i64, ad_row_stride as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(bank)
            .arg(sel)
            .arg(aq)
            .arg(ad)
            .arg(route_w)
            .arg(md)
            .arg(dst)
            .arg(&inf)
            .arg(&outf)
            .arg(&ns)
            .arg(&rb)
            .arg(&es)
            .arg(&ars)
            .arg(&adrs)
            .arg(&own);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn qmatvec_nvfp4_sel_into(
        &self,
        bank: &CudaSlice<u8>,
        sel: &CudaSlice<i32>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        n_sel: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
        expert_stride: usize,
        act_row_stride: usize,
        ad_row_stride: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        assert!(
            in_f % 64 == 0,
            "NVFP4 dp4a requires in_f % 64 == 0, got {in_f}"
        );
        if y.len() < n_sel * out_f || sel.len() < n_sel {
            return Err(format!(
                "NVFP4 sel output {} / sel {} shorter than {n_sel}x{out_f}",
                y.len(),
                sel.len()
            )
            .into());
        }
        // MEMRA_SEL_MR=1: 4-concurrent-row-groups twin — per row bit-identical (same 128-thread
        // striding + reduction). MEASURED SLOWER on the 188-SM card (40.8 vs 42.9 tok/s e2e,
        // 2026-08-21: 512-thread blocks trade occupancy for launch-tail savings and lose; the
        // sequential-rows variant was flat). Default stays the single-row form.
        // MEMRA_SEL_STREAM=1: 16-rows-per-block streaming twin with next-row register
        // prefetch (bit-identical per row; one group per thread, so in_f <= 4096 only).
        static MR: std::sync::OnceLock<u8> = std::sync::OnceLock::new();
        let mode = *MR.get_or_init(|| {
            if crate::tp::nvfp4_bank_v2_on() {
                3
            } else if std::env::var("MEMRA_SEL_STREAM").as_deref() == Ok("1") {
                2
            } else if std::env::var("MEMRA_SEL_MR").as_deref() == Ok("1") {
                1
            } else {
                0
            }
        });
        let mode = if mode == 2 && in_f > 4096 { 0 } else { mode };
        // v2s streaming twin (MEMRA_SEL_V2S=1 on top of the v2 bank): 8 contiguous rows per
        // block with next-row int4 prefetch; needs 16B-aligned rows (gate/up 2304B yes, down
        // 360B no -> single-row v2) and one slot per thread (in_f <= 4096).
        static V2S: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let v2s = mode == 3
            && *V2S.get_or_init(|| std::env::var("MEMRA_SEL_V2S").as_deref() == Ok("1"))
            && row_bytes % 16 == 0
            && in_f <= 4096;
        let f = match (mode, v2s) {
            (3, true) => self.func("qmatvec_nvfp4_dp4a_sel_v2s"),
            (3, false) => self.func("qmatvec_nvfp4_dp4a_sel_v2"),
            (2, _) => self.func("qmatvec_nvfp4_dp4a_sel_stream"),
            (1, _) => self.func("qmatvec_nvfp4_dp4a_sel_mr4"),
            _ => self.func("qmatvec_nvfp4_dp4a_sel"),
        };
        // Thread-fit block for narrow rows (the DOWN sweep: in_f=640 -> nsb=20 slots left
        // 108 of 128 threads idle AND thread-capped resident blocks). blockDim >= nsb keeps
        // thread g on slot g; the dropped threads contributed exact 0.0 partials to the
        // reduce, so the result bits are unchanged. Applies to the single-row forms only.
        let nsb = in_f >> 5;
        let fit_block: u32 = if (mode == 0 || mode == 3) && !v2s && nsb <= 32 {
            32
        } else if mode == 1 {
            512
        } else {
            128
        };
        let cfg = LaunchConfig {
            grid_dim: (
                if v2s {
                    (out_f as u32).div_ceil(8)
                } else {
                    match mode {
                        2 => (out_f as u32).div_ceil(16),
                        1 => (out_f as u32).div_ceil(4),
                        _ => out_f as u32,
                    }
                },
                n_sel as u32,
                1,
            ),
            block_dim: (fit_block, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, ns) = (in_f as i32, out_f as i32, n_sel as i32);
        let (rb, es, ars, adrs) = (
            row_bytes as i64,
            expert_stride as i64,
            act_row_stride as i64,
            ad_row_stride as i64,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(bank)
            .arg(sel)
            .arg(aq)
            .arg(ad)
            .arg(y)
            .arg(&inf)
            .arg(&outf)
            .arg(&ns)
            .arg(&rb)
            .arg(&es)
            .arg(&ars)
            .arg(&adrs);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Selected-experts batched twin of `silu_mul_scaled_q8_1`: [n_sel, n_per] rows, macros
    /// from device arrays indexed via sel. Per expert row bit-identical to the scalar kernel.
    /// `limit` = the step35 routed SwiGLU clamp (min(silu, limit) * clamp(up, +-limit)); None
    /// takes the plain SiLU kernel.
    #[allow(clippy::too_many_arguments)]
    pub fn silu_mul_scaled_q8_1_sel_into(
        &self,
        gate: &CudaSlice<f32>,
        up: &CudaSlice<f32>,
        gmac: &CudaSlice<f32>,
        umac: &CudaSlice<f32>,
        sel: &CudaSlice<i32>,
        limit: Option<f32>,
        out_q: &mut CudaSlice<i8>,
        out_d: &mut CudaSlice<f32>,
        n_per: usize,
        n_sel: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let n = n_per * n_sel;
        if n_per % 32 != 0 || out_q.len() < n || out_d.len() < n / 32 {
            return Err(format!(
                "silu sel geometry n_per={n_per} n_sel={n_sel} q={} d={}",
                out_q.len(),
                out_d.len()
            )
            .into());
        }
        if let Some(limit) = limit {
            if limit <= 1e-6 {
                return Err(format!(
                    "silu sel clamp limit {limit} is at or below the 1e-6 eps gate"
                )
                .into());
            }
            let f = self.func("silu_mul_scaled_q8_1_sel_clamp");
            let cfg = LaunchConfig::for_num_elems(n as u32);
            let (np, ns) = (n_per as i32, n_sel as i32);
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(gate)
                .arg(up)
                .arg(gmac)
                .arg(umac)
                .arg(sel)
                .arg(&limit)
                .arg(out_q)
                .arg(out_d)
                .arg(&np)
                .arg(&ns);
            unsafe {
                b.launch(cfg)?;
            }
            return Ok(());
        }
        let f = self.func("silu_mul_scaled_q8_1_sel");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let (np, ns) = (n_per as i32, n_sel as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate)
            .arg(up)
            .arg(gmac)
            .arg(umac)
            .arg(sel)
            .arg(out_q)
            .arg(out_d)
            .arg(&np)
            .arg(&ns);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn htod(&self, v: &[f32]) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        Ok(self.gpu.stream().clone_htod(v)?)
    }
    pub fn htod_i32(&self, v: &[i32]) -> Result<CudaSlice<i32>, Box<dyn std::error::Error>> {
        Ok(self.gpu.stream().clone_htod(v)?)
    }
    /// i8 upload (moe-devq8-check: synthetic q8_1 activation bytes).
    pub fn htod_i8(&self, v: &[i8]) -> Result<CudaSlice<i8>, Box<dyn std::error::Error>> {
        Ok(self.gpu.stream().clone_htod(v)?)
    }
    pub fn htod_u64(&self, v: &[u64]) -> Result<CudaSlice<u64>, Box<dyn std::error::Error>> {
        Ok(self.gpu.stream().clone_htod(v)?)
    }
    /// View twin of `dtoh` (lean-logits component 3: D2H one row of a [B, n_vocab] stack).
    pub fn dtoh_view(
        &self,
        d: &cudarc::driver::CudaView<f32>,
    ) -> Result<Vec<f32>, Box<dyn std::error::Error>> {
        let v = self.gpu.stream().clone_dtoh(d)?;
        self.gpu.stream().synchronize()?;
        Ok(v)
    }
    pub fn dtoh(&self, d: &CudaSlice<f32>) -> Result<Vec<f32>, Box<dyn std::error::Error>> {
        let v = self.gpu.stream().clone_dtoh(d)?;
        self.gpu.stream().synchronize()?;
        Ok(v)
    }
    /// Queue two f32 device-to-host copies on the compute stream, then establish one host
    /// boundary for both. Hy3's CPU/GPU expert split needs the router logits and the MoE input;
    /// issuing them together avoids a second stream synchronization in every trunk layer.
    pub fn dtoh_pair(
        &self,
        a: &CudaSlice<f32>,
        b: &CudaSlice<f32>,
    ) -> Result<(Vec<f32>, Vec<f32>), Box<dyn std::error::Error>> {
        let av = self.gpu.stream().clone_dtoh(a)?;
        let bv = self.gpu.stream().clone_dtoh(b)?;
        self.gpu.stream().synchronize()?;
        Ok((av, bv))
    }
    /// View-scoped twin of `dtoh_pair` for reusable capacity buffers whose inactive tail must not
    /// cross a shape-sensitive host boundary.
    pub fn dtoh_pair_views(
        &self,
        a: &cudarc::driver::CudaView<f32>,
        b: &cudarc::driver::CudaView<f32>,
    ) -> Result<(Vec<f32>, Vec<f32>), Box<dyn std::error::Error>> {
        let av = self.gpu.stream().clone_dtoh(a)?;
        let bv = self.gpu.stream().clone_dtoh(b)?;
        self.gpu.stream().synchronize()?;
        Ok((av, bv))
    }
    /// Device-to-host copy of an i32 buffer (fused-router sel_idx readback).
    pub fn dtoh_i32(&self, d: &CudaSlice<i32>) -> Result<Vec<i32>, Box<dyn std::error::Error>> {
        let v = self.gpu.stream().clone_dtoh(d)?;
        self.gpu.stream().synchronize()?;
        Ok(v)
    }
    /// Device-to-host copy of a u8 buffer (used to read back the quantized KV cache for validation).
    pub fn dtoh_u8(&self, d: &CudaSlice<u8>) -> Result<Vec<u8>, Box<dyn std::error::Error>> {
        let v = self.gpu.stream().clone_dtoh(d)?;
        self.gpu.stream().synchronize()?;
        Ok(v)
    }
    pub fn dtoh_u8_view(
        &self,
        d: &cudarc::driver::CudaView<u8>,
    ) -> Result<Vec<u8>, Box<dyn std::error::Error>> {
        let v = self.gpu.stream().clone_dtoh(d)?;
        self.gpu.stream().synchronize()?;
        Ok(v)
    }
    pub fn zeros(&self, n: usize) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let s = self.gpu.stream().alloc_zeros::<f32>(n)?;
        self.keep_if_capturing(&s);
        Ok(s)
    }

    /// GPU-resident greedy argmax (CUDA-GRAPH-PLAN Phase 1): logits[n_vocab] -> token id in a
    /// resident device u32 [1]. PARALLEL 2-pass (RANK1 LEVER): the old single-CTA scan (one 256-thread
    /// block on one SM over 248K logits) was memory-starved at ~426us/token. Now pass 1 fans NB=256
    /// blocks across the SMs to saturate HBM, pass 2 reduces the NB partials. Bit-identical to host
    /// `argmax` (smallest index on tie). The whole point is NOT to dtoh logits — only a [1] u32 is read
    /// back (or kept resident for graph replay). Returns the device token buffer.
    /// Softmax probability of the (already-argmaxed) token `tok` under `logits` — the spec-decode
    /// p-min confidence signal. 2-pass like the parallel argmax; returns a device [1] f32.
    pub fn prob_of_token_device(
        &self,
        logits: &CudaSlice<f32>,
        tok: &CudaSlice<u32>,
        n_vocab: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let nb = ARGMAX_NB;
        let mut part = self.alloc_uninit::<f32>(nb)?;
        let mut p = self.alloc_uninit::<f32>(1)?;
        let f1 = self.func("prob_of_token_partial_f32");
        let cfg1 = LaunchConfig {
            grid_dim: (nb as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let nv = n_vocab as i32;
        let __s_b1 = self.gpu.stream();
        let mut b1 = __s_b1.launch_builder(&f1);
        b1.arg(logits).arg(tok).arg(&mut part).arg(&nv);
        unsafe {
            b1.launch(cfg1)?;
        }
        let f2 = self.func("prob_of_token_final_f32");
        let cfg2 = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let nbi = nb as i32;
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&f2);
        b2.arg(&part).arg(&mut p).arg(&nbi);
        unsafe {
            b2.launch(cfg2)?;
        }
        Ok(p)
    }

    /// Like `prob_of_token_device` but writes into a PERSISTENT `p_out` buffer (stable pointer).
    /// Required for CUDA-graph capture of the draft chain: the captured prob kernels must write
    /// where the host reads the p-min confidence between replays. Same kernels, same math.
    /// Slot-addressed twin of `prob_of_token_device_into`: token read from `tok_all[tok_idx]`
    /// (a view at the slot), probability written to `p_out[p_idx]` — same two kernels, the
    /// pointers just land mid-buffer. Zero-sync (gemma confidence-adaptive draft depth).
    pub fn prob_of_token_device_col(
        &self,
        logits: &CudaSlice<f32>,
        tok_all: &CudaSlice<u32>,
        tok_idx: usize,
        p_out: &mut CudaSlice<f32>,
        p_idx: usize,
        n_vocab: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let tok_v = tok_all.slice(tok_idx..tok_idx + 1);
        let mut p_v = p_out.slice_mut(p_idx..p_idx + 1);
        let nb = ARGMAX_NB;
        let mut part = self.alloc_uninit::<f32>(nb)?;
        let f1 = self.func("prob_of_token_partial_f32");
        let cfg1 = LaunchConfig {
            grid_dim: (nb as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let nv = n_vocab as i32;
        let __s_b1 = self.gpu.stream();
        let mut b1 = __s_b1.launch_builder(&f1);
        b1.arg(logits).arg(&tok_v).arg(&mut part).arg(&nv);
        unsafe {
            b1.launch(cfg1)?;
        }
        let f2 = self.func("prob_of_token_final_f32");
        let cfg2 = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let nbi = nb as i32;
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&f2);
        b2.arg(&part).arg(&mut p_v).arg(&nbi);
        unsafe {
            b2.launch(cfg2)?;
        }
        Ok(())
    }

    pub fn prob_of_token_device_into(
        &self,
        logits: &CudaSlice<f32>,
        tok: &CudaSlice<u32>,
        p_out: &mut CudaSlice<f32>,
        n_vocab: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let nb = ARGMAX_NB;
        let mut part = self.alloc_uninit::<f32>(nb)?;
        let f1 = self.func("prob_of_token_partial_f32");
        let cfg1 = LaunchConfig {
            grid_dim: (nb as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let nv = n_vocab as i32;
        let __s_b1 = self.gpu.stream();
        let mut b1 = __s_b1.launch_builder(&f1);
        b1.arg(logits).arg(tok).arg(&mut part).arg(&nv);
        unsafe {
            b1.launch(cfg1)?;
        }
        let f2 = self.func("prob_of_token_final_f32");
        let cfg2 = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let nbi = nb as i32;
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&f2);
        b2.arg(&part).arg(p_out).arg(&nbi);
        unsafe {
            b2.launch(cfg2)?;
        }
        Ok(())
    }

    /// Token-graph chunk loop: hist[idx] = *tok; idx += 1 — device-indexed history append
    /// (graph-constant params, device-varying index). Capture-safe.
    pub fn u32_hist_append(
        &self,
        tok: &CudaSlice<u32>,
        hist: &mut CudaSlice<u32>,
        idx: &mut CudaSlice<i32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("u32_hist_append");
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(tok).arg(&mut *hist).arg(&mut *idx);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn argmax_token_device(
        &self,
        logits: &CudaSlice<f32>,
        n_vocab: usize,
    ) -> Result<CudaSlice<u32>, Box<dyn std::error::Error>> {
        let mut tok = unsafe { self.gpu.stream().alloc::<u32>(1)? };
        self.argmax_token_device_into(logits, &mut tok, n_vocab)?;
        Ok(tok)
    }
    /// Like `argmax_token_device` but writes into a PERSISTENT `tok` buffer (stable pointer) instead
    /// of allocating a fresh one. Required for CUDA-graph capture: the captured argmax must write the
    /// next token into the SAME device buffer the next replay's embed_gather reads, so the buffer
    /// pointer is baked once and the token id never round-trips to host inside steady state. The
    /// pass-1 partials scratch (`argmax_partials`) is also a resident stable-pointer buffer so both
    /// captured passes bake fixed addresses.
    pub fn argmax_token_device_into(
        &self,
        logits: &CudaSlice<f32>,
        tok: &mut CudaSlice<u32>,
        n_vocab: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let nb = ARGMAX_NB;
        let f1 = self.func("argmax_partial_f32");
        let f2 = self.func("argmax_final_f32");
        let mut guard = self.argmax_partials.lock().unwrap();
        if guard.is_none() {
            // allocate ONCE; under generate_graph this runs in the tracking-off prime window so the
            // buffers carry no cudarc events (illegal inside capture).
            let pv = self.gpu.stream().alloc_zeros::<f32>(nb)?;
            let pi = self.gpu.stream().alloc_zeros::<i32>(nb)?;
            *guard = Some((pv, pi));
        }
        let (part_v, part_i) = guard.as_mut().unwrap();
        let nv = n_vocab as i32;
        let nbi = nb as i32;
        // pass 1: NB blocks x 256 threads grid-stride scan -> per-block (val, idx) partials.
        let cfg1 = LaunchConfig {
            grid_dim: (nb as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b1 = self.gpu.stream();
        let mut b1 = __s_b1.launch_builder(&f1);
        b1.arg(logits).arg(&mut *part_v).arg(&mut *part_i).arg(&nv);
        unsafe {
            b1.launch(cfg1)?;
        }
        // pass 2: one block reduces NB partials -> token_out[0].
        let cfg2 = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&f2);
        b2.arg(&*part_v).arg(&*part_i).arg(tok).arg(&nbi);
        unsafe {
            b2.launch(cfg2)?;
        }
        Ok(())
    }
    /// Column-`col` device argmax over a stacked verify-logits buffer [t, n_vocab] (spec accept
    /// walk): toks[out_idx] = argmax(logits[col*n_vocab .. (col+1)*n_vocab]). SAME 2-pass kernels
    /// and tie-break contract as `argmax_token_device_into` (bit-identical to host argmax,
    /// argmax_gate-validated) — only the input pointer (a column view) and the output slot differ.
    /// Lets the accept walk read ONE [t] u32 instead of dtoh'ing the full [t, n_vocab] logits.
    pub fn argmax_token_device_col(
        &self,
        logits: &CudaSlice<f32>,
        col: usize,
        n_vocab: usize,
        toks: &mut CudaSlice<u32>,
        out_idx: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let nb = ARGMAX_NB;
        let f1 = self.func("argmax_partial_f32");
        let f2 = self.func("argmax_final_f32");
        let mut guard = self.argmax_partials.lock().unwrap();
        if guard.is_none() {
            let pv = self.gpu.stream().alloc_zeros::<f32>(nb)?;
            let pi = self.gpu.stream().alloc_zeros::<i32>(nb)?;
            *guard = Some((pv, pi));
        }
        let (part_v, part_i) = guard.as_mut().unwrap();
        let col_view = logits.slice(col * n_vocab..(col + 1) * n_vocab);
        let nv = n_vocab as i32;
        let nbi = nb as i32;
        let cfg1 = LaunchConfig {
            grid_dim: (nb as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b1 = self.gpu.stream();
        let mut b1 = __s_b1.launch_builder(&f1);
        b1.arg(&col_view)
            .arg(&mut *part_v)
            .arg(&mut *part_i)
            .arg(&nv);
        unsafe {
            b1.launch(cfg1)?;
        }
        let mut tok_view = toks.slice_mut(out_idx..out_idx + 1);
        let cfg2 = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&f2);
        b2.arg(&*part_v).arg(&*part_i).arg(&mut tok_view).arg(&nbi);
        unsafe {
            b2.launch(cfg2)?;
        }
        Ok(())
    }
    /// Read back a device u32 buffer (the spec accept walk's [t] per-column argmax tokens).
    pub fn htod_u32_v(&self, v: &[u32]) -> Result<CudaSlice<u32>, Box<dyn std::error::Error>> {
        Ok(self.gpu.stream().clone_htod(v)?)
    }
    pub fn dtoh_u64(&self, d: &CudaSlice<u64>) -> Result<Vec<u64>, Box<dyn std::error::Error>> {
        let v = self.gpu.stream().clone_dtoh(d)?;
        self.gpu.stream().synchronize()?;
        Ok(v)
    }

    pub fn dtoh_u32(&self, d: &CudaSlice<u32>) -> Result<Vec<u32>, Box<dyn std::error::Error>> {
        let v = self.gpu.stream().clone_dtoh(d)?;
        self.gpu.stream().synchronize()?;
        Ok(v)
    }
    /// Allocate a zeroed device u32 buffer (persistent spec-loop prediction slots).
    /// H2D into an EXISTING u32 buffer (stable pointer — the per-step grammar-mask upload:
    /// contents change every step, the address must not, so a captured graph can read it).
    pub fn htod_u32_into(
        &self,
        dst: &mut CudaSlice<u32>,
        src: &[u32],
    ) -> Result<(), Box<dyn std::error::Error>> {
        let mut view = dst.slice_mut(0..src.len());
        self.gpu.stream().memcpy_htod(src, &mut view)?;
        Ok(())
    }

    /// H2D into an existing i32 buffer. OPTIPIPE uses this to refresh a stage-local saved-len
    /// table without changing the device address its reconcile kernel consumes.
    pub fn htod_i32_into(
        &self,
        dst: &mut CudaSlice<i32>,
        src: &[i32],
    ) -> Result<(), Box<dyn std::error::Error>> {
        let mut view = dst.slice_mut(0..src.len());
        self.gpu.stream().memcpy_htod(src, &mut view)?;
        Ok(())
    }

    pub fn alloc_u32_zeroed(&self, n: usize) -> Result<CudaSlice<u32>, Box<dyn std::error::Error>> {
        let s = self.gpu.stream().alloc_zeros::<u32>(n)?;
        self.keep_if_capturing(&s);
        Ok(s)
    }
    /// embed_gather into a PERSISTENT `x_out` buffer (stable pointer) for CUDA-graph capture (the
    /// embed output starts the per-step kernel chain and must be at a fixed address across replays).
    pub fn embed_gather_device_into(
        &self,
        embd: &CudaSlice<u8>,
        token_d: &CudaSlice<u32>,
        x_out: &mut CudaSlice<f32>,
        n_embd: usize,
        qtype: i32,
        row_bytes: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("embed_gather_u32");
        let cfg = LaunchConfig {
            grid_dim: (((n_embd as u32 + 255) / 256).max(1), 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (ne, qt, rb) = (n_embd as i32, qtype, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(embd)
            .arg(token_d)
            .arg(x_out)
            .arg(&ne)
            .arg(&qt)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }
    /// Read a [1] i32 device counter (pos / seqlen) back to host. Tiny D2H + sync.
    pub fn dtoh_i32_one(&self, d: &CudaSlice<i32>) -> Result<i32, Box<dyn std::error::Error>> {
        let v = self.gpu.stream().clone_dtoh(d)?;
        self.gpu.stream().synchronize()?;
        Ok(v[0])
    }
    /// Set a [1] i32 device counter IN PLACE (keeps the buffer pointer stable — required for the
    /// graph-resident pos/seqlen counters whose addresses are baked into captured graphs). Restores
    /// the counter value after the throwaway capture warmups corrupt it.
    /// ASYNC i32 single-slot store (value rides the kernel arg — no host-memory transfer/sync).
    /// The graph-arc device-len counters use this; set_i32_one below is the SYNCING pageable
    /// copy (fine at stream-idle boundaries, poison mid-round).
    pub fn i32_set_k(
        &self,
        dst: &mut CudaSlice<i32>,
        v: i32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("i32_set_k");
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (1, 1, 1),
            shared_mem_bytes: 0,
        };
        let idx = 0i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(dst).arg(&v).arg(&idx);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn set_i32_one(
        &self,
        d: &mut CudaSlice<i32>,
        v: i32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        self.gpu.stream().memcpy_htod(&[v], d)?;
        Ok(())
    }
    /// Set a [1] u32 device buffer IN PLACE (stable pointer) — for the resident `token_d` counter
    /// during priming / capture-state restore.
    pub fn set_u32_one(
        &self,
        d: &mut CudaSlice<u32>,
        v: u32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        self.gpu.stream().memcpy_htod(&[v], d)?;
        Ok(())
    }
    /// Read back a [1] u32 device buffer (the argmax token). One tiny D2H + sync.
    pub fn dtoh_u32_one(&self, d: &CudaSlice<u32>) -> Result<u32, Box<dyn std::error::Error>> {
        let v = self.gpu.stream().clone_dtoh(d)?;
        self.gpu.stream().synchronize()?;
        Ok(v[0])
    }
    /// Upload raw bytes to a resident device u8 buffer (e.g. the embed table for device gather).
    pub fn upload_u8(&self, bytes: &[u8]) -> Result<CudaSlice<u8>, Box<dyn std::error::Error>> {
        Ok(self.gpu.stream().clone_htod(bytes)?)
    }
    /// Embed-from-device (CUDA-GRAPH-PLAN Phase 1): gather+dequant the row for the token id in
    /// `token_d[0]` from the resident embed table `embd` -> x_out[n_embd]. Bit-identical to host
    /// EmbedHost::gather (same per-dtype `deq`). No host round-trip of the token id.
    pub fn embed_gather_device(
        &self,
        embd: &CudaSlice<u8>,
        token_d: &CudaSlice<u32>,
        n_embd: usize,
        qtype: i32,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let f = self.func("embed_gather_u32");
        let mut x = self.alloc_uninit::<f32>(n_embd)?;
        let cfg = LaunchConfig {
            grid_dim: (((n_embd as u32 + 255) / 256).max(1), 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (ne, qt, rb) = (n_embd as i32, qtype, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(embd)
            .arg(token_d)
            .arg(&mut x)
            .arg(&ne)
            .arg(&qt)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(x)
    }

    /// T-token device embed gather (spec verify/replay): tokens uploaded as a tiny [T] u32 htod,
    /// rows dequanted on-device -> x[T, n_embd]. Replaces host per-row dequant + T*n_embd*4B htod
    /// (nsys: 84% of spec API time was HtoD). Bit-identical rows (same per-dtype deq).
    pub fn embed_gather_device_t(
        &self,
        embd: &CudaSlice<u8>,
        tokens: &[u32],
        n_embd: usize,
        qtype: i32,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let t = tokens.len();
        let tok_d = self.gpu.stream().clone_htod(tokens)?;
        let f = self.func("embed_gather_u32_t");
        let mut x = self.alloc_uninit::<f32>(t * n_embd)?;
        let cfg = LaunchConfig {
            grid_dim: (((n_embd as u32 + 255) / 256).max(1), t as u32, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (ne, qt, rb, ti) = (n_embd as i32, qtype, row_bytes as i64, t as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(embd)
            .arg(&tok_d)
            .arg(&mut x)
            .arg(&ne)
            .arg(&qt)
            .arg(&rb)
            .arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(x)
    }

    /// T-token embed gather from a DEVICE token buffer (round-stream stage c: the verify tokens
    /// are assembled on-device from the draft-chain pack slots; no host round trip). Same kernel
    /// as embed_gather_device_t — bit-identical rows.
    /// embed_gather over a token VIEW (spec round: tokens live in the round's batch buffer).
    pub fn embed_gather_device_tv(
        &self,
        embd: &CudaSlice<u8>,
        tok_v: &cudarc::driver::CudaView<u32>,
        t: usize,
        n_embd: usize,
        qtype: i32,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let f = self.func("embed_gather_u32_t");
        let mut x = self.alloc_uninit::<f32>(t * n_embd)?;
        let cfg = LaunchConfig {
            grid_dim: (((n_embd as u32 + 255) / 256).max(1), t as u32, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (ne, qt, rb, ti) = (n_embd as i32, qtype, row_bytes as i64, t as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(embd)
            .arg(tok_v)
            .arg(&mut x)
            .arg(&ne)
            .arg(&qt)
            .arg(&rb)
            .arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(x)
    }

    pub fn embed_gather_device_td(
        &self,
        embd: &CudaSlice<u8>,
        tok_d: &CudaSlice<u32>,
        t: usize,
        n_embd: usize,
        qtype: i32,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let f = self.func("embed_gather_u32_t");
        let mut x = self.alloc_uninit::<f32>(t * n_embd)?;
        let cfg = LaunchConfig {
            grid_dim: (((n_embd as u32 + 255) / 256).max(1), t as u32, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (ne, qt, rb, ti) = (n_embd as i32, qtype, row_bytes as i64, t as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(embd)
            .arg(tok_d)
            .arg(&mut x)
            .arg(&ne)
            .arg(&qt)
            .arg(&rb)
            .arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(x)
    }

    /// Uninitialized device buffer — SKIPS the memset that `alloc_zeros` always issues. Decode
    /// profile (nsys): ~1050 memsets/token = 6.5% of decode GPU time + ~half the launch count, the
    /// dominant contributor to the 19% inter-kernel idle gap and a blocker for clean CUDA-graph
    /// capture. Use ONLY for buffers a kernel FULLY overwrites (every element written, no `+=`).
    /// SAFETY: caller guarantees the producing kernel writes every element before any read.
    #[inline]
    /// Keep an allocation alive for the current capture (no-op when retain mode is off).
    fn keep_if_capturing<T: cudarc::driver::DeviceRepr + Send + 'static>(&self, s: &CudaSlice<T>) {
        if self
            .capture_keep_on
            .load(std::sync::atomic::Ordering::Relaxed)
        {
            self.capture_keep.lock().unwrap().push(Box::new(s.clone()));
        }
    }

    fn alloc_uninit<T: cudarc::driver::DeviceRepr + Send + 'static>(
        &self,
        n: usize,
    ) -> Result<CudaSlice<T>, Box<dyn std::error::Error>> {
        let mut s = unsafe { self.gpu.stream().alloc::<T>(n)? };
        // MEMRA_DEBUG_ZERO_ALLOCS=1 (task #14 defect hunt): memset EVERY engine allocation —
        // the global uninit-read discriminator (the prime-fn-scoped zeroing experiment could
        // not cover engine-internal buffers). Debug-only: massive launch overhead.
        {
            static Z: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
            if *Z.get_or_init(|| std::env::var("MEMRA_DEBUG_ZERO_ALLOCS").as_deref() == Ok("1")) {
                // raw D8 memset (T lacks ValidAsZeroBits in the generic bound)
                use cudarc::driver::DevicePtrMut;
                let n_bytes = s.len() * std::mem::size_of::<T>();
                let stream = self.gpu.stream();
                let (p_, _g) = s.device_ptr_mut(&stream);
                unsafe {
                    cudarc::driver::sys::cuMemsetD8Async(p_, 0, n_bytes, stream.cu_stream())
                        .result()?;
                }
            }
        }
        self.keep_if_capturing(&s);
        Ok(s)
    }

    /// Public f32 uninitialized scratch (see `alloc_uninit`). For decode/forward scratch a kernel
    /// fully overwrites. SAFETY: producing kernel must write every element before any read.
    /// Uninitialized q8_1 activation pair (int8 + per-32 scales) — the fa combine q8-emit
    /// consumers alloc through this (m=1 decode arms).
    pub fn uninit_q8_pair(
        &self,
        n: usize,
    ) -> Result<(CudaSlice<i8>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        Ok((
            self.alloc_uninit::<i8>(n)?,
            self.alloc_uninit::<f32>(n / 32)?,
        ))
    }

    pub fn uninit(&self, n: usize) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        self.alloc_uninit::<f32>(n)
    }

    /// i8 uninitialized scratch (same contract as `uninit`).
    pub fn alloc_i8_uninit(&self, n: usize) -> Result<CudaSlice<i8>, Box<dyn std::error::Error>> {
        self.alloc_uninit::<i8>(n)
    }

    /// RMSNorm: x[ncols,nrows] row-major, weight[ncols] -> dst. One block/row, 256 threads.
    /// gemma4: 3 rms_norms of the SAME input in one launch (one reduction, three weights).
    /// Per-output bit-identical to three rms_norm calls (verbatim reduction/scale chain).
    #[allow(clippy::too_many_arguments)]
    pub fn rms_norm3(
        &self,
        x: &CudaSlice<f32>,
        w0: &CudaSlice<f32>,
        w1: &CudaSlice<f32>,
        w2: &CudaSlice<f32>,
        d0: &mut CudaSlice<f32>,
        d1: &mut CudaSlice<f32>,
        d2: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("rms_norm3_f32");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, e) = (ncols as i32, eps);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x)
            .arg(w0)
            .arg(w1)
            .arg(w2)
            .arg(d0)
            .arg(d1)
            .arg(d2)
            .arg(&nc)
            .arg(&e);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// gemma4 fused q/k/v head norms (one launch, per-row rms_norm_f32-verbatim).
    #[allow(clippy::too_many_arguments)]
    /// True when the warp-per-row qkv norm would engage for (rows, ncols) — the emit lane
    /// piggybacks on the same conditions.
    pub fn qkvnorm_w_on_prefill(rows: usize, ncols: usize) -> bool {
        static WARP_ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *WARP_ON.get_or_init(|| {
            std::env::var("MEMRA_QKVNORM_W")
                .map(|v| v != "0")
                .unwrap_or(true)
        }) && ncols % 4 == 0
            && rows >= 64
    }

    /// w4 norm with bf16 V EMIT (31B glue lane): the v segment also writes its normed rows as
    /// bf16 (the FA V operand — bit-identical to a post-hoc f32_to_bf16). Prefill-depth only.
    #[allow(clippy::too_many_arguments)]
    pub fn rms_norm_qkv_w4b(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        wq: &CudaSlice<f32>,
        wk: &CudaSlice<f32>,
        wv: &CudaSlice<f32>,
        dq: &mut CudaSlice<f32>,
        dk: &mut CudaSlice<f32>,
        dv: &mut CudaSlice<f32>,
        dvb: &mut CudaSlice<u8>,
        ncols: usize,
        rq: usize,
        rk: usize,
        eps: f32,
        vf16: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        assert!(ncols % 4 == 0 && rq + 2 * rk >= 64);
        let f = self.func("rms_norm_qkv_w4b_f32");
        let rows = (rq + 2 * rk) as u32;
        let cfg = LaunchConfig {
            grid_dim: (rows.div_ceil(8), 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, rqi, rki, rvi, e) = (ncols as i32, rq as i32, rk as i32, rk as i32, eps);
        let vf = vf16 as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(v)
            .arg(wq)
            .arg(wk)
            .arg(wv)
            .arg(dq)
            .arg(dk)
            .arg(dv)
            .arg(&mut *dvb)
            .arg(&nc)
            .arg(&rqi)
            .arg(&rki)
            .arg(&rvi)
            .arg(&e)
            .arg(&vf);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn rms_norm_qkv(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        wq: &CudaSlice<f32>,
        wk: &CudaSlice<f32>,
        wv: &CudaSlice<f32>,
        dq: &mut CudaSlice<f32>,
        dk: &mut CudaSlice<f32>,
        dv: &mut CudaSlice<f32>,
        ncols: usize,
        rq: usize,
        rk: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        // Warp-per-row float4 twin (default; MEMRA_QKVNORM_W=0 reverts): the block-per-row form
        // spends 767us/launch on 17k+ 2KB rows at prefill depth (launch/reduce latency-bound,
        // ~92GB/s). Own numeric config (reduce order differs) — battery-gated.
        static WARP_ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let warp_on = *WARP_ON.get_or_init(|| {
            std::env::var("MEMRA_QKVNORM_W")
                .map(|v| v != "0")
                .unwrap_or(true)
        });
        // rows >= 64 keeps decode (nh + 2*nkv rows) on the block-tree kernel — decode/verify/
        // replay numerics are untouched on every model; only prefill depth takes the new config.
        if warp_on && ncols % 4 == 0 && rq + 2 * rk >= 64 {
            let f = self.func("rms_norm_qkv_w4_f32");
            let rows = (rq + 2 * rk) as u32;
            let cfg = LaunchConfig {
                grid_dim: (rows.div_ceil(8), 1, 1),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let (nc, rqi, rki, rvi, e) = (ncols as i32, rq as i32, rk as i32, rk as i32, eps);
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(q)
                .arg(k)
                .arg(v)
                .arg(wq)
                .arg(wk)
                .arg(wv)
                .arg(dq)
                .arg(dk)
                .arg(dv)
                .arg(&nc)
                .arg(&rqi)
                .arg(&rki)
                .arg(&rvi)
                .arg(&e);
            unsafe {
                b.launch(cfg)?;
            }
            return Ok(());
        }
        let f = self.func("rms_norm_qkv_f32");
        let grid = (rq + 2 * rk) as u32;
        let cfg = LaunchConfig {
            grid_dim: (grid, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, rqi, rki, e) = (ncols as i32, rq as i32, rk as i32, eps);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(v)
            .arg(wq)
            .arg(wk)
            .arg(wv)
            .arg(dq)
            .arg(dk)
            .arg(dv)
            .arg(&nc)
            .arg(&rqi)
            .arg(&rki)
            .arg(&e);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// gemma4 fused pair of rms_norms over two different inputs (same width).
    #[allow(clippy::too_many_arguments)]
    pub fn rms_norm2x(
        &self,
        a: &CudaSlice<f32>,
        bb: &CudaSlice<f32>,
        wa: &CudaSlice<f32>,
        wb: &CudaSlice<f32>,
        da: &mut CudaSlice<f32>,
        db: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("rms_norm2x_f32");
        let cfg = LaunchConfig {
            grid_dim: (2 * nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, nr, e) = (ncols as i32, nrows as i32, eps);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(a)
            .arg(bb)
            .arg(wa)
            .arg(wb)
            .arg(da)
            .arg(db)
            .arg(&nc)
            .arg(&nr)
            .arg(&e);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// gemma4 R4: in-place final-logit softcap y = cap*tanh(y/cap).
    pub fn softcap(
        &self,
        y: &mut CudaSlice<f32>,
        cap: f32,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("softcap_f32");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let ni = n as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(y).arg(&cap).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// gemma4 suppress-token mask: y[row][ids[j]] = -inf over t logits rows (fixed-arg launch —
    /// graph-capture safe; NOT monotonic like softcap, so it must run before any argmax).
    pub fn mask_ids_rows(
        &self,
        y: &mut CudaSlice<f32>,
        ids: &CudaSlice<i32>,
        n_ids: usize,
        n_vocab: usize,
        t: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("mask_ids_rows_f32");
        let cfg = LaunchConfig::for_num_elems((n_ids * t) as u32);
        let (ni, nv, ti) = (n_ids as i32, n_vocab as i32, t as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(y).arg(ids).arg(&ni).arg(&nv).arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// gemma4: res = (a+b)*c AND dst = rms_norm(res, w) in one launch.
    #[allow(clippy::too_many_arguments)]
    pub fn add_scale_rms_norm(
        &self,
        a: &CudaSlice<f32>,
        b_in: &CudaSlice<f32>,
        c: f32,
        w: &CudaSlice<f32>,
        res: &mut CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("add_scale_rms_norm_f32");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, e2) = (ncols as i32, eps);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(a)
            .arg(b_in)
            .arg(&c)
            .arg(w)
            .arg(res)
            .arg(dst)
            .arg(&nc)
            .arg(&e2);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// gemma4: res = (a+b)*c AND the next layer's attn_norm EMITTED q8_1 in one launch.
    /// Quantize epilogue bit-identical to quantize_q8_1 (the rms_norm_q8_1 form).
    #[allow(clippy::too_many_arguments)]
    pub fn add_scale_rms_norm_q8_1(
        &self,
        a: &CudaSlice<f32>,
        b_in: &CudaSlice<f32>,
        c: f32,
        w: &CudaSlice<f32>,
        res: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(CudaSlice<i8>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        let mut out_q = self.alloc_uninit::<i8>(nrows * ncols)?;
        let mut out_d = self.alloc_uninit::<f32>(nrows * (ncols / 32))?;
        let (nc, e2) = (ncols as i32, eps);
        if Self::pdl_on() && Self::pdl_wb_on() {
            {
                use cudarc::driver::{DevicePtr, DevicePtrMut};
                let s = &self.gpu.stream();
                let (pa, _g0) = a.device_ptr(s);
                let (pb, _g1) = b_in.device_ptr(s);
                let (pw, _g2) = w.device_ptr(s);
                let (pr, _g3) = res.device_ptr_mut(s);
                let (pq, _g4) = out_q.device_ptr_mut(s);
                let (pd, _g5) = out_d.device_ptr_mut(s);
                let mut ps = [
                    &pa as *const _ as *mut std::ffi::c_void,
                    &pb as *const _ as *mut _,
                    &c as *const _ as *mut _,
                    &pw as *const _ as *mut _,
                    &pr as *const _ as *mut _,
                    &pq as *const _ as *mut _,
                    &pd as *const _ as *mut _,
                    &nc as *const _ as *mut _,
                    &e2 as *const _ as *mut _,
                ];
                unsafe {
                    self.launch_pdl(
                        "add_scale_rms_norm_q8_1",
                        (nrows as u32, 1, 1),
                        (rms_block(), 1, 1),
                        &mut ps,
                    )?;
                }
            }
            return Ok((out_q, out_d));
        }
        let f = self.func("add_scale_rms_norm_q8_1");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(a)
            .arg(b_in)
            .arg(&c)
            .arg(w)
            .arg(res)
            .arg(&mut out_q)
            .arg(&mut out_d)
            .arg(&nc)
            .arg(&e2);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((out_q, out_d))
    }

    /// Slot-fed add_scale_rms_norm_q8_1 twin (alloc-free capture lane).
    #[allow(clippy::too_many_arguments)]
    pub fn add_scale_rms_norm_q8_1_into(
        &self,
        a: &CudaSlice<f32>,
        b_in: &CudaSlice<f32>,
        c: f32,
        w: &CudaSlice<f32>,
        res: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
        out_q: &mut CudaSlice<i8>,
        out_d: &mut CudaSlice<f32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        debug_assert!(out_q.len() >= nrows * ncols && out_d.len() >= nrows * (ncols / 32));
        let (nc, e2) = (ncols as i32, eps);
        if Self::pdl_on() && Self::pdl_wb_on() {
            use cudarc::driver::{DevicePtr, DevicePtrMut};
            let s = &self.gpu.stream();
            let (pa, _g0) = a.device_ptr(s);
            let (pb, _g1) = b_in.device_ptr(s);
            let (pw, _g2) = w.device_ptr(s);
            let (pr, _g3) = res.device_ptr_mut(s);
            let (pq, _g4) = out_q.device_ptr_mut(s);
            let (pd, _g5) = out_d.device_ptr_mut(s);
            let mut ps = [
                &pa as *const _ as *mut std::ffi::c_void,
                &pb as *const _ as *mut _,
                &c as *const _ as *mut _,
                &pw as *const _ as *mut _,
                &pr as *const _ as *mut _,
                &pq as *const _ as *mut _,
                &pd as *const _ as *mut _,
                &nc as *const _ as *mut _,
                &e2 as *const _ as *mut _,
            ];
            unsafe {
                self.launch_pdl(
                    "add_scale_rms_norm_q8_1",
                    (nrows as u32, 1, 1),
                    (rms_block(), 1, 1),
                    &mut ps,
                )?;
            }
            return Ok(());
        }
        let f = self.func("add_scale_rms_norm_q8_1");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(a)
            .arg(b_in)
            .arg(&c)
            .arg(w)
            .arg(res)
            .arg(&mut *out_q)
            .arg(&mut *out_d)
            .arg(&nc)
            .arg(&e2);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// E4B glue fusion: rms(a, wa) prologue + the add_scale_rms_norm_q8_1 program — one launch
    /// replaces the per-layer rms_norm_f32(y) + emit pair in the PLE tail.
    #[allow(clippy::too_many_arguments)]
    pub fn rms_pre_add_scale_rms_norm_q8_1(
        &self,
        a: &CudaSlice<f32>,
        wa: &CudaSlice<f32>,
        b_in: &CudaSlice<f32>,
        c: f32,
        w: &CudaSlice<f32>,
        res: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(CudaSlice<i8>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        let mut out_q = self.alloc_uninit::<i8>(nrows * ncols)?;
        let mut out_d = self.alloc_uninit::<f32>(nrows * (ncols / 32))?;
        let (nc, e2) = (ncols as i32, eps);
        if Self::pdl_on() {
            {
                use cudarc::driver::{DevicePtr, DevicePtrMut};
                let s = &self.gpu.stream();
                let (pa, _g0) = a.device_ptr(s);
                let (pwa, _g1) = wa.device_ptr(s);
                let (pb, _g2) = b_in.device_ptr(s);
                let (pw, _g3) = w.device_ptr(s);
                let (pr, _g4) = res.device_ptr_mut(s);
                let (pq, _g5) = out_q.device_ptr_mut(s);
                let (pd, _g6) = out_d.device_ptr_mut(s);
                let mut ps = [
                    &pa as *const _ as *mut std::ffi::c_void,
                    &pwa as *const _ as *mut _,
                    &pb as *const _ as *mut _,
                    &c as *const _ as *mut _,
                    &pw as *const _ as *mut _,
                    &pr as *const _ as *mut _,
                    &pq as *const _ as *mut _,
                    &pd as *const _ as *mut _,
                    &nc as *const _ as *mut _,
                    &e2 as *const _ as *mut _,
                ];
                unsafe {
                    self.launch_pdl(
                        "rms_pre_add_scale_rms_norm_q8_1",
                        (nrows as u32, 1, 1),
                        (rms_block(), 1, 1),
                        &mut ps,
                    )?;
                }
            }
            return Ok((out_q, out_d));
        }
        let f = self.func("rms_pre_add_scale_rms_norm_q8_1");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(a)
            .arg(wa)
            .arg(b_in)
            .arg(&c)
            .arg(w)
            .arg(res)
            .arg(&mut out_q)
            .arg(&mut out_d)
            .arg(&nc)
            .arg(&e2);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((out_q, out_d))
    }

    /// GELU(tanh)*up with the activation emitted q8_1 alongside f32 (glue-fusion lane): the
    /// consumer matmul rides matmul_pre, killing its standalone quantize_q8_1 launch.
    pub fn gelu_tanh_mul_q8_1(
        &self,
        gate: &CudaSlice<f32>,
        up: &cudarc::driver::CudaView<f32>,
        act: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
    ) -> Result<(CudaSlice<i8>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        debug_assert!(ncols % 128 == 0);
        let mut out_q = self.alloc_uninit::<i8>(nrows * ncols)?;
        let mut out_d = self.alloc_uninit::<f32>(nrows * (ncols / 32))?;
        let nc = ncols as i32;
        if Self::pdl_on() {
            {
                use cudarc::driver::{DevicePtr, DevicePtrMut};
                let s = &self.gpu.stream();
                let (pg, _g0) = gate.device_ptr(s);
                let (pu, _g1) = up.device_ptr(s);
                let (pact, _g2) = act.device_ptr_mut(s);
                let (pq, _g3) = out_q.device_ptr_mut(s);
                let (pd, _g4) = out_d.device_ptr_mut(s);
                let mut ps = [
                    &pg as *const _ as *mut std::ffi::c_void,
                    &pu as *const _ as *mut _,
                    &pact as *const _ as *mut _,
                    &pq as *const _ as *mut _,
                    &pd as *const _ as *mut _,
                    &nc as *const _ as *mut _,
                ];
                unsafe {
                    self.launch_pdl(
                        "gelu_tanh_mul_q8_1",
                        (nrows as u32, 1, 1),
                        (rms_block(), 1, 1),
                        &mut ps,
                    )?;
                }
            }
            return Ok((out_q, out_d));
        }
        let f = self.func("gelu_tanh_mul_q8_1");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate)
            .arg(up)
            .arg(act)
            .arg(&mut out_q)
            .arg(&mut out_d)
            .arg(&nc);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((out_q, out_d))
    }

    /// Slot-fed gelu_tanh_mul_q8_1 twin (alloc-free capture lane; incl. the PDL arm).
    #[allow(clippy::too_many_arguments)]
    pub fn gelu_tanh_mul_q8_1_into(
        &self,
        gate: &CudaSlice<f32>,
        up: &cudarc::driver::CudaView<f32>,
        act: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        out_q: &mut CudaSlice<i8>,
        out_d: &mut CudaSlice<f32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        debug_assert!(ncols % 128 == 0);
        debug_assert!(out_q.len() >= nrows * ncols && out_d.len() >= nrows * (ncols / 32));
        let nc = ncols as i32;
        if Self::pdl_on() {
            use cudarc::driver::{DevicePtr, DevicePtrMut};
            let s = &self.gpu.stream();
            let (pg, _g0) = gate.device_ptr(s);
            let (pu, _g1) = up.device_ptr(s);
            let (pact, _g2) = act.device_ptr_mut(s);
            let (pq, _g3) = out_q.device_ptr_mut(s);
            let (pd, _g4) = out_d.device_ptr_mut(s);
            let mut ps = [
                &pg as *const _ as *mut std::ffi::c_void,
                &pu as *const _ as *mut _,
                &pact as *const _ as *mut _,
                &pq as *const _ as *mut _,
                &pd as *const _ as *mut _,
                &nc as *const _ as *mut _,
            ];
            unsafe {
                self.launch_pdl(
                    "gelu_tanh_mul_q8_1",
                    (nrows as u32, 1, 1),
                    (rms_block(), 1, 1),
                    &mut ps,
                )?;
            }
            return Ok(());
        }
        let f = self.func("gelu_tanh_mul_q8_1");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate)
            .arg(up)
            .arg(&mut *act)
            .arg(&mut *out_q)
            .arg(&mut *out_d)
            .arg(&nc);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// gemma4: add + rms_norm3 with outputs 0/2 emitted q8_1 (zsh + moe_in) and 1 f32 (router).
    #[allow(clippy::too_many_arguments)]
    pub fn add_rms_norm3_q8z(
        &self,
        a: &CudaSlice<f32>,
        b_in: &CudaSlice<f32>,
        w0: &CudaSlice<f32>,
        w1: &CudaSlice<f32>,
        w2: &CudaSlice<f32>,
        res: &mut CudaSlice<f32>,
        out1: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<
        (
            (CudaSlice<i8>, CudaSlice<f32>),
            (CudaSlice<i8>, CudaSlice<f32>),
        ),
        Box<dyn std::error::Error>,
    > {
        let mut q0 = self.alloc_uninit::<i8>(nrows * ncols)?;
        let mut d0 = self.alloc_uninit::<f32>(nrows * (ncols / 32))?;
        let mut q2 = self.alloc_uninit::<i8>(nrows * ncols)?;
        let mut d2 = self.alloc_uninit::<f32>(nrows * (ncols / 32))?;
        let f = self.func("add_rms_norm3_q8z_f32");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, e2) = (ncols as i32, eps);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(a)
            .arg(b_in)
            .arg(w0)
            .arg(w1)
            .arg(w2)
            .arg(res)
            .arg(&mut q0)
            .arg(&mut d0)
            .arg(out1)
            .arg(&mut q2)
            .arg(&mut d2)
            .arg(&nc)
            .arg(&e2);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(((q0, d0), (q2, d2)))
    }

    /// gemma4: res = a+b AND the three rms_norms of res in one launch.
    #[allow(clippy::too_many_arguments)]
    pub fn add_rms_norm3(
        &self,
        a: &CudaSlice<f32>,
        b_in: &CudaSlice<f32>,
        w0: &CudaSlice<f32>,
        w1: &CudaSlice<f32>,
        w2: &CudaSlice<f32>,
        res: &mut CudaSlice<f32>,
        d0: &mut CudaSlice<f32>,
        d1: &mut CudaSlice<f32>,
        d2: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("add_rms_norm3_f32");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, e2) = (ncols as i32, eps);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(a)
            .arg(b_in)
            .arg(w0)
            .arg(w1)
            .arg(w2)
            .arg(res)
            .arg(d0)
            .arg(d1)
            .arg(d2)
            .arg(&nc)
            .arg(&e2);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// dst = (a + b) * c (residual add + layer scale, one launch).
    pub fn add_scale(
        &self,
        a: &CudaSlice<f32>,
        b_in: &CudaSlice<f32>,
        c: f32,
        dst: &mut CudaSlice<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("add_scale_f32");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let ni = n as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(a).arg(b_in).arg(&c).arg(dst).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Vision-tower LayerNorm (with bias) over [nrows, ncols] — lane/vision.
    pub fn layer_norm_bias(
        &self,
        x: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        b: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("layer_norm_bias_f32");
        let (nc, e) = (ncols as i32, eps);
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut lb = __s_b.launch_builder(&f);
        lb.arg(x).arg(w).arg(b).arg(&mut *dst).arg(&nc).arg(&e);
        unsafe {
            lb.launch(cfg)?;
        }
        Ok(())
    }

    /// gelu_pytorch_tanh elementwise (vision tower MLP activation).
    pub fn gelu_tanh(
        &self,
        x: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gelu_tanh_f32");
        let ni = n as i64;
        let cfg = LaunchConfig {
            grid_dim: (n.div_ceil(256) as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut lb = __s_b.launch_builder(&f);
        lb.arg(x).arg(&mut *dst).arg(&ni);
        unsafe {
            lb.launch(cfg)?;
        }
        Ok(())
    }

    /// In-place row softmax over [nrows, ncols] (bidirectional vision attention).
    pub fn row_softmax(
        &self,
        x: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("row_softmax_f32");
        let nc = ncols as i32;
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut lb = __s_b.launch_builder(&f);
        lb.arg(&mut *x).arg(&nc);
        unsafe {
            lb.launch(cfg)?;
        }
        Ok(())
    }

    pub fn rms_norm(
        &self,
        x: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let (nc, e) = (ncols as i32, eps);
        let kname = if Self::norm_ilp_on() {
            "rms_norm_f32_v2"
        } else {
            "rms_norm_f32"
        };
        if Self::pdl_on() && Self::pdl_wb_on() {
            use cudarc::driver::{DevicePtr, DevicePtrMut};
            let s = &self.gpu.stream();
            let (px, _g0) = x.device_ptr(s);
            let (pw, _g1) = w.device_ptr(s);
            let (pd, _g2) = dst.device_ptr_mut(s);
            let mut ps = [
                &px as *const _ as *mut std::ffi::c_void,
                &pw as *const _ as *mut _,
                &pd as *const _ as *mut _,
                &nc as *const _ as *mut _,
                &e as *const _ as *mut _,
            ];
            unsafe {
                self.launch_pdl(kname, (nrows as u32, 1, 1), (rms_block(), 1, 1), &mut ps)?;
            }
            return Ok(());
        }
        let f = self.func(kname);
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(w).arg(dst).arg(&nc).arg(&e);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// RMS-norm with blockDim=1024 — BIT-IDENTICAL to the fused `rms_norm_q8_1` and
    /// `add_rms_norm_q8_1` kernels' sum-of-squares reduction. The spec verify path MUST use this
    /// to match decode's FP accumulation order: the standard `rms_norm` at blockDim=256 has a
    /// different per-thread stride (ncols/256 partials vs ncols/1024 partials) and therefore a
    /// different shfl-tree reduction that can shift `scale = rsqrt(sum/n + eps)` by ULPs, causing
    /// divergence through the GDN scan and argmax flips on the 9B text prompt. The underlying
    /// `rms_norm_f32` kernel supports any blockDim (generic reduce with shared[32]).
    pub fn rms_norm_decode(
        &self,
        x: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func(if Self::norm_ilp_on() {
            "rms_norm_f32_v2"
        } else {
            "rms_norm_f32"
        });
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (1024, 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, e) = (ncols as i32, eps);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(w).arg(dst).arg(&nc).arg(&e);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// DECODE GLUE-FUSION LEVER: `z = rms_norm(x)*w` emitted DIRECTLY as q8_1 (no f32 `z` materialized,
    /// no standalone quantize_q8_1 launch). Returns (out_q [nrows*ncols i8], out_d [nrows*nblk f32])
    /// ready to feed matmul_pre. BIT-IDENTICAL to rms_norm + quantize_q8_1. ncols % 32 == 0.
    pub fn rms_norm_q8_1(
        &self,
        x: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(CudaSlice<i8>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        let nblk = ncols / 32;
        let mut q = self.alloc_uninit::<i8>(nrows * ncols)?;
        let mut d = self.alloc_uninit::<f32>(nrows * nblk)?;
        let (nc, e) = (ncols as i32, eps);
        if Self::pdl_on() {
            {
                use cudarc::driver::{DevicePtr, DevicePtrMut};
                let s = &self.gpu.stream();
                let (px, _g0) = x.device_ptr(s);
                let (pw, _g1) = w.device_ptr(s);
                let (pq, _g2) = q.device_ptr_mut(s);
                let (pd, _g3) = d.device_ptr_mut(s);
                let mut ps = [
                    &px as *const _ as *mut std::ffi::c_void,
                    &pw as *const _ as *mut _,
                    &pq as *const _ as *mut _,
                    &pd as *const _ as *mut _,
                    &nc as *const _ as *mut _,
                    &e as *const _ as *mut _,
                ];
                unsafe {
                    self.launch_pdl("rms_norm_q8_1", (nrows as u32, 1, 1), (1024, 1, 1), &mut ps)?;
                }
            }
            return Ok((q, d));
        }
        let f = self.func("rms_norm_q8_1");
        // 1024 threads: decode is nrows=1 -> ONE CTA; 32 warps hide the pass1->pass2 latency
        // (s[32] reduce already sized for 32 warps). Same shape math at any blockDim.
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (1024, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(w).arg(&mut q).arg(&mut d).arg(&nc).arg(&e);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((q, d))
    }

    /// Slot-fed rms_norm_q8_1 twin (alloc-free capture lane): identical launch (incl. the
    /// PDL arm), caller-owned outputs.
    pub fn rms_norm_q8_1_into(
        &self,
        x: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
        q: &mut CudaSlice<i8>,
        d: &mut CudaSlice<f32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let nblk = ncols / 32;
        debug_assert!(q.len() >= nrows * ncols && d.len() >= nrows * nblk);
        let (nc, e) = (ncols as i32, eps);
        if Self::pdl_on() {
            use cudarc::driver::{DevicePtr, DevicePtrMut};
            let s = &self.gpu.stream();
            let (px, _g0) = x.device_ptr(s);
            let (pw, _g1) = w.device_ptr(s);
            let (pq, _g2) = q.device_ptr_mut(s);
            let (pd, _g3) = d.device_ptr_mut(s);
            let mut ps = [
                &px as *const _ as *mut std::ffi::c_void,
                &pw as *const _ as *mut _,
                &pq as *const _ as *mut _,
                &pd as *const _ as *mut _,
                &nc as *const _ as *mut _,
                &e as *const _ as *mut _,
            ];
            unsafe {
                self.launch_pdl("rms_norm_q8_1", (nrows as u32, 1, 1), (1024, 1, 1), &mut ps)?;
            }
            return Ok(());
        }
        let f = self.func("rms_norm_q8_1");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (1024, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(w).arg(&mut *q).arg(&mut *d).arg(&nc).arg(&e);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Slot-fed quantize_q8_1 twin (alloc-free capture lane).
    pub fn quantize_q8_1_into(
        &self,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        q: &mut CudaSlice<i8>,
        d: &mut CudaSlice<f32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let nblk = in_f / 32;
        debug_assert!(q.len() >= m * in_f && d.len() >= m * nblk);
        let cfg = LaunchConfig::for_num_elems((m * in_f) as u32);
        let (inf, mi) = (in_f as i32, m as i32);
        if Self::pdl_on() && Self::pdl_wb_on() {
            use cudarc::driver::{DevicePtr, DevicePtrMut};
            let s = &self.gpu.stream();
            let (px, _g0) = x.device_ptr(s);
            let (pq, _g1) = q.device_ptr_mut(s);
            let (pd, _g2) = d.device_ptr_mut(s);
            let mut ps = [
                &px as *const _ as *mut std::ffi::c_void,
                &pq as *const _ as *mut _,
                &pd as *const _ as *mut _,
                &inf as *const _ as *mut _,
                &mi as *const _ as *mut _,
            ];
            unsafe {
                self.launch_pdl("quantize_q8_1", cfg.grid_dim, cfg.block_dim, &mut ps)?;
            }
            return Ok(());
        }
        let f = self.func("quantize_q8_1");
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(&mut *q).arg(&mut *d).arg(&inf).arg(&mi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// DECODE GLUE-FUSION LEVER: `res = a+b; z = rms_norm(res)*w` with z emitted as q8_1. `res` is
    /// still written (the post-ffn residual add reads it). Fuses add_rms_norm + quantize_q8_1.
    /// Returns (out_q, out_d) for matmul_pre. BIT-IDENTICAL. ncols % 32 == 0.
    pub fn add_rms_norm_q8_1(
        &self,
        a: &CudaSlice<f32>,
        b_in: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        res: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(CudaSlice<i8>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        let nblk = ncols / 32;
        let mut q = self.alloc_uninit::<i8>(nrows * ncols)?;
        let mut d = self.alloc_uninit::<f32>(nrows * nblk)?;
        let f = self.func("add_rms_norm_q8_1");
        // 1024 threads: same single-CTA-at-decode reasoning as rms_norm_q8_1.
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (1024, 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, e) = (ncols as i32, eps);
        let __s_bld = self.gpu.stream();
        let mut bld = __s_bld.launch_builder(&f);
        bld.arg(a)
            .arg(b_in)
            .arg(w)
            .arg(res)
            .arg(&mut q)
            .arg(&mut d)
            .arg(&nc)
            .arg(&e);
        unsafe {
            bld.launch(cfg)?;
        }
        Ok((q, d))
    }

    /// RANK3 LEVER (add+rmsnorm fuse): `res = a + b; dst = rms_norm(res) * w` in ONE launch. Fuses
    /// e.add(a,b,res) + e.rms_norm(res,w,dst), removing one launch + one HBM read of the residual per
    /// residual+norm pair. BIT-IDENTICAL to the two-kernel sequence (same IEEE add, same reduction).
    /// O-PROJ TAIL FUSION M2: mixed = a0+a1 composed in-register, then the VERBATIM
    /// add_rms_norm program. Raw UVA pointers for the join partials (persistent ws rows).
    #[allow(clippy::too_many_arguments)]
    pub fn join_add_rms_norm_raw(
        &self,
        a0_raw: u64,
        a1_raw: u64,
        x: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        res: &mut CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if a0_raw == 0 || a1_raw == 0 || x.len() < ncols || res.len() < ncols || dst.len() < ncols {
            return Err("join_add_rms_norm geometry".into());
        }
        let f = self.func("join_add_rms_norm_f32");
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, e) = (ncols as i32, eps);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(&a0_raw)
            .arg(&a1_raw)
            .arg(x)
            .arg(w)
            .arg(&mut *res)
            .arg(&mut *dst)
            .arg(&nc)
            .arg(&e);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn add_rms_norm(
        &self,
        a: &CudaSlice<f32>,
        b: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        res: &mut CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let (nc, e) = (ncols as i32, eps);
        let kname = if Self::norm_ilp_on() {
            "add_rms_norm_f32_v2"
        } else {
            "add_rms_norm_f32"
        };
        if Self::pdl_on() && Self::pdl_wb_on() {
            use cudarc::driver::{DevicePtr, DevicePtrMut};
            let s = &self.gpu.stream();
            let (pa, _g0) = a.device_ptr(s);
            let (pb, _g1) = b.device_ptr(s);
            let (pw, _g2) = w.device_ptr(s);
            let (pr, _g3) = res.device_ptr_mut(s);
            let (pd, _g4) = dst.device_ptr_mut(s);
            let mut ps = [
                &pa as *const _ as *mut std::ffi::c_void,
                &pb as *const _ as *mut _,
                &pw as *const _ as *mut _,
                &pr as *const _ as *mut _,
                &pd as *const _ as *mut _,
                &nc as *const _ as *mut _,
                &e as *const _ as *mut _,
            ];
            unsafe {
                self.launch_pdl(kname, (nrows as u32, 1, 1), (rms_block(), 1, 1), &mut ps)?;
            }
            return Ok(());
        }
        let f = self.func(kname);
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&f);
        b2.arg(a)
            .arg(b)
            .arg(w)
            .arg(&mut *res)
            .arg(&mut *dst)
            .arg(&nc)
            .arg(&e);
        unsafe {
            b2.launch(cfg)?;
        }
        Ok(())
    }

    /// E4B glue fusion: rms(a, wa) prologue + add_rms_norm — folds the post-attn norm into
    /// the tail entry (res = rms(a)*wa + b; dst = rms(res)*w).
    #[allow(clippy::too_many_arguments)]
    pub fn rms_pre_add_rms_norm(
        &self,
        a: &CudaSlice<f32>,
        wa: &CudaSlice<f32>,
        b: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        res: &mut CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("rms_pre_add_rms_norm_f32");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, e) = (ncols as i32, eps);
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&f);
        b2.arg(a)
            .arg(wa)
            .arg(b)
            .arg(w)
            .arg(&mut *res)
            .arg(&mut *dst)
            .arg(&nc)
            .arg(&e);
        unsafe {
            b2.launch(cfg)?;
        }
        Ok(())
    }

    /// wave-2 fold: rms(a,wa) + add + ffn-norm with zsh EMITTED q8_1 (fused2 consumes it).
    #[allow(clippy::too_many_arguments)]
    pub fn rms_pre_add_rms_norm_q8z(
        &self,
        a: &CudaSlice<f32>,
        wa: &CudaSlice<f32>,
        b: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        res: &mut CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(CudaSlice<i8>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        debug_assert!(ncols % 128 == 0);
        let mut out_q = self.alloc_uninit::<i8>(nrows * ncols)?;
        let mut out_d = self.alloc_uninit::<f32>(nrows * (ncols / 32))?;
        let (nc, e) = (ncols as i32, eps);
        if Self::pdl_on() {
            {
                use cudarc::driver::{DevicePtr, DevicePtrMut};
                let s = &self.gpu.stream();
                let (pa, _g0) = a.device_ptr(s);
                let (pwa, _g1) = wa.device_ptr(s);
                let (pb, _g2) = b.device_ptr(s);
                let (pw, _g3) = w.device_ptr(s);
                let (pr, _g4) = res.device_ptr_mut(s);
                let (pdst, _g5) = dst.device_ptr_mut(s);
                let (pq, _g6) = out_q.device_ptr_mut(s);
                let (pd, _g7) = out_d.device_ptr_mut(s);
                let mut ps = [
                    &pa as *const _ as *mut std::ffi::c_void,
                    &pwa as *const _ as *mut _,
                    &pb as *const _ as *mut _,
                    &pw as *const _ as *mut _,
                    &pr as *const _ as *mut _,
                    &pdst as *const _ as *mut _,
                    &pq as *const _ as *mut _,
                    &pd as *const _ as *mut _,
                    &nc as *const _ as *mut _,
                    &e as *const _ as *mut _,
                ];
                unsafe {
                    self.launch_pdl(
                        "rms_pre_add_rms_norm_q8z_f32",
                        (nrows as u32, 1, 1),
                        (rms_block(), 1, 1),
                        &mut ps,
                    )?;
                }
            }
            return Ok((out_q, out_d));
        }
        let f = self.func("rms_pre_add_rms_norm_q8z_f32");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&f);
        b2.arg(a)
            .arg(wa)
            .arg(b)
            .arg(w)
            .arg(&mut *res)
            .arg(&mut *dst)
            .arg(&mut out_q)
            .arg(&mut out_d)
            .arg(&nc)
            .arg(&e);
        unsafe {
            b2.launch(cfg)?;
        }
        Ok((out_q, out_d))
    }

    /// Slot-fed twin of `rms_pre_add_rms_norm_q8z` (gemma4 pn-fold, slotted/graph arm):
    /// identical kernel, caller-owned outputs, PLAIN launch only — the dc_slotted capture
    /// body must stay attribute-free (the fused2_into precedent).
    #[allow(clippy::too_many_arguments)]
    pub fn rms_pre_add_rms_norm_q8z_into(
        &self,
        a: &CudaSlice<f32>,
        wa: &CudaSlice<f32>,
        b: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        res: &mut CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
        out_q: &mut CudaSlice<i8>,
        out_d: &mut CudaSlice<f32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        debug_assert!(ncols % 128 == 0);
        let (nc, e) = (ncols as i32, eps);
        let f = self.func("rms_pre_add_rms_norm_q8z_f32");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b2 = __s_b.launch_builder(&f);
        b2.arg(a)
            .arg(wa)
            .arg(b)
            .arg(w)
            .arg(&mut *res)
            .arg(&mut *dst)
            .arg(&mut *out_q)
            .arg(&mut *out_d)
            .arg(&nc)
            .arg(&e);
        unsafe {
            b2.launch(cfg)?;
        }
        Ok(())
    }

    /// Slot-fed twin of `rms_pre_add_scale_rms_norm_q8_1` (gemma4 pn-fold exit, slotted
    /// arm): identical kernel, caller-owned outputs, PLAIN launch (capture-safe).
    #[allow(clippy::too_many_arguments)]
    pub fn rms_pre_add_scale_rms_norm_q8_1_into(
        &self,
        a: &CudaSlice<f32>,
        wa: &CudaSlice<f32>,
        b_in: &CudaSlice<f32>,
        c: f32,
        w: &CudaSlice<f32>,
        res: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
        out_q: &mut CudaSlice<i8>,
        out_d: &mut CudaSlice<f32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        debug_assert!(ncols % 128 == 0);
        let (nc, e2) = (ncols as i32, eps);
        let f = self.func("rms_pre_add_scale_rms_norm_q8_1");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b2 = __s_b.launch_builder(&f);
        b2.arg(a)
            .arg(wa)
            .arg(b_in)
            .arg(&c)
            .arg(w)
            .arg(&mut *res)
            .arg(&mut *out_q)
            .arg(&mut *out_d)
            .arg(&nc)
            .arg(&e2);
        unsafe {
            b2.launch(cfg)?;
        }
        Ok(())
    }

    /// gemma4 pn-fold seam (GAP-DIAGNOSIS verdict 7, the E4B glue backport): the dense
    /// decode/verify/slotted trio folds post_attn_norm into the tail entry
    /// (rms_pre_add_rms_norm[_q8z]) and post_ffw_norm into the residual exit
    /// (rms_pre_add_scale_rms_norm_q8_1). BITS-CHANGING vs the two-launch chain (the
    /// single-phase reduction's expansion rounding — E4B receipts); every arm moves
    /// together so decode == verify == graph parity holds BY CONSTRUCTION within either
    /// seam value. MEMRA_G4_PNFOLD=0 restores the unfused chain everywhere.
    pub fn g4_pnfold_on() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| {
            std::env::var("MEMRA_G4_PNFOLD")
                .map(|v| v != "0")
                .unwrap_or(true)
        })
    }

    /// wave-4b: OUT-dim concat of three Q4_0 tensors (same in_features; rows are independent
    /// blocks, so the concat is a D2D byte concat of the GGUF-layout planes). Returns None
    /// off-class (non-Q4_0, mismatched widths, or any tensor already rp-swapped in place).
    pub fn build_q4_out_concat3(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        w2: &crate::model::GpuTensor,
    ) -> Result<Option<crate::model::GpuTensor>, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        let part = |w: &GpuTensor| -> Option<(usize, usize)> {
            match w {
                GpuTensor::Quant {
                    qtype,
                    row_bytes,
                    rp,
                    ..
                } if *qtype == QT_Q4_0 && !*rp => Some((*row_bytes, w.out_features())),
                _ => None,
            }
        };
        let (Some((rb0, o0)), Some((rb1, o1)), Some((rb2, o2))) = (part(w0), part(w1), part(w2))
        else {
            return Ok(None);
        };
        if rb0 != rb1
            || rb0 != rb2
            || w0.in_features() != w1.in_features()
            || w0.in_features() != w2.in_features()
        {
            return Ok(None);
        }
        fn bytes_of(w: &crate::model::GpuTensor) -> &CudaSlice<u8> {
            match w {
                crate::model::GpuTensor::Quant { bytes, .. } => bytes,
                _ => unreachable!(),
            }
        }
        let (b0, b1, b2) = (bytes_of(w0), bytes_of(w1), bytes_of(w2));
        let total = rb0 * (o0 + o1 + o2);
        let mut cat = self.alloc_u8(total)?;
        self.copy_u8_into(&mut cat, 0, b0, rb0 * o0)?;
        self.copy_u8_into(&mut cat, rb0 * o0, b1, rb1 * o1)?;
        self.copy_u8_into(&mut cat, rb0 * (o0 + o1), b2, rb2 * o2)?;
        Ok(Some(GpuTensor::Quant {
            bytes: cat,
            qtype: QT_Q4_0,
            row_bytes: rb0,
            ne: vec![w0.in_features() as u64, (o0 + o1 + o2) as u64],
            scale: 1.0,
            rp: false,
            #[cfg(memra_cutlass)]
            cutlass: None,
            fp8: None,
            blk: None,
            rp4: None,
            f16: None,
        }))
    }

    /// FULL-WIDTH-ROPE CONTRACT for the fused rms_norm+qkv+rope kernels
    /// (lane/graph-s-key-exactness-20260819, probe O-10 follow-up).
    ///
    /// `rms_norm_qkv_rope_f32` / `_cat_f32` (`cu/kernels.cu`) and
    /// `rms_norm_qkv_rope_append_body` (`cu/flash_attn.cu`) take NO `n_dims`/`n_rot` argument.
    /// They compute `int half = ncols / 2` and rotate the FULL head width by construction — the
    /// standalone `rope_neox*` kernels take `n_dims` and early-return above it, these do not.
    ///
    /// Every call site today is a gemma-4 arm, where `n_rot == head_dim`, so nothing is wrong
    /// now. But this is the n_rot bug ONE FUSION away: fuse a partial-rotary arch onto these
    /// kernels (qwen3.5 = 64 rotary dims of a 256-wide head; step35 full-attn = 64 of 128) and
    /// 192 dims that must pass through unrotated get rotated silently — no error, no NaN, just a
    /// wrong model. The n_rot lane already paid for that class once, in the config derivation.
    ///
    /// So the fusions now take the layer's DERIVED rope width and refuse anything but full
    /// width. A future partial-rotary caller fails at its first launch with the geometry named
    /// instead of serving quietly wrong logits.
    fn full_width_rope_only(
        kernel: &str,
        n_rot: usize,
        head_dim: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if n_rot == head_dim {
            return Ok(());
        }
        Err(format!(
            "{kernel}: PARTIAL ROTARY REFUSED — n_rot {n_rot} != head_dim {head_dim}. This fused \
             rms_norm+qkv+rope kernel carries no n_dims parameter and rotates the full head \
             width (half = ncols/2), so it would rotate dims {n_rot}..{head_dim} that must pass \
             through unrotated. Use the split path (rms_norm_qkv + rope_neox/rope_neox2 with \
             n_dims={n_rot}), or add an n_dims early-return to the kernel and widen this guard."
        )
        .into())
    }

    /// wave-4b: the qkv-cat twin — one contiguous [rq+2*rk, hd] input from the concat matvec.
    /// `n_rot` is the layer's derived rotary width and MUST equal `head_dim`
    /// ([`Engine::full_width_rope_only`]).
    #[allow(clippy::too_many_arguments)]
    pub fn rms_norm_qkv_rope_cat(
        &self,
        qkv: &CudaSlice<f32>,
        wq: &CudaSlice<f32>,
        wk: &CudaSlice<f32>,
        wv: &CudaSlice<f32>,
        q: &mut CudaSlice<f32>,
        k: &mut CudaSlice<f32>,
        v: &mut CudaSlice<f32>,
        head_dim: usize,
        n_rot: usize,
        rq: usize,
        rk: usize,
        pos: &CudaSlice<i32>,
        nh_q: usize,
        nh_k: usize,
        base: f32,
        freq_scale: f32,
        ff: Option<&CudaSlice<f32>>,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        Self::full_width_rope_only("rms_norm_qkv_rope_cat", n_rot, head_dim)?;
        let rows = rq + rk + rk;
        let theta_scale = base.powf(-2.0 / head_dim as f32);
        let (nc, rqi, rki, nhq, nhk) = (
            head_dim as i32,
            rq as i32,
            rk as i32,
            nh_q as i32,
            nh_k as i32,
        );
        if Self::pdl_on() {
            use cudarc::driver::{DevicePtr, DevicePtrMut};
            let s = &self.gpu.stream();
            let (pqkv, _g0) = qkv.device_ptr(s);
            let (pwq, _g1) = wq.device_ptr(s);
            let (pwk, _g2) = wk.device_ptr(s);
            let (pwv, _g3) = wv.device_ptr(s);
            let (pq, _g4) = q.device_ptr_mut(s);
            let (pk, _g5) = k.device_ptr_mut(s);
            let (pv, _g6) = v.device_ptr_mut(s);
            let (ppos, _g7) = pos.device_ptr(s);
            let (pff, _g8) = match ff {
                Some(t) => {
                    let (p, g) = t.device_ptr(s);
                    (p, Some(g))
                }
                None => (0, None),
            };
            let mut ps = [
                &pqkv as *const _ as *mut std::ffi::c_void,
                &pwq as *const _ as *mut _,
                &pwk as *const _ as *mut _,
                &pwv as *const _ as *mut _,
                &pq as *const _ as *mut _,
                &pk as *const _ as *mut _,
                &pv as *const _ as *mut _,
                &nc as *const _ as *mut _,
                &rqi as *const _ as *mut _,
                &rki as *const _ as *mut _,
                &ppos as *const _ as *mut _,
                &nhq as *const _ as *mut _,
                &nhk as *const _ as *mut _,
                &theta_scale as *const _ as *mut _,
                &freq_scale as *const _ as *mut _,
                &pff as *const _ as *mut _,
                &eps as *const _ as *mut _,
            ];
            unsafe {
                self.launch_pdl(
                    "rms_norm_qkv_rope_cat_f32",
                    (rows as u32, 1, 1),
                    (rms_block(), 1, 1),
                    &mut ps,
                )?;
            }
            return Ok(());
        }
        let f = self.func("rms_norm_qkv_rope_cat_f32");
        let cfg = LaunchConfig {
            grid_dim: (rows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        match ff {
            Some(t) => {
                b.arg(qkv)
                    .arg(wq)
                    .arg(wk)
                    .arg(wv)
                    .arg(&mut *q)
                    .arg(&mut *k)
                    .arg(&mut *v)
                    .arg(&nc)
                    .arg(&rqi)
                    .arg(&rki)
                    .arg(pos)
                    .arg(&nhq)
                    .arg(&nhk)
                    .arg(&theta_scale)
                    .arg(&freq_scale)
                    .arg(t)
                    .arg(&eps);
                unsafe {
                    b.launch(cfg)?;
                }
            }
            None => {
                let null: u64 = 0;
                b.arg(qkv)
                    .arg(wq)
                    .arg(wk)
                    .arg(wv)
                    .arg(&mut *q)
                    .arg(&mut *k)
                    .arg(&mut *v)
                    .arg(&nc)
                    .arg(&rqi)
                    .arg(&rki)
                    .arg(pos)
                    .arg(&nhq)
                    .arg(&nhk)
                    .arg(&theta_scale)
                    .arg(&freq_scale)
                    .arg(&null)
                    .arg(&eps);
                unsafe {
                    b.launch(cfg)?;
                }
            }
        }
        Ok(())
    }

    /// wave-3 fold: rms_norm_qkv + rope_neox2 in ONE launch (n_dims == head_dim; ff nullable).
    /// `n_rot` is the layer's derived rotary width and MUST equal `head_dim`
    /// ([`Engine::full_width_rope_only`]).
    #[allow(clippy::too_many_arguments)]
    pub fn rms_norm_qkv_rope(
        &self,
        q0: &CudaSlice<f32>,
        k0: &CudaSlice<f32>,
        v0: &CudaSlice<f32>,
        wq: &CudaSlice<f32>,
        wk: &CudaSlice<f32>,
        wv: &CudaSlice<f32>,
        q: &mut CudaSlice<f32>,
        k: &mut CudaSlice<f32>,
        v: &mut CudaSlice<f32>,
        head_dim: usize,
        n_rot: usize,
        rq: usize,
        rk: usize,
        pos: &CudaSlice<i32>,
        nh_q: usize,
        nh_k: usize,
        base: f32,
        freq_scale: f32,
        ff: Option<&CudaSlice<f32>>,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        Self::full_width_rope_only("rms_norm_qkv_rope", n_rot, head_dim)?;
        let f = self.func("rms_norm_qkv_rope_f32");
        let rows = rq + rk + rk; // q rows + k rows + v rows (rk == rv)
        let cfg = LaunchConfig {
            grid_dim: (rows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let theta_scale = base.powf(-2.0 / head_dim as f32);
        let (nc, rqi, rki, nhq, nhk) = (
            head_dim as i32,
            rq as i32,
            rk as i32,
            nh_q as i32,
            nh_k as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        match ff {
            Some(t) => {
                b.arg(q0)
                    .arg(k0)
                    .arg(v0)
                    .arg(wq)
                    .arg(wk)
                    .arg(wv)
                    .arg(&mut *q)
                    .arg(&mut *k)
                    .arg(&mut *v)
                    .arg(&nc)
                    .arg(&rqi)
                    .arg(&rki)
                    .arg(pos)
                    .arg(&nhq)
                    .arg(&nhk)
                    .arg(&theta_scale)
                    .arg(&freq_scale)
                    .arg(t)
                    .arg(&eps);
                unsafe {
                    b.launch(cfg)?;
                }
            }
            None => {
                let null: u64 = 0;
                b.arg(q0)
                    .arg(k0)
                    .arg(v0)
                    .arg(wq)
                    .arg(wk)
                    .arg(wv)
                    .arg(&mut *q)
                    .arg(&mut *k)
                    .arg(&mut *v)
                    .arg(&nc)
                    .arg(&rqi)
                    .arg(&rki)
                    .arg(pos)
                    .arg(&nhq)
                    .arg(&nhk)
                    .arg(&theta_scale)
                    .arg(&freq_scale)
                    .arg(&null)
                    .arg(&eps);
                unsafe {
                    b.launch(cfg)?;
                }
            }
        }
        Ok(())
    }

    /// FUSED norm+rope+APPEND (m=1 decode, 2026-07-23): one launch replaces the
    /// rms_norm_qkv_rope + append_kv_quantized_dc pair. Kernel lives in the flash fatbins
    /// (format-flavored quant tail) — `g` must mirror the append path's flavor exactly.
    /// `n_rot` is the layer's derived rotary width and MUST equal `head_dim`
    /// ([`Engine::full_width_rope_only`]).
    #[allow(clippy::too_many_arguments)]
    pub fn rms_norm_qkv_rope_append_dc(
        &self,
        q0: &CudaSlice<f32>,
        k0: &CudaSlice<f32>,
        v0: &CudaSlice<f32>,
        wq: &CudaSlice<f32>,
        wk: &CudaSlice<f32>,
        wv: &CudaSlice<f32>,
        q: &mut CudaSlice<f32>,
        k: &mut CudaSlice<f32>,
        v: &mut CudaSlice<f32>,
        head_dim: usize,
        n_rot: usize,
        rq: usize,
        rk: usize,
        pos: &CudaSlice<i32>,
        nh_q: usize,
        nh_k: usize,
        base: f32,
        freq_scale: f32,
        ff: Option<&CudaSlice<f32>>,
        eps: f32,
        kc: &mut CudaSlice<u8>,
        vc: &mut CudaSlice<u8>,
        t_dev: &CudaSlice<i32>,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        g: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        Self::full_width_rope_only("rms_norm_qkv_rope_append_dc", n_rot, head_dim)?;
        let rows = rq + rk + rk;
        let theta_scale = base.powf(-2.0 / head_dim as f32);
        let (nc, rqi, rki, nhq, nhk) = (
            head_dim as i32,
            rq as i32,
            rk as i32,
            nh_q as i32,
            nh_k as i32,
        );
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        if Self::pdl_on() && Self::pdl_wb_on() {
            use cudarc::driver::{DevicePtr, DevicePtrMut};
            let s = &self.gpu.stream();
            let (p0, _a0) = q0.device_ptr(s);
            let (p1, _a1) = k0.device_ptr(s);
            let (p2, _a2) = v0.device_ptr(s);
            let (pwq, _a3) = wq.device_ptr(s);
            let (pwk, _a4) = wk.device_ptr(s);
            let (pwv, _a5) = wv.device_ptr(s);
            let (pq, _a6) = q.device_ptr_mut(s);
            let (pk, _a7) = k.device_ptr_mut(s);
            let (pv, _a8) = v.device_ptr_mut(s);
            let (pp, _a9) = pos.device_ptr(s);
            let pff: u64 = match ff {
                Some(t) => {
                    let (p, _gg) = t.device_ptr(s);
                    p as u64
                }
                None => 0,
            };
            let (pkc, _a10) = kc.device_ptr_mut(s);
            let (pvc, _a11) = vc.device_ptr_mut(s);
            let (pt, _a12) = t_dev.device_ptr(s);
            let mut ps = [
                &p0 as *const _ as *mut std::ffi::c_void,
                &p1 as *const _ as *mut _,
                &p2 as *const _ as *mut _,
                &pwq as *const _ as *mut _,
                &pwk as *const _ as *mut _,
                &pwv as *const _ as *mut _,
                &pq as *const _ as *mut _,
                &pk as *const _ as *mut _,
                &pv as *const _ as *mut _,
                &nc as *const _ as *mut _,
                &rqi as *const _ as *mut _,
                &rki as *const _ as *mut _,
                &pp as *const _ as *mut _,
                &nhq as *const _ as *mut _,
                &nhk as *const _ as *mut _,
                &theta_scale as *const _ as *mut _,
                &freq_scale as *const _ as *mut _,
                &pff as *const _ as *mut _,
                &eps as *const _ as *mut _,
                &pkc as *const _ as *mut _,
                &pvc as *const _ as *mut _,
                &pt as *const _ as *mut _,
                &ktb as *const _ as *mut _,
                &vtb as *const _ as *mut _,
            ];
            unsafe {
                self.launch_pdl_flash(
                    g,
                    "rms_norm_qkv_rope_append_dc_f32",
                    (rows as u32, 1, 1),
                    (rms_block(), 1, 1),
                    0,
                    &mut ps,
                )?;
            }
            return Ok(());
        }
        let f = if g {
            self.func_g("rms_norm_qkv_rope_append_dc_f32")
        } else {
            self.func("rms_norm_qkv_rope_append_dc_f32")
        };
        let cfg = LaunchConfig {
            grid_dim: (rows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        match ff {
            Some(t) => {
                b.arg(q0)
                    .arg(k0)
                    .arg(v0)
                    .arg(wq)
                    .arg(wk)
                    .arg(wv)
                    .arg(&mut *q)
                    .arg(&mut *k)
                    .arg(&mut *v)
                    .arg(&nc)
                    .arg(&rqi)
                    .arg(&rki)
                    .arg(pos)
                    .arg(&nhq)
                    .arg(&nhk)
                    .arg(&theta_scale)
                    .arg(&freq_scale)
                    .arg(t)
                    .arg(&eps)
                    .arg(&mut *kc)
                    .arg(&mut *vc)
                    .arg(t_dev)
                    .arg(&ktb)
                    .arg(&vtb);
                unsafe {
                    b.launch(cfg)?;
                }
            }
            None => {
                let null: u64 = 0;
                b.arg(q0)
                    .arg(k0)
                    .arg(v0)
                    .arg(wq)
                    .arg(wk)
                    .arg(wv)
                    .arg(&mut *q)
                    .arg(&mut *k)
                    .arg(&mut *v)
                    .arg(&nc)
                    .arg(&rqi)
                    .arg(&rki)
                    .arg(pos)
                    .arg(&nhq)
                    .arg(&nhk)
                    .arg(&theta_scale)
                    .arg(&freq_scale)
                    .arg(&null)
                    .arg(&eps)
                    .arg(&mut *kc)
                    .arg(&mut *vc)
                    .arg(t_dev)
                    .arg(&ktb)
                    .arg(&vtb);
                unsafe {
                    b.launch(cfg)?;
                }
            }
        }
        Ok(())
    }

    /// wave-2 fold: a + b with the sum emitted q8_1 alongside f32.
    /// Host-len twin of `rms_norm_qkv_rope_append_dc` (zoo-fusion arc): the eager decode
    /// arm tracks kv length on host (`kvl.len`), so the append slot rides the launch arg
    /// instead of the device counter. Kernel body is SHARED with the _dc entry (one
    /// inlined body — bit-identical to the rms_norm_qkv_rope + append pair it replaces,
    /// same law as the dc fold). `n_rot` is the layer's derived rotary width and MUST equal
    /// `head_dim` ([`Engine::full_width_rope_only`]).
    #[allow(clippy::too_many_arguments)]
    pub fn rms_norm_qkv_rope_append(
        &self,
        q0: &CudaSlice<f32>,
        k0: &CudaSlice<f32>,
        v0: &CudaSlice<f32>,
        wq: &CudaSlice<f32>,
        wk: &CudaSlice<f32>,
        wv: &CudaSlice<f32>,
        q: &mut CudaSlice<f32>,
        k: &mut CudaSlice<f32>,
        v: &mut CudaSlice<f32>,
        head_dim: usize,
        n_rot: usize,
        rq: usize,
        rk: usize,
        pos: &CudaSlice<i32>,
        nh_q: usize,
        nh_k: usize,
        base: f32,
        freq_scale: f32,
        ff: Option<&CudaSlice<f32>>,
        eps: f32,
        kc: &mut CudaSlice<u8>,
        vc: &mut CudaSlice<u8>,
        t: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        g: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        Self::full_width_rope_only("rms_norm_qkv_rope_append", n_rot, head_dim)?;
        let rows = rq + rk + rk;
        let theta_scale = base.powf(-2.0 / head_dim as f32);
        let (nc, rqi, rki, nhq, nhk) = (
            head_dim as i32,
            rq as i32,
            rk as i32,
            nh_q as i32,
            nh_k as i32,
        );
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let ti = t as i32;
        if Self::pdl_on() && Self::pdl_wb_on() {
            use cudarc::driver::{DevicePtr, DevicePtrMut};
            let s = &self.gpu.stream();
            let (p0, _a0) = q0.device_ptr(s);
            let (p1, _a1) = k0.device_ptr(s);
            let (p2, _a2) = v0.device_ptr(s);
            let (pwq, _a3) = wq.device_ptr(s);
            let (pwk, _a4) = wk.device_ptr(s);
            let (pwv, _a5) = wv.device_ptr(s);
            let (pq, _a6) = q.device_ptr_mut(s);
            let (pk, _a7) = k.device_ptr_mut(s);
            let (pv, _a8) = v.device_ptr_mut(s);
            let (pp, _a9) = pos.device_ptr(s);
            let pff: u64 = match ff {
                Some(t) => {
                    let (p, _gg) = t.device_ptr(s);
                    p as u64
                }
                None => 0,
            };
            let (pkc, _a10) = kc.device_ptr_mut(s);
            let (pvc, _a11) = vc.device_ptr_mut(s);
            let mut ps = [
                &p0 as *const _ as *mut std::ffi::c_void,
                &p1 as *const _ as *mut _,
                &p2 as *const _ as *mut _,
                &pwq as *const _ as *mut _,
                &pwk as *const _ as *mut _,
                &pwv as *const _ as *mut _,
                &pq as *const _ as *mut _,
                &pk as *const _ as *mut _,
                &pv as *const _ as *mut _,
                &nc as *const _ as *mut _,
                &rqi as *const _ as *mut _,
                &rki as *const _ as *mut _,
                &pp as *const _ as *mut _,
                &nhq as *const _ as *mut _,
                &nhk as *const _ as *mut _,
                &theta_scale as *const _ as *mut _,
                &freq_scale as *const _ as *mut _,
                &pff as *const _ as *mut _,
                &eps as *const _ as *mut _,
                &pkc as *const _ as *mut _,
                &pvc as *const _ as *mut _,
                &ti as *const _ as *mut _,
                &ktb as *const _ as *mut _,
                &vtb as *const _ as *mut _,
            ];
            unsafe {
                self.launch_pdl_flash(
                    g,
                    "rms_norm_qkv_rope_append_f32",
                    (rows as u32, 1, 1),
                    (rms_block(), 1, 1),
                    0,
                    &mut ps,
                )?;
            }
            return Ok(());
        }
        let f = if g {
            self.func_g("rms_norm_qkv_rope_append_f32")
        } else {
            self.func("rms_norm_qkv_rope_append_f32")
        };
        let cfg = LaunchConfig {
            grid_dim: (rows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        let null: u64 = 0;
        b.arg(q0)
            .arg(k0)
            .arg(v0)
            .arg(wq)
            .arg(wk)
            .arg(wv)
            .arg(&mut *q)
            .arg(&mut *k)
            .arg(&mut *v)
            .arg(&nc)
            .arg(&rqi)
            .arg(&rki)
            .arg(pos)
            .arg(&nhq)
            .arg(&nhk)
            .arg(&theta_scale)
            .arg(&freq_scale);
        match ff {
            Some(t) => {
                b.arg(t);
            }
            None => {
                b.arg(&null);
            }
        }
        b.arg(&eps)
            .arg(&mut *kc)
            .arg(&mut *vc)
            .arg(&ti)
            .arg(&ktb)
            .arg(&vtb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn add_q8_1(
        &self,
        a: &CudaSlice<f32>,
        b: &CudaSlice<f32>,
        res: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
    ) -> Result<(CudaSlice<i8>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        debug_assert!(ncols % 128 == 0);
        let mut out_q = self.alloc_uninit::<i8>(nrows * ncols)?;
        let mut out_d = self.alloc_uninit::<f32>(nrows * (ncols / 32))?;
        let f = self.func("add_q8_1_f32");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let nc = ncols as i32;
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&f);
        b2.arg(a)
            .arg(b)
            .arg(&mut *res)
            .arg(&mut out_q)
            .arg(&mut out_d)
            .arg(&nc);
        unsafe {
            b2.launch(cfg)?;
        }
        Ok((out_q, out_d))
    }

    /// E4B FFN-tail exit fusion (glue wave 5): resid = b + rms(a, wa) emitted f32 + q8_1 pair
    /// in ONE launch — replaces rms_norm(a,wa->sn) + add_q8_1(sn,b). Same rms_block() config
    /// as both parents (bit-identity: identical reduction + quad-walk quantize).
    pub fn rms_pre_add_q8_1(
        &self,
        a: &CudaSlice<f32>,
        wa: &CudaSlice<f32>,
        b: &CudaSlice<f32>,
        res: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(CudaSlice<i8>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        debug_assert!(ncols % 128 == 0);
        let mut out_q = self.alloc_uninit::<i8>(nrows * ncols)?;
        let mut out_d = self.alloc_uninit::<f32>(nrows * (ncols / 32))?;
        let f = self.func("rms_pre_add_q8_1_f32");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, ep) = (ncols as i32, eps);
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&f);
        b2.arg(a)
            .arg(wa)
            .arg(b)
            .arg(&mut *res)
            .arg(&mut out_q)
            .arg(&mut out_d)
            .arg(&nc)
            .arg(&ep);
        unsafe {
            b2.launch(cfg)?;
        }
        Ok((out_q, out_d))
    }

    /// L2 norm per row (head_dim), no weight.
    /// PREFILL l2 dispatch (round 27): the warp-per-row float4 v2 when the numeric-config
    /// seam allows (MEMRA_L2_V2, default ON, d_state==128 only); else the strided kernel.
    pub fn l2_v2_on(ncols: usize) -> bool {
        ncols == 128 && std::env::var("MEMRA_L2_V2").as_deref() != Ok("0")
    }

    pub fn l2_norm_pp(
        &self,
        x: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        dst16: Option<&mut CudaSlice<u8>>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if Self::l2_v2_on(ncols) {
            let f = self.func("l2_norm_pp_v2_f32");
            let rows_per_block = 8u32; // 256 threads = 8 warps = 8 rows
            let cfg = LaunchConfig {
                grid_dim: ((nrows as u32).div_ceil(rows_per_block), 1, 1),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let (nc, nr, e) = (ncols as i32, nrows as i32, eps);
            // mirror-fold: bf16 twin address by value (0 = skip; matches the nullable param)
            let d16: u64 = match dst16 {
                Some(d) => self.addr_u8(d),
                None => 0,
            };
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(x).arg(dst).arg(&d16).arg(&nc).arg(&nr).arg(&e);
            unsafe {
                b.launch(cfg)?;
            }
            return Ok(());
        }
        self.l2_norm(x, dst, ncols, nrows, eps)
    }

    pub fn l2_norm(
        &self,
        x: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("l2_norm_f32");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, e) = (ncols as i32, eps);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(dst).arg(&nc).arg(&e);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// L2-norm with blockDim=32 (warp-tree reduction) — BIT-IDENTICAL to gdn_prep_decode_f32's
    /// per-warp L2 norm. The verify path MUST use this to match decode's FP accumulation order:
    /// l2_norm at blockDim=256 produces a different shfl-tree reduction of the 128-element
    /// squared-sum (pairwise tree vs serial-4-then-warp-tree), causing ULP differences that
    /// propagate through gdn_scan and flip argmax on marginal logits.
    pub fn l2_norm_decode(
        &self,
        x: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("l2_norm_f32");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, e) = (ncols as i32, eps);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(dst).arg(&nc).arg(&e);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// RoPE NEOX in-place. x:[head_dim, n_heads, n_tokens], pos:[n_tokens].
    pub fn rope_neox(
        &self,
        x: &mut CudaSlice<f32>,
        pos: &CudaSlice<i32>,
        head_dim: usize,
        n_dims: usize,
        n_heads: usize,
        n_tokens: usize,
        freq_base: f32,
        freq_scale: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("rope_neox_f32");
        let theta_scale = (freq_base).powf(-2.0 / n_dims as f32);
        let grid = (n_heads * n_tokens) as u32;
        let cfg = LaunchConfig {
            grid_dim: (grid, 1, 1),
            block_dim: ((head_dim / 2) as u32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (hd, nd, nh) = (head_dim as i32, n_dims as i32, n_heads as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x)
            .arg(pos)
            .arg(&hd)
            .arg(&nd)
            .arg(&nh)
            .arg(&theta_scale)
            .arg(&freq_scale);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// RoPE NEOX with per-dim freq factors (gemma4 global layers, rope_freqs.weight [n_dims/2]).
    pub fn rope_neox_ff(
        &self,
        x: &mut CudaSlice<f32>,
        pos: &CudaSlice<i32>,
        head_dim: usize,
        n_dims: usize,
        n_heads: usize,
        n_tokens: usize,
        freq_base: f32,
        freq_scale: f32,
        ff: &CudaSlice<f32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("rope_neox_ff_f32");
        let theta_scale = (freq_base).powf(-2.0 / n_dims as f32);
        let grid = (n_heads * n_tokens) as u32;
        let cfg = LaunchConfig {
            grid_dim: (grid, 1, 1),
            block_dim: ((head_dim / 2) as u32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (hd, nd, nh) = (head_dim as i32, n_dims as i32, n_heads as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x)
            .arg(pos)
            .arg(&hd)
            .arg(&nd)
            .arg(&nh)
            .arg(&theta_scale)
            .arg(&freq_scale)
            .arg(ff);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// gemma4: rope q and k in one launch (per-row chain = rope_neox / rope_neox_ff verbatim).
    #[allow(clippy::too_many_arguments)]
    pub fn rope_neox2(
        &self,
        q: &mut CudaSlice<f32>,
        k: &mut CudaSlice<f32>,
        pos: &CudaSlice<i32>,
        head_dim: usize,
        n_dims: usize,
        nh_q: usize,
        nh_k: usize,
        n_tokens: usize,
        freq_base: f32,
        freq_scale: f32,
        ff: Option<&CudaSlice<f32>>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("rope_neox2_f32");
        let theta_scale = (freq_base).powf(-2.0 / n_dims as f32);
        let grid = ((nh_q + nh_k) * n_tokens) as u32;
        let cfg = LaunchConfig {
            grid_dim: (grid, 1, 1),
            block_dim: ((head_dim / 2) as u32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (hd, nd, nq, nk, nt) = (
            head_dim as i32,
            n_dims as i32,
            nh_q as i32,
            nh_k as i32,
            n_tokens as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(pos)
            .arg(&hd)
            .arg(&nd)
            .arg(&nq)
            .arg(&nk)
            .arg(&nt)
            .arg(&theta_scale)
            .arg(&freq_scale);
        match ff {
            Some(ffv) => {
                b.arg(ffv);
                unsafe {
                    b.launch(cfg)?;
                }
            }
            None => {
                let null: u64 = 0;
                b.arg(&null);
                unsafe {
                    b.launch(cfg)?;
                }
            }
        }
        Ok(())
    }

    /// gemma4 R1: dst = GELU_tanh(gate) * up.
    pub fn gelu_tanh_mul(
        &self,
        gate: &CudaSlice<f32>,
        up: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gelu_tanh_mul_f32");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let ni = n as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate).arg(up).arg(dst).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn silu_mul(
        &self,
        gate: &CudaSlice<f32>,
        up: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("silu_mul_f32");
        // float4 kernel: one thread per 4 elements (tail handled in-kernel)
        let cfg = LaunchConfig::for_num_elems((n as u32).div_ceil(4));
        let ni = n as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate).arg(up).arg(dst).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// SwiGLU twin using Memra's host-matching expf transcription.
    pub fn silu_mul_host_expf(
        &self,
        gate: &CudaSlice<f32>,
        up: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("silu_mul_host_expf_f32");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let ni = n as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate).arg(up).arg(dst).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Step routed-expert clamp twin using Memra's host-matching expf transcription.
    pub fn silu_clamped_mul_host_expf(
        &self,
        gate: &CudaSlice<f32>,
        up: &CudaSlice<f32>,
        limit: f32,
        dst: &mut CudaSlice<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if !limit.is_finite() || limit <= 0.0 {
            return Err(
                format!("Step routed-expert clamp limit must be positive, got {limit}").into(),
            );
        }
        let f = self.func("silu_clamped_mul_host_expf_f32");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let ni = n as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate).arg(up).arg(&limit).arg(dst).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// f16out twin of `silu_mul` (task #17): the epilogue also emits the fp16 GEMM operand
    /// for the down projection — kills the standalone convert pass. Bit-identical class.
    pub fn silu_mul_f16out(
        &self,
        gate: &CudaSlice<f32>,
        up: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        dst16: &mut CudaSlice<u8>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("silu_mul_f16out_f32");
        let cfg = LaunchConfig::for_num_elems((n as u32).div_ceil(4));
        let ni = n as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate).arg(up).arg(dst).arg(dst16).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// FFN SwiGLU epilogue fusion (RANK3 LEVER 2): `dst = silu(gate*gs) * (up*us)` in ONE launch,
    /// folding the per-tensor NVFP4 macro-scale (`gs`,`us`) that would otherwise be two separate
    /// `scale_inplace` launches on the gate/up matmul outputs. BIT-IDENTICAL to
    /// scale_inplace(gate,gs); scale_inplace(up,us); silu_mul(gate,up,dst) — identical float ops in
    /// identical order. For non-NVFP4 weights gs==us==1.0 -> identical to `silu_mul`. Net: -2
    /// launches per dense FFN layer (the gate+up post-matmul scales).
    pub fn silu_mul_scaled(
        &self,
        gate: &CudaSlice<f32>,
        up: &CudaSlice<f32>,
        gs: f32,
        us: f32,
        dst: &mut CudaSlice<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("silu_mul_scaled_f32");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let ni = n as i32;
        let (gsf, usf) = (gs, us);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate).arg(up).arg(&gsf).arg(&usf).arg(dst).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// swigluoai (MiniMax-M3 / GPT-OSS): clamped SwiGLU epilogue, math 1:1 vs llama.cpp
    /// ggml_cuda_op_swiglu_oai_single. `dst = swish_alpha(clamp(gate*gs)) * (1 + clamp(up*us))`.
    /// gs/us fold the NVFP4 macro-scales exactly like `silu_mul_scaled`.
    #[allow(clippy::too_many_arguments)]
    pub fn swigluoai_mul_scaled(
        &self,
        gate: &CudaSlice<f32>,
        up: &CudaSlice<f32>,
        gs: f32,
        us: f32,
        alpha: f32,
        limit: f32,
        dst: &mut CudaSlice<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("swigluoai_mul_scaled_f32");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let ni = n as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate)
            .arg(up)
            .arg(&gs)
            .arg(&us)
            .arg(&alpha)
            .arg(&limit)
            .arg(dst)
            .arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// RANK2 LEVER (q8_1 quant-fold): SwiGLU epilogue that EMITS the q8_1 quantization of `act`
    /// directly (aq int8 [n] + ad f32 [n/32]), so ffn_down's standalone `quantize_q8_1` launch is
    /// removed — the down-proj activation has one consumer, so the quant folds into the producer for
    /// free (no extra HBM read; no f32 `act` write). gs/us fold the gate/up NVFP4 macro-scales like
    /// `silu_mul_scaled`. BIT-IDENTICAL q8_1 to silu_mul_scaled(...) then quantize_q8_1(...). Only
    /// valid when ffn_down uses the q8_1 dp4a/mmvq path; the caller checks `uses_q8_1_fast(ffn_down)`.
    /// n must be a multiple of 32 (n_ff always is).
    pub fn silu_mul_scaled_q8_1(
        &self,
        gate: &CudaSlice<f32>,
        up: &CudaSlice<f32>,
        gs: f32,
        us: f32,
        n: usize,
    ) -> Result<(CudaSlice<i8>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        let f = self.func("silu_mul_scaled_q8_1");
        let nblk = n / 32;
        let mut aq = self.alloc_uninit::<i8>(n)?; // full-overwrite output
        let mut ad = self.alloc_uninit::<f32>(nblk)?; // full-overwrite output
        // WARP-PER-BLOCK kernel: one warp (32 lanes) per 32-block -> n threads total.
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let (gsf, usf, ni) = (gs, us, n as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate)
            .arg(up)
            .arg(&gsf)
            .arg(&usf)
            .arg(&mut aq)
            .arg(&mut ad)
            .arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((aq, ad))
    }

    pub fn add(
        &self,
        a: &CudaSlice<f32>,
        b_in: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("add_f32");
        // float4 kernel: one thread per 4 elements (tail handled in-kernel)
        let cfg = LaunchConfig::for_num_elems((n as u32).div_ceil(4));
        let ni = n as i32;
        let __s_bld = self.gpu.stream();
        let mut bld = __s_bld.launch_builder(&f);
        bld.arg(a).arg(b_in).arg(dst).arg(&ni);
        unsafe {
            bld.launch(cfg)?;
        }
        Ok(())
    }

    pub fn mul(
        &self,
        a: &CudaSlice<f32>,
        b_in: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("mul_f32");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let ni = n as i32;
        let __s_bld = self.gpu.stream();
        let mut bld = __s_bld.launch_builder(&f);
        bld.arg(a).arg(b_in).arg(dst).arg(&ni);
        unsafe {
            bld.launch(cfg)?;
        }
        Ok(())
    }

    /// Unified weight-tensor matmul: dispatches quant tensors to qmatvec (weights packed) and
    /// float tensors to cuBLASLt. y[m,out] = x[m,in] @ W[out,in]^T.
    pub fn matmul(
        &self,
        w: &crate::model::GpuTensor,
        x: &CudaSlice<f32>,
        m: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        let in_f = w.in_features();
        let out_f = w.out_features();
        // PREFILL (T>1) ROOT FIX: batched tensor-core int8 GEMM. Decodes each weight tile to int8
        // in smem ONCE and reuses across all tokens via mma — vs the dp4a matvec's per-token weight
        // re-read. Only the 4 daily-hot dtypes; m=1 decode keeps dp4a (it's bandwidth-bound, mma
        // gives nothing). Quantize the activation once here then call the GEMM.
        // m cutoff FIXED at 16: the m=4 MMA-verify A/B (2026-07-06, was MEMRA_GEMM_M) measured
        // NEGATIVE — the MMA tile grid starves at m=4 (BN=256 -> grid.y=1) and its FP order
        // shifted verify argmax at tight margins. Do not lower without re-running that battery.
        #[allow(non_snake_case)]
        // VERIFY-EXACT scope pushes the GEMM crossover out of reach (usize::MAX) — the
        // t>=16 dflash verify must ride the decode-exact batched class (parity law).
        let GEMM_M_THRESHOLD = if self.verify_exact_on() {
            usize::MAX
        } else {
            16usize
        };

        // PREFILL GEMM (m>=16). ACCURACY-FIRST dispatch (2026-06-28, prefill-gemm-beat-research wf
        // wllbyo6vc step 1): the int8 W4A8 GEMM (qmatvec_gemm, q8_1 activation, s32 accumulate) is
        // ACCURATE (prefill logit maxdiff 0.159, < dp4a 0.55) and the default. The FP4 W4A4 mxf4 path
        // (try_fp4_gemm) quantizes the ACTIVATION to e2m1 4-bit (8 magnitude levels) -> maxdiff 1.0
        // when combined — a real accuracy loss, NOT a math bug. So FP4-W4A4 is taken ONLY under the
        // explicit MEMRA_FP4 opt-in AND it must come SECOND (int8 W4A8 is the correct default for NVFP4).
        // The workflow plan rebuilds the FP4 path (kill per-K repack, widen K, deepen pipeline, TMA) to
        // be both fast AND accurate; until then NVFP4 prefill defaults to the accurate int8 GEMM.
        // TINY-OUT_F GUARD (2026-06-28, ncu trace): the tiling GEMM's grid is (ceil(out_f/BM=64),
        // ceil(m/BN=256)). For tiny out_f (ssm_beta/ssm_alpha out_f=num_v_heads~32), grid.x=1 -> only
        // ceil(m/256) CTAs (e.g. 2 for m=512) on 82 SMs = 0.39% SM throughput, 852us EACH (measured
        // worst offender). The dp4a path grids (out_f, m) = far more CTAs, filling the GPU. So route
        // out_f < 2*BM to dp4a (skip the tiling GEMM which structurally can't fill the SMs here).
        const GEMM_MIN_OUT_F: usize = 128; // 2*BM; below this the GEMM grid.x starves the 82 SMs
        // VENDORED llama MMQ prefill GEMMs. NVFP4 W4A8 is DEFAULT-ON (2026-07-05 flip: same int8
        // accuracy class as the int8 GEMM below at ~1.9x pp512, rp-loader coexists with the A6
        // repack; MEMRA_MMQ_W4A8=0 = escape hatch). W4A4 mxf4nvf4 + Q4_K/Q5_K stay behind MEMRA_MMQ=1.
        // The env policy lives in mmq_supports/qmatvec_mmq. Feeds raw f32 activation `x` (the
        // launcher quantizes internally). out_f>=MMQ_Y/2 keeps the tile grid from starving the SMs.
        // FP8-ACT PREFILL (MEMRA_PP_FP8=1, probe verdict 2026-07-08): F8-E4M3-origin projections
        // carry their raw e4m3 device bytes (the `fp8` operand stashed at load next to the Q8_0
        // re-encode) — cuBLASLt FP8 TN at 620-795 TF vs 47-72 TF for this class's int8 GEMM.
        // Weight side EXACT (checkpoint bytes); activation rides ONE per-batch e4m3 scale
        // (amax/448) folded with weight_scale in-GEMM. Prefill only; decode keeps Q8_0 untouched.
        if m >= GEMM_M_THRESHOLD {
            if let Some(y) = self.try_fp8_gemm(w, x, m)? {
                return Ok(y);
            }
            // PER-BLOCK FP8 MMQ (lane/fp8-mmq): the block-128 class try_fp8_gemm skips (cuBLASLt
            // takes no block grid on sm_120). Exact per block — the checkpoint's e4m3 bytes and its
            // f32 grid go into the tile unchanged. TWO SOURCES, TWO DEFAULTS: the load-time stash is
            // opt-in (MEMRA_FP8_MMQ=1), the native-resident QT_F8_E4M3_BLK grid is DEFAULT ON
            // (MEMRA_FP8_MMQ=0 reverts it to dequant-per-call) — see fp8_ffi.rs for why the same
            // tile defaults differently by operand source.
            if let Some(y) = self.try_fp8_blk_mmq(w, x, m)? {
                return Ok(y);
            }
            // FP16-mirror prefill (MEMRA_PP_F16=1, probe 2026-07-26: 3.2-3.7x the MMQ class).
            // Mirror presence IS the gate (only built under the env). Decode never reaches here.
            if let Some(y) = self.try_f16_gemm(w, x, m)? {
                return Ok(y);
            }
        }
        // F8-E4M3 BLOCK-128 (QT_F8_E4M3_BLK, lane/fp8-blk128-decode). TWO arms, split at the SAME
        // m threshold the rest of this method uses:
        //   * m >= threshold (prefill): dequant-per-call to the ARM B' Q8_0 slab and recurse, so
        //     prefill keeps the floor's kernels AND the floor's bits (try_e4m3_blk_prefill).
        //   * m <  threshold: the native per-block GEMV — m=1 decode and the m=2..15 verify tiers.
        //     grid.y=m runs the exact m=1 program per (token,row), so the decode-parity law holds
        //     across every tier by construction with no batched twin needed.
        //
        // NOT gated on `fast`: this dtype has no dp4a twin and no Stage-A f32-dequant oracle (the
        // generic `deq()` switch has no block-scale input), exactly as QT_F8_E4M3 has none, so
        // MEMRA_FAST=0 cannot route it anywhere else. Placed before every GEMM/MMQ arm below
        // because gemm_supports/mmq_supports/mmvq_supports all deliberately REFUSE this qtype —
        // reaching the generic tail would panic rather than produce wrong numbers, and this pair of
        // arms is what makes sure it never gets there.
        if let GpuTensor::Quant { qtype, .. } = w {
            if *qtype == QT_F8_E4M3_BLK {
                if m >= GEMM_M_THRESHOLD {
                    if let Some(y) = self.try_e4m3_blk_prefill(w, x, m)? {
                        return Ok(y);
                    }
                }
                let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
                if let Some(y) = self.try_e4m3_blk_pre(w, &aq, &ad, m)? {
                    return Ok(y);
                }
            }
        }
        if m >= GEMM_M_THRESHOLD && out_f >= GEMM_MIN_OUT_F && self.mmq_supports(w) {
            return self.qmatvec_mmq(w, x, m);
        }
        if m >= GEMM_M_THRESHOLD && out_f >= GEMM_MIN_OUT_F && self.gemm_supports(w) {
            let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
            return self.qmatvec_gemm(w, &aq, &ad, m);
        }
        // FP4 W4A4 only as an explicit speed/accuracy tradeoff opt-in, and only if the int8 GEMM
        // above didn't already handle this weight (e.g. NVFP4 with in_f%64!=0, or MEMRA_NO_GEMM set).
        if m >= GEMM_M_THRESHOLD {
            if let Some(y) = self.try_fp4_gemm(w, x, m, in_f, out_f)? {
                return Ok(y);
            }
        }
        // Stage-B fast int8 dp4a is the DEFAULT since 2026-07-08 (it has been the daily path
        // for weeks; the old opt-in flag was a silent-slow-path landmine). MEMRA_FAST=0 reverts
        // to Stage-A f32-dequant (the correctness oracle path).
        let fast = std::env::var("MEMRA_FAST").as_deref() != Ok("0");
        // PERF-3 decode-GEMV: m=1 warp-per-row MMVQ (MEMRA_MMVQ). The big decode matvecs reach
        // `matmul` directly (ffn_down, lm_head output, wo), so route them here too — not only the
        // matmul_pre siblings. qmatvec_mmvq_raw quantizes the activation internally (q8_1) like the
        // _fast paths; the NVFP4 macro-scale is applied by the `scale != 1.0` block below.
        if m == 1 && fast {
            if let GpuTensor::Quant {
                bytes,
                qtype,
                row_bytes,
                rp,
                rp4,
                scale,
                ..
            } = w
            {
                if self.mmvq_supports(*qtype) {
                    // NVFP4 macro-scale rides the kernel's fused epilogue arg (one launch total);
                    // non-NVFP4 has scale==1.0 so qmatvec_mmvq skips scale_inplace either way.
                    // Q4_0 split-plane mirror (rp4): the decode arm reads it via the _rp twins.
                    let (bytes, rp) = match rp4 {
                        Some(m4) => (m4, true),
                        None => (bytes, *rp),
                    };
                    let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
                    return self.qmatvec_mmvq(
                        bytes, &aq, &ad, m, in_f, out_f, *qtype, *row_bytes, *scale, rp,
                    );
                }
            }
        }
        // BATCHED weight-resident matvec for the m=2-4 band (the MTP/verify forward's ffn_down, wo, and
        // lm_head `output` reach `matmul` directly at m=T=2..4). Walks the weight ONCE, dp4a vs all m
        // activation columns -> 1 weight read for m tokens (vs grid.y=m re-reading m times below). Quant
        // the activation once here (q8_1) like the _fast paths; macro-scale applied via the scale!=1.0
        // block below. MEMRA_NO_BATCHED -> per-m path.
        //
        // DECODE-PARITY GATE (2026-07-07, the 9B synth K=3/4/6 spec FAIL root cause): the batched
        // kernels are bit-identical per (token,row) to MMVQ's 32-thread warp reduce, NOT to the
        // dp4a kernels' 128-thread two-level reduce. Without MEMRA_MMVQ the m=1 decode chain rides
        // dp4a, so a verify riding batched here has a DIFFERENT FP order than the decode it must
        // match bit-for-bit — greedy spec flips at tight-margin tokens (the old HANDOVER "ENV LAW:
        // FAST+MMVQ both required" footgun, closed here). Parity law: the m>1 kernel CLASS must be
        // a pure function of (dtype, env) equal to the m=1 class — batched iff MMVQ. Without MMVQ
        // the verify falls to the per-m grid.y=m dp4a path below (each column = the exact m=1
        // dp4a program). MEMRA_MMVQ=1 (the daily config) is dispatch-unchanged.
        if (2..=16).contains(&m)
            && fast
            && std::env::var("MEMRA_NO_BATCHED").is_err()
            && (m <= 4 || Self::b8_enabled())
        {
            // b16 tier (2026-07-11, spec K>7): Q4_0/Q6_K have base+_rp b16 kernels; Q8_0's
            // b16 exists only as the split-plane _rp twin, so it joins iff the q8rp mirror
            // is present (rp4) — the mirror pick below then routes to the _rp family.
            // QT_F8_E4M3 joins unconditionally (lane/rp-on-st): its b16 IS the base kernel,
            // because the native e4m3 row layout is already aligned and needs no mirror.
            // NVFP4/Q4_K/Q8_0 all join unconditionally now (lane/rp-on-st): each has base + _rp
            // b16 twins, so either residency layout has its aligned form at this width. Q8_0's
            // old `rp4.is_some()` precondition is GONE — the mirror is a bandwidth lever, not the
            // exact tier's admission ticket (it was refusing FP8-ST over 23.9 MiB of ssm_beta).
            let m_ok = m <= 8
                || matches!(w, GpuTensor::Quant { qtype, .. }
                if *qtype == QT_Q4_0 || *qtype == QT_Q6_K || *qtype == QT_F8_E4M3
                    || *qtype == QT_NVFP4 || *qtype == QT_Q4_K || *qtype == QT_Q5_K || *qtype == QT_Q8_0);
            if m_ok {
                if let GpuTensor::Quant {
                    bytes,
                    qtype,
                    row_bytes,
                    rp,
                    rp4,
                    ..
                } = w
                {
                    if self.batched_supports(*qtype) && self.mmvq_supports(*qtype) {
                        let (bytes, rp) = match rp4 {
                            Some(m4) => (m4, true),
                            None => (bytes, *rp),
                        };
                        let mcols = Self::batched_mcols(m);
                        let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
                        let mut y = self.qmatvec_mmvq_batched(
                            bytes, &aq, &ad, m, in_f, out_f, *qtype, *row_bytes, mcols, 1.0, rp,
                        )?;
                        if let GpuTensor::Quant { scale, .. } = w {
                            if *scale != 1.0 {
                                self.scale_inplace(&mut y, *scale, m * out_f)?;
                            }
                        }
                        return Ok(y);
                    }
                }
            }
        }
        // F8-E4M3 (MEMRA_ST_E4M3) catch-all for the m<16 band the arms above didn't take (m=9..15,
        // the K=8 verify tier; or m=2..8 under MEMRA_NO_BATCHED/MEMRA_B8=0): grid.y=m e4m3 mmvq —
        // the SAME per-(token,row) program as the m=1 decode launch (bit-identical by construction),
        // weight re-read m times (rare tier; exactness over bandwidth here). There is no _dp4a twin
        // for this dtype, so the generic match below must never see it under `fast`.
        if fast {
            if let GpuTensor::Quant {
                bytes,
                qtype,
                row_bytes,
                scale,
                ..
            } = w
            {
                if *qtype == QT_F8_E4M3 {
                    let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
                    return self.qmatvec_mmvq(
                        bytes, &aq, &ad, m, in_f, out_f, *qtype, *row_bytes, *scale, false,
                    );
                }
            }
        }
        let mut y = match w {
            GpuTensor::Quant {
                bytes,
                qtype,
                row_bytes,
                ..
            } if fast && *qtype == QT_Q8_0 => {
                self.qmatvec_q8_0_fast(bytes, x, m, in_f, out_f, *row_bytes)?
            }
            GpuTensor::Quant {
                bytes,
                qtype,
                row_bytes,
                ..
            } if fast && *qtype == QT_Q4_K => {
                self.qmatvec_q4_K_fast(bytes, x, m, in_f, out_f, *row_bytes)?
            }
            GpuTensor::Quant {
                bytes,
                qtype,
                row_bytes,
                ..
            } if fast && *qtype == QT_Q6_K => {
                self.qmatvec_q6_K_fast(bytes, x, m, in_f, out_f, *row_bytes)?
            }
            GpuTensor::Quant {
                bytes,
                qtype,
                row_bytes,
                ..
            } if fast && *qtype == QT_Q5_K => {
                self.qmatvec_q5_K_fast(bytes, x, m, in_f, out_f, *row_bytes)?
            }
            GpuTensor::Quant {
                bytes,
                qtype,
                row_bytes,
                ..
            } if fast && *qtype == QT_Q3_K => {
                self.qmatvec_q3_K_fast(bytes, x, m, in_f, out_f, *row_bytes)?
            }
            GpuTensor::Quant {
                bytes,
                qtype,
                row_bytes,
                rp,
                ..
            } if fast && *qtype == QT_NVFP4 => self.qmatvec_dp4a_named(
                if *rp {
                    "qmatvec_nvfp4_dp4a_rp"
                } else {
                    "qmatvec_nvfp4_dp4a"
                },
                &bytes.slice(0..bytes.len()),
                x,
                m,
                in_f,
                out_f,
                *row_bytes,
            )?,
            // IQ4_XS trunk fast path — DEFAULT ON since 2026-08-02 (MEMRA_IQ_FAST=0 reverts to
            // Stage-A; see iq_fast_enabled). The old opt-in default was the KAT-Coder decode
            // anomaly (research/kat-anomaly-20260802/).
            GpuTensor::Quant {
                bytes,
                qtype,
                row_bytes,
                ..
            } if fast && *qtype == QT_IQ4_XS && Self::iq_fast_enabled() => {
                self.qmatvec_iq4_XS_fast(bytes, x, m, in_f, out_f, *row_bytes)?
            }
            // B3: IQ3_S uses the Stage-A f32 dequant-in-kernel path. There is NO
            // qmatvec_iq3_s_dp4a kernel — do NOT add a `*qtype == QT_IQ3_S` fast guard here
            // without first writing the matching kernel, or func() will panic
            // "kernel ... not in any fatbin".
            GpuTensor::Quant {
                bytes,
                qtype,
                row_bytes,
                rp,
                ..
            } =>
            // Stage-A generic: repacked NVFP4 uses the device-side split-plane tag (the
            // deq(row,j) form cannot address the planes; same value/product order).
            {
                self.qmatvec(
                    bytes,
                    x,
                    m,
                    in_f,
                    out_f,
                    if *rp && *qtype == QT_NVFP4 {
                        QT_NVFP4_RP
                    } else {
                        *qtype
                    },
                    *row_bytes,
                )?
            }
            GpuTensor::Float { data, .. } => self.linear(x, data, m, in_f, out_f)?,
            // MEMRA_FULL_PREC bf16-resident weight: dequant-on-use to f32 scratch, then the same
            // cuBLASLt f32 GEMV as the Float arm.
            GpuTensor::FloatBf16 { data, .. } => {
                // DECODE-TIER ROWS FAST PATH (2 <= m <= 8, bf16-mmv class): the chunked
                // arm dequants the WHOLE weight to f32 scratch per call — 4.7 ms/call on
                // the 1.24 GB LM head (nsys: 8x591us bf16_to_f32 per batch tick / per
                // verify round). matvec_bf16_f32acc_x4_rows runs the t=1 decode head
                // program PER ROW (identical dot + reduce), so decode/verify tiers keep
                // the t=1 numeric class and skip the convert. Prefill (m>8) keeps GEMM.
                if (1..=32).contains(&m) && Self::bf16_mmv_on() && in_f % 8 == 0 {
                    let mut y = self.alloc_uninit::<f32>(m * out_f)?;
                    self.matvec_bf16_rows_into(data, x, &mut y, in_f, out_f, m)?;
                    y
                } else {
                    self.linear_bf16_chunked(x, data, m, in_f, out_f, false, None)?
                }
            }
        };
        // NVFP4 per-tensor macro-scale (post-matmul). scale==1.0 for all other quants/float -> no-op.
        if let GpuTensor::Quant { scale, .. } = w {
            if *scale != 1.0 {
                self.scale_inplace(&mut y, *scale, m * out_f)?;
            }
        }
        Ok(y)
    }

    /// True when `MEMRA_FAST=0`, i.e. the Stage-A f32 oracle is the requested arithmetic and every
    /// `matmul_pre` call will take the raw-f32 escape rather than the q8_1 pair.
    ///
    /// WHY THIS EXISTS AS ITS OWN PREDICATE: `uses_q8_1_fast` needs a weight, but the callers that
    /// have to DECIDE WHETHER TO MATERIALIZE an f32 activation sit one level above any weight — the
    /// gemma-4 decode trunk emits a q8_1 pair per layer and hands it down, so the f32 has to be
    /// produced (or not) by the loop that owns the residual. Cached in a OnceLock like every other
    /// `*_on()` flag here: this is read once per layer per token on the decode path, and the daily
    /// path must not pay an env lookup for a flag that is off.
    pub fn stage_a_raw_needed() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| std::env::var("MEMRA_FAST").as_deref() == Ok("0"))
    }

    /// True if `w` would take the int8-dp4a fast path under MEMRA_FAST (so its activation can be
    /// pre-quantized once and shared across sibling matmuls via `matmul_pre`).
    pub fn uses_q8_1_fast(&self, w: &crate::model::GpuTensor) -> bool {
        use crate::model::GpuTensor;
        if std::env::var("MEMRA_FAST").as_deref() == Ok("0") {
            return false;
        }
        match w {
            // QT_F8_E4M3_BLK is admitted for the same reason QT_F8_E4M3 is: its ONLY kernel class
            // takes the shared q8_1 activation, so callers may pre-quantize once and share it
            // across siblings. It is NOT admitted to any of the fused/dual epilogue doors those
            // siblings can then open (`q8_fused_params`, `e4m3_fused_params` and
            // `matmul_pre_dual_noscale` all match on their own qtype and refuse this one) — the
            // block class has no fused twin yet, so each of its projections takes its own launch.
            GpuTensor::Quant { qtype, .. } => {
                matches!(
                    *qtype,
                    QT_Q8_0
                        | QT_Q4_K
                        | QT_Q6_K
                        | QT_Q5_K
                        | QT_Q3_K
                        | QT_NVFP4
                        | QT_F8_E4M3
                        | QT_F8_E4M3_BLK
                        | QT_Q4_0
                ) || (*qtype == QT_IQ4_XS && Self::iq_fast_enabled())
            }
            GpuTensor::Float { .. } | GpuTensor::FloatBf16 { .. } => false,
        }
    }

    /// matmul with a PRE-QUANTIZED q8_1 activation (aq,ad from `quantize_q8_1`). Skips the
    /// per-matmul re-quantize so sibling matmuls that share an input (gate+up share `z`;
    /// q/k/v + wqkv/gate/beta/alpha share `h`) quantize ONCE. Caller MUST have checked
    /// `uses_q8_1_fast(w)`; falls back to plain `matmul` otherwise (Stage-A / Float / non-fast).
    pub fn matmul_pre(
        &self,
        w: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        x_fallback: &CudaSlice<f32>,
        m: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        // Every raw-f32 arm below (fp8/f16/MMQ/fp4) reads m*in_f from x_fallback. Callers that
        // pre-quantized and dropped the f32 input pass an EMPTY x_fallback (E4B's fusion port:
        // h = zeros(0)) — the length guard keeps those on the aq/ad GEMM instead of feeding a
        // 0-byte buffer to a convert kernel (illegal address -> cublasLt status 13; the E4B
        // rc=30013 dig, 2026-07-31).
        let x_raw_ok = x_fallback.len() >= m * w.in_features();
        // FP8-ACT PREFILL (MEMRA_PP_FP8=1): same arm as `matmul` — the fp8 operand needs the RAW
        // f32 activation (per-batch e4m3 quant differs from q8_1), so x_fallback not aq/ad.
        if m >= 16 && x_raw_ok && !self.verify_exact_on() {
            if let Some(y) = self.try_fp8_gemm(w, x_fallback, m)? {
                return Ok(y);
            }
            // PER-BLOCK FP8 MMQ — same arm as `matmul` (stash opt-in, native-resident default ON);
            // its own quantizer wants the RAW f32 activation, so x_fallback not aq/ad.
            if let Some(y) = self.try_fp8_blk_mmq(w, x_fallback, m)? {
                return Ok(y);
            }
            // FP16-mirror prefill (same arm as `matmul` — fp16 wants the RAW f32 activation).
            if let Some(y) = self.try_f16_gemm(w, x_fallback, m)? {
                return Ok(y);
            }
        }
        // BLOCK-128 e4m3 (QT_F8_E4M3_BLK) — the same two arms as `matmul`, split at the same m, and
        // placed at the same point in the order (after the prefill GEMM hooks, before every arm
        // that refuses this qtype). The prefill arm needs the RAW f32 activation for the Q8_0
        // dispatch it recurses into, so it takes x_fallback and is skipped when that is empty
        // (a pre-quantized caller that dropped its f32 input never runs at prefill m anyway).
        if m >= 16 && x_raw_ok && !self.verify_exact_on() {
            if let Some(y) = self.try_e4m3_blk_prefill(w, x_fallback, m)? {
                return Ok(y);
            }
        }
        if let Some(y) = self.try_e4m3_blk_pre(w, aq, ad, m)? {
            return Ok(y);
        }
        // VENDORED llama MMQ prefill GEMMs (NVFP4 W4A8 default-on; W4A4/k-quant behind MEMRA_MMQ=1
        // — policy in mmq_supports) — use the RAW f32 activation (their own internal quant:
        // q8_1 D4 for NVFP4 W4A8, FP8/UE4M3 for W4A4, q8_1 DS4 for Q4_K/Q5_K), so x_fallback not
        // aq/ad.
        if m >= 16
            && w.out_features() >= 128
            && self.mmq_supports(w)
            && !self.verify_exact_on()
            && x_raw_ok
        {
            return self.qmatvec_mmq(w, x_fallback, m);
        }
        // Stage-C FP4 prefill (MEMRA_FP4): native mxf4 GEMM needs the f32 activation (FP4-quant differs
        // from q8_1), so re-quantize from x_fallback rather than reuse aq/ad. NVFP4 only, m>=16.
        if m >= 16 && x_raw_ok && !self.verify_exact_on() {
            if let Some(y) =
                self.try_fp4_gemm(w, x_fallback, m, w.in_features(), w.out_features())?
            {
                return Ok(y);
            }
        }
        // Prefill GEMM root fix: if T>1 and the dtype has a GEMM kernel, batch via tensor cores
        // (reuses the already-quantized aq/ad — no extra quantize). m=1 falls through to dp4a.
        if m >= 16 && self.gemm_supports(w) && !self.verify_exact_on() {
            return self.qmatvec_gemm(w, aq, ad, m);
        }
        // THE SEVENTH ARM. Every raw-f32 arm above is guarded by `x_raw_ok`; this one — the
        // Stage-A / Float escape — was not, and it is the ONLY one `MEMRA_FAST=0` opens. So the
        // 2026-07-31 E4B rc=30013 fix (add the length guard, keep empty-fallback callers off the
        // raw-f32 arms) protected the six arms the FAST path can reach and missed the one the
        // ORACLE path reaches. Consequence before this guard: the gemma-4 decode arms, which pass
        // `e.zeros(0)` because their trunk emits only a q8_1 pair and never materializes an f32
        // attn-normed activation, fell through to Stage-A `qmatvec_f32` (cu/qmatvec.cu:5442),
        // which reads `m * in_f` floats out of a 0-byte allocation ->
        // CUDA_ERROR_ILLEGAL_ADDRESS at layer 0 of the first decode token. The fault is STICKY:
        // it poisons the context, so every LATER request in that process fails with an unrelated
        // message ("cache alloc failed: ...") and the true cause appears exactly once, in the
        // first failure. That is what made `MEMRA_FAST=0` — the reference named in
        // `tools/argmax-margin-gate.sh`'s own header — return an opaque HTTP 500 on the gemma
        // dense artifact and left the arm with no working truth instrument.
        //
        // Refuse loudly instead of reading out of bounds. A named error at the true call site is
        // strictly better than an illegal address surfacing later at an unrelated sync point, and
        // an oracle that cannot run must say so rather than corrupt the context it runs in.
        if !self.uses_q8_1_fast(w) {
            if !x_raw_ok {
                return Err(format!(
                    "matmul_pre: q8_1-fast is off for this weight but x_fallback holds {} f32 \
                     (need m*in_f = {}*{} = {}). This call site pre-quantized its activation and \
                     dropped the f32, so there is nothing to fall back to — pass the real f32 \
                     activation (see Engine::rms_norm_decode, which is bit-identical to \
                     rms_norm_q8_1's reduction) or keep the weight on the q8_1 path.",
                    x_fallback.len(),
                    m,
                    w.in_features(),
                    m * w.in_features()
                )
                .into());
            }
            return self.matmul(w, x_fallback, m);
        }
        let in_f = w.in_features();
        let out_f = w.out_features();
        let (bytes, qtype, row_bytes, scale, rp) = match w {
            GpuTensor::Quant {
                bytes,
                qtype,
                row_bytes,
                scale,
                rp,
                ..
            } => (bytes, *qtype, *row_bytes, *scale, *rp),
            _ => unreachable!("uses_q8_1_fast guaranteed Quant"),
        };
        // Q4_0 split-plane mirror: only the mmvq/batched decode arms read it (the _rp twins);
        // the dp4a/oracle tails below keep the raw GGUF bytes.
        let (mbytes, mrp) = match w {
            GpuTensor::Quant { rp4: Some(m4), .. } => (m4, true),
            _ => (bytes, rp),
        };
        // PERF-3 decode-GEMV: warp-per-row MMVQ for the m=1 decode arm, gated behind MEMRA_MMVQ.
        // Only the 4 daily-hot dtypes have an _mmvq kernel (Q8_0/Q4_K/Q6_K/NVFP4); Q5_K/Q3_K/IQ4_XS
        // keep _dp4a (the oracle/fallback). Bit-equivalent to _dp4a up to f32 reduction order.
        if m == 1 && self.mmvq_supports(qtype) {
            return self.qmatvec_mmvq(mbytes, aq, ad, m, in_f, out_f, qtype, row_bytes, scale, mrp);
        }
        // BATCHED weight-resident matvec for the m=2-4 band (the MTP/verify forward: full_attn_verify
        // and decode_step_t run their projections at m=T=k=2..4). The plain _dp4a path below launches
        // grid.y=m INDEPENDENT blocks per output row -> the weight row is re-read m times from HBM/L2.
        // The _b2/_b4 kernels walk the weight ONCE and dp4a vs all m activation columns, so m tokens
        // cost ~1 weight read instead of m (decode is weight-BW-bound). BIT-IDENTICAL per (token,row)
        // to the _mmvq path (32-thread warp reduce — NOT the dp4a 128-thread reduce below).
        // m=2 -> mcols=2; m∈{3,4} -> mcols=4; m∈{5..8} -> mcols=8 (kernel guards c>=m).
        // MEMRA_NO_BATCHED forces the per-m grid.y=m path (the A/B reference); MEMRA_B8=0 keeps
        // m=5..8 on the old per-m path (b8-tier-only seam).
        // DECODE-PARITY GATE (2026-07-07): batched iff mmvq_supports — see matmul's parity note.
        // Without MEMRA_MMVQ, m=1 decode rides dp4a (the arm below at m=1); the verify must ride
        // the SAME class per column (grid.y=m dp4a = the exact m=1 dp4a program per column).
        if (2..=16).contains(&m) && self.batched_supports(qtype) && self.mmvq_supports(qtype)
            && std::env::var("MEMRA_NO_BATCHED").is_err()
            && (m <= 4 || Self::b8_enabled())
            // b16 tier: every class routed here now has base + _rp b16 kernels (Q4_0/Q6_K
            // pre-existing; NVFP4/Q4_K/Q8_0-base/F8_E4M3 added lane/rp-on-st 2026-08-06), so
            // there is no mirror precondition left — `mrp` still selects the LAYOUT below.
            && (m <= 8 || qtype == QT_Q4_0 || qtype == QT_Q6_K || qtype == QT_NVFP4
                || qtype == QT_Q4_K || qtype == QT_Q5_K || qtype == QT_F8_E4M3 || qtype == QT_Q8_0)
        {
            let mcols = Self::batched_mcols(m);
            return self.qmatvec_mmvq_batched(
                mbytes, aq, ad, m, in_f, out_f, qtype, row_bytes, mcols, scale, mrp,
            );
        }
        // F8-E4M3 catch-all (m=9..15 / batched-disabled seams): grid.y=m e4m3 mmvq — this dtype
        // has NO _dp4a twin, and per (token,row) the mmvq body is the exact m=1 decode program.
        // Q4_0 joins the catch-all (2026-07-11): adaptive-K cap 8 makes verify t=9 reachable
        // for the first time (past the b8 tier) and Q4_0 has no dp4a twin either. The mirror
        // (mbytes/mrp) keeps the rp layout consistent with the m=1 decode program.
        if qtype == QT_F8_E4M3 || qtype == QT_Q4_0 {
            let (b2, r2) = if qtype == QT_Q4_0 {
                (mbytes, mrp)
            } else {
                (bytes, rp)
            };
            return self.qmatvec_mmvq(b2, aq, ad, m, in_f, out_f, qtype, row_bytes, scale, r2);
        }
        let name = match qtype {
            QT_Q8_0 => "qmatvec_q8_0_dp4a",
            QT_Q4_K => "qmatvec_q4_K_dp4a",
            QT_Q6_K => "qmatvec_q6_K_dp4a",
            QT_Q5_K => "qmatvec_q5_K_dp4a",
            QT_Q3_K => "qmatvec_q3_K_dp4a",
            QT_NVFP4 => {
                if rp {
                    "qmatvec_nvfp4_dp4a_rp"
                } else {
                    "qmatvec_nvfp4_dp4a"
                }
            }
            QT_IQ4_XS => "qmatvec_iq4_XS_dp4a",
            _ => unreachable!(),
        };
        let f = self.func(name);
        let mut y = self.alloc_uninit::<f32>(m * out_f)?; // full-overwrite GEMM output: skip memset
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, m as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi, rb) = (in_f as i32, out_f as i32, m as i32, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(bytes)
            .arg(aq)
            .arg(ad)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        if scale != 1.0 {
            self.scale_inplace(&mut y, scale, m * out_f)?;
        }
        Ok(y)
    }

    /// DECODE-EXACT matmul at any m: guarantees the SAME warp-per-row (MMVQ, 32-thread) FP
    /// accumulation order as the T=1 decode path for EVERY token row. The spec-decode verify MUST
    /// use this for linear-attn projections to be bit-identical to greedy decode. The dp4a kernel
    /// (128 threads, two-level reduction) used by `matmul`/`matmul_pre` at m>=5 has a different
    /// shfl-tree shape that produces ULP differences propagating through gdn_scan into argmax flips.
    /// The MMVQ kernel with grid.y=m already processes each row independently (same 32-thread warp
    /// reduce as m=1); this method just forces that path unconditionally.
    pub fn matmul_decode_exact(
        &self,
        w: &crate::model::GpuTensor,
        x: &CudaSlice<f32>,
        m: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        // FLOAT tensors (35B ssm_beta/ssm_alpha on every linear layer, F32 ne=[2048,32]): the
        // generic path is cuBLASLt, whose reduction splits are n-DEPENDENT — m=1 vs m=2 col-0
        // outputs differ in every bit (probe 2026-07-06: 32/32 bit-diff, maxdiff 3.5e-3), which
        // shifted 35B verify logits 0.26-0.56 vs eager and flipped greedy at tight margins (the
        // p3 spec FAIL). Decode-exact contract: per-COLUMN m=1 cuBLASLt calls — each column's
        // reduction is the exact kernel the T=1 decode path runs, so verify==decode bit-for-bit.
        // m<=10 here (K+2 verify tier), so the extra launches are a handful of 4us gemvs.
        if let GpuTensor::Float { data, .. } = w {
            return self.linear_decode_exact(x, data, m, w.in_features(), w.out_features());
        }
        // MEMRA_FULL_PREC bf16-resident weight: dequant-on-use, then the per-column decode-exact
        // float linear (same n-independent reduction contract as the Float arm above).
        if let GpuTensor::FloatBf16 { data, .. } = w {
            let (in_f, out_f) = (w.in_features(), w.out_features());
            // Rows fast path: per-row t=1 program (STRONGER than the chunked per-column
            // contract — the whole-weight f32 dequant disappears too).
            if (1..=32).contains(&m) && Self::bf16_mmv_on() && in_f % 8 == 0 {
                let mut y = self.alloc_uninit::<f32>(m * out_f)?;
                self.matvec_bf16_rows_into(data, x, &mut y, in_f, out_f, m)?;
                return Ok(y);
            }
            return self.linear_bf16_chunked(x, data, m, in_f, out_f, true, None);
        }
        if !self.uses_q8_1_fast(w) {
            return self.matmul(w, x, m);
        }
        let in_f = w.in_features();
        let out_f = w.out_features();
        let (bytes, qtype, row_bytes, scale, rp) = match w {
            GpuTensor::Quant {
                bytes,
                qtype,
                row_bytes,
                scale,
                rp,
                ..
            } => (bytes, *qtype, *row_bytes, *scale, *rp),
            _ => return self.matmul(w, x, m),
        };
        // Q4_0 split-plane mirror for the mmvq/batched arms below (dp4a tail = matmul_pre,
        // which does its own mirror pick).
        let (bytes, rp) = match w {
            GpuTensor::Quant { rp4: Some(m4), .. } => (m4, true),
            _ => (bytes, rp),
        };
        let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
        // BLOCK-128 e4m3 (QT_F8_E4M3_BLK): the same single kernel every other entry dispatches, so
        // the decode-exact contract needs nothing special — grid.y=m runs the m=1 program per
        // (token,row) by construction, which is exactly what this method exists to guarantee.
        if let Some(y) = self.try_e4m3_blk_pre(w, &aq, &ad, m)? {
            return Ok(y);
        }
        // Batched weight-resident matvec for m=2-8: BIT-IDENTICAL per (token,row) to MMVQ (exact
        // integer dp4a, same warp reduce — kernel-check gate rel=0.00e0), one weight read for m
        // tokens. The dispatch the divergence fix must avoid is dp4a's 128-thread two-level
        // reduce, NOT this. m=5..8 is the K=4..7 spec-verify tier (b8): pre-b8 T=5 fell to the
        // grid.y=m per-row MMVQ below = 5 full weight reads/launch — the measured 27B K=4 cliff.
        // DECODE-PARITY GATE (2026-07-07): batched (MMVQ-class order) only when the m=1 decode
        // chain rides MMVQ too — without MEMRA_MMVQ decode is dp4a, so the exact-contract here
        // must be per-column dp4a (matmul_pre fallthrough), not the MMVQ order.
        if (2..=16).contains(&m) && self.batched_supports(qtype) && self.mmvq_supports(qtype)
            && std::env::var("MEMRA_NO_BATCHED").is_err()
            && (m <= 4 || Self::b8_enabled())
            // Every b16 class has base + _rp twins after lane/rp-on-st (see matmul_pre's note):
            // no mirror precondition, `rp` selects the layout only.
            && (m <= 8 || qtype == QT_Q4_0 || qtype == QT_Q6_K || qtype == QT_F8_E4M3
                || qtype == QT_NVFP4 || qtype == QT_Q4_K || qtype == QT_Q5_K || qtype == QT_Q8_0)
        {
            let mcols = Self::batched_mcols(m);
            return self.qmatvec_mmvq_batched(
                bytes, &aq, &ad, m, in_f, out_f, qtype, row_bytes, mcols, scale, rp,
            );
        }
        if self.mmvq_supports(qtype) {
            // MMVQ at grid.y=m: each row is processed by its own warp independently — same 32-thread
            // accumulation + warp_reduce_sum as m=1 decode. Bit-identical per row.
            return self.qmatvec_mmvq(bytes, &aq, &ad, m, in_f, out_f, qtype, row_bytes, scale, rp);
        }
        // Fallback for non-MMVQ quant types (Q5_K, Q3_K): use dp4a (the only available kernel).
        // These types are not used in the 27B's linear-attn NVFP4+Q4_K layers.
        self.matmul_pre(w, &aq, &ad, x, m)
    }

    /// DECODE-EXACT matmul from a PRE-QUANTIZED q8_1 activation (batched-verify epilogue
    /// re-fuse, lane/vt-fixes fix 2, 2026-08-03): the EXACT `matmul_decode_exact` dispatch for
    /// q8_1-fast Quant tensors, with the caller's (aq, ad) replacing the internal
    /// `quantize_q8_1`. quantize_q8_1 is deterministic (same input bytes -> same q8 bytes), so
    /// sharing one quantize across sibling matmuls of the same activation — or consuming the
    /// q8 emitted by a fused epilogue (rms_norm_q8_1 / add_rms_norm_q8_1 /
    /// silu_mul_scaled_q8_1 / gated_rmsnorm_q8_1, all kernel-check-pinned bit-identical to
    /// their unfused chains) — cannot change any dispatched kernel's input bytes.
    /// Caller MUST guarantee `uses_q8_1_fast(w)` (the fused epilogues only exist on that path).
    pub fn matmul_decode_exact_pre(
        &self,
        w: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        debug_assert!(
            self.uses_q8_1_fast(w),
            "matmul_decode_exact_pre: caller must guarantee q8_1-fast"
        );
        // BLOCK-128 e4m3: same single kernel, all m — see matmul_decode_exact's note.
        if let Some(y) = self.try_e4m3_blk_pre(w, aq, ad, m)? {
            return Ok(y);
        }
        let in_f = w.in_features();
        let out_f = w.out_features();
        let (bytes, qtype, row_bytes, scale, rp) = match w {
            GpuTensor::Quant {
                bytes,
                qtype,
                row_bytes,
                scale,
                rp,
                ..
            } => (bytes, *qtype, *row_bytes, *scale, *rp),
            _ => {
                return Err(
                    "matmul_decode_exact_pre: Quant tensor required (q8_1-fast contract)".into(),
                );
            }
        };
        // Q4_0 split-plane mirror — same pick as matmul_decode_exact.
        let (bytes, rp) = match w {
            GpuTensor::Quant { rp4: Some(m4), .. } => (m4, true),
            _ => (bytes, rp),
        };
        // Dispatch mirror of matmul_decode_exact's q8_1-fast tail, condition for condition.
        if (2..=16).contains(&m)
            && self.batched_supports(qtype)
            && self.mmvq_supports(qtype)
            && std::env::var("MEMRA_NO_BATCHED").is_err()
            && (m <= 4 || Self::b8_enabled())
            && (m <= 8
                || qtype == QT_Q4_0
                || qtype == QT_Q6_K
                || qtype == QT_F8_E4M3
                || qtype == QT_NVFP4
                || qtype == QT_Q4_K
                || qtype == QT_Q5_K
                || qtype == QT_Q8_0)
        {
            let mcols = Self::batched_mcols(m);
            return self.qmatvec_mmvq_batched(
                bytes, aq, ad, m, in_f, out_f, qtype, row_bytes, mcols, scale, rp,
            );
        }
        if self.mmvq_supports(qtype) {
            return self.qmatvec_mmvq(bytes, aq, ad, m, in_f, out_f, qtype, row_bytes, scale, rp);
        }
        // Non-MMVQ quant types (Q5_K/Q3_K under MEMRA_MMVQ=0): dp4a via matmul_pre — the same
        // fallback matmul_decode_exact takes. m <= 16 on the verify tier never reads x_fallback.
        let x0 = self.zeros(0)?;
        self.matmul_pre(w, aq, ad, &x0, m)
    }

    /// DUAL gate+up batched matvec from a PRE-QUANTIZED activation, macro-scales DEFERRED
    /// (lane/vt-fixes fix 2): same eligibility as `matmul_decode_exact_dual`, but the caller's
    /// (aq, ad) replaces the internal quantize and the NVFP4 per-tensor scales are RETURNED
    /// instead of applied via two `scale_inplace` launches — the fused SwiGLU epilogue
    /// (`silu_mul_scaled_q8_1`) folds them, exactly like the m=1 decode chain does. Deferring
    /// is value-exact: `y[i]*s` inline in the epilogue is the same IEEE multiply scale_inplace
    /// would store (f32 store/load round-trips are exact). None -> caller falls back to the
    /// per-tensor path.
    pub fn matmul_decode_exact_dual_pre(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
    ) -> Result<Option<((CudaSlice<f32>, f32), (CudaSlice<f32>, f32))>, Box<dyn std::error::Error>>
    {
        use crate::model::GpuTensor;
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let on = *ON.get_or_init(|| {
            std::env::var("MEMRA_SPEC_DUAL_T")
                .map(|v| v != "0")
                .unwrap_or(true)
        });
        if !on
            || !(2..=7).contains(&m)
            || std::env::var("MEMRA_NO_BATCHED").is_ok()
            || !self.uses_q8_1_fast(w0)
            || !self.uses_q8_1_fast(w1)
        {
            return Ok(None);
        }
        // DECODE-PARITY GATE (lane/nvfp4-strict, 2026-08-05): batched iff MMVQ — the dual
        // kernels are the MMVQ warp-reduce family, and without MEMRA_MMVQ the m=1 decode
        // chain this verify must match bit-for-bit rides dp4a (see matmul_decode_exact's
        // note). The singles enforce this via `mmvq_supports`; the dual door skipped it.
        if !self.mmvq_supports(QT_NVFP4) {
            return Ok(None);
        }
        let (in_f, out_f) = (w0.in_features(), w0.out_features());
        if w1.in_features() != in_f || w1.out_features() != out_f {
            return Ok(None);
        }
        let (b0, b1, row_bytes, s0, s1, rp) = match (w0, w1) {
            (
                GpuTensor::Quant {
                    bytes: b0,
                    qtype: q0,
                    row_bytes: rb0,
                    scale: s0,
                    rp: rp0,
                    rp4: None,
                    ..
                },
                GpuTensor::Quant {
                    bytes: b1,
                    qtype: q1,
                    row_bytes: rb1,
                    scale: s1,
                    rp: rp1,
                    rp4: None,
                    ..
                },
            ) if *q0 == QT_NVFP4 && *q1 == QT_NVFP4 && rb0 == rb1 && rp0 == rp1 => {
                (b0, b1, *rb0, *s0, *s1, *rp0)
            }
            _ => return Ok(None),
        };
        // m=5..7: only the exact-width rp duals exist (vt-fixes fix 1b); GGUF layout keeps
        // the singles. The b8 dual (MCOLS=8 at m=5..8) measured FLAT and stays dead.
        if m > 4 && !(rp && Self::b8_enabled() && std::env::var("MEMRA_B567").as_deref() != Ok("0"))
        {
            return Ok(None);
        }
        let (y0, y1) =
            self.qmatvec_batched_dual_raw(b0, b1, aq, ad, m, in_f, out_f, row_bytes, rp)?;
        Ok(Some(((y0, s0), (y1, s1))))
    }

    /// GROUP-4 GDN-tuple BATCHED matvec (trunk-kernels slice C): the qwen35 linear-layer
    /// in-projection 4-tuple (wqkv / wqkv_gate / ssm_beta / ssm_alpha) from ONE pre-quantized
    /// activation in ONE launch. Blocks map to the concatenated row space; every out_f must be
    /// a multiple of 8 (rows_per_block) so each warp's row pair resolves to one tensor; per
    /// (tensor, token, row) the kernel body is `nvfp4_mmvq_batched_rp` VERBATIM with the
    /// tensor's macro-scale fused at the write (== the conditional scale_inplace pass,
    /// bit-identical) -> BIT-IDENTICAL to the four single launches. Split-plane rp NVFP4 only,
    /// m=2..16 (exact-width MCOLS at m=5..7 mirroring the B567 law; m>4 requires b8_enabled
    /// like the singles; m=9..=16 rides the b16 form — the E4 width lift that lets
    /// `matmul_nvfp4_fused3/4` delegate that class here, lane/orndecode2).
    /// None -> caller runs the four singles. MEMRA_TK_GDN_GROUP=0 rollback.
    pub fn matmul_decode_exact_group4_pre(
        &self,
        ws: [&crate::model::GpuTensor; 4],
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
    ) -> Result<Option<Vec<CudaSlice<f32>>>, Box<dyn std::error::Error>> {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let on = *ON.get_or_init(|| {
            std::env::var("MEMRA_TK_GDN_GROUP")
                .map(|v| v != "0")
                .unwrap_or(true)
        });
        self.matmul_decode_exact_group_pre(&ws, aq, ad, m, on, "GDN group4")
    }

    /// GROUP-3 twin for the qwen35 full-attention q/k/v triple (trunk-kernels slice D):
    /// the SAME group4 kernels with n3=0 (blocks never reach the fourth range; W3/y3 are
    /// never dereferenced) — per (tensor, token, row) bit-identical to the three singles
    /// exactly as the group4 door is to its four. MEMRA_TK_FA_GROUP=0 rollback.
    pub fn matmul_decode_exact_group3_pre(
        &self,
        ws: [&crate::model::GpuTensor; 3],
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
    ) -> Result<Option<Vec<CudaSlice<f32>>>, Box<dyn std::error::Error>> {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let on = *ON.get_or_init(|| {
            std::env::var("MEMRA_TK_FA_GROUP")
                .map(|v| v != "0")
                .unwrap_or(true)
        });
        self.matmul_decode_exact_group_pre(&ws, aq, ad, m, on, "FA group3")
    }

    /// Shared core of the group3/group4 doors: eligibility mirror of the singles' batched
    /// dispatch, then ONE `qmatvec_nvfp4_mmvq_group4_b*_rp` launch over the concatenated
    /// row space (3-tensor callers ride n3=0). Returns one output per input tensor.
    fn matmul_decode_exact_group_pre(
        &self,
        ws: &[&crate::model::GpuTensor],
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
        on: bool,
        tag: &'static str,
    ) -> Result<Option<Vec<CudaSlice<f32>>>, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        if !on
            || !(2..=16).contains(&m)
            || std::env::var("MEMRA_NO_BATCHED").is_ok()
            || (m > 4 && !Self::b8_enabled())
            || !self.mmvq_supports(QT_NVFP4)
            || !self.batched_supports(QT_NVFP4)
        {
            return Ok(None);
        }
        let in_f = ws[0].in_features();
        let mut parts: Vec<(&CudaSlice<u8>, usize, f32)> = Vec::with_capacity(4);
        for w in ws {
            if !self.uses_q8_1_fast(w) || w.in_features() != in_f {
                return Ok(None);
            }
            match w {
                GpuTensor::Quant {
                    bytes,
                    qtype,
                    scale,
                    rp: true,
                    rp4: None,
                    ..
                } if *qtype == QT_NVFP4 && w.out_features() % 8 == 0 => {
                    parts.push((bytes, w.out_features(), *scale));
                }
                _ => return Ok(None),
            }
        }
        // MCOLS tier mirrors the singles: batched_mcols + the B567 exact-width law at m=5..7.
        static B567: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let b567 = *B567.get_or_init(|| std::env::var("MEMRA_B567").as_deref() != Ok("0"));
        let mcols = if (5..=7).contains(&m) && b567 {
            m
        } else {
            Self::batched_mcols(m)
        };
        let kname: &'static str = match mcols {
            2 => "qmatvec_nvfp4_mmvq_group4_b2_rp",
            4 => "qmatvec_nvfp4_mmvq_group4_b4_rp",
            5 => "qmatvec_nvfp4_mmvq_group4_b5_rp",
            6 => "qmatvec_nvfp4_mmvq_group4_b6_rp",
            7 => "qmatvec_nvfp4_mmvq_group4_b7_rp",
            8 => "qmatvec_nvfp4_mmvq_group4_b8_rp",
            16 => "qmatvec_nvfp4_mmvq_group4_b16_rp",
            _ => return Ok(None),
        };
        // Engagement receipt PER DOOR (dead-arm lesson): one shared Once here suppressed
        // the second door's print on the slice-D battery — key the once-set by tag.
        if std::env::var("MEMRA_DEBUG").is_ok() {
            use std::sync::Mutex;
            static SEEN: Mutex<Vec<&'static str>> = Mutex::new(Vec::new());
            let mut seen = SEEN.lock().unwrap();
            if !seen.contains(&tag) {
                seen.push(tag);
                eprintln!("[memra] {tag} batched ENGAGED (m={m})");
            }
        }
        const ROWS_PER_BLOCK: u32 = 4; // matches MEMRA_MMVQ_ROWS in qmatvec.cu
        let rows_per_block = ROWS_PER_BLOCK * 2; // WROWS=2 in the group kernel
        let total: usize = parts.iter().map(|p| p.1).sum();
        let three = parts.len() == 3;
        let mut y0 = self.alloc_uninit::<f32>(m * parts[0].1)?;
        let mut y1 = self.alloc_uninit::<f32>(m * parts[1].1)?;
        let mut y2 = self.alloc_uninit::<f32>(m * parts[2].1)?;
        // 3-tensor callers: n3=0 means no block ever resolves to the fourth range — W3/y3
        // are never dereferenced; a 1-element dummy keeps the launch ABI without aliasing y0.
        let mut y3 = self.alloc_uninit::<f32>(if three { 1 } else { m * parts[3].1 })?;
        let cfg = LaunchConfig {
            grid_dim: ((total as u32 + rows_per_block - 1) / rows_per_block, 1, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (inf, mi) = (in_f as i32, m as i32);
        let (n0, n1, n2) = (parts[0].1 as i32, parts[1].1 as i32, parts[2].1 as i32);
        let n3 = if three { 0i32 } else { parts[3].1 as i32 };
        let (s0, s1, s2) = (parts[0].2, parts[1].2, parts[2].2);
        let s3 = if three { 1.0f32 } else { parts[3].2 };
        let w3 = if three { parts[0].0 } else { parts[3].0 };
        let f = self.func(kname);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(parts[0].0)
            .arg(parts[1].0)
            .arg(parts[2].0)
            .arg(w3)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&mut y2)
            .arg(&mut y3)
            .arg(&inf)
            .arg(&n0)
            .arg(&n1)
            .arg(&n2)
            .arg(&n3)
            .arg(&mi)
            .arg(&s0)
            .arg(&s1)
            .arg(&s2)
            .arg(&s3);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(Some(if three {
            vec![y0, y1, y2]
        } else {
            vec![y0, y1, y2, y3]
        }))
    }

    /// DUAL gate+up BATCHED matvec at verify t=2..8 (lane/verify-economics, 2026-08-02): ONE
    /// launch computes both FFN projections of a verify batch — same activation, same shape,
    /// blockIdx.y selects the tensor. Per (tensor, token, row) the kernel body is the single
    /// batched program on the SAME layout (split-plane rp: b2 rp / b4 rpr2 / b8 rpr2; GGUF:
    /// b2 base / b4 r2 / b8 r2) -> BIT-IDENTICAL to the two single `matmul_decode_exact`
    /// launches (kernel-check gates bitwise on both layouts; run-spec K=1..8 arbitrates e2e).
    /// The one activation quantize replaces two IDENTICAL quantizes of the same `x` (same
    /// kernel, same input -> same q8_1 bytes), and the two independent weight streams in one
    /// grid restore the memory-level parallelism the two-launch form loses to tail drain +
    /// launch gap (m=1 dual_mr2 precedent: DRAM 40% -> 47-50% on the 27B pair).
    /// `Some((y0, y1))` only when both tensors are NVFP4, the SAME layout (both rp or both
    /// GGUF, no rp4 mirror), identical (in_f, out_f, row_bytes), q8_1-fast, and m in 2..=4
    /// (the b2/b4 tiers = verify T for K=1..3, the profitable-K window — the b8 dual measured
    /// FLAT vs the rpsc singles x3 interleaved, research/verify-economics-20260802, and was
    /// killed per doctrine). None -> caller runs the two singles. MEMRA_SPEC_DUAL_T=0 rollback.
    pub fn matmul_decode_exact_dual(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        x: &CudaSlice<f32>,
        m: usize,
    ) -> Result<Option<(CudaSlice<f32>, CudaSlice<f32>)>, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let on = *ON.get_or_init(|| {
            std::env::var("MEMRA_SPEC_DUAL_T")
                .map(|v| v != "0")
                .unwrap_or(true)
        });
        if !on
            || !(2..=4).contains(&m)
            || std::env::var("MEMRA_NO_BATCHED").is_ok()
            || !self.uses_q8_1_fast(w0)
            || !self.uses_q8_1_fast(w1)
        {
            return Ok(None);
        }
        // DECODE-PARITY GATE (lane/nvfp4-strict, 2026-08-05): batched iff MMVQ — same law as
        // the singles' `batched_supports && mmvq_supports` check in matmul_decode_exact,
        // which this dual door bypassed. Without MEMRA_MMVQ the m=1 decode is dp4a; the
        // verify must ride the per-column dp4a class, not the MMVQ-family dual.
        if !self.mmvq_supports(QT_NVFP4) {
            return Ok(None);
        }
        let (in_f, out_f) = (w0.in_features(), w0.out_features());
        if w1.in_features() != in_f || w1.out_features() != out_f {
            return Ok(None);
        }
        let (b0, b1, row_bytes, s0, s1, rp) = match (w0, w1) {
            (
                GpuTensor::Quant {
                    bytes: b0,
                    qtype: q0,
                    row_bytes: rb0,
                    scale: s0,
                    rp: rp0,
                    rp4: None,
                    ..
                },
                GpuTensor::Quant {
                    bytes: b1,
                    qtype: q1,
                    row_bytes: rb1,
                    scale: s1,
                    rp: rp1,
                    rp4: None,
                    ..
                },
            ) if *q0 == QT_NVFP4 && *q1 == QT_NVFP4 && rb0 == rb1 && rp0 == rp1 => {
                (b0, b1, *rb0, *s0, *s1, *rp0)
            }
            _ => return Ok(None),
        };
        // Engagement receipt (MEMRA_DEBUG=1): the first dead-arm A/B lesson — a `rp: false`
        // gate silently no-op'd the whole experiment; prove the arm is live in the log.
        if std::env::var("MEMRA_DEBUG").is_ok() {
            static ONCE: std::sync::Once = std::sync::Once::new();
            ONCE.call_once(|| eprintln!("[memra] dual gate+up batched ENGAGED (m={m} rp={rp})"));
        }
        let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
        let (y0, y1) =
            self.qmatvec_batched_dual_raw(b0, b1, &aq, &ad, m, in_f, out_f, row_bytes, rp)?;
        let mut y0 = y0;
        let mut y1 = y1;
        if s0 != 1.0 {
            self.scale_inplace(&mut y0, s0, m * out_f)?;
        }
        if s1 != 1.0 {
            self.scale_inplace(&mut y1, s1, m * out_f)?;
        }
        Ok(Some((y0, y1)))
    }

    /// Launch body of the dual batched twins from raw NVFP4 weight bytes + a pre-quantized q8_1
    /// activation (kernel-check's bit-equivalence entry; matmul_decode_exact_dual's core).
    /// mcols tier = batched_mcols(m); macro-scale NOT applied. `rp` selects the split-plane
    /// twins (both buffers must be the repacked layout).
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_batched_dual_raw(
        &self,
        b0: &CudaSlice<u8>,
        b1: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
        rp: bool,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        const ROWS_PER_BLOCK: u32 = 4;
        let mcols = Self::batched_mcols(m);
        // EXACT-WIDTH duals at m=5..7 (vt-fixes fix 1b): rp-only; bit-identical to the two
        // b5/b6/b7 singles (blockIdx.y selects the tensor, same template body).
        let tiny_rp1 = rp
            && mcols == 4
            && out_f <= 128
            && std::env::var("MEMRA_NVFP4_AUX_DUAL").as_deref() != Ok("0");
        let (name, rows_per_block) = if tiny_rp1 {
            ("qmatvec_nvfp4_mmvq_dual_b4_rp", ROWS_PER_BLOCK)
        } else {
            match (mcols, rp, m) {
                (2, false, _) => ("qmatvec_nvfp4_mmvq_dual_b2", ROWS_PER_BLOCK),
                (4, false, _) => ("qmatvec_nvfp4_mmvq_dual_b4_r2", ROWS_PER_BLOCK * 2),
                (2, true, _) => ("qmatvec_nvfp4_mmvq_dual_b2_rp", ROWS_PER_BLOCK),
                (4, true, _) => ("qmatvec_nvfp4_mmvq_dual_b4_rpr2", ROWS_PER_BLOCK * 2),
                (8, true, 5) => ("qmatvec_nvfp4_mmvq_dual_b5_rpr2", ROWS_PER_BLOCK * 2),
                (8, true, 6) => ("qmatvec_nvfp4_mmvq_dual_b6_rpr2", ROWS_PER_BLOCK * 2),
                (8, true, 7) => ("qmatvec_nvfp4_mmvq_dual_b7_rpr2", ROWS_PER_BLOCK * 2),
                _ => {
                    return Err(
                        format!("qmatvec_batched_dual_raw: no dual kernel for m {m}").into(),
                    );
                }
            }
        };
        let f = self.func(name);
        let mut y0 = self.alloc_uninit::<f32>(m * out_f)?;
        let mut y1 = self.alloc_uninit::<f32>(m * out_f)?;
        let cfg = LaunchConfig {
            grid_dim: ((out_f as u32 + rows_per_block - 1) / rows_per_block, 2, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi, rb) = (in_f as i32, out_f as i32, m as i32, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((y0, y1))
    }

    /// Like `matmul_pre` but RETURNS THE RAW (un-macro-scaled) matmul output together with the
    /// per-tensor NVFP4 scale, instead of applying `scale_inplace` internally. Used by the fused
    /// SwiGLU epilogue (RANK3 LEVER 2) so the gate/up scales fold into one `silu_mul_scaled` launch.
    /// `Some((y_raw, scale))` only on the m==1 decode fast path (mmvq / dp4a) where the scale is a
    /// separate post-launch op we can defer; returns `None` for every other path (prefill GEMM, FP4
    /// GEMM, Stage-A, Float) so the caller falls back to the scaled `matmul_pre` + `silu_mul`.
    /// DUAL gate+up NVFP4 matvec (mm-fusion): ONE launch computes both projections (same
    /// activation, same shape) — grid.y selects the tensor. Bit-identical per element to two
    /// mr2 launches at m=1. Returns (gate_raw, up_raw) un-scaled (caller folds the two macro
    /// scales into the SwiGLU epilogue, same as the matmul_pre_noscale contract). None unless
    /// both tensors are NVFP4 q8_1-fast with identical (in_f, out_f, row_bytes) and m==1.
    pub fn matmul_pre_dual_noscale(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
    ) -> Result<Option<((CudaSlice<f32>, f32), (CudaSlice<f32>, f32))>, Box<dyn std::error::Error>>
    {
        use crate::model::GpuTensor;
        if m != 1 || !self.uses_q8_1_fast(w0) || !self.uses_q8_1_fast(w1) {
            return Ok(None);
        }
        // FP-ORDER LAW (lane/nvfp4-strict, 2026-08-05): every kernel this door can dispatch
        // (q8_0 fused2, nvfp4 dual_mr2) is the MMVQ family — 32-thread warp reduce. Without
        // MEMRA_MMVQ the m=1 singles ride dp4a (128-thread two-level reduce), so fusing here
        // would mix dispatch families across the pair — the exact class `q8_fused_params`
        // already refuses for Q8_0. The NVFP4 arm lacked this check, which is why
        // decode-batch-gate `--mode strict`'s equalizing env (MEMRA_MMVQ=0) never pinned
        // NVFP4 models: decode_step_h kept riding dual_mr2 while the batched body fell to
        // dp4a (gate1 maxdiff 1.639e-1 / gate2 step-8 divergence at the 2026-08-05 train
        // HEAD, research/nvfp4-strict-20260805/). Default env (MMVQ on) is dispatch-unchanged.
        if !self.mmvq_supports(QT_NVFP4) {
            return Ok(None);
        }
        let (in_f, out_f) = (w0.in_features(), w0.out_features());
        if w1.in_features() != in_f || w1.out_features() != out_f {
            return Ok(None);
        }
        // Q8_0 ARM (lane/q27-deepdive, 2026-08-05): the dense-FFN gate+up pair on a Q8_0 trunk fell
        // through this NVFP4-only gate to two `matmul_pre_noscale` launches — measured 128 of the
        // 1015 launches/token on q27-Q8_0 decode, the single largest un-fused class in the tick
        // (nsys `research/q27-deepdive-20260805/nsys/`). `q8_fused2_core` already serves the same
        // pair shape for the shared-expert gate/up, and its kernel body is `qmatvec_q8_0_mmvq`
        // VERBATIM per (tensor,row) -> BIT-IDENTICAL to the two separate launches. Q8_0 carries no
        // macro-scale (q8_fused_params requires scale==1.0), so the noscale contract is satisfied
        // by returning 1.0 for both: the SwiGLU epilogue's fold becomes the identity it already is
        // on this dtype today. Seam: MEMRA_Q8_FFN_FUSE2=0 rolls back to the two-launch pair.
        // rp4 guard: with MEMRA_Q8RP the singles route to the `_rp` split-plane twin over the
        // mirror buffer; the fused2 kernel has no `_rp` form, so fusing there would swap
        // dispatch families mid-model. Bail and let the two singles run (mirror lane unchanged).
        let no_mirror =
            |w: &crate::model::GpuTensor| !matches!(w, GpuTensor::Quant { rp4: Some(_), .. });
        if self.q8_ffn_fuse2_on()
            && no_mirror(w0)
            && no_mirror(w1)
            && let Some([p0, p1]) = self.q8_fused_params(&[w0, w1])
        {
            let (y0, y1) = self.q8_fused2_core(p0.0, p1.0, aq, ad, in_f, p0.1, p1.1, p0.2)?;
            return Ok(Some(((y0, 1.0), (y1, 1.0))));
        }
        // F8-E4M3 ARM (lane/fp8-decode-v1, 2026-08-05): with native e4m3 residency the FFN gate+up
        // pair (and the ssm beta+alpha dual, which routes through this same entry) fell through
        // both the NVFP4 gate below and the Q8_0 arm above to two `matmul_pre_noscale` launches —
        // native residency was UN-FUSING the trunk relative to the Q8_0 slab it replaces. The
        // fused2 kernel body is `qmatvec_e4m3_mmvq` VERBATIM per (tensor,row). Contract match:
        // `matmul_pre_noscale` on e4m3 launches with scale 1.0 and RETURNS the per-tensor
        // weight_scale for the caller to fold, so we pass ws=1.0 here and return (s0,s1) — same
        // bits, and the two macro-scale multiplies still fold into the SwiGLU epilogue.
        // MEMRA_E4M3_DUAL=0 rolls back to the two-launch pair.
        if let Some([p0, p1]) = self.e4m3_fused_params(&[w0, w1]) {
            let (y0, y1) =
                self.e4m3_fused2_core(p0.0, p1.0, aq, ad, in_f, p0.1, p1.1, p0.2, 1.0, 1.0)?;
            return Ok(Some(((y0, p0.3), (y1, p1.3))));
        }
        let (b0, q0, rb0, s0, rp0) = match w0 {
            GpuTensor::Quant {
                bytes,
                qtype,
                row_bytes,
                scale,
                rp,
                ..
            } => (bytes, *qtype, *row_bytes, *scale, *rp),
            _ => return Ok(None),
        };
        let (b1, q1, rb1, s1, rp1) = match w1 {
            GpuTensor::Quant {
                bytes,
                qtype,
                row_bytes,
                scale,
                rp,
                ..
            } => (bytes, *qtype, *row_bytes, *scale, *rp),
            _ => return Ok(None),
        };
        if q0 != QT_NVFP4 || q1 != QT_NVFP4 || rb0 != rb1 || rp0 != rp1 {
            return Ok(None);
        }
        const ROWS_PER_BLOCK: u32 = 4; // matches MEMRA_MMVQ_ROWS in qmatvec.cu
        const RPW: u32 = 2;
        let rows_per_block = ROWS_PER_BLOCK * RPW;
        let f = self.func(if rp0 {
            "qmatvec_nvfp4_mmvq_dual_mr2_rp"
        } else {
            "qmatvec_nvfp4_mmvq_dual_mr2"
        });
        let mut y0 = self.alloc_uninit::<f32>(out_f)?;
        let mut y1 = self.alloc_uninit::<f32>(out_f)?;
        let cfg = LaunchConfig {
            grid_dim: ((out_f as u32 + rows_per_block - 1) / rows_per_block, 2, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi, rb) = (in_f as i32, out_f as i32, 1i32, rb0 as i64);
        // noscale contract: the caller folds s0/s1 into the SwiGLU epilogue — the kernel's fused
        // yscale args stay 1.0 here (they exist for the single-tensor callers).
        let one = 1.0f32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&rb)
            .arg(&one)
            .arg(&one);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(Some(((y0, s0), (y1, s1))))
    }

    /// FUSED NVFP4 matvec TRIPLE with unequal out_f (rig-native decode increment 1,
    /// lane/rig-native-nvfp4): wq+wk+wv in ONE launch via the q8_0 fused2 block-offset
    /// recipe. Per (tensor,row,t) the kernel body is nvfp4_mmvq_multirow_rp VERBATIM ->
    /// bit-identical to three separate `matmul_pre` launches; yscales fold in-kernel exactly
    /// as the singles do. grid.y = m (the t-parallel verify rows ride the same launch).
    /// None when ineligible (not all rp NVFP4 / in_f mismatch / mmvq off) — callers fall
    /// back to the three singles.
    #[allow(clippy::too_many_arguments)]
    pub fn matmul_nvfp4_fused3(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        w2: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
    ) -> Result<Option<(CudaSlice<f32>, CudaSlice<f32>, CudaSlice<f32>)>, Box<dyn std::error::Error>>
    {
        use crate::model::GpuTensor;
        // m == 1 rides the original fused kernel; m = 2..=8 rides the BATCHED fused twin
        // (lane/samplat — the fused4_b8 pattern): the seg body is nvfp4_mmvq_batched_rp_sc
        // verbatim, weight rows read once for all m columns, bit-identical per
        // (tensor,row,column) to the three bN_rpsc singles. The old "at m>1 the fused
        // segments would re-read the weight per row" note described the grid.y=m lift,
        // which this twin deliberately is NOT.
        if !self.mmvq_supports(QT_NVFP4)
            || !self.uses_q8_1_fast(w0)
            || !self.uses_q8_1_fast(w1)
            || !self.uses_q8_1_fast(w2)
        {
            return Ok(None);
        }
        // m = 9..=16 (lane/orndecode2): the exact-16 tier's trio width rides the GROUP3
        // door — same family and bit-identity law as the fused4 delegate above.
        if (9..=16).contains(&m) {
            return Ok(
                match self.matmul_decode_exact_group3_pre([w0, w1, w2], aq, ad, m)? {
                    Some(mut ys) => {
                        let y2 = ys.pop().unwrap();
                        let y1 = ys.pop().unwrap();
                        let y0 = ys.pop().unwrap();
                        Some((y0, y1, y2))
                    }
                    None => None,
                },
            );
        }
        if !(1..=8).contains(&m) {
            return Ok(None);
        }
        if m > 1 {
            let in_f = w0.in_features();
            if std::env::var("MEMRA_NVFP4_FUSED3B").as_deref() == Ok("0")
                || !self.batched_supports(QT_NVFP4)
                || std::env::var("MEMRA_NO_BATCHED").is_ok()
                || (m > 4 && !Self::b8_enabled())
                || in_f % 512 != 0
                || in_f / 64 > 272
            {
                return Ok(None);
            }
        }
        let unpack = |w: &crate::model::GpuTensor| match w {
            GpuTensor::Quant {
                bytes,
                qtype,
                scale,
                rp,
                ..
            } if *qtype == QT_NVFP4 && *rp => Some((bytes as *const CudaSlice<u8>, *scale)),
            _ => None,
        };
        let (Some(p0), Some(p1), Some(p2)) = (unpack(w0), unpack(w1), unpack(w2)) else {
            return Ok(None);
        };
        let in_f = w0.in_features();
        if w1.in_features() != in_f || w2.in_features() != in_f {
            return Ok(None);
        }
        let (o0, o1, o2) = (w0.out_features(), w1.out_features(), w2.out_features());
        const ROWS_PER_BLOCK: u32 = 4; // MEMRA_MMVQ_ROWS
        const RPW: u32 = 2;
        let rows_pb = ROWS_PER_BLOCK * RPW;
        let nb = |o: usize| (o as u32).div_ceil(rows_pb);
        let mut y0 = self.alloc_uninit::<f32>(m * o0)?;
        let mut y1 = self.alloc_uninit::<f32>(m * o1)?;
        let mut y2 = self.alloc_uninit::<f32>(m * o2)?;
        let (inf, oi0, oi1, oi2, mi) = (in_f as i32, o0 as i32, o1 as i32, o2 as i32, m as i32);
        // SAFETY: the raw pointers come straight from the &GpuTensor borrows above and are
        // only dereferenced for the launch-arg build inside this call.
        let (b0, b1, b2) = unsafe { (&*p0.0, &*p1.0, &*p2.0) };
        if m > 1 {
            // batched twin has no in-kernel scale — refuse scale carriers (GGUF trunk = 1.0).
            if p0.1 != 1.0 || p1.1 != 1.0 || p2.1 != 1.0 {
                return Ok(None);
            }
            let f = self.func("qmatvec_nvfp4_mmvq_fused3_b8_rpsc");
            let cfg = LaunchConfig {
                grid_dim: (nb(o0) + nb(o1) + nb(o2), 1, 1),
                block_dim: (32, ROWS_PER_BLOCK, 1),
                shared_mem_bytes: 0,
            };
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(b0)
                .arg(b1)
                .arg(b2)
                .arg(aq)
                .arg(ad)
                .arg(&mut y0)
                .arg(&mut y1)
                .arg(&mut y2)
                .arg(&inf)
                .arg(&oi0)
                .arg(&oi1)
                .arg(&oi2)
                .arg(&mi);
            unsafe {
                b.launch(cfg)?;
            }
            return Ok(Some((y0, y1, y2)));
        }
        let f = self.func("qmatvec_nvfp4_mmvq_fused3_rp");
        let cfg = LaunchConfig {
            grid_dim: (nb(o0) + nb(o1) + nb(o2), m as u32, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(b2)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&mut y2)
            .arg(&inf)
            .arg(&oi0)
            .arg(&oi1)
            .arg(&oi2)
            .arg(&mi)
            .arg(&p0.1)
            .arg(&p1.1)
            .arg(&p2.1);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(Some((y0, y1, y2)))
    }

    /// fused2 twin of `matmul_nvfp4_fused3`, for MIXED-type weight groups: the gemma4
    /// dense NVFP4mix recipe keeps `attn_v` and `ffn_down` at Q8_0 (full-NVFP4 was
    /// measured garbage on this dense class), so its q/k/v trio and gate/up/down never
    /// satisfy an all-NVFP4 fused3 — the pairs that ARE uniformly NVFP4 (q,k and
    /// gate,up) fuse here instead. m==1 only, same law as fused3/fused4: per
    /// (tensor,row) the kernel seg body is VERBATIM, so the fusion is bit-identical to
    /// two separate launches. `MEMRA_NVFP4_FUSED2=0` is the rollback seam and the
    /// same-binary interleaved A/B arm.
    pub fn matmul_nvfp4_fused2(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
    ) -> Result<Option<(CudaSlice<f32>, CudaSlice<f32>)>, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        static FUSED2_OFF: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let off =
            *FUSED2_OFF.get_or_init(|| std::env::var("MEMRA_NVFP4_FUSED2").as_deref() == Ok("0"));
        // m==1 ONLY: at m>1 the singles ride the _b16 weight-once column program (one weight
        // read serves all m rows); the fused segments would re-read the weight per row.
        if off
            || m != 1
            || !self.mmvq_supports(QT_NVFP4)
            || !self.uses_q8_1_fast(w0)
            || !self.uses_q8_1_fast(w1)
        {
            return Ok(None);
        }
        let unpack = |w: &crate::model::GpuTensor| match w {
            GpuTensor::Quant {
                bytes,
                qtype,
                scale,
                rp,
                ..
            } if *qtype == QT_NVFP4 && *rp => Some((bytes as *const CudaSlice<u8>, *scale)),
            _ => None,
        };
        let (Some(p0), Some(p1)) = (unpack(w0), unpack(w1)) else {
            return Ok(None);
        };
        let in_f = w0.in_features();
        if w1.in_features() != in_f {
            return Ok(None);
        }
        let (o0, o1) = (w0.out_features(), w1.out_features());
        const ROWS_PER_BLOCK: u32 = 4; // MEMRA_MMVQ_ROWS
        const RPW: u32 = 2;
        let rows_pb = ROWS_PER_BLOCK * RPW;
        let nb = |o: usize| (o as u32).div_ceil(rows_pb);
        let f = self.func("qmatvec_nvfp4_mmvq_fused2_rp");
        let mut y0 = self.alloc_uninit::<f32>(m * o0)?;
        let mut y1 = self.alloc_uninit::<f32>(m * o1)?;
        let cfg = LaunchConfig {
            grid_dim: (nb(o0) + nb(o1), m as u32, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (inf, oi0, oi1, mi) = (in_f as i32, o0 as i32, o1 as i32, m as i32);
        // SAFETY: the raw pointers come straight from the &GpuTensor borrows above and are
        // only dereferenced for the launch-arg build inside this call.
        let (b0, b1) = unsafe { (&*p0.0, &*p1.0) };
        // PDL wave-B: the kernel carries MEMRA_PDL_ENTRY — launch overlaps the producer's
        // drain (quantize/norm emitting aq/ad). Same math, same order, bit-identical.
        if Self::pdl_on() && Self::pdl_mmvq_on() && Self::pdl_nvfp4q8_on() {
            {
                use cudarc::driver::{DevicePtr, DevicePtrMut};
                let s = &self.gpu.stream();
                let (pw0, _g0) = b0.device_ptr(s);
                let (pw1, _g1) = b1.device_ptr(s);
                let (paq, _g2) = aq.device_ptr(s);
                let (pad, _g3) = ad.device_ptr(s);
                let (py0, _g4) = y0.device_ptr_mut(s);
                let (py1, _g5) = y1.device_ptr_mut(s);
                let (s0, s1) = (p0.1, p1.1);
                let mut ps = [
                    &pw0 as *const _ as *mut std::ffi::c_void,
                    &pw1 as *const _ as *mut _,
                    &paq as *const _ as *mut _,
                    &pad as *const _ as *mut _,
                    &py0 as *const _ as *mut _,
                    &py1 as *const _ as *mut _,
                    &inf as *const _ as *mut _,
                    &oi0 as *const _ as *mut _,
                    &oi1 as *const _ as *mut _,
                    &mi as *const _ as *mut _,
                    &s0 as *const _ as *mut _,
                    &s1 as *const _ as *mut _,
                ];
                unsafe {
                    self.launch_pdl(
                        "qmatvec_nvfp4_mmvq_fused2_rp",
                        cfg.grid_dim,
                        cfg.block_dim,
                        &mut ps,
                    )?;
                }
            }
            return Ok(Some((y0, y1)));
        }
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&inf)
            .arg(&oi0)
            .arg(&oi1)
            .arg(&mi)
            .arg(&p0.1)
            .arg(&p1.1);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(Some((y0, y1)))
    }

    /// Slot-fed fused2 twin (alloc-free capture lane): identical launch to
    /// `matmul_nvfp4_fused2`, caller-owned outputs — the gemma4 dc_slotted graph body
    /// needs zero mem nodes, so the allocating wrapper can't serve it. Returns false
    /// (decline) on any non-NVFP4/rp pair; the caller chains or refuses.
    pub fn matmul_nvfp4_fused2_into(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        y0: &mut CudaSlice<f32>,
        y1: &mut CudaSlice<f32>,
    ) -> Result<bool, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        static FUSED2_OFF: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let off =
            *FUSED2_OFF.get_or_init(|| std::env::var("MEMRA_NVFP4_FUSED2").as_deref() == Ok("0"));
        if off
            || !self.mmvq_supports(QT_NVFP4)
            || !self.uses_q8_1_fast(w0)
            || !self.uses_q8_1_fast(w1)
        {
            return Ok(false);
        }
        let unpack = |w: &crate::model::GpuTensor| match w {
            GpuTensor::Quant {
                bytes,
                qtype,
                scale,
                rp,
                ..
            } if *qtype == QT_NVFP4 && *rp => Some((bytes as *const CudaSlice<u8>, *scale)),
            _ => None,
        };
        let (Some(p0), Some(p1)) = (unpack(w0), unpack(w1)) else {
            return Ok(false);
        };
        let in_f = w0.in_features();
        if w1.in_features() != in_f {
            return Ok(false);
        }
        let (o0, o1) = (w0.out_features(), w1.out_features());
        if y0.len() < o0 || y1.len() < o1 {
            return Ok(false);
        }
        const ROWS_PER_BLOCK: u32 = 4; // MEMRA_MMVQ_ROWS
        const RPW: u32 = 2;
        let rows_pb = ROWS_PER_BLOCK * RPW;
        let nb = |o: usize| (o as u32).div_ceil(rows_pb);
        let f = self.func("qmatvec_nvfp4_mmvq_fused2_rp");
        let cfg = LaunchConfig {
            grid_dim: (nb(o0) + nb(o1), 1, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (inf, oi0, oi1, mi) = (in_f as i32, o0 as i32, o1 as i32, 1i32);
        // SAFETY: the raw pointers come straight from the &GpuTensor borrows above and are
        // only dereferenced for the launch-arg build inside this call.
        let (b0, b1) = unsafe { (&*p0.0, &*p1.0) };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(aq)
            .arg(ad)
            .arg(&mut *y0)
            .arg(&mut *y1)
            .arg(&inf)
            .arg(&oi0)
            .arg(&oi1)
            .arg(&mi)
            .arg(&p0.1)
            .arg(&p1.1);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(true)
    }

    /// fused4 twin of `matmul_nvfp4_fused3`: the Linear-mixer projection quartet
    /// (wqkv + wqkv_gate + ssm_beta + ssm_alpha) in one launch, m==1 only. Per
    /// (tensor,row) the kernel body is nvfp4_mmvq_multirow_rp VERBATIM — bit-identical
    /// to four separate launches (rig-native decode increment 2, RIG-NATIVE-DECODE.md).
    #[allow(clippy::type_complexity)]
    pub fn matmul_nvfp4_fused4(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        w2: &crate::model::GpuTensor,
        w3: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
    ) -> Result<
        Option<(
            CudaSlice<f32>,
            CudaSlice<f32>,
            CudaSlice<f32>,
            CudaSlice<f32>,
        )>,
        Box<dyn std::error::Error>,
    > {
        use crate::model::GpuTensor;
        // MEMRA_NVFP4_FUSED4=0: rollback seam + the same-binary interleaved A/B arm.
        // m == 1 rides the original fused kernel; m = 2..=8 rides the BATCHED fused twin
        // (lane/samplat, 2026-08-21): same quartet-in-one-launch shape, seg body =
        // nvfp4_mmvq_batched_rp_sc verbatim (weight rows read once for all m columns) —
        // bit-identical per (tensor,row,column) to the four bN_rpsc singles it replaces.
        // Admission mirrors the singles' batched gates below.
        if std::env::var("MEMRA_NVFP4_FUSED4").as_deref() == Ok("0")
            || !self.mmvq_supports(QT_NVFP4)
            || !self.uses_q8_1_fast(w0)
            || !self.uses_q8_1_fast(w1)
            || !self.uses_q8_1_fast(w2)
            || !self.uses_q8_1_fast(w3)
        {
            return Ok(None);
        }
        // m = 9..=16 (lane/orndecode2, the exact-16 tier's trunk): the rp-sc seg twins stop
        // at 8; this width class rides the GROUP4 door instead — nvfp4_mmvq_batched_rp<16,_>
        // body, the SAME family as the b16_rp singles the tier would otherwise launch four
        // times, bit-identical per (tensor,token,row) incl. the fused write-side scale.
        if (9..=16).contains(&m) {
            return Ok(
                match self.matmul_decode_exact_group4_pre([w0, w1, w2, w3], aq, ad, m)? {
                    Some(mut ys) => {
                        let y3 = ys.pop().unwrap();
                        let y2 = ys.pop().unwrap();
                        let y1 = ys.pop().unwrap();
                        let y0 = ys.pop().unwrap();
                        Some((y0, y1, y2, y3))
                    }
                    None => None,
                },
            );
        }
        if !(1..=8).contains(&m) {
            return Ok(None);
        }
        if m > 1 {
            // the batched-twin gates: the bN_rpsc program this must stay byte-identical to
            // (matmul_pre's batched arm), plus the rp-sc dispatch shape requirements.
            let in_f = w0.in_features();
            if !self.batched_supports(QT_NVFP4)
                || std::env::var("MEMRA_NO_BATCHED").is_ok()
                || (m > 4 && !Self::b8_enabled())
                || in_f % 512 != 0
                || in_f / 64 > 272
            {
                return Ok(None);
            }
        }
        let unpack = |w: &crate::model::GpuTensor| match w {
            GpuTensor::Quant {
                bytes,
                qtype,
                scale,
                rp,
                ..
            } if *qtype == QT_NVFP4 && *rp => Some((bytes as *const CudaSlice<u8>, *scale)),
            _ => None,
        };
        let (Some(p0), Some(p1), Some(p2), Some(p3)) =
            (unpack(w0), unpack(w1), unpack(w2), unpack(w3))
        else {
            return Ok(None);
        };
        let in_f = w0.in_features();
        if w1.in_features() != in_f || w2.in_features() != in_f || w3.in_features() != in_f {
            return Ok(None);
        }
        let (o0, o1, o2, o3) = (
            w0.out_features(),
            w1.out_features(),
            w2.out_features(),
            w3.out_features(),
        );
        const ROWS_PER_BLOCK: u32 = 4; // MEMRA_MMVQ_ROWS
        const RPW: u32 = 2;
        let rows_pb = ROWS_PER_BLOCK * RPW;
        let nb = |o: usize| (o as u32).div_ceil(rows_pb);
        let mut y0 = self.alloc_uninit::<f32>(m * o0)?;
        let mut y1 = self.alloc_uninit::<f32>(m * o1)?;
        let mut y2 = self.alloc_uninit::<f32>(m * o2)?;
        let mut y3 = self.alloc_uninit::<f32>(m * o3)?;
        let (inf, oi0, oi1, oi2, oi3, mi) = (
            in_f as i32,
            o0 as i32,
            o1 as i32,
            o2 as i32,
            o3 as i32,
            m as i32,
        );
        // SAFETY: the raw pointers come straight from the &GpuTensor borrows above and are
        // only dereferenced for the launch-arg build inside this call.
        let (b0, b1, b2, b3) = unsafe { (&*p0.0, &*p1.0, &*p2.0, &*p3.0) };
        if m > 1 {
            // Batched fused twin: no in-kernel scale (the bN_rpsc program has none) — refuse
            // scale-carrying tensors so the singles path keeps them (GGUF trunk scales are 1.0).
            if p0.1 != 1.0 || p1.1 != 1.0 || p2.1 != 1.0 || p3.1 != 1.0 {
                return Ok(None);
            }
            let f = self.func("qmatvec_nvfp4_mmvq_fused4_b8_rpsc");
            let cfg = LaunchConfig {
                grid_dim: (nb(o0) + nb(o1) + nb(o2) + nb(o3), 1, 1),
                block_dim: (32, ROWS_PER_BLOCK, 1),
                shared_mem_bytes: 0,
            };
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(b0)
                .arg(b1)
                .arg(b2)
                .arg(b3)
                .arg(aq)
                .arg(ad)
                .arg(&mut y0)
                .arg(&mut y1)
                .arg(&mut y2)
                .arg(&mut y3)
                .arg(&inf)
                .arg(&oi0)
                .arg(&oi1)
                .arg(&oi2)
                .arg(&oi3)
                .arg(&mi);
            unsafe {
                b.launch(cfg)?;
            }
            return Ok(Some((y0, y1, y2, y3)));
        }
        let f = self.func("qmatvec_nvfp4_mmvq_fused4_rp");
        let cfg = LaunchConfig {
            grid_dim: (nb(o0) + nb(o1) + nb(o2) + nb(o3), m as u32, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(b2)
            .arg(b3)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&mut y2)
            .arg(&mut y3)
            .arg(&inf)
            .arg(&oi0)
            .arg(&oi1)
            .arg(&oi2)
            .arg(&oi3)
            .arg(&mi)
            .arg(&p0.1)
            .arg(&p1.1)
            .arg(&p2.1)
            .arg(&p3.1);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(Some((y0, y1, y2, y3)))
    }

    /// FUSED Q8_0 m=1 matvec PAIR with UNEQUAL out_f (trunk launch-fusion, 2026-07-05). Folds two
    /// same-input q8_0 projections (35B trunk: wqkv+wqkv_gate 8192/4096, gate_shexp+up_shexp
    /// 512/512) into ONE launch via a block-offset split (blocks [0,nb0) -> w0, rest -> w1) — the
    /// dual-mr2 recipe with the same-out_f restriction lifted. Per (tensor,row) the kernel body is
    /// qmatvec_q8_0_mmvq VERBATIM -> BIT-IDENTICAL to two separate m=1 launches. Returns None when
    /// ineligible (not both Q8_0 / in_f mismatch / MEMRA_MMVQ off / MEMRA_Q8_DUAL=0) — caller falls
    /// back to the per-tensor path.
    pub fn matmul_q8_fused2(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
    ) -> Result<Option<(CudaSlice<f32>, CudaSlice<f32>)>, Box<dyn std::error::Error>> {
        // e4m3 twin (lane/fp8-decode-v1): this entry is the trunk's generic m=1 pair door
        // (wqkv+wqkv_gate, ssm_beta+alpha, gate_shexp+up_shexp), so admitting QT_F8_E4M3 here
        // fuses the NATIVE-RESIDENCY FP8 trunk at every existing call site with no call-site
        // change. Scale is folded in-kernel per range -> the returned buffers are already scaled,
        // exactly like the per-tensor `matmul_pre` e4m3 dispatch this replaces.
        if let Some([p0, p1]) = self.e4m3_fused_params(&[w0, w1]) {
            return Ok(Some(self.e4m3_fused2_core(
                p0.0,
                p1.0,
                aq,
                ad,
                w0.in_features(),
                p0.1,
                p1.1,
                p0.2,
                p0.3,
                p1.3,
            )?));
        }
        let Some([p0, p1]) = self.q8_fused_params(&[w0, w1]) else {
            return Ok(None);
        };
        Ok(Some(self.q8_fused2_core(
            p0.0,
            p1.0,
            aq,
            ad,
            w0.in_features(),
            p0.1,
            p1.1,
            p0.2,
        )?))
    }

    #[allow(clippy::too_many_arguments)]
    fn q8_fused2_core(
        &self,
        b0: &CudaSlice<u8>,
        b1: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        in_f: usize,
        out0: usize,
        out1: usize,
        row_bytes: usize,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        const ROWS_PER_BLOCK: u32 = 4; // matches MEMRA_MMVQ_ROWS in qmatvec.cu
        let nb0 = (out0 as u32).div_ceil(ROWS_PER_BLOCK);
        let nb1 = (out1 as u32).div_ceil(ROWS_PER_BLOCK);
        let f = self.func("qmatvec_q8_0_mmvq_fused2");
        let mut y0 = self.alloc_uninit::<f32>(out0)?;
        let mut y1 = self.alloc_uninit::<f32>(out1)?;
        let cfg = LaunchConfig {
            grid_dim: (nb0 + nb1, 1, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (inf, o0, o1, rbl) = (in_f as i32, out0 as i32, out1 as i32, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&inf)
            .arg(&o0)
            .arg(&o1)
            .arg(&rbl);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((y0, y1))
    }

    /// f32-activation entry for the fused2 pair: quantizes x to q8_1 ONCE then runs the fused
    /// launch — replaces two `matmul(w, x, 1)` calls that would each re-quantize the same x
    /// (35B shared-expert gate+up per MoE layer per token). Same bits: quantize_q8_1 is
    /// deterministic, the fused body is the MMVQ kernel verbatim. None when ineligible (the
    /// callers' m==1-under-MEMRA_FAST dispatch would take MMVQ; anything else falls back).
    pub fn matmul_q8_fused2_x(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        x: &CudaSlice<f32>,
    ) -> Result<Option<(CudaSlice<f32>, CudaSlice<f32>)>, Box<dyn std::error::Error>> {
        if !self.uses_q8_1_fast(w0) || !self.uses_q8_1_fast(w1) {
            return Ok(None);
        }
        if let Some([p0, p1]) = self.e4m3_fused_params(&[w0, w1]) {
            let (aq, ad) = self.quantize_q8_1(x, 1, w0.in_features())?;
            return Ok(Some(self.e4m3_fused2_core(
                p0.0,
                p1.0,
                &aq,
                &ad,
                w0.in_features(),
                p0.1,
                p1.1,
                p0.2,
                p0.3,
                p1.3,
            )?));
        }
        let Some([p0, p1]) = self.q8_fused_params(&[w0, w1]) else {
            return Ok(None);
        };
        let (aq, ad) = self.quantize_q8_1(x, 1, w0.in_features())?;
        Ok(Some(self.q8_fused2_core(
            p0.0,
            p1.0,
            &aq,
            &ad,
            w0.in_features(),
            p0.1,
            p1.1,
            p0.2,
        )?))
    }

    /// Test entry for the kernel_check gate: launch the fused2 kernel from raw weight bytes,
    /// quantizing the f32 activation internally (mirrors qmatvec_mmvq_raw; no env gating).
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_q8_fused2_raw(
        &self,
        b0: &CudaSlice<u8>,
        b1: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        in_f: usize,
        out0: usize,
        out1: usize,
        row_bytes: usize,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        let (aq, ad) = self.quantize_q8_1(x, 1, in_f)?;
        self.q8_fused2_core(b0, b1, &aq, &ad, in_f, out0, out1, row_bytes)
    }

    /// FUSED Q8_0 m=1 matvec TRIPLE (wq+wk+wv on the 35B full-attn layers: out_f 8192/512/512).
    /// Same block-offset recipe as `matmul_q8_fused2` with three ranges. BIT-IDENTICAL per
    /// (tensor,row) to three separate m=1 MMVQ launches.
    /// FUSED Q4_0 m=1 TRIPLE (gemma q/k/v — same quantized input; per (tensor,row) chain
    /// identical to the mr2 kernel). Returns None unless all three are Q4_0 with equal in_f.
    pub fn matmul_q4_fused3(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        w2: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
    ) -> Result<Option<(CudaSlice<f32>, CudaSlice<f32>, CudaSlice<f32>)>, Box<dyn std::error::Error>>
    {
        use crate::model::GpuTensor;
        let q4 = |w: &GpuTensor| -> Option<(usize, usize)> {
            match w {
                GpuTensor::Quant {
                    qtype, row_bytes, ..
                } if *qtype == QT_Q4_0 => Some((*row_bytes, w.out_features())),
                _ => None,
            }
        };
        let (Some((rb0, o0)), Some((rb1, o1)), Some((rb2, o2))) = (q4(w0), q4(w1), q4(w2)) else {
            return Ok(None);
        };
        if w0.in_features() != w1.in_features() || w0.in_features() != w2.in_features() {
            return Ok(None);
        }
        // Effective (bytes, rp) per tensor: mirror (rp4) OR the in-place swap (rp flag,
        // bytes already split). Mixed layouts cannot share one fused launch -> fall back to
        // the separate matvecs (each routes its own rp).
        fn eff(w: &GpuTensor) -> (&CudaSlice<u8>, bool) {
            match w {
                GpuTensor::Quant { bytes, rp4, rp, .. } => match rp4 {
                    Some(m) => (m, true),
                    None => (bytes, *rp),
                },
                _ => unreachable!(),
            }
        }
        let ((b0, rp0), (b1, rp1), (b2, rp2)) = (eff(w0), eff(w1), eff(w2));
        if rp0 != rp1 || rp1 != rp2 {
            return Ok(None);
        }
        let rp = rp0;
        let rpb: u32 = 4;
        // mr1 (one row/warp, 2026-07-14): follows the singles' MEMRA_Q40_MR default — the
        // fused t=1 kernels were left on mr2 when the singles flipped (DRAM-duty map:
        // fused3 57% / fused2 86%; small qkv segments starve under mr2's half grid).
        let mr1 = rp && Self::q40_mr1_on();
        let nb = |o: usize| {
            if mr1 {
                (o as u32).div_ceil(rpb)
            } else {
                (o as u32).div_ceil(2).div_ceil(rpb)
            }
        };
        let grid = nb(o0) + nb(o1) + nb(o2);
        let mut y0 = self.alloc_uninit::<f32>(o0)?;
        let mut y1 = self.alloc_uninit::<f32>(o1)?;
        let mut y2 = self.alloc_uninit::<f32>(o2)?;
        let f = self.func(if mr1 {
            "qmatvec_q4_0_mmvq_fused3_mr1_rp"
        } else if rp {
            "qmatvec_q4_0_mmvq_fused3_rp"
        } else {
            "qmatvec_q4_0_mmvq_fused3"
        });
        let cfg = LaunchConfig {
            grid_dim: (grid, 1, 1),
            block_dim: (32, rpb, 1),
            shared_mem_bytes: 0,
        };
        let inf = w0.in_features() as i32;
        let (oo0, oo1, oo2) = (o0 as i32, o1 as i32, o2 as i32);
        let (r0, r1, r2) = (rb0 as i64, rb1 as i64, rb2 as i64);
        // PDL wave-A (2026-07-23): the mr1 kernel carries MEMRA_PDL_ENTRY; only that
        // variant may take the programmatic-serialization launch.
        if mr1 && Self::pdl_on() && Self::pdl_mmvq_on() {
            {
                use cudarc::driver::{DevicePtr, DevicePtrMut};
                let s = &self.gpu.stream();
                let (p0, _g0) = b0.device_ptr(s);
                let (p1, _g1) = b1.device_ptr(s);
                let (p2, _g2) = b2.device_ptr(s);
                let (paq, _g3) = aq.device_ptr(s);
                let (pad, _g4) = ad.device_ptr(s);
                let (py0, _g5) = y0.device_ptr_mut(s);
                let (py1, _g6) = y1.device_ptr_mut(s);
                let (py2, _g7) = y2.device_ptr_mut(s);
                let mut ps = [
                    &p0 as *const _ as *mut std::ffi::c_void,
                    &p1 as *const _ as *mut _,
                    &p2 as *const _ as *mut _,
                    &paq as *const _ as *mut _,
                    &pad as *const _ as *mut _,
                    &py0 as *const _ as *mut _,
                    &py1 as *const _ as *mut _,
                    &py2 as *const _ as *mut _,
                    &inf as *const _ as *mut _,
                    &oo0 as *const _ as *mut _,
                    &oo1 as *const _ as *mut _,
                    &oo2 as *const _ as *mut _,
                    &r0 as *const _ as *mut _,
                    &r1 as *const _ as *mut _,
                    &r2 as *const _ as *mut _,
                ];
                unsafe {
                    self.launch_pdl(
                        "qmatvec_q4_0_mmvq_fused3_mr1_rp",
                        (grid, 1, 1),
                        (32, rpb, 1),
                        &mut ps,
                    )?;
                }
            }
            return Ok(Some((y0, y1, y2)));
        }
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(b2)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&mut y2)
            .arg(&inf)
            .arg(&oo0)
            .arg(&oo1)
            .arg(&oo2)
            .arg(&r0)
            .arg(&r1)
            .arg(&r2);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(Some((y0, y1, y2)))
    }

    /// Slot-fed fused3 twin (alloc-free capture lane): identical launch, caller-owned outputs.
    /// Returns Ok(false) when the fused path is unavailable (caller falls back).
    #[allow(clippy::too_many_arguments)]
    pub fn matmul_q4_fused3_into(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        w2: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        y0: &mut CudaSlice<f32>,
        y1: &mut CudaSlice<f32>,
        y2: &mut CudaSlice<f32>,
    ) -> Result<bool, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        let q4 = |w: &GpuTensor| -> Option<(usize, usize)> {
            match w {
                GpuTensor::Quant {
                    qtype, row_bytes, ..
                } if *qtype == QT_Q4_0 => Some((*row_bytes, w.out_features())),
                _ => None,
            }
        };
        let (Some((rb0, o0)), Some((rb1, o1)), Some((rb2, o2))) = (q4(w0), q4(w1), q4(w2)) else {
            return Ok(false);
        };
        if w0.in_features() != w1.in_features() || w0.in_features() != w2.in_features() {
            return Ok(false);
        }
        fn eff(w: &GpuTensor) -> (&CudaSlice<u8>, bool) {
            match w {
                GpuTensor::Quant { bytes, rp4, rp, .. } => match rp4 {
                    Some(m) => (m, true),
                    None => (bytes, *rp),
                },
                _ => unreachable!(),
            }
        }
        let ((b0, rp0), (b1, rp1), (b2, rp2)) = (eff(w0), eff(w1), eff(w2));
        if rp0 != rp1 || rp1 != rp2 {
            return Ok(false);
        }
        let rp = rp0;
        let rpb: u32 = 4;
        let mr1 = rp && Self::q40_mr1_on();
        let nb = |o: usize| {
            if mr1 {
                (o as u32).div_ceil(rpb)
            } else {
                (o as u32).div_ceil(2).div_ceil(rpb)
            }
        };
        let grid = nb(o0) + nb(o1) + nb(o2);
        debug_assert!(y0.len() >= o0 && y1.len() >= o1 && y2.len() >= o2);
        let f = self.func(if mr1 {
            "qmatvec_q4_0_mmvq_fused3_mr1_rp"
        } else if rp {
            "qmatvec_q4_0_mmvq_fused3_rp"
        } else {
            "qmatvec_q4_0_mmvq_fused3"
        });
        let cfg = LaunchConfig {
            grid_dim: (grid, 1, 1),
            block_dim: (32, rpb, 1),
            shared_mem_bytes: 0,
        };
        let inf = w0.in_features() as i32;
        let (oo0, oo1, oo2) = (o0 as i32, o1 as i32, o2 as i32);
        let (r0, r1, r2) = (rb0 as i64, rb1 as i64, rb2 as i64);
        // PDL wave-A: identical to the owned twin (capture-lane parity).
        if mr1 && Self::pdl_on() && Self::pdl_mmvq_on() {
            use cudarc::driver::{DevicePtr, DevicePtrMut};
            let s = &self.gpu.stream();
            let (p0, _g0) = b0.device_ptr(s);
            let (p1, _g1) = b1.device_ptr(s);
            let (p2, _g2) = b2.device_ptr(s);
            let (paq, _g3) = aq.device_ptr(s);
            let (pad, _g4) = ad.device_ptr(s);
            let (py0, _g5) = y0.device_ptr_mut(s);
            let (py1, _g6) = y1.device_ptr_mut(s);
            let (py2, _g7) = y2.device_ptr_mut(s);
            let mut ps = [
                &p0 as *const _ as *mut std::ffi::c_void,
                &p1 as *const _ as *mut _,
                &p2 as *const _ as *mut _,
                &paq as *const _ as *mut _,
                &pad as *const _ as *mut _,
                &py0 as *const _ as *mut _,
                &py1 as *const _ as *mut _,
                &py2 as *const _ as *mut _,
                &inf as *const _ as *mut _,
                &oo0 as *const _ as *mut _,
                &oo1 as *const _ as *mut _,
                &oo2 as *const _ as *mut _,
                &r0 as *const _ as *mut _,
                &r1 as *const _ as *mut _,
                &r2 as *const _ as *mut _,
            ];
            unsafe {
                self.launch_pdl(
                    "qmatvec_q4_0_mmvq_fused3_mr1_rp",
                    (grid, 1, 1),
                    (32, rpb, 1),
                    &mut ps,
                )?;
            }
            return Ok(true);
        }
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(b2)
            .arg(aq)
            .arg(ad)
            .arg(&mut *y0)
            .arg(&mut *y1)
            .arg(&mut *y2)
            .arg(&inf)
            .arg(&oo0)
            .arg(&oo1)
            .arg(&oo2)
            .arg(&r0)
            .arg(&r1)
            .arg(&r2);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(true)
    }

    /// FUSED Q4_0 m=1 PAIR (gemma shared gate+up).
    pub fn matmul_q4_fused2(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
    ) -> Result<Option<(CudaSlice<f32>, CudaSlice<f32>)>, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        let q4 = |w: &GpuTensor| -> Option<(usize, usize)> {
            match w {
                GpuTensor::Quant {
                    qtype, row_bytes, ..
                } if *qtype == QT_Q4_0 => Some((*row_bytes, w.out_features())),
                _ => None,
            }
        };
        let (Some((rb0, o0)), Some((rb1, o1))) = (q4(w0), q4(w1)) else {
            return Ok(None);
        };
        if w0.in_features() != w1.in_features() {
            return Ok(None);
        }
        // Effective (bytes, rp) per tensor (mirror or in-place swap); mixed -> separate matvecs.
        fn eff(w: &GpuTensor) -> (&CudaSlice<u8>, bool) {
            match w {
                GpuTensor::Quant { bytes, rp4, rp, .. } => match rp4 {
                    Some(m) => (m, true),
                    None => (bytes, *rp),
                },
                _ => unreachable!(),
            }
        }
        let ((b0, rp0), (b1, rp1)) = (eff(w0), eff(w1));
        if rp0 != rp1 {
            return Ok(None);
        }
        let rp = rp0;
        let rpb: u32 = 4;
        // mr1 twin — see matmul_q4_fused3.
        let mr1 = rp && Self::q40_mr1_on();
        let nb = |o: usize| {
            if mr1 {
                (o as u32).div_ceil(rpb)
            } else {
                (o as u32).div_ceil(2).div_ceil(rpb)
            }
        };
        let grid = nb(o0) + nb(o1);
        let mut y0 = self.alloc_uninit::<f32>(o0)?;
        let mut y1 = self.alloc_uninit::<f32>(o1)?;
        let f = self.func(if mr1 {
            "qmatvec_q4_0_mmvq_fused2_mr1_rp"
        } else if rp {
            "qmatvec_q4_0_mmvq_fused2_rp"
        } else {
            "qmatvec_q4_0_mmvq_fused2"
        });
        let cfg = LaunchConfig {
            grid_dim: (grid, 1, 1),
            block_dim: (32, rpb, 1),
            shared_mem_bytes: 0,
        };
        let inf = w0.in_features() as i32;
        let (oo0, oo1) = (o0 as i32, o1 as i32);
        let (r0, r1) = (rb0 as i64, rb1 as i64);
        // PDL wave-A: mr1 kernel carries MEMRA_PDL_ENTRY.
        if mr1 && Self::pdl_on() && Self::pdl_mmvq_on() {
            {
                use cudarc::driver::{DevicePtr, DevicePtrMut};
                let s = &self.gpu.stream();
                let (p0, _g0) = b0.device_ptr(s);
                let (p1, _g1) = b1.device_ptr(s);
                let (paq, _g2) = aq.device_ptr(s);
                let (pad, _g3) = ad.device_ptr(s);
                let (py0, _g4) = y0.device_ptr_mut(s);
                let (py1, _g5) = y1.device_ptr_mut(s);
                let mut ps = [
                    &p0 as *const _ as *mut std::ffi::c_void,
                    &p1 as *const _ as *mut _,
                    &paq as *const _ as *mut _,
                    &pad as *const _ as *mut _,
                    &py0 as *const _ as *mut _,
                    &py1 as *const _ as *mut _,
                    &inf as *const _ as *mut _,
                    &oo0 as *const _ as *mut _,
                    &oo1 as *const _ as *mut _,
                    &r0 as *const _ as *mut _,
                    &r1 as *const _ as *mut _,
                ];
                unsafe {
                    self.launch_pdl(
                        "qmatvec_q4_0_mmvq_fused2_mr1_rp",
                        (grid, 1, 1),
                        (32, rpb, 1),
                        &mut ps,
                    )?;
                }
            }
            return Ok(Some((y0, y1)));
        }
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&inf)
            .arg(&oo0)
            .arg(&oo1)
            .arg(&r0)
            .arg(&r1);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(Some((y0, y1)))
    }

    /// Slot-fed fused2 twin (alloc-free capture lane): identical launch, caller-owned outputs.
    pub fn matmul_q4_fused2_into(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        y0: &mut CudaSlice<f32>,
        y1: &mut CudaSlice<f32>,
    ) -> Result<bool, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        let q4 = |w: &GpuTensor| -> Option<(usize, usize)> {
            match w {
                GpuTensor::Quant {
                    qtype, row_bytes, ..
                } if *qtype == QT_Q4_0 => Some((*row_bytes, w.out_features())),
                _ => None,
            }
        };
        let (Some((rb0, o0)), Some((rb1, o1))) = (q4(w0), q4(w1)) else {
            return Ok(false);
        };
        if w0.in_features() != w1.in_features() {
            return Ok(false);
        }
        fn eff(w: &GpuTensor) -> (&CudaSlice<u8>, bool) {
            match w {
                GpuTensor::Quant { bytes, rp4, rp, .. } => match rp4 {
                    Some(m) => (m, true),
                    None => (bytes, *rp),
                },
                _ => unreachable!(),
            }
        }
        let ((b0, rp0), (b1, rp1)) = (eff(w0), eff(w1));
        if rp0 != rp1 {
            return Ok(false);
        }
        let rp = rp0;
        let rpb: u32 = 4;
        let mr1 = rp && Self::q40_mr1_on();
        let nb = |o: usize| {
            if mr1 {
                (o as u32).div_ceil(rpb)
            } else {
                (o as u32).div_ceil(2).div_ceil(rpb)
            }
        };
        let grid = nb(o0) + nb(o1);
        debug_assert!(y0.len() >= o0 && y1.len() >= o1);
        let f = self.func(if mr1 {
            "qmatvec_q4_0_mmvq_fused2_mr1_rp"
        } else if rp {
            "qmatvec_q4_0_mmvq_fused2_rp"
        } else {
            "qmatvec_q4_0_mmvq_fused2"
        });
        let cfg = LaunchConfig {
            grid_dim: (grid, 1, 1),
            block_dim: (32, rpb, 1),
            shared_mem_bytes: 0,
        };
        let inf = w0.in_features() as i32;
        let (oo0, oo1) = (o0 as i32, o1 as i32);
        let (r0, r1) = (rb0 as i64, rb1 as i64);
        // PDL wave-A: identical to the owned twin (capture-lane parity).
        if mr1 && Self::pdl_on() && Self::pdl_mmvq_on() {
            use cudarc::driver::{DevicePtr, DevicePtrMut};
            let s = &self.gpu.stream();
            let (p0, _g0) = b0.device_ptr(s);
            let (p1, _g1) = b1.device_ptr(s);
            let (paq, _g2) = aq.device_ptr(s);
            let (pad, _g3) = ad.device_ptr(s);
            let (py0, _g4) = y0.device_ptr_mut(s);
            let (py1, _g5) = y1.device_ptr_mut(s);
            let mut ps = [
                &p0 as *const _ as *mut std::ffi::c_void,
                &p1 as *const _ as *mut _,
                &paq as *const _ as *mut _,
                &pad as *const _ as *mut _,
                &py0 as *const _ as *mut _,
                &py1 as *const _ as *mut _,
                &inf as *const _ as *mut _,
                &oo0 as *const _ as *mut _,
                &oo1 as *const _ as *mut _,
                &r0 as *const _ as *mut _,
                &r1 as *const _ as *mut _,
            ];
            unsafe {
                self.launch_pdl(
                    "qmatvec_q4_0_mmvq_fused2_mr1_rp",
                    (grid, 1, 1),
                    (32, rpb, 1),
                    &mut ps,
                )?;
            }
            return Ok(true);
        }
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(aq)
            .arg(ad)
            .arg(&mut *y0)
            .arg(&mut *y1)
            .arg(&inf)
            .arg(&oo0)
            .arg(&oo1)
            .arg(&r0)
            .arg(&r1);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(true)
    }

    /// BATCHED fused2 (2026-07-13, megakernel-microcosm probe): gate+up b-tier matvecs in
    /// ONE segmented-grid launch — the up segment fills SMs as the gate segment drains
    /// (the per-launch tail waves behind the 6x-falsified b-tier plateau). Bit-identical
    /// per row to two mr2_rp launches. rp layout required; m in 2..=8 (b16 has no twin).
    pub fn matmul_q4_fused2_batched(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
    ) -> Result<Option<(CudaSlice<f32>, CudaSlice<f32>)>, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        if m < 2 || m > 8 {
            return Ok(None);
        }
        let q4 = |w: &GpuTensor| -> Option<(usize, usize)> {
            match w {
                GpuTensor::Quant {
                    qtype, row_bytes, ..
                } if *qtype == QT_Q4_0 => Some((*row_bytes, w.out_features())),
                _ => None,
            }
        };
        let (Some((rb0, o0)), Some((_rb1, o1))) = (q4(w0), q4(w1)) else {
            return Ok(None);
        };
        if w0.in_features() != w1.in_features() {
            return Ok(None);
        }
        fn eff(w: &GpuTensor) -> (&CudaSlice<u8>, bool) {
            match w {
                GpuTensor::Quant { bytes, rp4, rp, .. } => match rp4 {
                    Some(mr) => (mr, true),
                    None => (bytes, *rp),
                },
                _ => unreachable!(),
            }
        }
        let ((b0, rp0), (b1, rp1)) = (eff(w0), eff(w1));
        if !rp0 || !rp1 {
            return Ok(None);
        }
        let mcols = Self::batched_mcols(m);
        let rpb: u32 = 4;
        let nb = |o: usize| (o as u32).div_ceil(2 * rpb);
        let grid = nb(o0) + nb(o1);
        let mut y0 = self.alloc_uninit::<f32>(m * o0)?;
        let mut y1 = self.alloc_uninit::<f32>(m * o1)?;
        let f = self.func(match mcols {
            2 => "qmatvec_q4_0_mmvq_b2_f2_rp",
            4 => "qmatvec_q4_0_mmvq_b4_f2_rp",
            _ => "qmatvec_q4_0_mmvq_b8_f2_rp",
        });
        let cfg = LaunchConfig {
            grid_dim: (grid, 1, 1),
            block_dim: (32, rpb, 1),
            shared_mem_bytes: 0,
        };
        let inf = w0.in_features() as i32;
        let (oo0, oo1, mi) = (o0 as i32, o1 as i32, m as i32);
        let rb = rb0 as i64;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&inf)
            .arg(&oo0)
            .arg(&oo1)
            .arg(&mi)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(Some((y0, y1)))
    }

    /// BATCHED fused3 (see matmul_q4_fused2_batched): three-segment single launch for the
    /// verify qkv triple. Same-in_f q4_0 rp tensors, m in 2..=8. Bit-identical per row.
    #[allow(clippy::too_many_arguments)]
    pub fn matmul_q4_fused3_batched(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        w2: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
    ) -> Result<Option<(CudaSlice<f32>, CudaSlice<f32>, CudaSlice<f32>)>, Box<dyn std::error::Error>>
    {
        use crate::model::GpuTensor;
        if m < 2 || m > 8 {
            return Ok(None);
        }
        let q4 = |w: &GpuTensor| -> Option<usize> {
            match w {
                GpuTensor::Quant { qtype, .. } if *qtype == QT_Q4_0 => Some(w.out_features()),
                _ => None,
            }
        };
        let (Some(o0), Some(o1), Some(o2)) = (q4(w0), q4(w1), q4(w2)) else {
            return Ok(None);
        };
        if w0.in_features() != w1.in_features() || w0.in_features() != w2.in_features() {
            return Ok(None);
        }
        fn eff(w: &GpuTensor) -> (&CudaSlice<u8>, bool) {
            match w {
                GpuTensor::Quant { bytes, rp4, rp, .. } => match rp4 {
                    Some(mr) => (mr, true),
                    None => (bytes, *rp),
                },
                _ => unreachable!(),
            }
        }
        let ((b0, rp0), (b1, rp1), (b2, rp2)) = (eff(w0), eff(w1), eff(w2));
        if !rp0 || !rp1 || !rp2 {
            return Ok(None);
        }
        let mcols = Self::batched_mcols(m);
        let rpb: u32 = 4;
        let nb = |o: usize| (o as u32).div_ceil(2 * rpb);
        let grid = nb(o0) + nb(o1) + nb(o2);
        let mut y0 = self.alloc_uninit::<f32>(m * o0)?;
        let mut y1 = self.alloc_uninit::<f32>(m * o1)?;
        let mut y2 = self.alloc_uninit::<f32>(m * o2)?;
        let f = self.func(match mcols {
            2 => "qmatvec_q4_0_mmvq_b2_f3_rp",
            4 => "qmatvec_q4_0_mmvq_b4_f3_rp",
            _ => "qmatvec_q4_0_mmvq_b8_f3_rp",
        });
        let cfg = LaunchConfig {
            grid_dim: (grid, 1, 1),
            block_dim: (32, rpb, 1),
            shared_mem_bytes: 0,
        };
        let inf = w0.in_features() as i32;
        let (oo0, oo1, oo2, mi) = (o0 as i32, o1 as i32, o2 as i32, m as i32);
        let rb = 0i64;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(b2)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&mut y2)
            .arg(&inf)
            .arg(&oo0)
            .arg(&oo1)
            .arg(&oo2)
            .arg(&mi)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(Some((y0, y1, y2)))
    }

    pub fn matmul_q8_fused3(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        w2: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
    ) -> Result<Option<(CudaSlice<f32>, CudaSlice<f32>, CudaSlice<f32>)>, Box<dyn std::error::Error>>
    {
        // e4m3 twin (lane/fp8-decode-v1): the full-attn wq/wk/wv triple — on the NV-27B those three
        // are per-tensor FP8, so native residency without this arm meant three separate launches.
        if let Some([p0, p1, p2]) = self.e4m3_fused_params(&[w0, w1, w2]) {
            return Ok(Some(self.e4m3_fused3_core(
                p0.0,
                p1.0,
                p2.0,
                aq,
                ad,
                w0.in_features(),
                p0.1,
                p1.1,
                p2.1,
                p0.2,
                p0.3,
                p1.3,
                p2.3,
            )?));
        }
        let Some([p0, p1, p2]) = self.q8_fused_params(&[w0, w1, w2]) else {
            return Ok(None);
        };
        Ok(Some(self.q8_fused3_core(
            p0.0,
            p1.0,
            p2.0,
            aq,
            ad,
            w0.in_features(),
            p0.1,
            p1.1,
            p2.1,
            p0.2,
        )?))
    }

    #[allow(clippy::too_many_arguments)]
    fn q8_fused3_core(
        &self,
        b0: &CudaSlice<u8>,
        b1: &CudaSlice<u8>,
        b2: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        in_f: usize,
        out0: usize,
        out1: usize,
        out2: usize,
        row_bytes: usize,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        const ROWS_PER_BLOCK: u32 = 4;
        let nb0 = (out0 as u32).div_ceil(ROWS_PER_BLOCK);
        let nb1 = (out1 as u32).div_ceil(ROWS_PER_BLOCK);
        let nb2 = (out2 as u32).div_ceil(ROWS_PER_BLOCK);
        let f = self.func("qmatvec_q8_0_mmvq_fused3");
        let mut y0 = self.alloc_uninit::<f32>(out0)?;
        let mut y1 = self.alloc_uninit::<f32>(out1)?;
        let mut y2 = self.alloc_uninit::<f32>(out2)?;
        let cfg = LaunchConfig {
            grid_dim: (nb0 + nb1 + nb2, 1, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (inf, o0, o1, o2, rbl) = (
            in_f as i32,
            out0 as i32,
            out1 as i32,
            out2 as i32,
            row_bytes as i64,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(b2)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&mut y2)
            .arg(&inf)
            .arg(&o0)
            .arg(&o1)
            .arg(&o2)
            .arg(&rbl);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((y0, y1, y2))
    }

    /// Test entry for the kernel_check gate: fused3 from raw weight bytes (internal q8_1 quant).
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_q8_fused3_raw(
        &self,
        b0: &CudaSlice<u8>,
        b1: &CudaSlice<u8>,
        b2: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        in_f: usize,
        out0: usize,
        out1: usize,
        out2: usize,
        row_bytes: usize,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        let (aq, ad) = self.quantize_q8_1(x, 1, in_f)?;
        self.q8_fused3_core(b0, b1, b2, &aq, &ad, in_f, out0, out1, out2, row_bytes)
    }

    /// BATCHED twin of `matmul_q8_fused2` for the verify t=2-4 tier (MEMRA_SPEC_FUSED_T call
    /// sites, lane/close35b): ONE launch computes both same-input Q8_0 projections for m tokens.
    /// Per (tensor,token,row) the kernel body is q8_0_mmvq_batched VERBATIM with the identical
    /// row mapping (Q8_0's batched_variant is always "base") -> BIT-IDENTICAL to the two
    /// per-tensor _b2/_b4 launches `matmul_decode_exact` dispatches at m=2-4, with the caller's
    /// single shared q8_1 activation replacing two per-call re-quantizes (quantize_q8_1 is
    /// deterministic -> same bytes). None when ineligible (m outside 2..=4 / not both Q8_0 /
    /// in_f mismatch / MEMRA_MMVQ=0 / MEMRA_Q8_DUAL=0 / MEMRA_NO_BATCHED set — the last keeps
    /// dispatch parity: without batched kernels decode-exact runs grid.y=m MMVQ, and the fused
    /// twin must not introduce a batched program the reference path would not run).
    pub fn matmul_q8_fused2_t(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
    ) -> Result<Option<(CudaSlice<f32>, CudaSlice<f32>)>, Box<dyn std::error::Error>> {
        // m<=8 (lane/q27-deepdive, 2026-08-05): was 2..=4 (the verify tier's mcols 2/4). The
        // serving tick's mcols-8 tier now has its fused2_b8 wrapper, so c=5..8 batched decode
        // fuses too — same template body, still bit-identical to the two _b8 launches.
        if !(2..=8).contains(&m) || std::env::var("MEMRA_NO_BATCHED").is_ok() {
            return Ok(None);
        }
        // e4m3 twin: MEMRA_B8 parity — without it m=5..8 e4m3 decode runs the per-m grid.y=m path,
        // so the fused b8 launch would introduce a batched program the reference path would not run.
        if let Some([p0, p1]) = self.e4m3_fused_params(&[w0, w1]) {
            if m > 4 && !Self::b8_enabled() {
                return Ok(None);
            }
            return Ok(Some(self.e4m3_fused2_t_core(
                p0.0,
                p1.0,
                aq,
                ad,
                m,
                w0.in_features(),
                p0.1,
                p1.1,
                p0.2,
                p0.3,
                p1.3,
            )?));
        }
        let Some([p0, p1]) = self.q8_fused_params(&[w0, w1]) else {
            return Ok(None);
        };
        Ok(Some(self.q8_fused2_t_core(
            p0.0,
            p1.0,
            aq,
            ad,
            m,
            w0.in_features(),
            p0.1,
            p1.1,
            p0.2,
        )?))
    }

    #[allow(clippy::too_many_arguments)]
    fn q8_fused2_t_core(
        &self,
        b0: &CudaSlice<u8>,
        b1: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out0: usize,
        out1: usize,
        row_bytes: usize,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        const ROWS_PER_BLOCK: u32 = 4; // matches MEMRA_MMVQ_ROWS in qmatvec.cu
        let nb0 = (out0 as u32).div_ceil(ROWS_PER_BLOCK);
        let nb1 = (out1 as u32).div_ceil(ROWS_PER_BLOCK);
        let f = self.func(match Self::batched_mcols(m) {
            2 => "qmatvec_q8_0_mmvq_fused2_b2",
            4 => "qmatvec_q8_0_mmvq_fused2_b4",
            // b8 = the SERVING tier (lane/q27-deepdive): c=5..8 batched decode.
            _ => "qmatvec_q8_0_mmvq_fused2_b8",
        });
        let mut y0 = self.alloc_uninit::<f32>(m * out0)?;
        let mut y1 = self.alloc_uninit::<f32>(m * out1)?;
        let cfg = LaunchConfig {
            grid_dim: (nb0 + nb1, 1, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (inf, o0, o1, mi, rbl) = (
            in_f as i32,
            out0 as i32,
            out1 as i32,
            m as i32,
            row_bytes as i64,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&inf)
            .arg(&o0)
            .arg(&o1)
            .arg(&mi)
            .arg(&rbl);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((y0, y1))
    }

    /// Test entry for the kernel_check gate: fused2 batched from raw weight bytes (internal
    /// q8_1 quant of the [m, in_f] activation), no env gating.
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_q8_fused2_t_raw(
        &self,
        b0: &CudaSlice<u8>,
        b1: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out0: usize,
        out1: usize,
        row_bytes: usize,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
        self.q8_fused2_t_core(b0, b1, &aq, &ad, m, in_f, out0, out1, row_bytes)
    }

    /// BATCHED twin of `matmul_q8_fused3` (wq+wk+wv at verify t=2-4). Same contract as
    /// `matmul_q8_fused2_t` with three ranges.
    #[allow(clippy::too_many_arguments)]
    pub fn matmul_q8_fused3_t(
        &self,
        w0: &crate::model::GpuTensor,
        w1: &crate::model::GpuTensor,
        w2: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
    ) -> Result<Option<(CudaSlice<f32>, CudaSlice<f32>, CudaSlice<f32>)>, Box<dyn std::error::Error>>
    {
        if !(2..=4).contains(&m) || std::env::var("MEMRA_NO_BATCHED").is_ok() {
            return Ok(None);
        }
        if let Some([p0, p1, p2]) = self.e4m3_fused_params(&[w0, w1, w2]) {
            return Ok(Some(self.e4m3_fused3_t_core(
                p0.0,
                p1.0,
                p2.0,
                aq,
                ad,
                m,
                w0.in_features(),
                p0.1,
                p1.1,
                p2.1,
                p0.2,
                p0.3,
                p1.3,
                p2.3,
            )?));
        }
        let Some([p0, p1, p2]) = self.q8_fused_params(&[w0, w1, w2]) else {
            return Ok(None);
        };
        Ok(Some(self.q8_fused3_t_core(
            p0.0,
            p1.0,
            p2.0,
            aq,
            ad,
            m,
            w0.in_features(),
            p0.1,
            p1.1,
            p2.1,
            p0.2,
        )?))
    }

    #[allow(clippy::too_many_arguments)]
    fn q8_fused3_t_core(
        &self,
        b0: &CudaSlice<u8>,
        b1: &CudaSlice<u8>,
        b2: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out0: usize,
        out1: usize,
        out2: usize,
        row_bytes: usize,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        const ROWS_PER_BLOCK: u32 = 4;
        let nb0 = (out0 as u32).div_ceil(ROWS_PER_BLOCK);
        let nb1 = (out1 as u32).div_ceil(ROWS_PER_BLOCK);
        let nb2 = (out2 as u32).div_ceil(ROWS_PER_BLOCK);
        let f = self.func(if Self::batched_mcols(m) == 2 {
            "qmatvec_q8_0_mmvq_fused3_b2"
        } else {
            "qmatvec_q8_0_mmvq_fused3_b4"
        });
        let mut y0 = self.alloc_uninit::<f32>(m * out0)?;
        let mut y1 = self.alloc_uninit::<f32>(m * out1)?;
        let mut y2 = self.alloc_uninit::<f32>(m * out2)?;
        let cfg = LaunchConfig {
            grid_dim: (nb0 + nb1 + nb2, 1, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (inf, o0, o1, o2, mi, rbl) = (
            in_f as i32,
            out0 as i32,
            out1 as i32,
            out2 as i32,
            m as i32,
            row_bytes as i64,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(b2)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&mut y2)
            .arg(&inf)
            .arg(&o0)
            .arg(&o1)
            .arg(&o2)
            .arg(&mi)
            .arg(&rbl);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((y0, y1, y2))
    }

    /// Test entry for the kernel_check gate: fused3 batched from raw weight bytes.
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_q8_fused3_t_raw(
        &self,
        b0: &CudaSlice<u8>,
        b1: &CudaSlice<u8>,
        b2: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out0: usize,
        out1: usize,
        out2: usize,
        row_bytes: usize,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
        self.q8_fused3_t_core(b0, b1, b2, &aq, &ad, m, in_f, out0, out1, out2, row_bytes)
    }

    /// Rollback seam for the Q8_0 dense-FFN gate+up fusion arm in `matmul_pre_dual_noscale`
    /// (lane/q27-deepdive, 2026-08-05). Default ON; `MEMRA_Q8_FFN_FUSE2=0` restores the
    /// two-`matmul_pre_noscale` pair. Read once — the dispatch must not vary within a run.
    pub fn q8_ffn_fuse2_on(&self) -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| std::env::var("MEMRA_Q8_FFN_FUSE2").as_deref() != Ok("0"))
    }

    /// Eligibility + param extraction for the fused q8_0 launches: every tensor must be Quant Q8_0
    /// with macro-scale 1.0 (always true for GGUF q8_0; only NVFP4 carries scale) and share w[0]'s
    /// in_f (q8_0 row_bytes is a pure function of in_f, so equal in_f => equal row_bytes). MEMRA_MMVQ
    /// must be on: the fused body is the MMVQ kernel; without it decode m=1 runs dp4a and fusing
    /// would mix dispatch families (FP-order law). MEMRA_Q8_DUAL=0 = rollback seam.
    #[allow(clippy::type_complexity)]
    fn q8_fused_params<'w, const N: usize>(
        &self,
        ws: &[&'w crate::model::GpuTensor; N],
    ) -> Option<[(&'w CudaSlice<u8>, usize, usize); N]> {
        use crate::model::GpuTensor;
        if std::env::var("MEMRA_MMVQ").as_deref() == Ok("0") {
            return None;
        }
        if std::env::var("MEMRA_Q8_DUAL").is_ok_and(|v| v == "0") {
            return None;
        }
        let in_f = ws[0].in_features();
        let mut out: [Option<(&CudaSlice<u8>, usize, usize)>; N] = [None; N];
        for (i, w) in ws.iter().enumerate() {
            match w {
                GpuTensor::Quant {
                    bytes,
                    qtype,
                    row_bytes,
                    scale,
                    ..
                } if *qtype == QT_Q8_0 && *scale == 1.0 && w.in_features() == in_f => {
                    out[i] = Some((bytes, w.out_features(), *row_bytes))
                }
                _ => return None,
            }
        }
        Some(out.map(|o| o.unwrap()))
    }

    /// Rollback seam for the F8-E4M3 launch-fusion arm (lane/fp8-decode-v1, 2026-08-05).
    /// Default ON; `MEMRA_E4M3_DUAL=0` restores the per-tensor m=1/batched launches.
    pub fn e4m3_dual_on(&self) -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| std::env::var("MEMRA_E4M3_DUAL").as_deref() != Ok("0"))
    }

    /// Eligibility + param extraction for the FUSED e4m3 launches — the QT_F8_E4M3 twin of
    /// `q8_fused_params`. Differences that are inherent to the dtype, not policy:
    ///   * each tensor carries its OWN per-tensor `weight_scale` (returned as the 4th field);
    ///     Q8_0 hard-requires scale==1.0 because it has no macro-scale at all.
    ///   * no MEMRA_MMVQ gate: `mmvq_supports` exempts QT_F8_E4M3 (the e4m3 mmvq family is that
    ///     dtype's ONLY int8-act kernel class), so the per-tensor fallback these fused kernels
    ///     replace is ALWAYS the same mmvq body under every env — the FP-order law holds.
    ///   * `row_bytes == in_f` is asserted rather than derived: the native-residency load arm keeps
    ///     the checkpoint's raw [out_f, in_f] rows, and a re-encoded slab must never reach here.
    /// Rejects any split-plane mirror (`rp`/`rp4`): there is no `_rp` e4m3 fused form, so fusing
    /// there would swap dispatch families mid-model. MEMRA_E4M3_DUAL=0 = rollback seam.
    #[allow(clippy::type_complexity)]
    fn e4m3_fused_params<'w, const N: usize>(
        &self,
        ws: &[&'w crate::model::GpuTensor; N],
    ) -> Option<[(&'w CudaSlice<u8>, usize, usize, f32); N]> {
        use crate::model::GpuTensor;
        if !self.e4m3_dual_on() {
            return None;
        }
        let in_f = ws[0].in_features();
        let mut out: [Option<(&CudaSlice<u8>, usize, usize, f32)>; N] = [None; N];
        for (i, w) in ws.iter().enumerate() {
            match w {
                GpuTensor::Quant {
                    bytes,
                    qtype,
                    row_bytes,
                    scale,
                    rp,
                    rp4,
                    ..
                } if *qtype == QT_F8_E4M3
                    && w.in_features() == in_f
                    && *row_bytes == in_f
                    && !*rp
                    && rp4.is_none() =>
                {
                    out[i] = Some((bytes, w.out_features(), *row_bytes, *scale))
                }
                _ => return None,
            }
        }
        Some(out.map(|o| o.unwrap()))
    }

    /// FUSED e4m3 m=1 PAIR. Block-offset split (`qmatvec_e4m3_mmvq_fused2`), per-tensor
    /// weight_scale folded at the write like the single-tensor `qmatvec_e4m3_mmvq` — so per
    /// (tensor,row) this is BIT-IDENTICAL to two separate m=1 launches, scale included.
    #[allow(clippy::too_many_arguments)]
    fn e4m3_fused2_core(
        &self,
        b0: &CudaSlice<u8>,
        b1: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        in_f: usize,
        out0: usize,
        out1: usize,
        row_bytes: usize,
        ws0: f32,
        ws1: f32,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        const ROWS_PER_BLOCK: u32 = 4; // matches MEMRA_MMVQ_ROWS in qmatvec.cu
        let nb0 = (out0 as u32).div_ceil(ROWS_PER_BLOCK);
        let nb1 = (out1 as u32).div_ceil(ROWS_PER_BLOCK);
        let f = self.func("qmatvec_e4m3_mmvq_fused2");
        let mut y0 = self.alloc_uninit::<f32>(out0)?;
        let mut y1 = self.alloc_uninit::<f32>(out1)?;
        let cfg = LaunchConfig {
            grid_dim: (nb0 + nb1, 1, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (inf, o0, o1, rbl) = (in_f as i32, out0 as i32, out1 as i32, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&inf)
            .arg(&o0)
            .arg(&o1)
            .arg(&rbl)
            .arg(&ws0)
            .arg(&ws1);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((y0, y1))
    }

    /// FUSED e4m3 m=1 TRIPLE (`qmatvec_e4m3_mmvq_fused3`). Same contract as the pair.
    #[allow(clippy::too_many_arguments)]
    fn e4m3_fused3_core(
        &self,
        b0: &CudaSlice<u8>,
        b1: &CudaSlice<u8>,
        b2: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        in_f: usize,
        out0: usize,
        out1: usize,
        out2: usize,
        row_bytes: usize,
        ws0: f32,
        ws1: f32,
        ws2: f32,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        const ROWS_PER_BLOCK: u32 = 4;
        let nb0 = (out0 as u32).div_ceil(ROWS_PER_BLOCK);
        let nb1 = (out1 as u32).div_ceil(ROWS_PER_BLOCK);
        let nb2 = (out2 as u32).div_ceil(ROWS_PER_BLOCK);
        let f = self.func("qmatvec_e4m3_mmvq_fused3");
        let mut y0 = self.alloc_uninit::<f32>(out0)?;
        let mut y1 = self.alloc_uninit::<f32>(out1)?;
        let mut y2 = self.alloc_uninit::<f32>(out2)?;
        let cfg = LaunchConfig {
            grid_dim: (nb0 + nb1 + nb2, 1, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (inf, o0, o1, o2, rbl) = (
            in_f as i32,
            out0 as i32,
            out1 as i32,
            out2 as i32,
            row_bytes as i64,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(b2)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&mut y2)
            .arg(&inf)
            .arg(&o0)
            .arg(&o1)
            .arg(&o2)
            .arg(&rbl)
            .arg(&ws0)
            .arg(&ws1)
            .arg(&ws2);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((y0, y1, y2))
    }

    /// BATCHED FUSED e4m3 pair (m=2..8). The batched kernels carry no `ws` arg (every batched
    /// kernel in the tree is scale-free), so each output takes its own `scale_inplace` — the
    /// SAME post-op the per-tensor batched dispatch applies, hence still bit-identical.
    #[allow(clippy::too_many_arguments)]
    fn e4m3_fused2_t_core(
        &self,
        b0: &CudaSlice<u8>,
        b1: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out0: usize,
        out1: usize,
        row_bytes: usize,
        ws0: f32,
        ws1: f32,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        const ROWS_PER_BLOCK: u32 = 4;
        let nb0 = (out0 as u32).div_ceil(ROWS_PER_BLOCK);
        let nb1 = (out1 as u32).div_ceil(ROWS_PER_BLOCK);
        let f = self.func(match Self::batched_mcols(m) {
            2 => "qmatvec_e4m3_mmvq_fused2_b2",
            4 => "qmatvec_e4m3_mmvq_fused2_b4",
            _ => "qmatvec_e4m3_mmvq_fused2_b8",
        });
        let mut y0 = self.alloc_uninit::<f32>(m * out0)?;
        let mut y1 = self.alloc_uninit::<f32>(m * out1)?;
        let cfg = LaunchConfig {
            grid_dim: (nb0 + nb1, 1, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (inf, o0, o1, mi, rbl) = (
            in_f as i32,
            out0 as i32,
            out1 as i32,
            m as i32,
            row_bytes as i64,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&inf)
            .arg(&o0)
            .arg(&o1)
            .arg(&mi)
            .arg(&rbl);
        unsafe {
            b.launch(cfg)?;
        }
        if ws0 != 1.0 {
            self.scale_inplace(&mut y0, ws0, m * out0)?;
        }
        if ws1 != 1.0 {
            self.scale_inplace(&mut y1, ws1, m * out1)?;
        }
        Ok((y0, y1))
    }

    /// BATCHED FUSED e4m3 triple (m=2..4). Same contract as the batched pair.
    #[allow(clippy::too_many_arguments)]
    fn e4m3_fused3_t_core(
        &self,
        b0: &CudaSlice<u8>,
        b1: &CudaSlice<u8>,
        b2: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out0: usize,
        out1: usize,
        out2: usize,
        row_bytes: usize,
        ws0: f32,
        ws1: f32,
        ws2: f32,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        const ROWS_PER_BLOCK: u32 = 4;
        let nb0 = (out0 as u32).div_ceil(ROWS_PER_BLOCK);
        let nb1 = (out1 as u32).div_ceil(ROWS_PER_BLOCK);
        let nb2 = (out2 as u32).div_ceil(ROWS_PER_BLOCK);
        let f = self.func(if Self::batched_mcols(m) == 2 {
            "qmatvec_e4m3_mmvq_fused3_b2"
        } else {
            "qmatvec_e4m3_mmvq_fused3_b4"
        });
        let mut y0 = self.alloc_uninit::<f32>(m * out0)?;
        let mut y1 = self.alloc_uninit::<f32>(m * out1)?;
        let mut y2 = self.alloc_uninit::<f32>(m * out2)?;
        let cfg = LaunchConfig {
            grid_dim: (nb0 + nb1 + nb2, 1, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (inf, o0, o1, o2, mi, rbl) = (
            in_f as i32,
            out0 as i32,
            out1 as i32,
            out2 as i32,
            m as i32,
            row_bytes as i64,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(b0)
            .arg(b1)
            .arg(b2)
            .arg(aq)
            .arg(ad)
            .arg(&mut y0)
            .arg(&mut y1)
            .arg(&mut y2)
            .arg(&inf)
            .arg(&o0)
            .arg(&o1)
            .arg(&o2)
            .arg(&mi)
            .arg(&rbl);
        unsafe {
            b.launch(cfg)?;
        }
        if ws0 != 1.0 {
            self.scale_inplace(&mut y0, ws0, m * out0)?;
        }
        if ws1 != 1.0 {
            self.scale_inplace(&mut y1, ws1, m * out1)?;
        }
        if ws2 != 1.0 {
            self.scale_inplace(&mut y2, ws2, m * out2)?;
        }
        Ok((y0, y1, y2))
    }

    /// BLOCK-128 e4m3 MMVQ launcher (`qmatvec_e4m3_blk_mmvq`, lane/fp8-blk128-decode 2026-08-05).
    /// The per-block-dequant twin of `qmatvec_mmvq`'s QT_F8_E4M3 arm: same grid/block decomposition
    /// (warp per output row, ROWS_PER_BLOCK warps per block, grid.y = m), same q8_1 activation, but
    /// the weight scale is a resident [rows, cols] f32 grid read per k128 block inside the kernel
    /// instead of one scalar folded at the write. It cannot share `qmatvec_mmvq`'s body because
    /// that launcher's arg list is fixed at (bytes, aq, ad, y, in_f, out_f, m, row_bytes [, scale]).
    ///
    /// `mr` and `rp` have no analogue here (no split-plane e4m3 layout exists), so there is exactly
    /// one kernel and no name table — a shape this cannot serve must be refused at LOAD, not here.
    pub fn qmatvec_e4m3_blk_mmvq(
        &self,
        bytes: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        scales: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
        scale_cols: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let mut y = self.alloc_uninit::<f32>(m * out_f)?; // full-overwrite output: skip memset
        self.qmatvec_e4m3_blk_mmvq_into(
            bytes, aq, ad, scales, m, in_f, out_f, row_bytes, scale_cols, &mut y,
        )?;
        Ok(y)
    }

    /// Slot-fed twin of `qmatvec_e4m3_blk_mmvq` (caller-owned output; the alloc-free capture lane).
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_e4m3_blk_mmvq_into(
        &self,
        bytes: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        scales: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
        scale_cols: usize,
        y: &mut CudaSlice<f32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        const ROWS_PER_BLOCK: u32 = 4; // matches MEMRA_MMVQ_ROWS in qmatvec.cu
        let f = self.func("qmatvec_e4m3_blk_mmvq");
        let cfg = LaunchConfig {
            grid_dim: ((out_f as u32).div_ceil(ROWS_PER_BLOCK), m as u32, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1), // warp-per-row
            shared_mem_bytes: 0,                // warp-only reduce
        };
        let (inf, outf, mi, rb, sc) = (
            in_f as i32,
            out_f as i32,
            m as i32,
            row_bytes as i64,
            scale_cols as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(bytes)
            .arg(aq)
            .arg(ad)
            .arg(scales)
            .arg(&mut *y)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&rb)
            .arg(&sc);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// BLOCK-128 e4m3 BATCHED matvec (lane/rp-on-st, 2026-08-06): the weight-read-once twin of
    /// `qmatvec_e4m3_blk_mmvq` for m=2..16. Per (token,row) BIT-IDENTICAL to the grid.y=m launch
    /// (same fmaf chain, same per-k32 `s * ad` fold, same warp reduce), so it inherits the
    /// decode-exactness contract while reading the weight ONCE for up to `mcols` columns instead
    /// of `m` times. `mcols` must be one of {2,4,8,16} and satisfy `mcols >= m`.
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_e4m3_blk_mmvq_batched(
        &self,
        bytes: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        scales: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
        scale_cols: usize,
        mcols: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        const ROWS_PER_BLOCK: u32 = 4; // matches MEMRA_MMVQ_ROWS in qmatvec.cu
        debug_assert!(mcols >= m, "blk batched: mcols {mcols} < m {m}");
        let name = match mcols {
            2 => "qmatvec_e4m3_blk_mmvq_b2",
            4 => "qmatvec_e4m3_blk_mmvq_b4",
            8 => "qmatvec_e4m3_blk_mmvq_b8",
            16 => "qmatvec_e4m3_blk_mmvq_b16",
            _ => {
                return Err(
                    format!("qmatvec_e4m3_blk_mmvq_batched: no kernel for mcols {mcols}").into(),
                );
            }
        };
        let mut y = self.alloc_uninit::<f32>(m * out_f)?;
        let f = self.func(name);
        let cfg = LaunchConfig {
            grid_dim: ((out_f as u32).div_ceil(ROWS_PER_BLOCK), 1, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi, rb, sc) = (
            in_f as i32,
            out_f as i32,
            m as i32,
            row_bytes as i64,
            scale_cols as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(bytes)
            .arg(aq)
            .arg(ad)
            .arg(scales)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&rb)
            .arg(&sc);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(y)
    }

    /// Test entry for the kernel_check exactness gate: the block-128 e4m3 batched MMVQ from raw
    /// bytes with an internal q8_1 quantize (mirrors `qmatvec_batched_raw`).
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_e4m3_blk_batched_raw(
        &self,
        bytes: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        scales: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
        scale_cols: usize,
        mcols: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
        self.qmatvec_e4m3_blk_mmvq_batched(
            bytes, &aq, &ad, scales, m, in_f, out_f, row_bytes, scale_cols, mcols,
        )
    }

    /// Test entry for the kernel_check exactness gate: the block-128 e4m3 MMVQ from raw bytes with
    /// an internal q8_1 quantize (mirrors `qmatvec_mmvq_raw`).
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_e4m3_blk_mmvq_raw(
        &self,
        bytes: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        scales: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
        scale_cols: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
        self.qmatvec_e4m3_blk_mmvq(
            bytes, &aq, &ad, scales, m, in_f, out_f, row_bytes, scale_cols,
        )
    }

    /// Test entries for the kernel_check bit-parity gate: fused e4m3 launches from raw weight
    /// bytes with internal q8_1 quantize, no env gating (mirrors `qmatvec_q8_fused*_raw`).
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_e4m3_fused2_raw(
        &self,
        b0: &CudaSlice<u8>,
        b1: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        in_f: usize,
        out0: usize,
        out1: usize,
        row_bytes: usize,
        ws0: f32,
        ws1: f32,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        let (aq, ad) = self.quantize_q8_1(x, 1, in_f)?;
        self.e4m3_fused2_core(b0, b1, &aq, &ad, in_f, out0, out1, row_bytes, ws0, ws1)
    }

    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_e4m3_fused3_raw(
        &self,
        b0: &CudaSlice<u8>,
        b1: &CudaSlice<u8>,
        b2: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        in_f: usize,
        out0: usize,
        out1: usize,
        out2: usize,
        row_bytes: usize,
        ws0: f32,
        ws1: f32,
        ws2: f32,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        let (aq, ad) = self.quantize_q8_1(x, 1, in_f)?;
        self.e4m3_fused3_core(
            b0, b1, b2, &aq, &ad, in_f, out0, out1, out2, row_bytes, ws0, ws1, ws2,
        )
    }

    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_e4m3_fused2_t_raw(
        &self,
        b0: &CudaSlice<u8>,
        b1: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out0: usize,
        out1: usize,
        row_bytes: usize,
        ws0: f32,
        ws1: f32,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
        self.e4m3_fused2_t_core(b0, b1, &aq, &ad, m, in_f, out0, out1, row_bytes, ws0, ws1)
    }

    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_e4m3_fused3_t_raw(
        &self,
        b0: &CudaSlice<u8>,
        b1: &CudaSlice<u8>,
        b2: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out0: usize,
        out1: usize,
        out2: usize,
        row_bytes: usize,
        ws0: f32,
        ws1: f32,
        ws2: f32,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
        self.e4m3_fused3_t_core(
            b0, b1, b2, &aq, &ad, m, in_f, out0, out1, out2, row_bytes, ws0, ws1, ws2,
        )
    }

    /// THE single dispatch point for `QT_F8_E4M3_BLK` from a PRE-QUANTIZED q8_1 activation
    /// (lane/fp8-blk128-decode). Every `matmul_pre`-family entry calls this first, so the block-128
    /// class has exactly ONE code path across `matmul`, `matmul_pre`, `matmul_pre_noscale`,
    /// `matmul_decode_exact` and `matmul_decode_exact_pre` — the same kernel at the same grid for
    /// every m, which is what makes verify == decode bit-for-bit at every tier for free.
    ///
    /// Returns None for any other qtype (the caller continues its normal dispatch). The `blk: Some`
    /// pattern is part of the match, not an unwrap: qtype and grid presence are set together in the
    /// one residency arm that builds this tensor, and a qtype-without-grid would be a construction
    /// bug — better to fall through and hit a loud refusal than to unwrap a None here.
    fn try_e4m3_blk_pre(
        &self,
        w: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
    ) -> Result<Option<CudaSlice<f32>>, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        if let GpuTensor::Quant {
            bytes,
            qtype,
            row_bytes,
            blk: Some(g),
            ..
        } = w
        {
            if *qtype == QT_F8_E4M3_BLK {
                // BATCHED tier m=2..16 (lane/rp-on-st): weight read ONCE for up to mcols columns
                // instead of m grid.y re-reads. Bit-identical per (token,row) to the grid.y=m form
                // below, so the decode-exactness contract is preserved at every width. Gated by
                // the same seams the other batched families honor (MEMRA_NO_BATCHED, MEMRA_B8) so
                // one rollback door covers every dtype's batched tier.
                if (2..=16).contains(&m)
                    && std::env::var("MEMRA_NO_BATCHED").is_err()
                    && (m <= 4 || Self::b8_enabled())
                {
                    let mcols = Self::batched_mcols(m);
                    return Ok(Some(self.qmatvec_e4m3_blk_mmvq_batched(
                        bytes,
                        aq,
                        ad,
                        &g.scales,
                        m,
                        w.in_features(),
                        w.out_features(),
                        *row_bytes,
                        g.cols,
                        mcols,
                    )?));
                }
                return Ok(Some(self.qmatvec_e4m3_blk_mmvq(
                    bytes,
                    aq,
                    ad,
                    &g.scales,
                    m,
                    w.in_features(),
                    w.out_features(),
                    *row_bytes,
                    g.cols,
                )?));
            }
        }
        Ok(None)
    }

    /// PREFILL (m >= GEMM_M_THRESHOLD) for `QT_F8_E4M3_BLK` — DEQUANT-PER-CALL to the Q8_0 slab
    /// this class's residency replaced, then the ordinary Q8_0 prefill dispatch on the transient.
    ///
    /// WHY THIS EXISTS AT ALL, i.e. the regression it prevents: the decode kernel is a warp-per-row
    /// GEMV. At grid.y=m it re-reads the whole weight once PER TOKEN, so letting a 512-token prefill
    /// chunk reach it would be a ~500x weight-traffic blowup on the single most bandwidth-bound part
    /// of the forward. Native residency is a DECODE win and must not be paid for in prefill, so
    /// prefill keeps the floor's arithmetic and the floor's kernels.
    ///
    /// WHY DEQUANT-PER-CALL rather than a second resident slab: a resident slab is dual residency —
    /// it gives back the entire 1.0-vs-1.0625 B/weight win this lane exists to capture (and then
    /// some, since the e4m3 copy stays too). The transient costs one linear device pass per
    /// (projection, prefill call) and frees immediately.
    ///
    /// NUMERICALLY IT IS THE FLOOR, EXACTLY: `fp8_blk_dequant_q8_0` is the merged ARM B' kernel,
    /// gate-proven BYTE-IDENTICAL to the host dequant+re-encode (kernel-check `fp8-blk-gpu`). So the
    /// slab these bytes form is bit-for-bit the slab the `MEMRA_ST_E4M3_BLK=0` arm makes resident,
    /// and every prefill kernel downstream sees identical input — prefill logits under this lane are
    /// bit-identical to prefill logits under the floor, which is what makes the decode A/B a clean
    /// single-variable comparison instead of a two-variable one.
    ///
    /// WHAT IT COSTS, MEASURED, AND WHY THAT COST IS MOSTLY STRUCTURAL (27B block-128 ckpt, pp512,
    /// this rig = RTX 5090 Laptop, ~896 GB/s GDDR7). This arm makes prefill move the weight THREE
    /// times instead of once: read 6.88 GB of e4m3, write 7.31 GB of Q8_0, then the MMQ reads that
    /// 7.31 GB back. The two extra passes are 14.19 GB = 15.8 ms at this card's roofline against a
    /// ~332 ms pp512, i.e. **~-4.5% pp is a floor no kernel tuning can remove** — only deleting the
    /// dequant can. Measured: the dequant kernel costs 27.9 ms/pass (nsys, 208 projections) after
    /// the 2026-08-05 vector rewrite (was 66.5 ms at one byte per thread), and e2e pp512 is
    /// 1451.4 vs the slab arm's 1541.6 tok/s = -5.8% (N=3 interleaved pairs). So ~1.3pp of the
    /// -5.8% is residual kernel inefficiency and ~4.5pp is the extra traffic itself.
    ///
    /// SO THE DEQUANT IS NO LONGER THE DEFAULT ROUTE — it is the FALLBACK. The per-block FP8 MMQ
    /// tile (`try_fp8_blk_mmq`) consumes the resident e4m3 bytes + grid DIRECTLY, deleting both extra
    /// passes, and since 2026-08-05 it runs FIRST and by default for the native-resident source
    /// (`fp8_blk_mmq_native_enabled`; `MEMRA_FP8_MMQ=0` is the seam back to this dequant). On paper
    /// the trade was unassumable — lane/fp8-mmq-v2 measured that tile at 0.85-1.09x the Q8_0 MMQ
    /// floor GEMM-only, so it swapped a -4.5% traffic cost for a 0-to-15% GEMM cost of unknown sign.
    /// Measured on the 27B (3 arms interleaved, N=3, research/fp8blk-20260805/VERDICT.md): slab
    /// 1540.5 / this dequant 1449.1 / the tile 1553.3 tok/s, min(tile) > max(slab). The tile wins
    /// because v2's denominator had its slab already resident while this class's floor must build it
    /// every call; same tile, opposite sign, because the question changed.
    ///
    /// THIS ARM STILL RUNS, and is not dead code: every `try_fp8_blk_mmq` precondition (in_f % 16,
    /// grid dims vs shape, per-tensor scale == 1.0, the e4m3-NaN scan) refuses by falling through to
    /// here, so a checkpoint the tile cannot take keeps exact prefill on the floor's own bits rather
    /// than losing the class. It is also what `MEMRA_FP8_MMQ=0` reverts to.
    fn try_e4m3_blk_prefill(
        &self,
        w: &crate::model::GpuTensor,
        x: &CudaSlice<f32>,
        m: usize,
    ) -> Result<Option<CudaSlice<f32>>, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        let GpuTensor::Quant {
            bytes,
            qtype,
            blk: Some(g),
            ..
        } = w
        else {
            return Ok(None);
        };
        if *qtype != QT_F8_E4M3_BLK {
            return Ok(None);
        }
        // NO-DEQUANT ROUTE, THE DEFAULT (MEMRA_FP8_MMQ=0 reverts): the per-block MMQ tile eats the
        // resident e4m3 bytes and grid as-is, so neither extra weight pass happens. Its own
        // preconditions (in_f % 16, grid dims, scale == 1.0, no e4m3 NaN code) can refuse — fall
        // through to the dequant below when they do, never silently produce nothing.
        if let Some(y) = self.try_fp8_blk_mmq(w, x, m)? {
            return Ok(Some(y));
        }
        let (in_f, out_f) = (w.in_features(), w.out_features());
        let slab = self.fp8_blk_dequant_q8_0_dev(bytes, &g.scales, out_f, in_f)?;
        let tmp = GpuTensor::Quant {
            bytes: slab,
            qtype: QT_Q8_0,
            row_bytes: in_f / 32 * 34,
            ne: vec![in_f as u64, out_f as u64],
            scale: 1.0,
            rp: false,
            #[cfg(memra_cutlass)]
            cutlass: None,
            fp8: None,
            blk: None,
            f16: None,
            rp4: None,
        };
        // Recursion terminates: `tmp` is QT_Q8_0 with `blk: None`, so it cannot re-enter this arm.
        Ok(Some(self.matmul(&tmp, x, m)?))
    }

    pub fn matmul_pre_noscale(
        &self,
        w: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
    ) -> Result<Option<(CudaSlice<f32>, f32)>, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        // BLOCK-128 e4m3: every scale factor is folded inside the kernel per k128, so the
        // "separable post-op scale" this entry exists to defer is 1.0 — return it explicitly
        // rather than let the tail below refuse and cost the caller a re-dispatch.
        if m == 1 {
            if let Some(y) = self.try_e4m3_blk_pre(w, aq, ad, m)? {
                return Ok(Some((y, 1.0)));
            }
        }
        // Only the m==1 fast path applies the scale as a separable post-op; bail everywhere else.
        if m != 1 || !self.uses_q8_1_fast(w) {
            return Ok(None);
        }
        let in_f = w.in_features();
        let out_f = w.out_features();
        let (bytes, qtype, row_bytes, scale, rp) = match w {
            GpuTensor::Quant {
                bytes,
                qtype,
                row_bytes,
                scale,
                rp,
                ..
            } => (bytes, *qtype, *row_bytes, *scale, *rp),
            _ => return Ok(None),
        };
        // MMVQ warp-per-row (scale==1.0 passed -> kernel skips its internal scale; we return scale).
        if self.mmvq_supports(qtype) {
            // Q4_0 split-plane mirror (dp4a fallback below keeps the raw GGUF bytes).
            let (mbytes, mrp) = match w {
                GpuTensor::Quant { rp4: Some(m4), .. } => (m4, true),
                _ => (bytes, rp),
            };
            let y = self.qmatvec_mmvq(
                mbytes, aq, ad, m, in_f, out_f, qtype, row_bytes, /*scale*/ 1.0, mrp,
            )?;
            return Ok(Some((y, scale)));
        }
        // dp4a fallback: same launch as matmul_pre but WITHOUT the post scale_inplace.
        let name = match qtype {
            QT_Q8_0 => "qmatvec_q8_0_dp4a",
            QT_Q4_K => "qmatvec_q4_K_dp4a",
            QT_Q6_K => "qmatvec_q6_K_dp4a",
            QT_Q5_K => "qmatvec_q5_K_dp4a",
            QT_Q3_K => "qmatvec_q3_K_dp4a",
            QT_NVFP4 => {
                if rp {
                    "qmatvec_nvfp4_dp4a_rp"
                } else {
                    "qmatvec_nvfp4_dp4a"
                }
            }
            QT_IQ4_XS => "qmatvec_iq4_XS_dp4a",
            _ => return Ok(None),
        };
        let f = self.func(name);
        let mut y = self.alloc_uninit::<f32>(m * out_f)?;
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, m as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi, rb) = (in_f as i32, out_f as i32, m as i32, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(bytes)
            .arg(aq)
            .arg(ad)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(Some((y, scale)))
    }

    /// True if `qtype` has a warp-per-row MMVQ decode kernel AND MEMRA_MMVQ is set. Only the 4
    /// daily-hot dtypes (Q8_0, Q4_K, Q6_K, NVFP4) — others keep the _dp4a matvec (oracle/fallback).
    pub fn mmvq_supports(&self, qtype: i32) -> bool {
        // DEFAULT ON since 2026-07-08 (MEMRA_MMVQ=0 reverts to the _dp4a matvec class).
        // QT_F8_E4M3 is exempt from the MEMRA_MMVQ=0 escape: the e4m3 mmvq family is that dtype's
        // ONLY int8-act kernel class (there is no _dp4a twin), so its m=1/verify/batched dispatch
        // is a pure function of the dtype — the decode-parity law holds under every env.
        if qtype == QT_F8_E4M3 {
            return true;
        }
        if std::env::var("MEMRA_MMVQ").as_deref() == Ok("0") {
            return false;
        }
        matches!(
            qtype,
            QT_Q8_0 | QT_Q4_K | QT_Q5_K | QT_Q6_K | QT_NVFP4 | QT_Q4_0
        )
    }

    /// PERF-3 warp-per-row MMVQ launcher (decode m=1 hot path). block=(32,ROWS_PER_BLOCK,1):
    /// one warp owns one output row, warp-only __shfl reduction (no smem barrier). Bit-equivalent
    /// to qmatvec_*_dp4a up to f32 reduction order. Pre-quantized q8_1 activation (aq,ad). NVFP4
    /// per-tensor macro-scale applied post (scale==1.0 for other dtypes -> no-op).
    pub fn qmatvec_mmvq(
        &self,
        bytes: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        qtype: i32,
        row_bytes: usize,
        scale: f32,
        rp: bool,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let mut y = self.alloc_uninit::<f32>(m * out_f)?; // full-overwrite GEMM output: skip memset
        self.qmatvec_mmvq_into(
            bytes, aq, ad, m, in_f, out_f, qtype, row_bytes, scale, rp, &mut y,
        )?;
        Ok(y)
    }

    /// Slot-fed MMVQ twin (alloc-free capture lane): full policy body, caller-owned output.
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_mmvq_into(
        &self,
        bytes: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        qtype: i32,
        row_bytes: usize,
        scale: f32,
        rp: bool,
        y: &mut CudaSlice<f32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        debug_assert!(y.len() >= m * out_f);
        const ROWS_PER_BLOCK: u32 = 4; // matches MEMRA_MMVQ_ROWS in qmatvec.cu
        // SMALL-SHAPE GRID FILL (H100 lane, 2026-07-26 microbench: attn qkv out_f=2048 =
        // 0.97 waves at the 4-warp block -> 66% of peak). The g2 twin (2 warps/block)
        // doubles the grid when the 4-warp launch would be sub-wave; per-row program
        // identical -> bit-identical. MEMRA_Q80_G2=0 reverts.
        if qtype == QT_Q8_0
            && rp
            && m == 1
            && out_f >= 64
            && (out_f as u32).div_ceil(ROWS_PER_BLOCK) < 4 * self.sm_count() as u32
            && {
                static G2: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
                *G2.get_or_init(|| std::env::var("MEMRA_Q80_G2").as_deref() != Ok("0"))
            }
        {
            let f = self.func("qmatvec_q8_0_mmvq_rp_g2");
            let cfg = LaunchConfig {
                grid_dim: ((out_f as u32).div_ceil(2), 1, 1),
                block_dim: (32, 2, 1),
                shared_mem_bytes: 0,
            };
            let (inf, outf, mi, rb) = (in_f as i32, out_f as i32, 1i32, row_bytes as i64);
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(bytes)
                .arg(aq)
                .arg(ad)
                .arg(&mut *y)
                .arg(&inf)
                .arg(&outf)
                .arg(&mi)
                .arg(&rb);
            unsafe {
                b.launch(cfg)?;
            }
            if scale != 1.0 {
                self.scale_inplace(y, scale, out_f)?;
            }
            return Ok(());
        }
        // Multi-row-per-warp (mr2) policy, fixed since the 2026-07 sweeps (the MEMRA_MMVQ_MR
        // override + mr4 kernel were retired 2026-07-08 — mr4 regressed on register pressure and
        // crashed under rp; q4_K/q6_K mr2 measured flat, "no gain = no change"):
        //   NVFP4 m=1 -> mr2 (clean +1-2% on 9B: RPW acc chains hide the weight-load latency
        //     that pins the single-row kernel at 30-46% DRAM). Bit-identical per row.
        //   Q5_K m=1 -> mr2 (2026-07-05: the FR-Spec trimmed draft head is Q5_K 32768 rows = 8%
        //     of the 27B p3 spec wall; latency-bound like the other k-quants pre-fix).
        //   Q4_K/Q6_K m=1 -> single-row (mr2 measured +0.7% / flat — weight-bandwidth-bound).
        let mut mr: u32 = if m == 1 && (qtype == QT_NVFP4 || qtype == QT_Q5_K) {
            2
        } else {
            1
        };
        // Q4_0 mr (gemma trunk): DEFAULT 1 since 2026-07-13 (MEMRA_Q40_MR=2 reverts) — the
        // mr1 rp twin doubles the block count and wins the tail-quantization/latency battle
        // on every gemma model (E4B +3.75%: 198.9 vs 191.7; 26B +0.7%; 31B +0.9%; N=2-3
        // valid-window interleaved, bit-identical per row — same dot program).
        if m == 1 && qtype == QT_Q4_0 {
            static Q40MR: std::sync::OnceLock<u32> = std::sync::OnceLock::new();
            // shape policy PROBED NEGATIVE (2026-07-13): tall-only mr1 197.2 vs
            // mr1-everywhere 198.7 — mr1 wins wide-output shapes too; arm removed.
            mr = *Q40MR.get_or_init(|| {
                std::env::var("MEMRA_Q40_MR")
                    .ok()
                    .and_then(|v| v.parse().ok())
                    .unwrap_or(1)
            });
        }
        // q5issue lane (2026-07-08): MEMRA_Q5K_ISSUE swaps the q5_K m=1 mmvq kernels for the
        // issue-reduced `_il` bodies (uint4 header/qh/qs loads + branchless scale decode —
        // cuts ~34 LDG.U16 + ~5 LDG.U8 + a warp-divergent scale branch per 32-elem group-row
        // to 5 LDG.128). Bit-identical per (token,row) to the reference kernels.
        // `1` = shape-aware policy (N=3 clock-locked micro-bench, mem P0, synthetic real shapes):
        //   out_f <= 65536 (trunk/frspec regime): il at the default mr — mr2_il -9.5%/-10.5%
        //     on 4096x4096/4096x8192, -3.1% on the 32768 frspec head vs the mr2-ref default;
        //   out_f > 65536 (the 248320-row 27B lm_head, already ~97% of the mem wall): mr2_il
        //     REGRESSES +22% there but mr1_il wins -2.1% vs the mr2-ref default -> force mr=1.
        // `2` = force il at the current mr for EVERY shape (A/B probe seam). Default OFF.
        let q5_mode = std::env::var("MEMRA_Q5K_ISSUE").ok();
        let q5_force = q5_mode.as_deref() == Some("2");
        // DEFAULT ON since 2026-07-08 (MEMRA_Q5K_ISSUE=0 reverts): +1.8% 9B plain e2e N=3
        // (128.2 -> 130.4), 27B flat (its big head is already at the mem wall), all gates green.
        let q5_il = qtype == QT_Q5_K
            && m == 1
            && (q5_force || q5_mode.as_deref().map(|v| v != "0").unwrap_or(true));
        if q5_il && !q5_force && out_f > 65536 {
            mr = 1;
        }
        // Q4_0 split-plane rp: mr2 default; MEMRA_Q40_MR=1 reaches the mr1 rp twin
        // (2026-07-13 — the tall-input/short-output tail-quantization probe).
        if qtype == QT_Q4_0 && rp && mr != 1 {
            mr = 2;
        }
        // Q8_0 rp (H100 lane): mr1 default — the q4_0 mr2 recipe MEASURED NEGATIVE on H100
        // (2026-07-26 N=3: mr1 186.2 vs mr2 171.5 tok/s; halving the grid on 132 SMs costs
        // more than 2-row ILP buys). mr2 kernel stays behind MEMRA_Q80_MR=2 for the corpus.
        if qtype == QT_Q8_0 && rp {
            static Q80MR: std::sync::OnceLock<u32> = std::sync::OnceLock::new();
            mr = *Q80MR.get_or_init(|| {
                std::env::var("MEMRA_Q80_MR")
                    .ok()
                    .and_then(|v| v.parse().ok())
                    .unwrap_or(1)
            });
        }
        let name = match (qtype, mr, rp) {
            (QT_NVFP4, 2, false) => "qmatvec_nvfp4_mmvq_mr2",
            (QT_NVFP4, 2, true) => "qmatvec_nvfp4_mmvq_mr2_rp",
            (QT_NVFP4, _, true) => "qmatvec_nvfp4_mmvq_rp",
            (QT_Q4_0, 1, true) => "qmatvec_q4_0_mmvq_rp",
            (QT_Q4_0, _, true) => "qmatvec_q4_0_mmvq_mr2_rp",
            (QT_Q5_K, 2, _) => {
                if q5_il {
                    "qmatvec_q5_K_mmvq_mr2_il"
                } else {
                    "qmatvec_q5_K_mmvq_mr2"
                }
            }
            (QT_Q8_0, 2, true) => "qmatvec_q8_0_mmvq_mr2_rp",
            // rpca (cp.async-staged weight ring): MEASURED NEGATIVE on H100 for Q8_0
            // (2026-07-26 N=3: 181.8 vs plain rp 185.5 — the smem round-trip exceeds the
            // latency it hides for 8-bit direct-dp4a; the NVFP4 win case overlaps table
            // decode with half the bytes). OPT-IN via MEMRA_Q80_CA=1 for the corpus.
            (QT_Q8_0, _, true)
                if in_f % 1024 == 0 && {
                    static CA: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
                    *CA.get_or_init(|| std::env::var("MEMRA_Q80_CA").as_deref() == Ok("1"))
                } =>
            {
                "qmatvec_q8_0_mmvq_rpca"
            }
            (QT_Q8_0, _, true) => "qmatvec_q8_0_mmvq_rp",
            (QT_Q8_0, _, _) => "qmatvec_q8_0_mmvq",
            // K-quant split-plane twins (H100 K-quant coalescing fix, 2026-08-01): the rp4
            // mirror routes here; GGUF layout keeps the plain kernels. rp bytes MUST never
            // reach a GGUF-layout kernel or vice versa.
            (QT_Q4_K, _, true) => "qmatvec_q4_K_mmvq_rp",
            (QT_Q6_K, _, true) => "qmatvec_q6_K_mmvq_rp",
            (QT_Q4_K, _, _) => "qmatvec_q4_K_mmvq",
            (QT_Q4_0, 2, false) => "qmatvec_q4_0_mmvq_mr2",
            (QT_Q4_0, _, false) => "qmatvec_q4_0_mmvq",
            (QT_Q5_K, _, _) => {
                if q5_il {
                    "qmatvec_q5_K_mmvq_il"
                } else {
                    "qmatvec_q5_K_mmvq"
                }
            }
            (QT_Q6_K, _, _) => "qmatvec_q6_K_mmvq",
            (QT_NVFP4, _, false) => "qmatvec_nvfp4_mmvq",
            (QT_F8_E4M3, _, _) => "qmatvec_e4m3_mmvq",
            _ => panic!("qmatvec_mmvq: qtype {qtype} has no MMVQ kernel"),
        };
        let f = self.func(name);
        // each block still has ROWS_PER_BLOCK warps; with mr rows/warp it covers ROWS_PER_BLOCK*mr rows.
        let rows_per_block = ROWS_PER_BLOCK * mr;
        let cfg = LaunchConfig {
            grid_dim: (
                (out_f as u32 + rows_per_block - 1) / rows_per_block,
                m as u32,
                1,
            ),
            block_dim: (32, ROWS_PER_BLOCK, 1), // warp-per-row (x mr rows each)
            shared_mem_bytes: 0,                // warp-only reduce at m=1
        };
        let (inf, outf, mi, rb) = (in_f as i32, out_f as i32, m as i32, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        // NVFP4 + e4m3 mmvq kernels take the macro-scale as a fused epilogue arg (applied at the
        // write — bit-identical to the old separate scale_inplace pass, minus one launch per matvec:
        // 53 scale launches/token on the 9B; for e4m3 the scale is the checkpoint's per-tensor f32
        // weight_scale). Other mmvq kernels keep the 8-arg signature.
        if qtype == QT_NVFP4 || qtype == QT_F8_E4M3 {
            // PDL wave-B: the nvfp4 mr2_rp single (gemma wo / generic rp singles) joins
            // the wave-A launch class — 9-arg flavor (fused macro-scale epilogue).
            if Self::pdl_on()
                && Self::pdl_mmvq_on()
                && Self::pdl_nvfp4q8_on()
                && name == "qmatvec_nvfp4_mmvq_mr2_rp"
            {
                use cudarc::driver::{DevicePtr, DevicePtrMut};
                let s = &self.gpu.stream();
                let (pw, _g0) = bytes.device_ptr(s);
                let (paq, _g1) = aq.device_ptr(s);
                let (pad, _g2) = ad.device_ptr(s);
                let (py, _g3) = y.device_ptr_mut(s);
                let mut ps = [
                    &pw as *const _ as *mut std::ffi::c_void,
                    &paq as *const _ as *mut _,
                    &pad as *const _ as *mut _,
                    &py as *const _ as *mut _,
                    &inf as *const _ as *mut _,
                    &outf as *const _ as *mut _,
                    &mi as *const _ as *mut _,
                    &rb as *const _ as *mut _,
                    &scale as *const _ as *mut _,
                ];
                unsafe {
                    self.launch_pdl(name, cfg.grid_dim, cfg.block_dim, &mut ps)?;
                }
                return Ok(());
            }
            b.arg(bytes)
                .arg(aq)
                .arg(ad)
                .arg(&mut *y)
                .arg(&inf)
                .arg(&outf)
                .arg(&mi)
                .arg(&rb)
                .arg(&scale);
            unsafe {
                b.launch(cfg)?;
            }
        } else if Self::pdl_on()
            && Self::pdl_mmvq_on()
            && (matches!(
                name,
                "qmatvec_q4_0_mmvq_rp" | "qmatvec_q6_K_mmvq" | "qmatvec_q6_K_mmvq_rp"
            ) || (Self::pdl_nvfp4q8_on()
                && matches!(name, "qmatvec_q8_0_mmvq_rp" | "qmatvec_q8_0_mmvq_mr2_rp")))
        {
            // PDL wave-A (2026-07-23): the two decode-hot single-matvec kernels carry
            // MEMRA_PDL_ENTRY — grid launches while the producer drains. ONLY the marked
            // names may take this launch (unmarked kernels would read unordered).
            {
                use cudarc::driver::{DevicePtr, DevicePtrMut};
                let s = &self.gpu.stream();
                let (pw, _g0) = bytes.device_ptr(s);
                let (paq, _g1) = aq.device_ptr(s);
                let (pad, _g2) = ad.device_ptr(s);
                let (py, _g3) = y.device_ptr_mut(s);
                let mut ps = [
                    &pw as *const _ as *mut std::ffi::c_void,
                    &paq as *const _ as *mut _,
                    &pad as *const _ as *mut _,
                    &py as *const _ as *mut _,
                    &inf as *const _ as *mut _,
                    &outf as *const _ as *mut _,
                    &mi as *const _ as *mut _,
                    &rb as *const _ as *mut _,
                ];
                unsafe {
                    self.launch_pdl(name, cfg.grid_dim, cfg.block_dim, &mut ps)?;
                }
            }
            if scale != 1.0 {
                self.scale_inplace(y, scale, m * out_f)?;
            }
        } else {
            b.arg(bytes)
                .arg(aq)
                .arg(ad)
                .arg(&mut *y)
                .arg(&inf)
                .arg(&outf)
                .arg(&mi)
                .arg(&rb);
            unsafe {
                b.launch(cfg)?;
            }
            if scale != 1.0 {
                self.scale_inplace(y, scale, m * out_f)?;
            }
        }
        Ok(())
    }

    /// Test entry for the kernel_check bit-equivalence gate: run the warp-per-row MMVQ directly
    /// from raw weight bytes (quantize the f32 activation `x` to q8_1 internally). NVFP4 per-tensor
    /// macro-scale is NOT applied (caller compares bare, like qmatvec_*_fast). Mirrors qmatvec_gemm_raw.
    pub fn qmatvec_mmvq_raw(
        &self,
        bytes: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        qtype: i32,
        row_bytes: usize,
        rp: bool,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
        self.qmatvec_mmvq(bytes, &aq, &ad, m, in_f, out_f, qtype, row_bytes, 1.0, rp)
    }

    /// True if `qtype` has a batched weight-resident (`_b2`/`_b4`) matvec kernel. These mirror the
    /// `_mmvq` kernels but iterate the m token columns INSIDE one warp/row, so the weight bytes leave
    /// HBM/L2 once for m tokens (vs grid.y=m re-reading m times). The 5 daily-hot dtypes have them.
    pub fn batched_supports(&self, qtype: i32) -> bool {
        matches!(
            qtype,
            QT_Q8_0 | QT_Q4_K | QT_Q5_K | QT_Q6_K | QT_NVFP4 | QT_F8_E4M3 | QT_Q4_0
        )
    }

    /// IQ4_XS trunk fast seam: MEMRA_IQ_FAST=0 reverts non-expert IQ4_XS matmuls to the Stage-A
    /// f32 oracle path. Default ON since 2026-08-02 (research/kat-anomaly-20260802/): the old
    /// opt-in default left every IQ4_XS-trunk artifact (KAT-Coder IQ4_XS: attn_qkv/attn_gate/
    /// ssm_out/shexp, ~0.52GB re-read per decode tick) on the oracle kernel — decode 106.7 ->
    /// 193.4 tok/s (x5 interleaved), pp512 228 -> 697, same bytes, via qmatvec_iq4_XS_dp4a. The
    /// supported artifacts carry IQ4_XS only in EXPERT banks (their own dispatch, not this seam),
    /// so this admission is dispatch-unchanged for every non-IQ4_XS-trunk model.
    pub fn iq_fast_enabled() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| {
            std::env::var("MEMRA_IQ_FAST")
                .map(|v| v != "0")
                .unwrap_or(true)
        })
    }

    /// b8 tier seam: MEMRA_B8=0 keeps m=5..8 on the per-m grid.y=m path (m=2..4 batched dispatch
    /// unaffected). Default ON — the K=4..7 spec-verify weight-read-once fix.
    pub fn b8_enabled() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| std::env::var("MEMRA_B8").map(|v| v != "0").unwrap_or(true))
    }

    /// Compile-time column batch for a runtime m: 2 -> b2, 3..4 -> b4, 5..8 -> b8.
    pub fn batched_mcols(m: usize) -> usize {
        if m == 2 {
            2
        } else if m <= 4 {
            4
        } else if m <= 8 {
            8
        } else {
            16
        }
    }

    /// Kernel name for the batched matvec of `(qtype, mcols)`. mcols ∈ {2,4,8}. The b8 tier is the
    /// K=4..7 spec-verify fix (T=5..8): pre-b8 those T fell to grid.y=m per-row MMVQ = m full
    /// weight reads/launch — the measured 27B K=4 cliff (101 -> 73 tok/s at p3 despite acceptance
    /// holding 54%). One b8 launch reads the weight ONCE for up to 8 columns (c >= m masked).
    fn batched_kernel_name(qtype: i32, mcols: usize) -> Option<&'static str> {
        Some(match (qtype, mcols) {
            (QT_Q8_0, 2) => "qmatvec_q8_0_mmvq_b2",
            (QT_Q8_0, 4) => "qmatvec_q8_0_mmvq_b4",
            (QT_Q8_0, 8) => "qmatvec_q8_0_mmvq_b8",
            // b16 now has BOTH forms (lane/rp-on-st, 2026-08-06). It used to be rp-ONLY, which
            // made the q8rp mirror the exact-16 tier's admission ticket for any model carrying a
            // single Q8_0 matmul — measured as the FP8-ST refusal (`L0.ssm_beta qtype=0
            // rp4=false`, 96 t / 23.9 MiB = 0.143% of resident weight). The mirror stays a
            // BANDWIDTH lever on Q8_0-dominant GGUFs; it is no longer a correctness prerequisite.
            (QT_Q8_0, 16) => "qmatvec_q8_0_mmvq_b16",
            (QT_Q4_K, 2) => "qmatvec_q4_K_mmvq_b2",
            (QT_Q4_K, 4) => "qmatvec_q4_K_mmvq_b4",
            (QT_Q4_K, 8) => "qmatvec_q4_K_mmvq_b8",
            // b16 base + _rp (lane/rp-on-st): the 9B NVFP4 GGUF's blocker — real NVFP4 GGUFs keep
            // Q4_K attention next to NVFP4 MLP, and the tier's predicate is an ALL.
            (QT_Q4_K, 16) => "qmatvec_q4_K_mmvq_b16",
            (QT_Q5_K, 2) => "qmatvec_q5_K_mmvq_b2",
            (QT_Q5_K, 4) => "qmatvec_q5_K_mmvq_b4",
            (QT_Q5_K, 8) => "qmatvec_q5_K_mmvq_b8",
            // b16 base only (lane/rp-on-st): Q5_K has no rp twins at any width, so there is
            // nothing to mirror. Named by the diagnostic as `L0.wqkv_gate qtype=3` on the 9B.
            (QT_Q5_K, 16) => "qmatvec_q5_K_mmvq_b16",
            (QT_Q6_K, 2) => "qmatvec_q6_K_mmvq_b2",
            (QT_Q6_K, 4) => "qmatvec_q6_K_mmvq_b4",
            (QT_Q6_K, 8) => "qmatvec_q6_K_mmvq_b8",
            (QT_Q6_K, 16) => "qmatvec_q6_K_mmvq_b16",
            (QT_NVFP4, 2) => "qmatvec_nvfp4_mmvq_b2",
            (QT_NVFP4, 4) => "qmatvec_nvfp4_mmvq_b4",
            (QT_NVFP4, 8) => "qmatvec_nvfp4_mmvq_b8",
            // b16 (lane/rp-on-st): no mirror needed — NVFP4's 36 B/k32 block is already the
            // aligned form its own kernel walks. Unlocks the exact-16 tier for every NVFP4 model
            // AND for the mixed FP8-ST artifact, whose 193 NVFP4 tensors were refusing it.
            (QT_NVFP4, 16) => "qmatvec_nvfp4_mmvq_b16",
            (QT_F8_E4M3, 2) => "qmatvec_e4m3_mmvq_b2",
            (QT_F8_E4M3, 4) => "qmatvec_e4m3_mmvq_b4",
            (QT_F8_E4M3, 8) => "qmatvec_e4m3_mmvq_b8",
            // b16 tier (lane/rp-on-st): e4m3 needs NO split-plane mirror to reach it — its native
            // row-major layout is already 32B-aligned per k32 block, so the base kernel IS the
            // aligned form. Contrast Q8_0, whose b16 exists only as the `_rp` twin (hence q8rp).
            (QT_F8_E4M3, 16) => "qmatvec_e4m3_mmvq_b16",
            (QT_Q4_0, 2) => "qmatvec_q4_0_mmvq_b2",
            (QT_Q4_0, 4) => "qmatvec_q4_0_mmvq_b4",
            (QT_Q4_0, 8) => "qmatvec_q4_0_mmvq_b8",
            (QT_Q4_0, 16) => "qmatvec_q4_0_mmvq_b16",
            _ => return None,
        })
    }

    /// BATCHED weight-tile-resident matvec from a PRE-QUANTIZED q8_1 activation (the m=2-8 verify/MTP
    /// win). One warp walks the weight row ONCE, dp4a vs all m activation columns -> weight HBM/L2
    /// traffic 1x for m tokens (vs grid.y=m re-reading it m times). `mcols` ∈ {2,4,8} is the
    /// compile-time batch; m must be <= mcols (the c >= m columns are masked in-kernel). y is
    /// [m, out_f] token-major. NVFP4 per-tensor macro-scale applied post
    /// (scale==1.0 for other dtypes -> no-op). BIT-IDENTICAL per (token,row) to qmatvec_*_mmvq.
    ///
    /// NVFP4 VARIANT DISPATCH: the batched NVFP4 kernel measured memory-LATENCY bound on the real
    /// 27B verify (ncu --set full, 12 steady launches: long_scoreboard 18-30 stalls/issue vs <=1.7
    /// for every other reason, DRAM only 41-51% active, lg_throttle 0.7, L1 hit 94% — ONE 6-LDG
    /// weight wavefront in flight per warp is the binding constraint, NOT bandwidth and NOT the
    /// column-unroll break). Two exactness-free fixes, chosen PER SHAPE from the DRAM-cold 8-copy
    /// msweep on all six 27B shapes (2026-07-03):
    ///   `pf` = next-g weight-prefetch double-buffer (48 regs, occupancy intact) — wins everywhere
    ///          it applies for b4 (-3..-14%), never loses;
    ///   `r2` = two rows/warp (67 regs -> 7 resident blocks/SM) — the bigger win (-8.5..-30%) but
    ///          wave-quantization-sensitive: with the grid halved to ceil(out_f/8) blocks, a
    ///          fractional straggler wave (waves in ~1.05-1.5) costs a full extra latency round on
    ///          a latency-bound kernel (27B ffn_down 640 blocks / 574 resident = 1.11 waves: +17%),
    ///          while <=1 wave (9B ffn_down 0.89: -30%) or >=2 waves (tail amortized; qkv 2.2:
    ///          -8.5%, ffn_gate 3.8: -12.5%) win. For b2, r2 wins on DEEP k-loops (in_f>=6144:
    ///          -8..-19%) where the 2-col body starves weight MLP hardest; pf measured negative.
    /// b4: r2 when waves(out_f) <= 1 (and grid fills >=half the SMs) or >= 2, else pf.
    /// b2: in_f>=6144 -> r2, else base.
    /// MEMRA_MMVQ_BV=base|pf|r2|pfr2 forces one variant everywhere (A/B + rollback seam).
    /// All variants BIT-IDENTICAL per (token,row): same dp4a order, scales, adg factor, reduce —
    /// only load issue time and the row->warp mapping change (kernel-check gates all of them).
    /// `rp` = the weight buffer is the A6 SPLIT-PLANE repacked layout (NVFP4 only): the same
    /// wave-aware auto rule applies, mapped onto the `_rp` twins (rp/rpr2/rpr2w8 mirror
    /// pf/r2/r2w8 — regs 44/67/64 land in the same residency classes).
    /// The variant the batched dispatch will pick for this (shape, m, mcols, layout) — exposed so
    /// gates can distinguish bit-identical variants (bit-bad==0 required) from the k-split family
    /// (deterministic but k-reduce-order-shifted: rel<1e-3 + run-to-run bit-identity required).
    /// Device SM count (cached) — grid-fill policy input.
    pub fn sm_count(&self) -> i32 {
        static SMS: std::sync::OnceLock<i32> = std::sync::OnceLock::new();
        *SMS.get_or_init(|| {
            use cudarc::driver::sys::CUdevice_attribute_enum as A;
            self.gpu
                .ctx
                .attribute(A::CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT)
                .unwrap_or(82)
        })
    }

    pub fn batched_variant(
        &self,
        _m: usize,
        in_f: usize,
        out_f: usize,
        qtype: i32,
        row_bytes: usize,
        mcols: usize,
        rp: bool,
    ) -> &'static str {
        // Q8_0 never joined the auto variant machinery (on sm_120 its only batched shapes
        // were tiny aux tensors). On Q8_0-trunk models the layout is the whole game: the
        // split-plane mirror (rp) routes to the _rp twins (H100 coalescing fix, 2026-07-26);
        // GGUF layout stays "base". rp bytes MUST never reach the base kernel or vice versa.
        if qtype == QT_Q8_0 {
            return if rp { "rp" } else { "base" };
        }
        static BV: std::sync::OnceLock<&'static str> = std::sync::OnceLock::new();
        let bv = *BV.get_or_init(|| match std::env::var("MEMRA_MMVQ_BV").as_deref() {
            Ok("base") => "base",
            Ok("pf") => "pf",
            Ok("r2") => "r2",
            Ok("r2w8") => "r2w8",
            Ok("pfr2") => "pfr2",
            Ok("ca") => "ca",
            Ok("car2") => "car2",
            // rp* = SPLIT-PLANE REPACKED layout kernels (A6 prototype): W must already be the
            // repacked buffer (msweep MSWEEP_RP harness) — never valid on GGUF-layout weights.
            Ok("rp") => "rp",
            Ok("rpr2") => "rpr2",
            Ok("rpr2w8") => "rpr2w8",
            // rpca* = cp.async software-pipelined split-plane (2026-07-05): hides the _rp
            // long_scoreboard load stall. rp-layout only; b4/b2 (no b8 twin).
            Ok("rpca") => "rpca",
            Ok("rpcar2") => "rpcar2",
            // 2026-07-06 m-small latency arc: rpsc = rpr2 + per-warp smem scale prestage (kills
            // the scale-plane global dependency, zero reg growth); rpms/rpmsc = m-split x2
            // across warp pairs (2x blocks of rpr2, column halves per warp, BIT-identical to
            // _rp); rpks/rpksc = k-split x2 (fastest microbench cells but k-reduce-order-shifted:
            // run-spec self-consistency FAILED on the 27B daily driver — verify logits must be
            // bit-identical to the decode path — measurement corpus ONLY, never auto).
            Ok("rpsc") => "rpsc",
            Ok("rpms") => "rpms",
            Ok("rpmsc") => "rpmsc",
            Ok("rpks") => "rpks",
            Ok("rpksc") => "rpksc",
            _ => "auto",
        });
        // cp.async ring variants need 16B-aligned rows (in_f%256==0 -> (in_f/64)*36 % 16 == 0)
        // and whole 32-group warp iterations (nsb%32==0 <=> in_f%1024==0). All 27B/9B trunk
        // shapes qualify; anything else falls back to the register variants.
        let ca_ok = qtype == QT_NVFP4 && (row_bytes % 16 == 0) && (in_f % 1024 == 0);
        // rpsc: smem scale plane fits (nsb64 <= 272) + int4-aligned staging (nsb64 % 4 == 0).
        // rpks/rpksc: half-plane staging alignment needs nsb64 % 8 == 0 (in_f % 512 == 0).
        // MEMRA_KS=0 removes the 2026-07-06 rpsc/rpks/rpksc entries from AUTO (rollback seam;
        // forced MEMRA_MMVQ_BV values still work).
        static KS_ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let ks_on = *KS_ON.get_or_init(|| std::env::var("MEMRA_KS").as_deref() != Ok("0"));
        let sc_ok = ks_on && qtype == QT_NVFP4 && (in_f % 256 == 0) && (in_f / 64 <= 272);
        let ks_ok = ks_on && qtype == QT_NVFP4 && (in_f % 512 == 0) && (in_f / 64 <= 272);
        static SMS: std::sync::OnceLock<i32> = std::sync::OnceLock::new();
        let sms = *SMS.get_or_init(|| {
            use cudarc::driver::sys::CUdevice_attribute_enum as A;
            self.gpu
                .ctx
                .attribute(A::CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT)
                .unwrap_or(82)
        });
        // k-quant r2 port (2026-07-04): q4_K/q5_K/q6_K have _r2/_r2w8 twins. ncu on the DRAM-cold
        // 9B msweep showed q4_K/q5_K b4 memory-latency bound like NVFP4 pre-fix (long_scoreboard
        // 19.6/16.4 per issue, DRAM 47.7/38.2%, L2 weight hit ~13%); q6_K lm_head is the exception
        // at DRAM 90-91% = wall-bound (yet r2 still wins -8%: deeper MLP raises achieved DRAM).
        // No _pf port (a k-quant group stages 10+ words vs NVFP4's 5 — register cost outweighs;
        // r2 covers the same MLP) and no rp (GGUF layout only). Q8_0 stays base: its only real
        // batched shapes are the tiny out_f=32 ssm_alpha/beta (8-block grids never fill one SM).
        // AUTO RULE = the measured winners table (differs from NVFP4's!):
        //   r2w8 NEVER in auto — the reg squeeze (72 -> 64 regs = stack spill) loses to unbounded
        //     r2 on every measured k-quant cell, incl. the wave-crossing lm_heads (q6_K 1316 vs
        //     r2 1258us) — kernels kept behind the force seam for the corpus;
        //   q4_K: r2 whenever the halved grid fills the SMs (blocks >= 4*SMs), INCLUDING the
        //     1.05-2.0 straggler window where NVFP4's r2 lost (qkv 1.78 waves: r2 -15% here; the
        //     k-quant base kernel leaves more latency on the table than a straggler wave costs);
        //   q5_K/q6_K: r2 only at waves >= 2 (the 248320-row lm_heads, 48+ waves: q6_K -8%, q5_K
        //     -2%); mid shapes measured base-or-flat (q5_K qkv 49.1 base vs 49.7 r2, attn_gate
        //     flat, attn_k base) — the 5/6-bit two-stream unpack makes r2's staging pricier.
        //   b2 same table with 8-row blocks: q4_K r2 when filled (-3..-22% all measured shapes),
        //     q5_K/q6_K r2 at waves >= 2 (27B lm_head -2.9%; 9B q6_K flat, harmless).
        let kq_r2 = matches!(qtype, QT_Q4_K | QT_Q5_K | QT_Q6_K);
        // MEMRA_KQ_BV=base|r2|r2w8 forces the k-quant variant WITHOUT touching the NVFP4 dispatch
        // (MEMRA_MMVQ_BV is global — an interleaved k-quant-only e2e A/B needs this narrower seam).
        static KQBV: std::sync::OnceLock<&'static str> = std::sync::OnceLock::new();
        let kq_bv = *KQBV.get_or_init(|| match std::env::var("MEMRA_KQ_BV").as_deref() {
            Ok("base") => "base",
            Ok("r2") => "r2",
            Ok("r2w8") => "r2w8",
            _ => "auto",
        });
        let variant: &'static str = if qtype == QT_Q4_0 {
            // Q4_0 r2 (gemma verify trunk, 2026-07-10): shared activation loads + the
            // row-independent ones-sum computed once per (col,group) for 2 rows. Same
            // fill rule as q4_K: r2 when the halved grid still fills the SMs.
            static Q40BV: std::sync::OnceLock<&'static str> = std::sync::OnceLock::new();
            let q40 = *Q40BV.get_or_init(|| match std::env::var("MEMRA_Q40_BV").as_deref() {
                // ms/sm/la = force-only measurement seams (ALL FLAT/NEGATIVE 2026-07-13,
                // never auto): m-split flat (nvcc keeps 72 regs); smem-slab −11% (staging
                // + syncs cost more than the stalls, bank-pad made no difference);
                // register load-ahead flat (nvcc already reorders). The b-tier limiter
                // is still unidentified — see the jsonl row.
                Ok("base") => "base",
                Ok("r2") => "r2",
                Ok("ms") => "ms",
                Ok("sm") => "sm",
                Ok("la") => "la",
                _ => "auto",
            });
            let v = if q40 != "auto" {
                q40
            } else if (out_f as u32).div_ceil(8) >= 4 * sms as u32 {
                "r2"
            } else {
                "base"
            };
            // split-plane mirror twins (2026-07-10): same fill rule, _rp names.
            // (m-split r2 pair twin PROBED FLAT 2026-07-13 — nvcc kept 72 regs either way
            // and the limiter is the per-column activation load chain (long_scoreboard
            // 42.5%), not occupancy; arm killed per doctrine, jsonl row is the record.)
            if rp {
                match v {
                    "ms" => "r2ms_rp",
                    "sm" => "r2sm_rp",
                    "la" => "r2la_rp",
                    "r2" => "r2_rp",
                    _ => "rp",
                }
            } else if matches!(v, "ms" | "sm" | "la") {
                "r2"
            } else {
                v
            }
        } else if qtype != QT_NVFP4 && !kq_r2 {
            "base"
        } else if kq_r2 && rp {
            // K-quant split-plane mirror (2026-08-01): only the plain _rp batched twins are
            // compiled for q4_K/q6_K — rp is a LAYOUT, it must survive every heuristic
            // (split-plane bytes through a GGUF-layout kernel = NaN). q5_K never mirrors.
            "rp"
        } else if kq_r2 {
            // k-quant r2w8 only exists at b4 (b2_r2 already 8-resident; b8 has no w8 twin) ->
            // mcols != 4 forced r2w8 falls to unbounded r2.
            if kq_bv != "auto" {
                if kq_bv == "r2w8" && mcols != 4 {
                    "r2"
                } else {
                    kq_bv
                }
            } else if bv != "auto" {
                match bv {
                    "r2" | "pfr2" | "rpr2" | "car2" => "r2",
                    "r2w8" | "rpr2w8" => {
                        if mcols != 4 {
                            "r2"
                        } else {
                            "r2w8"
                        }
                    }
                    _ => "base", // base/pf/ca/rp forced -> base (no such k-quant kernels)
                }
            } else {
                let blocks = (out_f + 7) / 8;
                let waves = blocks as f64 / (7 * sms as usize) as f64;
                let filled = blocks >= 4 * sms as usize;
                let use_r2 = if qtype == QT_Q4_K {
                    filled
                } else {
                    waves >= 2.0
                };
                if use_r2 { "r2" } else { "base" }
            }
        } else if bv != "auto" {
            // r2w8 only exists for b4/b8 (the b2_r2 kernel is already 8-blocks-resident at 60 regs).
            // ca/car2 need the alignment gate AND have no b8 twins; pfr2 has no b8 twin either —
            // unsupported (shape, mcols) combos fall back to pf/r2.
            // On rp buffers, forced legacy names map to their rp twins (layout law).
            let v = if bv == "r2w8" && mcols == 2 {
                "r2"
            } else if bv == "ca" && (!ca_ok || mcols == 8) {
                "pf"
            } else if bv == "car2" && (!ca_ok || mcols == 8) {
                "r2"
            } else if bv == "pfr2" && mcols == 8 {
                "r2"
            } else if (bv == "rpr2w8" || bv == "rpr2") && mcols == 2 {
                "rpr2"
            }
            // rpca* has no b8 twin (falls to rpr2w8/rpr2); needs the ca alignment gate.
            else if (bv == "rpca" || bv == "rpcar2") && (!ca_ok || mcols == 8) {
                if mcols == 8 { "rpr2w8" } else { "rpr2" }
            } else if bv == "rpcar2" && mcols == 2 {
                "rpca"
            }
            // rpsc/rpmsc/rpks* gate on smem-fit + alignment; fall to rpr2 outside it
            // (rpms has no smem and no alignment need — always valid on rp buffers).
            else if (bv == "rpsc" || bv == "rpmsc") && !sc_ok {
                "rpr2"
            } else if (bv == "rpks" || bv == "rpksc") && !ks_ok {
                "rpr2"
            } else {
                bv
            };
            if rp {
                match v {
                    "base" | "pf" | "ca" | "rp" => "rp",
                    "r2" | "pfr2" | "car2" | "rpr2" => "rpr2",
                    "r2w8" | "rpr2w8" => {
                        if mcols == 2 {
                            "rpr2"
                        } else {
                            "rpr2w8"
                        }
                    }
                    other => other, // rpca/rpcar2/rpsc/rpks/rpksc pass through (already rp-layout)
                }
            } else {
                v
            }
        } else if mcols == 8 {
            // b8 AUTO (2026-07-06 m-small latency arc, rtx6000 DRAM-cold rp msweep m=5/6/8 all five
            // 27B shapes): rpsc — the rpr2w8 schedule with the warp's scale rows prestaged to
            // smem, leaving ONE global dependency (the quant stream) in the k-loop at zero reg
            // growth. BIT-identical to rpr2w8 and wins or ties EVERY b8 cell: ffn_gate m5
            // 50.7->46.9 m8 64.1->57.1 (-11%), qkv m8 34.6->33.0, ssm_out m8 29.7->28.8,
            // attn_gate m8 26.9->26.1, ffn_down m5 58.2->56.9. The faster split-grid twins are
            // OUT: rpksc (k-split, ffn_down m5 -21%) broke run-spec self-consistency (k-reduce
            // order shifts verify argmax at tie margins — verify must stay bit-identical to the
            // m=1 decode chain); rpmsc (m-split, bit-identical) measured NEGATIVE everywhere
            // (twin warp's duplicated weight stream: ffn_down m5 85.7 vs 56.9).
            if rp {
                if sc_ok { "rpsc" } else { "rpr2w8" }
            } else {
                "r2w8"
            }
        } else if mcols >= 4 {
            // r2 runs 7 resident blocks/SM (67 regs); its __launch_bounds__(128,8) twin `r2w8`
            // (64 regs) runs 8. grid = ceil(out_f/8) for both. rp twins land in the same
            // residency classes (rp 44 regs ~ pf-class occupancy, rpr2 67, rpr2w8 64).
            let blocks = (out_f + 7) / 8;
            let r7 = 7 * sms as usize;
            let r8 = 8 * sms as usize;
            let waves = blocks as f64 / r7 as f64;
            let filled = blocks >= 4 * sms as usize;
            // 2026-07-06 m-small latency arc: b4 keeps the wave rule (rpms/rpmsc measured
            // flat-to-negative at m=3/4 on every shape — the m-split twin duplicates the weight
            // stream; rpsc b4 also negative on r2-class picks, ffn_down m4 51.1 vs 46.5).
            if filled && blocks.div_ceil(r8) < blocks.div_ceil(r7) {
                // the extra residency drops the INTEGER wave count -> the straggler wave a
                // latency-bound kernel pays in full disappears (ffn_down 1.11 -> 0.98 waves:
                // 112.5 -> 81.6us, beats pf 90.1; qkv 2.23 -> 1.95: 58.1 -> 51.1).
                if rp { "rpr2w8" } else { "r2w8" }
            } else if waves >= 2.0 || (waves <= 1.0 && filled) {
                // tail amortized (>=2 waves) or single wave: unbounded r2 (no reg-squeeze tax —
                // gate/up 81.1 vs 83.9 bounded, attn_q 61.0 vs 63.4).
                if rp { "rpr2" } else { "r2" }
            } else {
                // fractional straggler-wave window with no crossing, or grid too small to fill
                // the SMs (tiny out_f<=1024 shapes want max row-parallelism): prefetch variant
                // (rp = the r1 split-plane twin — measured the attn_gate winner, 35.4 vs pf 36.4).
                if rp { "rp" } else { "pf" }
            }
        } else if in_f >= 6144 {
            // b2 deep-k (2026-07-06): every new twin measured flat-to-negative here (rpms 44.1
            // vs rpr2 40.8 ffn_down; rpsc 43.6; the winning rpks is banned on k-order) — rpr2
            // stays.
            if rp { "rpr2" } else { "r2" }
        } else if rp {
            // b2 shallow-k: qkv (out_f=10240, 0.97 waves at 7-resident) is the one measured cell
            // where the r2-schedule scale-prestage twin beats the r1 rp pick (24.7 vs 28.9us
            // -15%); the wider (ffn_gate 1.65 waves) and smaller (attn_gate 0.58) shapes LOSE
            // (41.8 vs 38.2 / 16.6 vs 14.6) — gate on the single-wave window.
            let waves = ((out_f + 7) / 8) as f64 / (7 * sms as usize) as f64;
            if sc_ok && waves >= 0.9 && waves <= 1.1 {
                "rpsc"
            } else {
                "rp"
            }
        } else {
            "base"
        };
        variant
    }

    pub fn qmatvec_mmvq_batched(
        &self,
        bytes: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        qtype: i32,
        row_bytes: usize,
        mcols: usize,
        scale: f32,
        rp: bool,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        const ROWS_PER_BLOCK: u32 = 4;
        // TUNE SEAM (H100 lane): MEMRA_BVAR forces the batched-variant pick for the whole
        // process — the auto heuristics were tuned on sm_120 (82 SMs / 858 GB/s) and the
        // sm_90a re-tune sweeps this seam empirically. Layout variants stay safe: an rp
        // weight keeps its rp-layout kernel family regardless of the override.
        let forced: Option<&'static str> = {
            static V: std::sync::OnceLock<Option<String>> = std::sync::OnceLock::new();
            V.get_or_init(|| std::env::var("MEMRA_BVAR").ok())
                .as_deref()
                .map(|s| Box::leak(s.to_string().into_boxed_str()) as &'static str)
        };
        let variant = match forced {
            Some(v) if !rp || v.contains("rp") => v,
            _ => self.batched_variant(m, in_f, out_f, qtype, row_bytes, mcols, rp),
        };
        let base_name = Self::batched_kernel_name(qtype, mcols).ok_or_else(|| {
            format!("qmatvec_mmvq_batched: no kernel for qtype {qtype} mcols {mcols}")
        })?;
        // b16 tier (t=9..16 verify): only base/_rp b16 kernels are compiled — the b2..b8
        // per-shape perf variants (r2/pf/...) do not apply at this width. rp is a LAYOUT,
        // not a perf variant: it must survive (base kernel on split-plane bytes = NaN).
        let variant = if mcols == 16 {
            if rp { "rp" } else { "base" }
        } else {
            variant
        };
        // EXACT-WIDTH b5/b6/b7 twins (lane/vt-fixes fix 1, 2026-08-03): the b8 kernels
        // allocate acc[WROWS][8] at ANY m, so T=5..7 verify paid the full 8-wide register
        // tax — the measured T=4->5 cliff. The same template at MCOLS=m runs the identical
        // per-(token,row) chain (columns c >= m never execute in either form) ->
        // BIT-IDENTICAL to the b8 launch. NVFP4 split-plane only (the sm_120 default trunk);
        // covers both b8 auto schedules (rpsc, rpr2w8). MEMRA_B567=0 rollback.
        static B567: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let b567 = *B567.get_or_init(|| std::env::var("MEMRA_B567").as_deref() != Ok("0"));
        if b567
            && qtype == QT_NVFP4
            && rp
            && mcols == 8
            && (5..=7).contains(&m)
            && matches!(variant, "rpsc" | "rpr2w8")
        {
            let f = self.func(&format!("qmatvec_nvfp4_mmvq_b{m}_{variant}"));
            let rows_per_block = ROWS_PER_BLOCK * 2; // r2-class schedules: 2 rows/warp
            let mut y = self.alloc_uninit::<f32>(m * out_f)?;
            let cfg = LaunchConfig {
                grid_dim: ((out_f as u32 + rows_per_block - 1) / rows_per_block, 1, 1),
                block_dim: (32, ROWS_PER_BLOCK, 1),
                shared_mem_bytes: 0,
            };
            let (inf, outf, mi, rb) = (in_f as i32, out_f as i32, m as i32, row_bytes as i64);
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(bytes)
                .arg(aq)
                .arg(ad)
                .arg(&mut y)
                .arg(&inf)
                .arg(&outf)
                .arg(&mi)
                .arg(&rb);
            unsafe {
                b.launch(cfg)?;
            }
            if scale != 1.0 {
                self.scale_inplace(&mut y, scale, m * out_f)?;
            }
            return Ok(y);
        }
        let (name, rows_per_block): (std::borrow::Cow<'static, str>, u32) = match variant {
            "base" => (base_name.into(), ROWS_PER_BLOCK),
            "pf" => (format!("{base_name}_pf").into(), ROWS_PER_BLOCK),
            "ca" => (format!("{base_name}_ca").into(), ROWS_PER_BLOCK),
            "rp" => (format!("{base_name}_rp").into(), ROWS_PER_BLOCK),
            "rpca" => (format!("{base_name}_rpca").into(), ROWS_PER_BLOCK), // 1 row/warp cp.async
            // split families: 2 warp-pairs x 2 rows = 4 rows/block (the k-range or column set
            // splits across the pair's two warps; grid.x doubles vs rpr2 at the same regs).
            "rpks" => (format!("{base_name}_rpks").into(), ROWS_PER_BLOCK),
            "rpksc" => (format!("{base_name}_rpksc").into(), ROWS_PER_BLOCK),
            "rpms" => (format!("{base_name}_rpms").into(), ROWS_PER_BLOCK),
            "rpmsc" => (format!("{base_name}_rpmsc").into(), ROWS_PER_BLOCK),
            "r2ms_rp" => (format!("{base_name}_r2ms_rp").into(), ROWS_PER_BLOCK),
            "r2sm_rp" => (format!("{base_name}_r2sm_rp").into(), ROWS_PER_BLOCK * 2),
            "r2la_rp" => (format!("{base_name}_r2la_rp").into(), ROWS_PER_BLOCK * 2),
            v => (format!("{base_name}_{v}").into(), ROWS_PER_BLOCK * 2), // r2-class: 2 rows/warp
        };
        debug_assert!(
            !rp || name.contains("_rp"),
            "rp weight dispatched to a GGUF-layout kernel"
        );
        let f = self.func(&name);
        let mut y = self.alloc_uninit::<f32>(m * out_f)?;
        // r2sm_rp: [MCOLS][32 blk][8 int] activation slab + [MCOLS][32] f32 scales.
        let smem = if name.contains("_r2sm_rp") {
            (mcols * 32 * 9 * 4 + mcols * 32 * 4) as u32
        } else {
            0
        };
        let cfg = LaunchConfig {
            grid_dim: ((out_f as u32 + rows_per_block - 1) / rows_per_block, 1, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: smem,
        };
        let (inf, outf, mi, rb) = (in_f as i32, out_f as i32, m as i32, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(bytes)
            .arg(aq)
            .arg(ad)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        if scale != 1.0 {
            self.scale_inplace(&mut y, scale, m * out_f)?;
        }
        Ok(y)
    }

    /// BATCHED weight-tile-resident matvec from raw weight bytes (quantizes the f32 activation `x` to
    /// q8_1 internally; macro-scale NOT applied — caller compares bare, like qmatvec_*_fast). For the
    /// kernel_check bit-equivalence gate. `mcols` ∈ {2,4,8}. Works for Q8_0/Q4_K/Q5_K/Q6_K/NVFP4.
    pub fn qmatvec_batched_raw(
        &self,
        bytes: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        qtype: i32,
        row_bytes: usize,
        mcols: usize,
        rp: bool,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
        self.qmatvec_mmvq_batched(
            bytes, &aq, &ad, m, in_f, out_f, qtype, row_bytes, mcols, 1.0, rp,
        )
    }

    /// Back-compat NVFP4-only batched raw launcher (used by older gates). Delegates to the generic one.
    pub fn qmatvec_nvfp4_batched_raw(
        &self,
        bytes: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
        mcols: usize,
        rp: bool,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        self.qmatvec_batched_raw(bytes, x, m, in_f, out_f, QT_NVFP4, row_bytes, mcols, rp)
    }

    /// Stage-C FP4 gate (MEMRA_FP4): if `w` is an NVFP4 weight with in_f%64==0, run the native mxf4
    /// block-scale GEMM and apply the per-tensor macro-scale, returning Some(y). Else None (caller
    /// falls through to the int8 GEMM / dp4a). Strict opt-in over the proven int8 path; m>=16 only.
    fn try_fp4_gemm(
        &self,
        w: &crate::model::GpuTensor,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
    ) -> Result<Option<CudaSlice<f32>>, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        if cfg!(memra_portable_cuda) {
            return Ok(None);
        }
        // MEMRA_FP4 reaches qmatvec_gemm_nvfp4_fp4, which cu/qmatvec_gemm.cu:1234 omits on a
        // portable build (the mxf4 block-scale MMA is sm_120a-only). Refuse at the door.
        if std::env::var("MEMRA_FP4").is_ok() {
            refuse_portable_force("MEMRA_FP4", "the sm_120a mxf4 block-scale MMA");
        }
        if std::env::var("MEMRA_FP4").is_err() {
            return Ok(None);
        }
        // CUTLASS prefill branch (m>=128 + MEMRA_FP4_CUTLASS + a repacked CutlassWeight present): route
        // to the CUTLASS sm120 NVFP4 GEMM, folding the per-tensor macro-scale into the epilogue alpha
        // (1/scale) — no post-matmul scale_inplace. Decode (m<128) and the m∈[16,128) middle band keep
        // the hand-roll below: CUTLASS's 128-row M-tile wastes work under 128.
        // The hand-roll applies the per-tensor macro-scale as a POST-matmul MULTIPLY (scale_inplace(y,
        // scale)); CUTLASS's epilogue does D = alpha * (A@B^T), so alpha == scale reproduces it exactly
        // (NOT 1/scale — the plan sketch had this inverted; the kernel_check arm gates it). scale==1.0
        // for the common no-macro-scale case.
        #[cfg(memra_cutlass)]
        if m >= 128 && std::env::var("MEMRA_FP4_CUTLASS").is_ok() {
            if let GpuTensor::Quant {
                bytes,
                qtype,
                scale,
                row_bytes,
                cutlass,
                ..
            } = w
            {
                if *qtype == QT_NVFP4 && in_f % 64 == 0 {
                    if let Some(cw) = cutlass {
                        // Resident fast path: load-time-repacked B + swizzled SFB (no per-call repack).
                        let y = self.cutlass_fp4_gemm(
                            &cw.b_packed,
                            &cw.sfb_swizzled,
                            x,
                            *scale,
                            m,
                            out_f,
                            in_f,
                        )?;
                        return Ok(Some(y));
                    } else if std::env::var("MEMRA_FP4_CUTLASS_OTF").is_ok() {
                        // On-the-fly repack (MEMRA_FP4_CUTLASS_OTF): de-interleave + swizzle the B operand
                        // from raw bytes per prefill call. No resident doubling of the NVFP4 weight VRAM
                        // (the load-time repack ~doubles it) — needed for models that don't fit the
                        // resident path (e.g. the 27B on 24GB). Slower (per-call repack) but argmax-exact.
                        let (b_packed, sfb_sw) =
                            self.build_cutlass_weight(bytes, out_f, in_f, *row_bytes)?;
                        let y =
                            self.cutlass_fp4_gemm(&b_packed, &sfb_sw, x, *scale, m, out_f, in_f)?;
                        return Ok(Some(y));
                    }
                }
            }
        }
        if let GpuTensor::Quant {
            bytes,
            qtype,
            row_bytes,
            scale,
            rp,
            ..
        } = w
        {
            // A6: the hand-rolled W4A4 mxf4 GEMM reads 36B GGUF blocks — no rp port (MEMRA_FP4 is
            // an opt-in accuracy tradeoff); repacked tensors fall through to the int8 GEMM.
            if *qtype == QT_NVFP4 && in_f % 64 == 0 && !*rp {
                let y =
                    self.qmatvec_gemm_nvfp4_fp4(bytes, x, m, in_f, out_f, *row_bytes, *scale)?;
                return Ok(Some(y));
            }
        }
        Ok(None)
    }

    /// rms_norm + fused fp16 twin (task #14): f32 output verbatim `rms_norm` + the fp16
    /// copy the f16-mirror GEMM group would otherwise produce with a standalone convert
    /// launch. BIT-IDENTICAL end-to-end (same reduction, same __float2half values).
    pub fn rms_norm_f16out(
        &self,
        x: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        dst16: &mut CudaSlice<u8>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("rms_norm_f16out_f32");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, e) = (ncols as i32, eps);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(w).arg(dst).arg(dst16).arg(&nc).arg(&e);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// add+norm(+f16out) fusion for the prefill trunk (round 28; add_rms_norm precedent —
    /// bit-identical to add_f32 -> rms_norm_f16out). block_dim matches rms_norm_f16out's.
    #[allow(clippy::too_many_arguments)]
    pub fn add_rms_norm_f16out(
        &self,
        a: &CudaSlice<f32>,
        b: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        res: &mut CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        dst16: &mut CudaSlice<u8>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("add_rms_norm_f16out_f32");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (rms_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, e) = (ncols as i32, eps);
        let __s_lb = self.gpu.stream();
        let mut lb = __s_lb.launch_builder(&f);
        lb.arg(a)
            .arg(b)
            .arg(w)
            .arg(res)
            .arg(dst)
            .arg(dst16)
            .arg(&nc)
            .arg(&e);
        unsafe {
            lb.launch(cfg)?;
        }
        Ok(())
    }

    /// matmul_group with a PRE-EMITTED fp16 activation (task #14: the producer norm fused
    /// the convert). Mirror-less members fall back to `matmul` on the f32 activation.
    pub fn matmul_group_xh(
        &self,
        ws: &[&crate::model::GpuTensor],
        x: &CudaSlice<f32>,
        xh: &CudaSlice<u8>,
        m: usize,
    ) -> Result<Vec<CudaSlice<f32>>, Box<dyn std::error::Error>> {
        let mut out = Vec::with_capacity(ws.len());
        let in_f = ws[0].in_features();
        for w in ws {
            if w.in_features() == in_f && m >= 16 && !self.verify_exact_on() {
                if let Some(y) = self.try_f16_gemm_pre(w, xh, m)? {
                    out.push(y);
                    continue;
                }
            }
            out.push(self.matmul(w, x, m)?);
        }
        Ok(out)
    }

    /// task #14 pad-proofing: zero beta/g_log at rows >= len_d[0] (pads become identity
    /// GDN steps). Layouts [T, H].
    pub fn gdn_pad_mask(
        &self,
        beta: &mut CudaSlice<f32>,
        g_log: &mut CudaSlice<f32>,
        len_d: &CudaSlice<i32>,
        h: usize,
        t: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gdn_pad_mask_f32");
        let cfg = LaunchConfig::for_num_elems((t * h) as u32);
        let (hi, ti) = (h as i32, t as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(beta).arg(g_log).arg(len_d).arg(&hi).arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// task #14 pad-proofing: dst[ncols] = src row (len_d[0]-1) — device-indexed last-row
    /// gather for the padded prime graph's h_seed/hlast.
    pub fn row_gather_dev(
        &self,
        src: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        len_d: &CudaSlice<i32>,
        ncols: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("row_gather_dev_f32");
        let cfg = LaunchConfig::for_num_elems(ncols as u32);
        let nc = ncols as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(src).arg(dst).arg(len_d).arg(&nc);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Grouped matmul: several weights consuming ONE activation (hybrid layers: the GDN
    /// 4-tuple wqkv/gate/beta/alpha, attention q/k/v, ffn gate/up). Semantics identical to
    /// calling `matmul` per weight; the f16-mirror arm converts the activation ONCE for the
    /// whole group instead of once per GEMM (the standalone converts were ~250 launches/prime
    /// of small-kernel gap fuel — nsys 2026-07-26). Any member without a mirror (or with a
    /// different in_f) falls back to its own `matmul` — behavior unchanged.
    pub fn matmul_group(
        &self,
        ws: &[&crate::model::GpuTensor],
        x: &CudaSlice<f32>,
        m: usize,
    ) -> Result<Vec<CudaSlice<f32>>, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        let mut out = Vec::with_capacity(ws.len());
        let any_mirror = ws
            .iter()
            .any(|w| matches!(w, GpuTensor::Quant { f16: Some(_), .. }));
        if m >= 16 && any_mirror && !self.verify_exact_on() {
            let in_f = ws[0].in_features();
            let xh = self.f16_act(x, m * in_f, in_f)?;
            for w in ws {
                if w.in_features() == in_f {
                    if let Some(y) = self.try_f16_gemm_pre(w, &xh, m)? {
                        out.push(y);
                        continue;
                    }
                }
                out.push(self.matmul(w, x, m)?);
            }
            return Ok(out);
        }
        for w in ws {
            out.push(self.matmul(w, x, m)?);
        }
        Ok(out)
    }

    /// Cross-request grouped matmul (task #13): run ONE projection group over the
    /// CONCATENATION of several sequences' activations (m = sum of per-seq rows — the
    /// GEMM-batch win vLLM gets from continuous batching), then split each output back
    /// into per-seq buffers. Zero view plumbing: gather/scatter are stream-ordered D2D
    /// copies (~us at prime sizes). NUMERIC CONFIG NOTE: a GEMM at m=sum tiles K
    /// differently than per-seq GEMMs — argmax-gated like every prefill GEMM change.
    pub fn matmul_group_multi(
        &self,
        ws: &[&crate::model::GpuTensor],
        xs: &[&CudaSlice<f32>],
        ms: &[usize],
    ) -> Result<Vec<Vec<CudaSlice<f32>>>, Box<dyn std::error::Error>> {
        assert_eq!(xs.len(), ms.len());
        let in_f = ws[0].in_features();
        let total: usize = ms.iter().sum();
        let mut xcat = self.uninit(total * in_f)?;
        let mut off = 0usize;
        for (x, &m) in xs.iter().zip(ms) {
            self.copy_into(&mut xcat, off * in_f, x, m * in_f)?;
            off += m;
        }
        let ys = self.matmul_group(ws, &xcat, total)?;
        let mut out: Vec<Vec<CudaSlice<f32>>> = (0..xs.len()).map(|_| Vec::new()).collect();
        for (w, y) in ws.iter().zip(ys) {
            let out_f = w.out_features();
            let mut off = 0usize;
            for (s, &m) in ms.iter().enumerate() {
                let mut ys_s = self.uninit(m * out_f)?;
                let src = y.slice(off * out_f..(off + m) * out_f);
                self.gpu.stream().memcpy_dtod(&src, &mut ys_s)?;
                out[s].push(ys_s);
                off += m;
            }
        }
        Ok(out)
    }

    /// True if `w`'s qtype has a batched tensor-core GEMM kernel (the prefill T>1 root fix).
    /// Only the 4 daily-hot dtypes: Q8_0, Q4_K, Q6_K, NVFP4. NVFP4 needs in_f % 64 == 0.
    /// DEFAULT-ON (2026-06-28): measured pp512 9B-NVFP4 = 1413 tok/s WITH this GEMM vs 298 with the
    /// dp4a fallback (4.7x) AND MORE accurate (prefill logit maxdiff 0.159 vs dp4a 0.55, both argmax
    /// MATCH). The int8 tensor-core GEMM is unconditional (its historical MEMRA_GEMM opt-in gate
    /// shipped with Phase 0 — mma + smem swizzle + cp.async — and was removed). Prefill-only
    /// (m>=GEMM_M_THRESHOLD); m=1 decode keeps dp4a/MMVQ (this returns true but matmul only calls it
    /// at m>=threshold). Portable CUDA targets always use the correctness fallback; on sm_120a,
    /// MEMRA_NO_GEMM forces that same dp4a fallback (the bit-reference).
    pub fn gemm_supports(&self, w: &crate::model::GpuTensor) -> bool {
        use crate::model::GpuTensor;
        if !legacy_quant_gemm_allowed(
            cfg!(memra_portable_cuda),
            cfg!(memra_hopper_mma),
            std::env::var_os("MEMRA_NO_GEMM").is_some(),
        ) {
            return false;
        }
        match w {
            GpuTensor::Quant { qtype, .. } => {
                matches!(*qtype, QT_Q8_0 | QT_Q4_K | QT_Q6_K | QT_Q5_K | QT_Q4_0)
                    || (*qtype == QT_NVFP4 && w.in_features() % 64 == 0)
            }
            GpuTensor::Float { .. } | GpuTensor::FloatBf16 { .. } => false,
        }
    }

    /// Batched tensor-core int8 GEMM with a PRE-QUANTIZED q8_1 activation (aq,ad). The prefill
    /// (T>1) root fix: decode each weight 32-block to int8 in shared memory ONCE per (row-tile,
    /// K-step) and reuse it across all BN tokens via mma.sync.m16n8k32.s8 — amortizing the weight
    /// read/decode N-fold (vs the dp4a matvec's per-token re-read). s32 accumulate is exact vs
    /// dp4a; only the final f32 block-scale rounding differs. Caller MUST have checked
    /// `gemm_supports(w)`. y[m,out] token-major. NVFP4 per-tensor macro-scale applied post.
    pub fn qmatvec_gemm(
        &self,
        w: &crate::model::GpuTensor,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        use crate::model::GpuTensor;
        let in_f = w.in_features();
        let out_f = w.out_features();
        let (bytes, qtype, row_bytes, scale, rp) = match w {
            GpuTensor::Quant {
                bytes,
                qtype,
                row_bytes,
                scale,
                rp,
                ..
            } => (bytes, *qtype, *row_bytes, *scale, *rp),
            _ => unreachable!("gemm_supports guaranteed Quant"),
        };
        // wgmma arm (sm_90a, task 8): the m64n64k32 warpgroup kernel reads the rp4 split-plane
        // mirror AS-IS (qplane rows = its A operand, the half dplane its scales) and the same
        // (aq, ad) activation planes. Same numeric class as the mma kernel below (exact s32 per
        // 32-block, one f32 scale fold per block, ascending K) — argmax/tolerance gated like
        // every prefill GEMM, not bit-gated. MEMRA_WGMMA=0 restores the portable kernel.
        if cfg!(memra_hopper_mma) && qtype == QT_Q8_0 && out_f % 64 == 0 && wgmma_gemm_enabled() {
            if let GpuTensor::Quant { rp4: Some(m4), .. } = w {
                let mut y = self.qmatvec_gemm_q8_0_wgmma_raw(m4, aq, ad, m, in_f, out_f)?;
                if scale != 1.0 {
                    self.scale_inplace(&mut y, scale, m * out_f)?;
                }
                return Ok(y);
            }
        }
        let name = match qtype {
            QT_Q8_0 => "qmatvec_gemm_q8_0",
            QT_Q4_K => "qmatvec_gemm_q4_K",
            QT_Q4_0 => {
                if rp {
                    "qmatvec_gemm_q4_0_rp"
                } else {
                    "qmatvec_gemm_q4_0"
                }
            }
            QT_Q5_K => "qmatvec_gemm_q5_K",
            QT_Q6_K => "qmatvec_gemm_q6_K",
            QT_NVFP4 => {
                if rp {
                    "qmatvec_gemm_nvfp4_rp"
                } else {
                    "qmatvec_gemm_nvfp4"
                }
            }
            _ => unreachable!(),
        };
        let f = self.func(name);
        let mut y = self.alloc_uninit::<f32>(m * out_f)?; // full-overwrite GEMM output: skip memset
        // CTA tile MUST match the .cu per-kernel tile. MMQ-PORT: kernel1 (Q8_0/Q4_K/Q5_K) runs llama's
        // 128x128 SQUARE tile (K1_BM=128 x K1_BN=128, 8 warps); kernel2 (Q6_K/NVFP4) keeps 64x256, 4 warps
        // (the macro BM/BN in the .cu). Grid dims are selected by qtype so each launches its own tile.
        let is_k1 = matches!(qtype, QT_Q8_0 | QT_Q4_K | QT_Q5_K | QT_Q4_0);
        // TUNE SEAM: MEMRA_GEMM_K1_LAUNCH overrides kernel1's launch tile to match a -D-swept fatbin.
        let k1_tile = if is_k1 {
            k1_launch_override().unwrap_or((128, 128, 8))
        } else {
            (128, 128, 8)
        };
        let (bm, bn): (u32, u32) = if is_k1 {
            (k1_tile.0, k1_tile.1)
        } else {
            (64, 256)
        };
        let warps: u32 = if is_k1 {
            k1_tile.2
        } else {
            match qtype {
                QT_NVFP4 => 8,
                _ => 4,
            }
        };
        let cfg = LaunchConfig {
            grid_dim: ((out_f as u32 + bm - 1) / bm, (m as u32 + bn - 1) / bn, 1),
            block_dim: (32, warps, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi, rb) = (in_f as i32, out_f as i32, m as i32, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(bytes)
            .arg(aq)
            .arg(ad)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        if scale != 1.0 {
            self.scale_inplace(&mut y, scale, m * out_f)?;
        }
        Ok(y)
    }

    /// Test entry: run the GEMM directly from raw weight bytes + qtype (no GpuTensor). Quantizes
    /// the f32 activation `x` to q8_1 internally then launches the tensor-core GEMM. NVFP4 per-tensor
    /// macro-scale is NOT applied here (caller passes it separately, like the dp4a path). Used by
    /// kernel_check for the bit-equivalence gate vs qmatvec_*_dp4a.
    pub fn qmatvec_gemm_raw(
        &self,
        bytes: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        qtype: i32,
        row_bytes: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let (aq, ad) = self.quantize_q8_1(x, m, in_f)?;
        let name = match qtype {
            QT_Q8_0 => "qmatvec_gemm_q8_0",
            QT_Q4_K => "qmatvec_gemm_q4_K",
            QT_Q4_0 => "qmatvec_gemm_q4_0",
            QT_Q5_K => "qmatvec_gemm_q5_K",
            QT_Q6_K => "qmatvec_gemm_q6_K",
            QT_NVFP4 => "qmatvec_gemm_nvfp4",
            QT_NVFP4_RP => "qmatvec_gemm_nvfp4_rp",
            _ => panic!("qmatvec_gemm_raw: qtype {qtype} has no GEMM kernel"),
        };
        let f = self.func(name);
        let mut y = self.alloc_uninit::<f32>(m * out_f)?; // full-overwrite GEMM output: skip memset
        // MMQ-PORT: kernel1 (Q8_0/Q4_K/Q5_K) = llama 128x128 tile, 8 warps; kernel2 (Q6_K/NVFP4) = 64x256,
        // 4/8 warps. Grid tile per qtype (must match the .cu K1_BM/K1_BN vs BM/BN). KEEP IN SYNC w/ qmatvec_gemm.
        let is_k1 = matches!(qtype, QT_Q8_0 | QT_Q4_K | QT_Q5_K | QT_Q4_0);
        // TUNE SEAM: MEMRA_GEMM_K1_LAUNCH overrides kernel1's launch tile to match a -D-swept fatbin.
        let k1_tile = if is_k1 {
            k1_launch_override().unwrap_or((128, 128, 8))
        } else {
            (128, 128, 8)
        };
        let (bm, bn): (u32, u32) = if is_k1 {
            (k1_tile.0, k1_tile.1)
        } else {
            (64, 256)
        };
        let warps: u32 = if is_k1 {
            k1_tile.2
        } else {
            match qtype {
                QT_NVFP4 | QT_NVFP4_RP => 8,
                _ => 4,
            }
        };
        let cfg = LaunchConfig {
            grid_dim: ((out_f as u32 + bm - 1) / bm, (m as u32 + bn - 1) / bn, 1),
            block_dim: (32, warps, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi, rb) = (in_f as i32, out_f as i32, m as i32, row_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(bytes)
            .arg(&aq)
            .arg(&ad)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi)
            .arg(&rb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(y)
    }

    /// H100 warpgroup GEMM raw entry (task 8): launch `qmatvec_gemm_q8_0_wgmma` on an rp4
    /// split-plane mirror + pre-quantized (aq, ad) activation planes. One warpgroup (128 thr)
    /// owns a 64x64 C tile; grid (out_f/64, ceil(m/64)). out_f % 64 == 0 REQUIRED (row loads
    /// and dplane scale reads are unguarded); the token edge is guarded in-kernel.
    /// Standalone harness verdict (tools/bench_q8_gemm_wgmma.cu, 4096x4096x512): rel 1.6e-05
    /// vs CPU ref, 179us vs the portable mma kernel's 688us (3.84x, unpipelined).
    pub fn qmatvec_gemm_q8_0_wgmma_raw(
        &self,
        rp4: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        assert!(
            out_f % 64 == 0 && in_f % 32 == 0,
            "wgmma GEMM needs out_f%64==0, in_f%32==0"
        );
        let f = self.func("qmatvec_gemm_q8_0_wgmma");
        let mut y = self.alloc_uninit::<f32>(m * out_f)?; // full-overwrite GEMM output
        let cfg = LaunchConfig {
            grid_dim: ((out_f / 64) as u32, (m as u32).div_ceil(64), 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, mi) = (in_f as i32, out_f as i32, m as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(rp4)
            .arg(aq)
            .arg(ad)
            .arg(&mut y)
            .arg(&inf)
            .arg(&outf)
            .arg(&mi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(y)
    }

    /// y[i] *= s. NVFP4 per-tensor macro-scale broadcast over the whole output.
    pub fn scale_inplace(
        &self,
        y: &mut CudaSlice<f32>,
        s: f32,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("scale_f32");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let (sf, ni) = (s, n as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(y).arg(&sf).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// MEMRA_FULL_PREC dequant-on-use: expand a bf16-resident weight (`GpuTensor::FloatBf16`, raw
    /// bf16 bytes) to a transient f32 scratch of `n` elements, which then feeds the existing f32
    /// cuBLASLt GEMV. The scratch is freed when the caller drops it, so peak VRAM = resident bf16
    /// weights + ONE (largest) weight's f32 expansion + activations. SLOW IS FINE (research mode).
    pub fn bf16_to_f32(
        &self,
        data: &cudarc::driver::CudaView<'_, u8>,
        n: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let mut out = self.alloc_uninit::<f32>(n)?;
        let f = self.func("bf16_to_f32");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let ni = n as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(data).arg(&mut out).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(out)
    }

    /// Chunked bf16 linear (MEMRA_FULL_PREC): y[m,out] = x @ W_bf16^T with the f32 dequant scratch
    /// bounded to CHUNK_ROWS rows (256MB at in_f=4096) instead of the whole weight — the 4GB
    /// lm_head expansion OOM'd the 24GB budget. Row-chunking partitions OUTPUT rows; each row's
    /// dot is computed by the identical kernel on identical bytes, so per-(token,row) results are
    /// bit-identical to the unchunked form. `exact` selects linear_decode_exact (per-column m=1
    /// calls, the spec-verify contract) vs plain linear.
    fn linear_bf16_chunked(
        &self,
        x: &CudaSlice<f32>,
        data: &CudaSlice<u8>,
        m: usize,
        in_f: usize,
        out_f: usize,
        exact: bool,
        canonical_chunk_rows: Option<usize>,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        // MEMRA_STEP_TP_TIMING=1: cumulative cost of the per-call BF16->F32 expansion class
        // (alloc + convert kernel + f32 cuBLASLt = ~5x weight traffic). Prints every 1024 calls.
        // The stream sync per call perturbs wall; diagnostic only, never in a receipts run.
        static EXP_NS: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(0);
        static EXP_CALLS: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(0);
        static EXP_WBYTES: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(0);
        let timing = std::env::var("MEMRA_STEP_TP_TIMING").as_deref() == Ok("1");
        let started = timing.then(std::time::Instant::now);
        let result =
            self.linear_bf16_chunked_inner(x, data, m, in_f, out_f, exact, canonical_chunk_rows);
        if let Some(started) = started {
            use std::sync::atomic::Ordering;
            self.stream().synchronize()?;
            let ns = EXP_NS.fetch_add(started.elapsed().as_nanos() as u64, Ordering::Relaxed)
                + started.elapsed().as_nanos() as u64;
            let wb = EXP_WBYTES.fetch_add((in_f * out_f * 2) as u64, Ordering::Relaxed)
                + (in_f * out_f * 2) as u64;
            let calls = EXP_CALLS.fetch_add(1, Ordering::Relaxed) + 1;
            if calls % 1024 == 0 {
                eprintln!(
                    "[bf16-expand-timing] calls={calls} total_ms={:.1} avg_us={:.1} \
                     weight_gb={:.2}",
                    ns as f64 / 1.0e6,
                    ns as f64 / calls as f64 / 1.0e3,
                    wb as f64 / 1.0e9,
                );
            }
        }
        result
    }

    /// MEMRA_BF16_MMV=1: decode-time (m=1) BF16-resident matvec door. Numeric class change vs
    /// the expansion path (single-kernel deterministic tree reduce vs f32 cuBLASLt), so it is
    /// default OFF and gated by the run-gen argmax gate + boot battery like the other
    /// numeric-class doors (DEV_ROUTES precedent).
    pub(crate) fn bf16_mmv_on() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| std::env::var("MEMRA_BF16_MMV").as_deref() == Ok("1"))
    }

    /// One-block-per-row BF16 matvec: y[out_f] = W_bf16[out_f, in_f] @ x[in_f], f32 accumulate.
    /// Weights read once as raw bf16 (same bits<<16 expansion contract as `deq`'s QT_BF16 arm).
    fn matvec_bf16(
        &self,
        data: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        in_f: usize,
        out_f: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        if data.len() != in_f * out_f * 2 || x.len() < in_f || in_f % 8 != 0 {
            return Err(format!(
                "matvec_bf16 geometry bytes={} x={} in={in_f} out={out_f}",
                data.len(),
                x.len()
            )
            .into());
        }
        let mut y = self.alloc_uninit::<f32>(out_f)?;
        let f = self.func("matvec_bf16_f32acc");
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, 1, 1),
            block_dim: (mmv_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let ini = in_f as i32;
        let __s_bld = self.gpu.stream();
        let mut bld = __s_bld.launch_builder(&f);
        bld.arg(data).arg(x).arg(&mut y).arg(&ini);
        unsafe {
            bld.launch(cfg)?;
        }
        Ok(y)
    }

    /// Fused QK rms-norm + neox rope (t=1): one launch per rank replaces two rms_norm
    /// launches, a position upload, and the rope launch; the position is read directly from
    /// the caller's device counter (UVA). Numeric-class door (see the kernel doc).
    #[allow(clippy::too_many_arguments)]
    /// FUSION #1: qk norms + rope + dcw KV append + last-block len inc, one launch
    /// (replaces qk_norm_rope_into + append_kv_quantized_dcw + inc_i32 on the dcw path).
    /// Bit-identical to the split kernels; requires head_dim == 128 and
    /// kv_dim_v == kv_dim_k == nh_k * head_dim (caller-guarded fallback otherwise).
    #[allow(clippy::too_many_arguments)]
    /// T-ROW twin of `qk_norm_rope_append_inc_dcw` over a per-row session table (six u64
    /// words per row: K plane, V plane, len_ptr, base_ptr, done_ctr, pos_ptr). Raw q/k/v
    /// come from the [t, dim] tcol slabs; roped q lands in the [t, nh_q*head_dim] slab.
    /// Per-(row, head) block program == the t=1 kernel — bit-identical per row.
    #[allow(clippy::too_many_arguments)]
    pub fn qk_norm_rope_append_inc_dcw_rows(
        &self,
        q_raw_t: &CudaSlice<f32>,
        k_raw_t: &CudaSlice<f32>,
        v_raw_t: &CudaSlice<f32>,
        qw: &CudaSlice<f32>,
        kw: &CudaSlice<f32>,
        q_out_t: &mut CudaSlice<f32>,
        k_out_t: &mut CudaSlice<f32>,
        tab: &CudaSlice<u64>,
        pos_t: &CudaSlice<i32>,
        same_session: bool,
        t: usize,
        kv_dim_k: usize,
        kv_dim_v: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        head_dim: usize,
        n_dims: usize,
        nh_q: usize,
        nh_k: usize,
        eps: f32,
        freq_base: f32,
        freq_scale: f32,
        ff: Option<&CudaSlice<f32>>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if head_dim != 128
            || kv_dim_v != kv_dim_k
            || kv_dim_k != nh_k * head_dim
            || t == 0
            || t > 32
            || tab.len() < t * 6
            || pos_t.len() < t
            || q_raw_t.len() < t * nh_q * head_dim
            || k_raw_t.len() < t * nh_k * head_dim
            || v_raw_t.len() < t * kv_dim_v
            || q_out_t.len() < t * nh_q * head_dim
            || k_out_t.len() < t * nh_k * head_dim
        {
            return Err(format!(
                "qk_norm_rope_append_inc_rows geometry head_dim={head_dim} t={t} \
                 nh_q={nh_q} nh_k={nh_k}"
            )
            .into());
        }
        let f = self.func("qk_norm_rope_append_inc_dcw_rows");
        let same_t: i32 = if same_session { t as i32 } else { 0 };
        let theta_scale = (freq_base).powf(-2.0 / n_dims as f32);
        let cfg = LaunchConfig {
            grid_dim: ((nh_q + nh_k) as u32, 1, t as u32),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (kvk, kvv) = (kv_dim_k as i32, kv_dim_v as i32);
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let (hd, nd, nq, nk) = (head_dim as i32, n_dims as i32, nh_q as i32, nh_k as i32);
        let null: u64 = 0;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q_raw_t)
            .arg(k_raw_t)
            .arg(v_raw_t)
            .arg(qw)
            .arg(kw)
            .arg(q_out_t)
            .arg(k_out_t)
            .arg(tab)
            .arg(pos_t)
            .arg(&same_t)
            .arg(&kvk)
            .arg(&kvv)
            .arg(&ktb)
            .arg(&vtb)
            .arg(&hd)
            .arg(&nd)
            .arg(&nq)
            .arg(&nk)
            .arg(&eps)
            .arg(&theta_scale)
            .arg(&freq_scale);
        match ff {
            Some(freqs) => {
                b.arg(freqs);
            }
            None => {
                b.arg(&null);
            }
        }
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn qk_norm_rope_append_inc_dcw(
        &self,
        q_raw: &CudaSlice<f32>,
        k_raw: &CudaSlice<f32>,
        v_raw: &CudaSlice<f32>,
        qw: &CudaSlice<f32>,
        kw: &CudaSlice<f32>,
        q_out: &mut CudaSlice<f32>,
        k_out: &mut CudaSlice<f32>,
        pos: &CudaSlice<i32>,
        k_plane: &mut CudaSlice<u8>,
        v_plane: &mut CudaSlice<u8>,
        // Shared ref by the planes_and_counters_mut split-borrow contract; the kernel is the
        // (single) writer, exactly like the split append+inc pair it replaces.
        len_dev: &CudaSlice<i32>,
        base_dev: Option<&CudaSlice<i32>>,
        done_ctr: &mut CudaSlice<u32>,
        kv_dim_k: usize,
        kv_dim_v: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        head_dim: usize,
        n_dims: usize,
        nh_q: usize,
        nh_k: usize,
        eps: f32,
        freq_base: f32,
        freq_scale: f32,
        ff: Option<&CudaSlice<f32>>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if head_dim != 128
            || kv_dim_v != kv_dim_k
            || kv_dim_k != nh_k * head_dim
            || q_raw.len() < nh_q * head_dim
            || k_raw.len() < nh_k * head_dim
            || v_raw.len() < kv_dim_v
            || q_out.len() < nh_q * head_dim
            || k_out.len() < nh_k * head_dim
            || pos.is_empty()
            || done_ctr.is_empty()
        {
            return Err(format!(
                "qk_norm_rope_append_inc geometry head_dim={head_dim} nh_q={nh_q} nh_k={nh_k}                  kv_k={kv_dim_k} kv_v={kv_dim_v}"
            )
            .into());
        }
        let f = self.func("qk_norm_rope_append_inc_dcw");
        let theta_scale = (freq_base).powf(-2.0 / n_dims as f32);
        let cfg = LaunchConfig {
            grid_dim: ((nh_q + nh_k) as u32, 1, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (kvk, kvv) = (kv_dim_k as i32, kv_dim_v as i32);
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let (hd, nd, nq) = (head_dim as i32, n_dims as i32, nh_q as i32);
        let null: u64 = 0;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q_raw)
            .arg(k_raw)
            .arg(v_raw)
            .arg(qw)
            .arg(kw)
            .arg(q_out)
            .arg(k_out)
            .arg(pos)
            .arg(&mut *k_plane)
            .arg(&mut *v_plane)
            .arg(len_dev);
        match base_dev {
            Some(base) => {
                b.arg(base);
            }
            None => {
                b.arg(&null);
            }
        }
        b.arg(&mut *done_ctr)
            .arg(&kvk)
            .arg(&kvv)
            .arg(&ktb)
            .arg(&vtb)
            .arg(&hd)
            .arg(&nd)
            .arg(&nq)
            .arg(&eps)
            .arg(&theta_scale)
            .arg(&freq_scale);
        match ff {
            Some(freqs) => {
                b.arg(freqs);
            }
            None => {
                b.arg(&null);
            }
        }
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn qk_norm_rope_into(
        &self,
        q_raw: &CudaSlice<f32>,
        k_raw: &CudaSlice<f32>,
        qw: &CudaSlice<f32>,
        kw: &CudaSlice<f32>,
        q_out: &mut CudaSlice<f32>,
        k_out: &mut CudaSlice<f32>,
        pos: &CudaSlice<i32>,
        head_dim: usize,
        n_dims: usize,
        nh_q: usize,
        nh_k: usize,
        eps: f32,
        freq_base: f32,
        freq_scale: f32,
        ff: Option<&CudaSlice<f32>>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if head_dim > 512
            || q_raw.len() < nh_q * head_dim
            || k_raw.len() < nh_k * head_dim
            || q_out.len() < nh_q * head_dim
            || k_out.len() < nh_k * head_dim
            || qw.len() < head_dim
            || kw.len() < head_dim
            || pos.is_empty()
        {
            return Err(format!(
                "qk_norm_rope geometry head_dim={head_dim} nh_q={nh_q} nh_k={nh_k}"
            )
            .into());
        }
        let f = self.func("qk_norm_rope_f32");
        let theta_scale = (freq_base).powf(-2.0 / n_dims as f32);
        let cfg = LaunchConfig {
            grid_dim: ((nh_q + nh_k) as u32, 1, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (hd, nd, nq) = (head_dim as i32, n_dims as i32, nh_q as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q_raw)
            .arg(k_raw)
            .arg(qw)
            .arg(kw)
            .arg(q_out)
            .arg(k_out)
            .arg(pos)
            .arg(&hd)
            .arg(&nd)
            .arg(&nq)
            .arg(&eps)
            .arg(&theta_scale)
            .arg(&freq_scale);
        match ff {
            Some(ffv) => {
                b.arg(ffv);
                unsafe {
                    b.launch(cfg)?;
                }
            }
            None => {
                let null: u64 = 0;
                b.arg(&null);
                unsafe {
                    b.launch(cfg)?;
                }
            }
        }
        Ok(())
    }

    /// Four-block F32 matvec with in-order block accumulation (see the kernel doc): one
    /// launch computes a rank's whole O partial from its four canonical column blocks.
    #[allow(clippy::too_many_arguments)]
    pub fn matvec_f32_b4_into(
        &self,
        w: [&CudaSlice<f32>; 4],
        x: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        block_cols: usize,
        out_f: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if block_cols % 4 != 0
            || x.len() < 4 * block_cols
            || y.len() < out_f
            || w.iter().any(|w| w.len() != out_f * block_cols)
        {
            return Err(format!(
                "matvec_f32_b4 geometry block_cols={block_cols} out={out_f} x={}",
                x.len()
            )
            .into());
        }
        let f = self.func("matvec_f32_b4");
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, 1, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (bc, of) = (block_cols as i32, out_f as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(w[0])
            .arg(w[1])
            .arg(w[2])
            .arg(w[3])
            .arg(x)
            .arg(y)
            .arg(&bc)
            .arg(&of);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Sequential weighted row-sum: y[i] = sum_p w[p] * x[p*width+i] in row order — the exact
    /// per-element FP chain of a zero-reset plus n_rows sequential axpy launches.
    pub fn axpy_rows_seq_into(
        &self,
        x: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        width: usize,
        n_rows: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if x.len() < n_rows * width || w.len() < n_rows || y.len() < width {
            return Err(format!(
                "axpy_rows_seq geometry x={} w={} y={} width={width} rows={n_rows}",
                x.len(),
                w.len(),
                y.len()
            )
            .into());
        }
        let f = self.func("axpy_rows_seq_f32");
        let cfg = LaunchConfig::for_num_elems(width as u32);
        let (wi, nr) = (width as i32, n_rows as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(w).arg(y).arg(&wi).arg(&nr);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Row-offset twin of `axpy_rows_seq_md_into` (spec verify t-column combine): the
    /// accumulation runs over rows [row0, row0+n_rows) of a taller partial slab — the
    /// exact sequential FP chain of the base kernel over that window.
    #[allow(clippy::too_many_arguments)]
    pub fn axpy_rows_seq_md_off_into(
        &self,
        x: &CudaSlice<f32>,
        w_route: &CudaSlice<f32>,
        md: &CudaSlice<f32>,
        sel: &CudaSlice<i32>,
        y: &mut CudaSlice<f32>,
        width: usize,
        n_rows: usize,
        row0: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if x.len() < (row0 + n_rows) * width
            || w_route.len() < row0 + n_rows
            || sel.len() < row0 + n_rows
            || y.len() < width
        {
            return Err(format!(
                "axpy_rows_seq_md_off geometry x={} w={} sel={} y={} width={width} \
                 rows={n_rows} row0={row0}",
                x.len(),
                w_route.len(),
                sel.len(),
                y.len()
            )
            .into());
        }
        let f = self.func("axpy_rows_seq_md_off_f32");
        let cfg = LaunchConfig::for_num_elems(width as u32);
        let (wi, nr, r0) = (width as i32, n_rows as i32, row0 as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x)
            .arg(w_route)
            .arg(md)
            .arg(sel)
            .arg(y)
            .arg(&wi)
            .arg(&nr)
            .arg(&r0);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// T-COLUMN twin of `qmatvec_nvfp4_sel_gu_into` (spec verify, MEMRA_TCOL_FFN):
    /// 2*n_sel_col selection pairs over TWO activation rows (pair t reads row
    /// t/n_sel_col). Per-(pair,row) FP program == the t=1 gu kernel: each column's
    /// outputs are bit-equal to its own t=1 launch.
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_nvfp4_sel_gu_tcol_into(
        &self,
        gate_bank: &CudaSlice<u8>,
        up_bank: &CudaSlice<u8>,
        sel: &CudaSlice<i32>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        yg: &mut CudaSlice<f32>,
        yu: &mut CudaSlice<f32>,
        n_sel: usize,
        n_sel_col: usize,
        in_f: usize,
        out_f: usize,
        row_bytes: usize,
        expert_stride: usize,
        act_row_stride: usize,
        ad_row_stride: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        assert!(in_f % 64 == 0, "NVFP4 dp4a requires in_f % 64 == 0");
        if yg.len() < n_sel * out_f
            || yu.len() < n_sel * out_f
            || sel.len() < n_sel
            || n_sel_col == 0
            || n_sel % n_sel_col != 0
        {
            return Err("NVFP4 gu tcol geometry".into());
        }
        let f = self.func("qmatvec_nvfp4_dp4a_sel_v2_gu_tcol");
        let cfg = LaunchConfig {
            grid_dim: ((2 * out_f) as u32, n_sel as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, outf, ns, nsc) = (in_f as i32, out_f as i32, n_sel as i32, n_sel_col as i32);
        let (rb, es) = (row_bytes as i64, expert_stride as i64);
        let (ars, adrs) = (act_row_stride as i64, ad_row_stride as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate_bank)
            .arg(up_bank)
            .arg(sel)
            .arg(aq)
            .arg(ad)
            .arg(yg)
            .arg(yu)
            .arg(&inf)
            .arg(&outf)
            .arg(&ns)
            .arg(&rb)
            .arg(&es)
            .arg(&ars)
            .arg(&adrs)
            .arg(&nsc);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Device-routed twin of `axpy_rows_seq_into`: the per-row weight folds in-kernel
    /// (w_route[p] * md[sel[p]] — the same single f32 multiply the host fold performs).
    #[allow(clippy::too_many_arguments)]
    pub fn axpy_rows_seq_md_into(
        &self,
        x: &CudaSlice<f32>,
        w_route: &CudaSlice<f32>,
        md: &CudaSlice<f32>,
        sel: &CudaSlice<i32>,
        y: &mut CudaSlice<f32>,
        width: usize,
        n_rows: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if x.len() < n_rows * width
            || w_route.len() < n_rows
            || sel.len() < n_rows
            || y.len() < width
        {
            return Err(format!(
                "axpy_rows_seq_md geometry x={} w={} sel={} y={} width={width} rows={n_rows}",
                x.len(),
                w_route.len(),
                sel.len(),
                y.len()
            )
            .into());
        }
        let f = self.func("axpy_rows_seq_md_f32");
        let cfg = LaunchConfig::for_num_elems(width as u32);
        let (wi, nr) = (width as i32, n_rows as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x)
            .arg(w_route)
            .arg(md)
            .arg(sel)
            .arg(y)
            .arg(&wi)
            .arg(&nr);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// BF16 twin of `matvec_f32_qkv_into` (weights as raw checkpoint bf16 bytes).
    #[allow(clippy::too_many_arguments)]
    /// T-COLUMN twin of `matvec_bf16_qkvg_into` (spec verify): weights read once, T input
    /// columns accumulated with per-column FP order identical to the t=1 kernel. Outputs
    /// land column-major-of-rows: yq[c*out_q + row] etc.
    #[allow(clippy::too_many_arguments)]
    pub fn matvec_bf16_qkvg_tcol_into(
        &self,
        wq: &CudaSlice<u8>,
        wk: &CudaSlice<u8>,
        wv: &CudaSlice<u8>,
        wg: &CudaSlice<u8>,
        x_t: &CudaSlice<f32>,
        yq: &mut CudaSlice<f32>,
        yk: &mut CudaSlice<f32>,
        yv: &mut CudaSlice<f32>,
        yg: &mut CudaSlice<f32>,
        in_f: usize,
        out_q: usize,
        out_kv: usize,
        out_g: usize,
        t: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if t == 0
            || t > 8
            || in_f % 8 != 0
            || x_t.len() < t * in_f
            || yq.len() < t * out_q
            || yk.len() < t * out_kv
            || yv.len() < t * out_kv
            || (out_g > 0 && yg.len() < t * out_g)
        {
            return Err("matvec_bf16_qkvg_tcol geometry".into());
        }
        let grid = out_q + 2 * out_kv + out_g;
        let cfg = LaunchConfig {
            grid_dim: (grid as u32, 1, 1),
            block_dim: (mmv_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (ini, oq, okv, og, ti) = (
            in_f as i32,
            out_q as i32,
            out_kv as i32,
            out_g as i32,
            t as i32,
        );
        let __s_b = self.gpu.stream();
        // One runtime-T program for every live width. The compile-time 2/4/8 twins are
        // retained in the fatbin as research controls, but dispatching them by the current
        // batch width changes kernels inside a request when peers arrive or retire. That is
        // a load-history numeric-program switch, and their pre-twin TOKFP receipts did not
        // qualify it (Hermes `64fa2b55baf0d887`).
        let f = self.func("matvec_bf16_qkvg_tcol");
        let mut b = __s_b.launch_builder(&f);
        b.arg(wq)
            .arg(wk)
            .arg(wv)
            .arg(wg)
            .arg(x_t)
            .arg(yq)
            .arg(yk)
            .arg(yv)
            .arg(yg)
            .arg(&ini)
            .arg(&oq)
            .arg(&okv)
            .arg(&og)
            .arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn matvec_bf16_qkvg_into(
        &self,
        wq: &CudaSlice<u8>,
        wk: &CudaSlice<u8>,
        wv: &CudaSlice<u8>,
        wg: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        yq: &mut CudaSlice<f32>,
        yk: &mut CudaSlice<f32>,
        yv: &mut CudaSlice<f32>,
        yg: &mut CudaSlice<f32>,
        in_f: usize,
        out_q: usize,
        out_kv: usize,
        out_g: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if in_f % 8 != 0
            || wq.len() != out_q * in_f * 2
            || wk.len() != out_kv * in_f * 2
            || wv.len() != out_kv * in_f * 2
            || wg.len() < out_g * in_f * 2
            || x.len() < in_f
            || yq.len() < out_q
            || yk.len() < out_kv
            || yv.len() < out_kv
            || (out_g > 0 && yg.len() < out_g)
        {
            return Err(format!(
                "fused bf16 QKV geometry in={in_f} out_q={out_q} out_kv={out_kv} out_g={out_g}"
            )
            .into());
        }
        let f = self.func("matvec_bf16_qkvg");
        let cfg = LaunchConfig {
            grid_dim: ((out_q + 2 * out_kv + out_g) as u32, 1, 1),
            block_dim: (mmv_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (inf, oq, okv, og) = (in_f as i32, out_q as i32, out_kv as i32, out_g as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(wq)
            .arg(wk)
            .arg(wv)
            .arg(wg)
            .arg(x)
            .arg(yq)
            .arg(yk)
            .arg(yv)
            .arg(yg)
            .arg(&inf)
            .arg(&oq)
            .arg(&okv)
            .arg(&og);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// BF16 twin of `matvec_f32_b4_into` (weights as raw checkpoint bf16 bytes).
    pub fn matvec_bf16_b4_into(
        &self,
        w: [&CudaSlice<u8>; 4],
        x: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        block_cols: usize,
        out_f: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if block_cols % 8 != 0
            || x.len() < 4 * block_cols
            || y.len() < out_f
            || w.iter().any(|w| w.len() != out_f * block_cols * 2)
        {
            return Err(format!(
                "bf16 b4 geometry block_cols={block_cols} out={out_f} x={}",
                x.len()
            )
            .into());
        }
        // MEMRA_B4_X2=1: the #2b grid-halving twin — half the blocks, two rows each,
        // bit-identical per row (the second row's stream hides the first's reduce tail).
        static B4_X2: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let x2 = *B4_X2.get_or_init(|| std::env::var("MEMRA_B4_X2").as_deref() == Ok("1"));
        let f = self.func(if x2 {
            "matvec_bf16_b4_x2"
        } else {
            "matvec_bf16_b4"
        });
        let grid = if x2 { out_f.div_ceil(2) } else { out_f };
        let cfg = LaunchConfig {
            grid_dim: (grid as u32, 1, 1),
            block_dim: (mmv_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (bc, of) = (block_cols as i32, out_f as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(w[0])
            .arg(w[1])
            .arg(w[2])
            .arg(w[3])
            .arg(x)
            .arg(y)
            .arg(&bc)
            .arg(&of);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// T-COLUMN twin of `matvec_bf16_b4_into` (spec verify o_proj): weights read once, T
    /// gated rows (each 4*block_cols wide) accumulated with per-column FP order identical
    /// to the t=1 kernel. Outputs land y[c*out_f + row]. Same blockDim as the t=1 launch —
    /// the shared-memory reduce order depends on it. Refuses under MEMRA_B4_X2 (different
    /// t=1 program).
    pub fn matvec_bf16_b4_tcol_into(
        &self,
        w: [&CudaSlice<u8>; 4],
        x_t: &CudaSlice<f32>,
        y_t: &mut CudaSlice<f32>,
        block_cols: usize,
        out_f: usize,
        t: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if block_cols % 8 != 0
            || t == 0
            || t > 8
            || x_t.len() < t * 4 * block_cols
            || y_t.len() < t * out_f
            || w.iter().any(|w| w.len() != out_f * block_cols * 2)
        {
            return Err(format!(
                "bf16 b4 tcol geometry block_cols={block_cols} out={out_f} t={t} x={}",
                x_t.len()
            )
            .into());
        }
        if std::env::var("MEMRA_B4_X2").as_deref() == Ok("1") {
            return Err(
                "b4 tcol verify is qualified against the plain b4 kernel only \
                        (MEMRA_B4_X2=1 is a different t=1 program)"
                    .into(),
            );
        }
        // Keep one runtime-T program at every live width. Compile-time twins remain research
        // controls only; selecting them from the changing batch width switches programs
        // mid-request.
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, 1, 1),
            block_dim: (mmv_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (bc, of, ti) = (block_cols as i32, out_f as i32, t as i32);
        let __s_b = self.gpu.stream();
        let f = self.func("matvec_bf16_b4_tcol");
        let mut b = __s_b.launch_builder(&f);
        b.arg(w[0])
            .arg(w[1])
            .arg(w[2])
            .arg(w[3])
            .arg(x_t)
            .arg(y_t)
            .arg(&bc)
            .arg(&of)
            .arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// `matvec_bf16` writing into a caller-owned output (persistent-workspace form).
    /// q8_0 row bytes for an `in_f`-wide weight row: one 34-byte block per 32 weights.
    pub fn q8_0_row_bytes(in_f: usize) -> usize {
        in_f / 32 * 34
    }

    /// Encode a resident bf16 weight slab into its q8_0 mirror (MEMRA_STEP_TP_W8). Runs once
    /// per matrix at load; the block program is the one `quant_K_block` writes for the KV
    /// cache, so the two formats cannot drift apart.
    pub fn encode_q8_0_from_bf16(
        &self,
        w_bf16: &CudaSlice<u8>,
        out: &mut CudaSlice<u8>,
        in_f: usize,
        out_f: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if in_f % 32 != 0
            || w_bf16.len() < in_f * out_f * 2
            || out.len() < out_f * Self::q8_0_row_bytes(in_f)
        {
            return Err(format!(
                "encode_q8_0_from_bf16 geometry in={in_f} out={out_f} src={} dst={}",
                w_bf16.len(),
                out.len()
            )
            .into());
        }
        let f = self.func("encode_q8_0_rows_from_bf16");
        // Flat 1D grid of (row, 32-block) pairs, 4 pairs per block: rows on grid.y would cap
        // at 65535 and the LM head has 128896 rows.
        const PAIRS_PER_BLOCK: u32 = 4;
        let pairs = (out_f * (in_f / 32)) as u64;
        let cfg = LaunchConfig {
            grid_dim: ((pairs.div_ceil(PAIRS_PER_BLOCK as u64)) as u32, 1, 1),
            block_dim: (32, PAIRS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (ini, outi) = (in_f as i32, out_f as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(w_bf16).arg(out).arg(&ini).arg(&outi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Fused q8_0 QKV against a q8_1 activation (MEMRA_STEP_TP_W8): one launch over the
    /// stacked q/k/v rows, each row running the exact `qmatvec_q8_0_mmvq_rp` per-row program.
    /// Bit-identical to three per-matrix mmvq calls; it exists because those three launches
    /// plus the activation quantize measured SLOWER than the bf16 fused kernel.
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_q8_0_qkv_rp_into(
        &self,
        wq: &CudaSlice<u8>,
        wk: &CudaSlice<u8>,
        wv: &CudaSlice<u8>,
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        yq: &mut CudaSlice<f32>,
        yk: &mut CudaSlice<f32>,
        yv: &mut CudaSlice<f32>,
        in_f: usize,
        out_q: usize,
        out_kv: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        const ROWS_PER_BLOCK: u32 = 4; // matches MEMRA_MMVQ_ROWS in qmatvec.cu
        let rows = out_q + 2 * out_kv;
        let nblk = in_f / 32;
        if in_f % 32 != 0
            || aq.len() < in_f
            || ad.len() < nblk
            || yq.len() < out_q
            || yk.len() < out_kv
            || yv.len() < out_kv
            || wq.len() < out_q * nblk * 34
            || wk.len() < out_kv * nblk * 34
            || wv.len() < out_kv * nblk * 34
        {
            return Err(
                format!("q8_0 qkv rp geometry in={in_f} out_q={out_q} out_kv={out_kv}").into(),
            );
        }
        let f = self.func("qmatvec_q8_0_qkv_rp");
        let cfg = LaunchConfig {
            grid_dim: ((rows as u32).div_ceil(ROWS_PER_BLOCK), 1, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (ini, oq, okv) = (in_f as i32, out_q as i32, out_kv as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(wq)
            .arg(wk)
            .arg(wv)
            .arg(aq)
            .arg(ad)
            .arg(yq)
            .arg(yk)
            .arg(yv)
            .arg(&ini)
            .arg(&oq)
            .arg(&okv);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Fused q8_0 O projection over the four HEAD_SPLIT blocks (MEMRA_STEP_TP_W8): one
    /// launch, one warp per output row, per-block reduce then add — the same shape
    /// `matvec_bf16_b4` uses, against a q8_1 activation.
    #[allow(clippy::too_many_arguments)]
    pub fn qmatvec_q8_0_b4_rp_into(
        &self,
        w: [&CudaSlice<u8>; 4],
        aq: &CudaSlice<i8>,
        ad: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        block_cols: usize,
        out_f: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        const ROWS_PER_BLOCK: u32 = 4; // matches MEMRA_MMVQ_ROWS in qmatvec.cu
        let nblk = block_cols / 32;
        if block_cols % 32 != 0
            || aq.len() < 4 * block_cols
            || ad.len() < 4 * nblk
            || y.len() < out_f
            || w.iter().any(|p| p.len() < out_f * nblk * 34)
        {
            return Err(format!("q8_0 b4 rp geometry block_cols={block_cols} out={out_f}").into());
        }
        let f = self.func("qmatvec_q8_0_b4_rp");
        let cfg = LaunchConfig {
            grid_dim: ((out_f as u32).div_ceil(ROWS_PER_BLOCK), 1, 1),
            block_dim: (32, ROWS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (bc, of) = (block_cols as i32, out_f as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(w[0])
            .arg(w[1])
            .arg(w[2])
            .arg(w[3])
            .arg(aq)
            .arg(ad)
            .arg(y)
            .arg(&bc)
            .arg(&of);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Get-or-build this bf16 weight's q8_0 mirror and run the GEMV through it. Returns
    /// `None` when the shape has no mirror form, so the caller falls back to bf16.
    fn matvec_bf16_via_q8_mirror(
        &self,
        data: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        in_f: usize,
        out_f: usize,
    ) -> Result<Option<()>, Box<dyn std::error::Error>> {
        use cudarc::driver::DevicePtr;
        let key = {
            let s = self.gpu.stream();
            let (p, _g) = data.device_ptr(&s);
            p as u64
        };
        {
            let mut mirrors = self
                .w8_mirrors
                .lock()
                .map_err(|_| "w8 mirror map is poisoned")?;
            if !mirrors.contains_key(&key) {
                let mut interleaved = self.alloc_u8_uninit(out_f * Self::q8_0_row_bytes(in_f))?;
                self.encode_q8_0_from_bf16(data, &mut interleaved, in_f, out_f)?;
                let planar = self.build_q8_rp4_raw(&interleaved, in_f, out_f)?;
                mirrors.insert(key, planar);
                // Which weights this half actually covers is not obvious from the call graph:
                // the head and the shared expert may reach the GPU through the rows fast path
                // or the fused dual-silu launcher instead of here. One line per mirror answers
                // that without a profiler (the hybrid half measured +0.1% and this is how we
                // find out whether it even fired).
                if std::env::var("MEMRA_W8_TRACE").as_deref() == Ok("1") {
                    eprintln!(
                        "[w8-mirror] built in_f={in_f} out_f={out_f} mirrors={}",
                        mirrors.len()
                    );
                }
            }
        }
        let nblk = in_f / 32;
        {
            let mut act = self.w8_act.lock().map_err(|_| "w8 act map is poisoned")?;
            if !act.contains_key(&in_f) {
                let aq = self.alloc_uninit::<i8>(in_f)?;
                let ad = self.alloc_uninit::<f32>(nblk)?;
                act.insert(in_f, (aq, ad));
            }
            let (aq, ad) = act.get_mut(&in_f).expect("just inserted");
            self.quantize_q8_1_into(x, 1, in_f, aq, ad)?;
        }
        let mirrors = self
            .w8_mirrors
            .lock()
            .map_err(|_| "w8 mirror map is poisoned")?;
        let act = self.w8_act.lock().map_err(|_| "w8 act map is poisoned")?;
        let mirror = mirrors.get(&key).expect("built above");
        let (aq, ad) = act.get(&in_f).expect("built above");
        self.qmatvec_mmvq_into(
            mirror,
            aq,
            ad,
            1,
            in_f,
            out_f,
            QT_Q8_0,
            Self::q8_0_row_bytes(in_f),
            1.0,
            true,
            y,
        )?;
        Ok(Some(()))
    }

    pub fn matvec_bf16_into(
        &self,
        data: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        in_f: usize,
        out_f: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if data.len() != in_f * out_f * 2 || x.len() < in_f || in_f % 8 != 0 || y.len() < out_f {
            return Err(format!(
                "matvec_bf16_into geometry bytes={} x={} y={} in={in_f} out={out_f}",
                data.len(),
                x.len(),
                y.len()
            )
            .into());
        }
        // MEMRA_STEP_TP_W8, hybrid half: route this GEMV through a q8_0 mirror of the same
        // weight. Covers exactly the bf16 GEMVs that are NOT in a TP resident bank — the LM
        // head (324.4 -> 163.7 us measured), the shared-expert down rows (13.0 -> 5.6 us) and
        // the dense-FFN layers. Same numeric class as the QKV/o_proj arms (int8 dp4a with
        // per-32 scales), so it rides the same argmax acceptance; the bf16 slab stays resident
        // for prefill. The mirror builds on first use and is keyed by the slab's pointer.
        if step_tp_w8_on() && in_f % 32 == 0 && out_f >= 64 {
            if let Some(()) = self.matvec_bf16_via_q8_mirror(data, x, y, in_f, out_f)? {
                return Ok(());
            }
        }
        // MEMRA_DOWN_X4=1 (short-row shapes, in_f<=2048): four sequential rows per
        // block, exact f32acc per-row program — cures the 1-iteration latency
        // starvation (shexp down measured 420GB/s at in_f=1280).
        static X4: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let x4 = *X4.get_or_init(|| std::env::var("MEMRA_DOWN_X4").as_deref() == Ok("1"))
            && in_f <= 2048;
        if x4 {
            let f = self.func("matvec_bf16_f32acc_x4");
            let cfg = LaunchConfig {
                grid_dim: (out_f.div_ceil(4) as u32, 1, 1),
                block_dim: (mmv_block(), 1, 1),
                shared_mem_bytes: 0,
            };
            let (ini, outi) = (in_f as i32, out_f as i32);
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(data).arg(x).arg(y).arg(&ini).arg(&outi);
            unsafe {
                b.launch(cfg)?;
            }
            return Ok(());
        }
        let f = self.func("matvec_bf16_f32acc");
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, 1, 1),
            block_dim: (mmv_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let ini = in_f as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(data).arg(x).arg(y).arg(&ini);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// `matvec_bf16_into` over a WEIGHT VIEW (row-range slice of a bf16 tensor): the head-split
    /// door feeds each device its half of the lm-head rows. Same kernel, same per-row program.
    pub fn matvec_bf16_view_into(
        &self,
        data: &cudarc::driver::CudaView<'_, u8>,
        x: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        in_f: usize,
        out_f: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if data.len() != in_f * out_f * 2 || x.len() < in_f || in_f % 8 != 0 || y.len() < out_f {
            return Err(format!(
                "matvec_bf16_view_into geometry bytes={} x={} y={} in={in_f} out={out_f}",
                data.len(),
                x.len(),
                y.len()
            )
            .into());
        }
        let f = self.func("matvec_bf16_f32acc");
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, 1, 1),
            block_dim: (mmv_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let ini = in_f as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(data).arg(x).arg(y).arg(&ini);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// `matvec_bf16_into` with a RAW u64 output pointer (UVA — the dev1-shexp down row
    /// lands root-resident over P2P). Same kernel, same per-row program: bit-identical.
    pub fn matvec_bf16_raw_out(
        &self,
        w: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        y_raw: u64,
        in_f: usize,
        out_f: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if w.len() != in_f * out_f * 2 || x.len() < in_f || in_f % 8 != 0 || y_raw == 0 {
            return Err("matvec_bf16_raw_out geometry".into());
        }
        let f = self.func("matvec_bf16_f32acc");
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, 1, 1),
            block_dim: (mmv_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let ini = in_f as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(w).arg(x).arg(&y_raw).arg(&ini);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// MOE TAIL FUSION M1: dst = (a + b) + sh*scale[0] in one launch (sh/scale as RAW
    /// UVA pointers so the caller passes persistent-static rows without holding locks).
    /// Exact per-element sequence of the split add + add_scaled_rows pair.
    pub fn add3_raw(
        &self,
        a: &CudaSlice<f32>,
        b: &CudaSlice<f32>,
        sh_raw: u64,
        scale_raw: u64,
        dst: &mut CudaSlice<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if a.len() < n || b.len() < n || dst.len() < n || sh_raw == 0 || scale_raw == 0 {
            return Err("add3_raw geometry".into());
        }
        let f = self.func("add3_f32");
        let cfg = LaunchConfig {
            grid_dim: ((n as u32).div_ceil(256), 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let ni = n as i32;
        let __s_b = self.gpu.stream();
        let mut bld = __s_b.launch_builder(&f);
        bld.arg(a)
            .arg(b)
            .arg(&sh_raw)
            .arg(&scale_raw)
            .arg(dst)
            .arg(&ni);
        unsafe {
            bld.launch(cfg)?;
        }
        Ok(())
    }

    /// FUSION #2e: shexp down matvec + scaled accumulate (dst[r] += dot_r * scale[0]),
    /// one launch replacing matvec_bf16_into + the ownership copy + add_scaled_rows.
    pub fn matvec_bf16_down_addscale_into(
        &self,
        w: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        scale: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        in_f: usize,
        out_f: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if w.len() != in_f * out_f * 2
            || x.len() < in_f
            || in_f % 8 != 0
            || dst.len() < out_f
            || scale.is_empty()
        {
            return Err("matvec_bf16_down_addscale geometry".into());
        }
        let f = self.func("matvec_bf16_down_addscale");
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, 1, 1),
            block_dim: (mmv_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let ini = in_f as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(w).arg(x).arg(scale).arg(dst).arg(&ini);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// FUSION #2b: shexp dual matvec + SwiGLU act, one launch (bit-identical to
    /// matvec_bf16_dual_into + ffn_act_lim at gs=us=1; limit=None takes plain silu).
    /// T-ROW twin of `matvec_bf16_dual_silu_into` (per-row program identical).
    #[allow(clippy::too_many_arguments)]
    pub fn matvec_bf16_dual_silu_rows_into(
        &self,
        wg: &CudaSlice<u8>,
        wu: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        act: &mut CudaSlice<f32>,
        in_f: usize,
        out_f: usize,
        limit: Option<f32>,
        t: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if x.len() < t * in_f || act.len() < t * out_f || t == 0 || t > 32 {
            return Err("matvec_bf16_dual_silu_rows geometry".into());
        }
        let f = self.func("matvec_bf16_dual_silu_rows");
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, t as u32, 1),
            block_dim: (mmv_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (ini, outi) = (in_f as i32, out_f as i32);
        let lim = limit.unwrap_or(0.0);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(wg)
            .arg(wu)
            .arg(x)
            .arg(&mut *act)
            .arg(&ini)
            .arg(&outi)
            .arg(&lim);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// T-ROW twin of the bf16 f32acc-x4 matvec (per-row program identical).
    pub fn matvec_bf16_rows_into(
        &self,
        w: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        in_f: usize,
        out_f: usize,
        t: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if x.len() < t * in_f || y.len() < t * out_f || t == 0 || t > 32 || in_f % 8 != 0 {
            return Err("matvec_bf16_rows geometry".into());
        }
        // MEMRA_STEP_TP_W8: the LM head reaches the device HERE, not through
        // matvec_bf16_into — the W8 trace showed the hybrid half building mirrors only for
        // in_f=1280 out_f=4096 (the shared-expert down rows, which SHEXP_OVERLAP already
        // hides, hence its +0.1%). Route the t=1 decode row through the q8 mirror; wider t
        // (the verify walk) keeps bf16 so the prefill class is untouched.
        if t == 1 && step_tp_w8_on() && in_f % 32 == 0 && out_f >= 64 {
            if let Some(()) = self.matvec_bf16_via_q8_mirror(w, x, y, in_f, out_f)? {
                return Ok(());
            }
        }
        let f = self.func("matvec_bf16_f32acc_x4_rows");
        let cfg = LaunchConfig {
            grid_dim: (out_f.div_ceil(4) as u32, t as u32, 1),
            block_dim: (mmv_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (ini, outi) = (in_f as i32, out_f as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(w).arg(x).arg(&mut *y).arg(&ini).arg(&outi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn matvec_bf16_dual_silu_into(
        &self,
        wg: &CudaSlice<u8>,
        wu: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        act: &mut CudaSlice<f32>,
        in_f: usize,
        out_f: usize,
        limit: Option<f32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if wg.len() != in_f * out_f * 2
            || wu.len() != in_f * out_f * 2
            || x.len() < in_f
            || in_f % 8 != 0
            || act.len() < out_f
        {
            return Err("matvec_bf16_dual_silu geometry".into());
        }
        let f = self.func("matvec_bf16_dual_silu");
        let cfg = LaunchConfig {
            grid_dim: (out_f as u32, 1, 1),
            block_dim: (mmv_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (ini, outi) = (in_f as i32, out_f as i32);
        let lim = limit.unwrap_or(0.0);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(wg)
            .arg(wu)
            .arg(x)
            .arg(act)
            .arg(&ini)
            .arg(&outi)
            .arg(&lim);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// `matvec_bf16_dual_into` over WEIGHT VIEWS (row-range slices): the shexp row-split
    /// door feeds each device its half of the gate/up rows. Same kernel, same per-row program.
    #[allow(clippy::too_many_arguments)]
    pub fn matvec_bf16_dual_view_into(
        &self,
        wg: &cudarc::driver::CudaView<'_, u8>,
        wu: &cudarc::driver::CudaView<'_, u8>,
        x: &CudaSlice<f32>,
        yg: &mut CudaSlice<f32>,
        yu: &mut CudaSlice<f32>,
        in_f: usize,
        out_f: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if wg.len() != in_f * out_f * 2
            || wu.len() != in_f * out_f * 2
            || x.len() < in_f
            || in_f % 8 != 0
            || yg.len() < out_f
            || yu.len() < out_f
        {
            return Err(format!(
                "matvec_bf16_dual_view_into geometry wg={} wu={} x={} in={in_f} out={out_f}",
                wg.len(),
                wu.len(),
                x.len()
            )
            .into());
        }
        let f = self.func("matvec_bf16_dual");
        let cfg = LaunchConfig {
            grid_dim: ((2 * out_f) as u32, 1, 1),
            block_dim: (mmv_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (ini, outi) = (in_f as i32, out_f as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(wg)
            .arg(wu)
            .arg(x)
            .arg(yg)
            .arg(yu)
            .arg(&ini)
            .arg(&outi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// `matvec_bf16_dual` writing into caller-owned outputs (persistent-workspace form).
    #[allow(clippy::too_many_arguments)]
    pub fn matvec_bf16_dual_into(
        &self,
        wg: &CudaSlice<u8>,
        wu: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        yg: &mut CudaSlice<f32>,
        yu: &mut CudaSlice<f32>,
        in_f: usize,
        out_f: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if wg.len() != in_f * out_f * 2
            || wu.len() != in_f * out_f * 2
            || x.len() < in_f
            || in_f % 8 != 0
            || yg.len() < out_f
            || yu.len() < out_f
        {
            return Err(format!(
                "matvec_bf16_dual_into geometry wg={} wu={} x={} in={in_f} out={out_f}",
                wg.len(),
                wu.len(),
                x.len()
            )
            .into());
        }
        let f = self.func("matvec_bf16_dual");
        let cfg = LaunchConfig {
            grid_dim: ((2 * out_f) as u32, 1, 1),
            block_dim: (mmv_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (ini, outi) = (in_f as i32, out_f as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(wg)
            .arg(wu)
            .arg(x)
            .arg(yg)
            .arg(yu)
            .arg(&ini)
            .arg(&outi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Dual bf16 matvec: gate/up (same shape) from one shared input in one launch. Per row
    /// bit-identical to two `matvec_bf16` launches. Returns (gate, up).
    pub(crate) fn matvec_bf16_dual(
        &self,
        wg: &CudaSlice<u8>,
        wu: &CudaSlice<u8>,
        x: &CudaSlice<f32>,
        in_f: usize,
        out_f: usize,
    ) -> Result<(CudaSlice<f32>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        if wg.len() != in_f * out_f * 2
            || wu.len() != in_f * out_f * 2
            || x.len() < in_f
            || in_f % 8 != 0
        {
            return Err(format!(
                "matvec_bf16_dual geometry wg={} wu={} x={} in={in_f} out={out_f}",
                wg.len(),
                wu.len(),
                x.len()
            )
            .into());
        }
        let mut yg = self.alloc_uninit::<f32>(out_f)?;
        let mut yu = self.alloc_uninit::<f32>(out_f)?;
        let f = self.func("matvec_bf16_dual");
        let cfg = LaunchConfig {
            grid_dim: ((2 * out_f) as u32, 1, 1),
            block_dim: (mmv_block(), 1, 1),
            shared_mem_bytes: 0,
        };
        let (ini, outi) = (in_f as i32, out_f as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(wg)
            .arg(wu)
            .arg(x)
            .arg(&mut yg)
            .arg(&mut yu)
            .arg(&ini)
            .arg(&outi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((yg, yu))
    }

    #[allow(clippy::too_many_arguments)]
    fn linear_bf16_chunked_inner(
        &self,
        x: &CudaSlice<f32>,
        data: &CudaSlice<u8>,
        m: usize,
        in_f: usize,
        out_f: usize,
        exact: bool,
        canonical_chunk_rows: Option<usize>,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        const CHUNK_BYTES: usize = 256 << 20;
        // canonical_chunk_rows callers are the step TP resident ranks whose cross-topology
        // equality program is defined in cuBLASLt chunk shapes — the door leaves them alone.
        if m == 1
            && !exact
            && canonical_chunk_rows.is_none()
            && in_f % 8 == 0
            && Self::bf16_mmv_on()
        {
            return self.matvec_bf16(data, x, in_f, out_f);
        }
        let row_bytes = in_f
            .checked_mul(std::mem::size_of::<f32>())
            .ok_or("BF16 chunk row byte count overflow")?;
        if row_bytes == 0 || out_f == 0 {
            return Err("BF16 chunk dimensions must be nonzero".into());
        }
        let max_chunk_rows = (CHUNK_BYTES / row_bytes).max(1).min(out_f);
        let chunk_rows = match canonical_chunk_rows {
            Some(rows) if rows == 0 => {
                return Err("canonical BF16 chunk rows must be nonzero".into());
            }
            Some(rows) if rows > max_chunk_rows => {
                return Err(format!(
                    "canonical BF16 chunk rows {rows} exceed the {max_chunk_rows}-row scratch limit"
                )
                .into());
            }
            Some(rows) if out_f % rows != 0 => {
                return Err(format!(
                    "BF16 output width {out_f} is not divisible by canonical {rows}-row chunks"
                )
                .into());
            }
            Some(rows) => rows,
            None => max_chunk_rows,
        };
        if chunk_rows >= out_f {
            let wf32 = self.bf16_to_f32(&data.slice(0..in_f * out_f * 2), in_f * out_f)?;
            return if exact {
                self.linear_decode_exact(x, &wf32, m, in_f, out_f)
            } else {
                self.linear(x, &wf32, m, in_f, out_f)
            };
        }
        let mut y = self.alloc_uninit::<f32>(m * out_f)?;
        let mut r0 = 0usize;
        while r0 < out_f {
            let rows = chunk_rows.min(out_f - r0);
            let wslice = data.slice(r0 * in_f * 2..(r0 + rows) * in_f * 2);
            let wf32 = self.bf16_to_f32(&wslice, in_f * rows)?;
            let yc = if exact {
                self.linear_decode_exact(x, &wf32, m, in_f, rows)?
            } else {
                self.linear(x, &wf32, m, in_f, rows)?
            };
            // scatter [m, rows] into y[m, out_f] at column offset r0 (m is tiny in decode/verify)
            for mi in 0..m {
                let src = yc.slice(mi * rows..(mi + 1) * rows);
                let mut dst = y.slice_mut(mi * out_f + r0..mi * out_f + r0 + rows);
                self.gpu.stream().memcpy_dtod(&src, &mut dst)?;
            }
            r0 += rows;
        }
        Ok(y)
    }

    /// Execute an already resident BF16 projection. This is the model-faithful substrate used by
    /// Step tensor-parallel correctness ranks; it preserves checkpoint bytes and the existing
    /// chunked BF16 numerical program instead of re-encoding the weight.
    pub fn linear_bf16_resident(
        &self,
        x: &CudaSlice<f32>,
        data: &CudaSlice<u8>,
        m: usize,
        in_f: usize,
        out_f: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        if data.len() != in_f * out_f * 2 {
            return Err(format!("resident BF16 bytes {} != {out_f}x{in_f}x2", data.len()).into());
        }
        self.linear_bf16_chunked(x, data, m, in_f, out_f, false, None)
    }

    /// Execute a resident BF16 projection as fixed-width output-row chunks.
    ///
    /// Tensor-parallel ranks use this to give TP1/TP2/TP4/TP8 the same cuBLASLt problem shape
    /// for every checkpoint row. Callers must derive `canonical_chunk_rows` from the registered
    /// model topology rather than the active rank count.
    pub fn linear_bf16_resident_canonical_rows(
        &self,
        x: &CudaSlice<f32>,
        data: &CudaSlice<u8>,
        m: usize,
        in_f: usize,
        out_f: usize,
        canonical_chunk_rows: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        if data.len() != in_f * out_f * 2 {
            return Err(format!("resident BF16 bytes {} != {out_f}x{in_f}x2", data.len()).into());
        }
        self.linear_bf16_chunked(x, data, m, in_f, out_f, false, Some(canonical_chunk_rows))
    }

    /// Execute a load-time F32 mirror with the same fixed output-row chunks as the BF16 path.
    ///
    /// Expanding the checkpoint bytes once changes residency, not arithmetic: every cuBLASLt
    /// call receives the same F32 values and problem shape as `linear_bf16_chunked`.
    pub fn linear_f32_resident_canonical_rows(
        &self,
        x: &CudaSlice<f32>,
        data: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        canonical_chunk_rows: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        self.linear_f32_resident_canonical_rows_inner(
            x,
            data,
            m,
            in_f,
            out_f,
            canonical_chunk_rows,
            false,
        )
    }

    /// Execute fixed output-row chunks and assemble them with one strided placement per chunk.
    ///
    /// The projection shapes and values are identical to
    /// [`Self::linear_f32_resident_canonical_rows`]. Only the byte-preserving output layout step
    /// changes, replacing one device copy per token with one placement kernel per output chunk.
    pub fn linear_f32_resident_canonical_rows_strided(
        &self,
        x: &CudaSlice<f32>,
        data: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        canonical_chunk_rows: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        self.linear_f32_resident_canonical_rows_inner(
            x,
            data,
            m,
            in_f,
            out_f,
            canonical_chunk_rows,
            true,
        )
    }

    fn linear_f32_resident_canonical_rows_inner(
        &self,
        x: &CudaSlice<f32>,
        data: &CudaSlice<f32>,
        m: usize,
        in_f: usize,
        out_f: usize,
        canonical_chunk_rows: usize,
        strided_output: bool,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        if data.len() != in_f * out_f {
            return Err(format!("resident F32 values {} != {out_f}x{in_f}", data.len()).into());
        }
        if canonical_chunk_rows == 0
            || canonical_chunk_rows > out_f
            || out_f % canonical_chunk_rows != 0
        {
            return Err(format!(
                "invalid canonical F32 chunk rows {canonical_chunk_rows} for output width {out_f}"
            )
            .into());
        }
        if canonical_chunk_rows == out_f {
            return self.linear(x, data, m, in_f, out_f);
        }

        let mut y = self.alloc_uninit::<f32>(m * out_f)?;
        let input = x.slice(0..x.len());
        for r0 in (0..out_f).step_by(canonical_chunk_rows) {
            let weights = data.slice(r0 * in_f..(r0 + canonical_chunk_rows) * in_f);
            if m == 1 {
                let mut destination = y.slice_mut(r0..r0 + canonical_chunk_rows);
                self.linear_device_into(
                    &input,
                    &weights,
                    &mut destination,
                    1,
                    in_f,
                    canonical_chunk_rows,
                )?;
                continue;
            }
            let chunk = self.linear_device(&input, &weights, m, in_f, canonical_chunk_rows)?;
            if strided_output {
                self.place_rows_strided(&chunk, &mut y, canonical_chunk_rows, m, out_f, r0)?;
            } else {
                for token in 0..m {
                    let source = chunk
                        .slice(token * canonical_chunk_rows..(token + 1) * canonical_chunk_rows);
                    let mut destination =
                        y.slice_mut(token * out_f + r0..token * out_f + r0 + canonical_chunk_rows);
                    self.gpu.stream().memcpy_dtod(&source, &mut destination)?;
                }
            }
        }
        Ok(y)
    }

    /// One-token twin of `linear_f32_resident_canonical_rows` writing into a caller-owned
    /// output. Same cuBLASLt calls, values, and chunk order as the allocating variant at
    /// `m == 1`; only the output residency changes (persistent workspace instead of a fresh
    /// allocation per call). This is the projection substrate of the v2 Step TP decode driver.
    pub fn linear_f32_resident_canonical_rows_t1_into(
        &self,
        x: &CudaSlice<f32>,
        data: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        in_f: usize,
        out_f: usize,
        canonical_chunk_rows: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if data.len() != in_f * out_f {
            return Err(format!("resident F32 values {} != {out_f}x{in_f}", data.len()).into());
        }
        if y.len() != out_f || x.len() != in_f {
            return Err(format!(
                "resident F32 t1 shapes x={} y={} != in {in_f} out {out_f}",
                x.len(),
                y.len()
            )
            .into());
        }
        if canonical_chunk_rows == 0
            || canonical_chunk_rows > out_f
            || out_f % canonical_chunk_rows != 0
        {
            return Err(format!(
                "invalid canonical F32 chunk rows {canonical_chunk_rows} for output width {out_f}"
            )
            .into());
        }
        let input = x.slice(0..x.len());
        for r0 in (0..out_f).step_by(canonical_chunk_rows) {
            let weights = data.slice(r0 * in_f..(r0 + canonical_chunk_rows) * in_f);
            let mut destination = y.slice_mut(r0..r0 + canonical_chunk_rows);
            self.linear_device_into(
                &input,
                &weights,
                &mut destination,
                1,
                in_f,
                canonical_chunk_rows,
            )?;
        }
        Ok(())
    }

    /// One-token view-to-view linear into a caller-owned destination — the `linear` twin
    /// without the allocation, for workspace-resident operands.
    pub fn linear_t1_into(
        &self,
        x: &cudarc::driver::CudaView<'_, f32>,
        w: &cudarc::driver::CudaView<'_, f32>,
        y: &mut cudarc::driver::CudaViewMut<'_, f32>,
        in_f: usize,
        out_f: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        self.linear_device_into(x, w, y, 1, in_f, out_f)
    }

    /// On-device linear: y[m,out] = x[m,in] @ W[out,in]^T, weights row-major [out,in] (ggml).
    /// cuBLASLt col-major mapping (see memra_runtime::Gpu::linear_f32 for the derivation).
    /// DECODE-EXACT float linear: per-column m=1 cuBLASLt calls. cuBLASLt's reduction split is
    /// n-dependent (lt_ndep probe: m=1 vs m=2 col0 differs every bit), so spec-verify batches
    /// must not batch float matmuls the T=1 decode chain runs at m=1. Used by the small-t MoE
    /// router/shexp sites and matmul_decode_exact's Float arm.
    pub fn linear_decode_exact(
        &self,
        x: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        m_tokens: usize,
        in_f: usize,
        out_f: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        if m_tokens == 1 {
            return self.linear(x, w, 1, in_f, out_f);
        }
        let xv = self.view(x, m_tokens * in_f);
        let mut y = self.alloc_uninit::<f32>(m_tokens * out_f)?;
        for t in 0..m_tokens {
            let row = xv.slice(t * in_f..(t + 1) * in_f);
            let mut xr = self.alloc_uninit::<f32>(in_f)?;
            self.copy_view_into(&mut xr, 0, &row, in_f)?;
            let yr = self.linear(&xr, w, 1, in_f, out_f)?;
            self.copy_into(&mut y, t * out_f, &yr, out_f)?;
        }
        Ok(y)
    }

    pub fn linear(
        &self,
        x: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        m_tokens: usize,
        in_f: usize,
        out_f: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        self.linear_device(x, w, m_tokens, in_f, out_f)
    }

    fn linear_device<I>(
        &self,
        x: &I,
        w: &I,
        m_tokens: usize,
        in_f: usize,
        out_f: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>>
    where
        I: cudarc::driver::DevicePtr<f32>,
    {
        let mut c = self.alloc_uninit::<f32>(m_tokens * out_f)?; // cuBLASLt beta=0: C fully written
        self.linear_device_into(x, w, &mut c, m_tokens, in_f, out_f)?;
        Ok(c)
    }

    fn linear_device_into<I, O>(
        &self,
        x: &I,
        w: &I,
        c: &mut O,
        m_tokens: usize,
        in_f: usize,
        out_f: usize,
    ) -> Result<(), Box<dyn std::error::Error>>
    where
        I: cudarc::driver::DevicePtr<f32>,
        O: cudarc::driver::DevicePtrMut<f32>,
    {
        use cudarc::cublaslt::{Matmul, MatmulConfig};
        let cfg = MatmulConfig {
            transa: true,
            transb: false,
            transc: false,
            m: out_f as u64,
            n: m_tokens as u64,
            k: in_f as u64,
            alpha: 1.0,
            lda: in_f as i64,
            ldb: in_f as i64,
            beta: 0.0,
            ldc: out_f as i64,
            stride_a: None,
            stride_b: None,
            stride_c: None,
            stride_bias: None,
            batch_size: None,
        };
        let blas = self.gpu.blas();
        unsafe {
            blas.matmul(cfg, w, x, c, None, None)?;
        }
        Ok(())
    }

    /// Naive SDPA. Q:[head_dim,n_head,T], K/V:[head_dim,n_head_kv,T_kv] -> O:[head_dim,n_head,T].
    ///
    /// LONG-CTX DISPATCH (lane/hermes-perf-fixes, 2026-08-23): the smem kernel's `T_kv*4`
    /// dynamic shared memory exceeds the 48KB launch bound past T_kv=12288 — the plain
    /// full-attn sibling of the DFlash2 B2 crash the windowed layers fixed with
    /// `sdpa_naive_w_lo`. Past the bound this transparently takes the byte-identical
    /// gmem-scores twin (`sdpa_naive_gmem`, kernel_check-pinned) instead of returning the
    /// launch error mid-request.
    pub fn sdpa_naive(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        causal: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if t_kv * 4 > SDPA_NAIVE_SMEM_MAX {
            return self.sdpa_naive_gmem(
                q, k, v, o, head_dim, n_head, n_head_kv, t, t_kv, scale, causal,
            );
        }
        let f = self.func("sdpa_naive_f32");
        let cfg = LaunchConfig {
            grid_dim: (n_head as u32, t as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: (t_kv * 4) as u32,
        };
        let (hd, nh, nhkv, ti, tkvi, cz) = (
            head_dim as i32,
            n_head as i32,
            n_head_kv as i32,
            t as i32,
            t_kv as i32,
            causal as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(v)
            .arg(o)
            .arg(&hd)
            .arg(&nh)
            .arg(&nhkv)
            .arg(&ti)
            .arg(&tkvi)
            .arg(&scale)
            .arg(&cz);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Global-memory-scores twin of [`Self::sdpa_naive`] (lane/hermes-perf-fixes, 2026-08-23).
    /// Same kernel body with the per-(head, query) scores row in a device workspace instead
    /// of dynamic shared memory: identical loop structure and reduction order, so the output
    /// is BYTE-IDENTICAL to the smem kernel wherever both launch (kernel_check
    /// `sdpa_naive_gmem` pins bit-identity plus the >12k arm where the smem kernel MUST
    /// fail). O(n_head * T * T_kv * 4) workspace — fine for the tall-KV block shapes that
    /// hit the bound (dspark/dflash full-attn: T <= block size), guarded so a square
    /// T==T_kv caller cannot silently allocate tens of GB.
    #[allow(clippy::too_many_arguments)]
    pub fn sdpa_naive_gmem(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        causal: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let ws_len = n_head
            .checked_mul(t)
            .and_then(|x| x.checked_mul(t_kv))
            .ok_or("sdpa_naive_gmem: scores workspace size overflow")?;
        let ws_bytes = ws_len
            .checked_mul(std::mem::size_of::<f32>())
            .ok_or("sdpa_naive_gmem: scores workspace byte count overflow")?;
        if ws_bytes > SDPA_NAIVE_GMEM_WS_MAX {
            return Err(format!(
                "sdpa_naive_gmem: scores workspace {ws_bytes} bytes (heads {n_head} x T {t} x \
                 T_kv {t_kv}) exceeds the {SDPA_NAIVE_GMEM_WS_MAX}-byte guard — this shape \
                 needs a tiled/flash kernel, not the naive oracle"
            )
            .into());
        }
        let mut scores = self.uninit(ws_len)?;
        let f = self.func("sdpa_naive_gmem_f32");
        let cfg = LaunchConfig {
            grid_dim: (n_head as u32, t as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (hd, nh, nhkv, ti, tkvi, cz) = (
            head_dim as i32,
            n_head as i32,
            n_head_kv as i32,
            t as i32,
            t_kv as i32,
            causal as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(v)
            .arg(o)
            .arg(&mut scores)
            .arg(&hd)
            .arg(&nh)
            .arg(&nhkv)
            .arg(&ti)
            .arg(&tkvi)
            .arg(&scale)
            .arg(&cz);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Island twin (lane/gemma-vision): causal + sliding-window attention with
    /// bidirectional image islands. `span_id` labels each absolute kv position
    /// (-1 text, >=0 island id); same-island keys are visible unconditionally,
    /// reproducing the reference's non-causal image batch. window 0 = no window.
    #[allow(clippy::too_many_arguments)]
    pub fn sdpa_naive_island(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        o: &mut CudaSlice<f32>,
        span_id: &CudaSlice<i32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        window: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("sdpa_naive_island_f32");
        let cfg = LaunchConfig {
            grid_dim: (n_head as u32, t as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: (t_kv * 4) as u32,
        };
        let (hd, nh, nhkv, ti, tkvi, wi) = (
            head_dim as i32,
            n_head as i32,
            n_head_kv as i32,
            t as i32,
            t_kv as i32,
            window as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(v)
            .arg(o)
            .arg(span_id)
            .arg(&hd)
            .arg(&nh)
            .arg(&nhkv)
            .arg(&ti)
            .arg(&tkvi)
            .arg(&scale)
            .arg(&wi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Windowed sdpa_naive twin (gemma4 R6): masks keys older than q_pos-(window-1).
    #[allow(clippy::too_many_arguments)]
    pub fn sdpa_naive_w(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        causal: bool,
        window: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("sdpa_naive_w_f32");
        let cfg = LaunchConfig {
            grid_dim: (n_head as u32, t as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: (t_kv * 4) as u32,
        };
        let (hd, nh, nhkv, ti, tkvi, cz, wi) = (
            head_dim as i32,
            n_head as i32,
            n_head_kv as i32,
            t as i32,
            t_kv as i32,
            causal as i32,
            window as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(v)
            .arg(o)
            .arg(&hd)
            .arg(&nh)
            .arg(&nhkv)
            .arg(&ti)
            .arg(&tkvi)
            .arg(&scale)
            .arg(&cz)
            .arg(&wi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Lo-clipped windowed sdpa_naive twin (lane/dflash2-longctx, DFLASH2-EVAL §10.6(c)).
    /// Same mask law as `sdpa_naive_w`, but keys below every query's window floor are never
    /// read: kv_lo = max(0, (t_kv - t) + 1 - window) — the oldest key visible to the OLDEST
    /// query row (q_pos = t_kv - t). Dynamic shared memory shrinks from t_kv*4 bytes (which
    /// blows the 48KB launch bound at ~12k rows — the B2 ctx crash) to (t_kv - kv_lo)*4 =
    /// (window - 1 + t)*4, and the key scan drops from O(t_kv) to O(window + t). Output is
    /// byte-identical to `sdpa_naive_w` (masked keys contribute exact zeros to same-order
    /// reductions; kernel_check `sdpa_naive_w_lo` pins it). window == 0 (no window) keeps
    /// kv_lo = 0 and is then shape-identical to the legacy kernel, including its bound.
    #[allow(clippy::too_many_arguments)]
    pub fn sdpa_naive_w_lo(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        causal: bool,
        window: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let kv_lo = if window > 0 {
            (t_kv - t + 1).saturating_sub(window)
        } else {
            0
        };
        let smem = (t_kv - kv_lo) * 4;
        if smem > 48 * 1024 {
            return Err(format!(
                "sdpa_naive_w_lo: window {window} + T {t} rows need {smem} bytes of dynamic \
                 shared memory (> 48KB launch bound) — this kernel clips the OLD side only; \
                 a window this wide needs the multi-pass long-ctx kernel"
            )
            .into());
        }
        let f = self.func("sdpa_naive_w_lo_f32");
        let cfg = LaunchConfig {
            grid_dim: (n_head as u32, t as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: smem as u32,
        };
        let (hd, nh, nhkv, ti, tkvi, cz, wi, lo) = (
            head_dim as i32,
            n_head as i32,
            n_head_kv as i32,
            t as i32,
            t_kv as i32,
            causal as i32,
            window as i32,
            kv_lo as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(v)
            .arg(o)
            .arg(&hd)
            .arg(&nh)
            .arg(&nhkv)
            .arg(&ti)
            .arg(&tkvi)
            .arg(&scale)
            .arg(&cz)
            .arg(&wi)
            .arg(&lo);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// SDPA where K/V are CudaViews into a resident KV cache (decode hot path, no host round-trip).
    pub fn sdpa_naive_view(
        &self,
        q: &CudaSlice<f32>,
        k: &cudarc::driver::CudaView<f32>,
        v: &cudarc::driver::CudaView<f32>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        causal: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("sdpa_naive_f32");
        let cfg = LaunchConfig {
            grid_dim: (n_head as u32, t as u32, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: (t_kv * 4) as u32,
        };
        let (hd, nh, nhkv, ti, tkvi, cz) = (
            head_dim as i32,
            n_head as i32,
            n_head_kv as i32,
            t as i32,
            t_kv as i32,
            causal as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(v)
            .arg(o)
            .arg(&hd)
            .arg(&nh)
            .arg(&nhkv)
            .arg(&ti)
            .arg(&tkvi)
            .arg(&scale)
            .arg(&cz);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Correctness fallback for quantized resident K/V views. Dequantizes K and V once into f32
    /// workspaces, then calls `sdpa_naive`. This is an explicit API: the optimized prefill view
    /// dispatch remains unchanged, so callers can use it as a reference or compatibility path.
    /// Dequant a quantized KV view into caller-owned f32 buffers (one grid-stride launch).
    /// `g` picks the kf8vf8-module stamp for e4m3 caches (same flag contract as fa_decode/
    /// fa_prefill_view). Used by the E4B shared-KV prefill arms (2026-07-31) to feed the
    /// f32 fa_prefill_w / fa_prefill_hd512 twins from the target layer's quantized rows.
    #[allow(clippy::too_many_arguments)]
    pub fn fa_dequant_kv_view_f32(
        &self,
        k: &cudarc::driver::CudaView<u8>,
        v: &cudarc::driver::CudaView<u8>,
        kf: &mut CudaSlice<f32>,
        vf: &mut CudaSlice<f32>,
        kv_dim_k: usize,
        kv_dim_v: usize,
        t_kv: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        g: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = if g {
            self.func_g("fa_dequant_kv_ws_f32")
        } else {
            self.func("fa_dequant_kv_ws_f32")
        };
        let total = (t_kv * (kv_dim_k + kv_dim_v)) as u64;
        let nblk = ((total + 255) / 256).min(65535 * 16) as u32;
        let cfg = LaunchConfig {
            grid_dim: (nblk.max(1), 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (kdk, kdv, tkvi) = (kv_dim_k as i32, kv_dim_v as i32, t_kv as i32);
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(k)
            .arg(v)
            .arg(&mut *kf)
            .arg(&mut *vf)
            .arg(&kdk)
            .arg(&kdv)
            .arg(&tkvi)
            .arg(&ktb)
            .arg(&vtb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    #[allow(clippy::too_many_arguments)]
    pub fn sdpa_naive_quantized_view(
        &self,
        q: &CudaSlice<f32>,
        k: &cudarc::driver::CudaView<u8>,
        v: &cudarc::driver::CudaView<u8>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        causal: bool,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let kv_dim = n_head_kv * head_dim;
        let mut kf = self.uninit(t_kv * kv_dim)?;
        let mut vf = self.uninit(t_kv * kv_dim)?;
        let f = self.func("fa_dequant_kv_ws_f32");
        let total = (2 * t_kv * kv_dim) as u64;
        let nblk = ((total + 255) / 256).min(65535 * 16) as u32;
        let cfg = LaunchConfig {
            grid_dim: (nblk.max(1), 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (kv_dim_i, t_kv_i) = (kv_dim as i32, t_kv as i32);
        let (k_tok_bytes_i, v_tok_bytes_i) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(k)
            .arg(v)
            .arg(&mut kf)
            .arg(&mut vf)
            .arg(&kv_dim_i)
            .arg(&kv_dim_i)
            .arg(&t_kv_i)
            .arg(&k_tok_bytes_i)
            .arg(&v_tok_bytes_i);
        unsafe { b.launch(cfg)? };
        self.sdpa_naive(
            q, &kf, &vf, o, head_dim, n_head, n_head_kv, t, t_kv, scale, causal,
        )
    }

    /// WINDOWED twin of `sdpa_naive_quantized_view` (step35 SWA prefill): dequant the KV byte
    /// view into f32 workspaces with the SAME `fa_dequant_kv_ws_f32` launch, then run
    /// `sdpa_naive_w` instead of `sdpa_naive`. `window == 0` is the unwindowed form (the kernel
    /// treats a non-positive window as "no window mask"), so this is a strict superset of the
    /// unwindowed function above and produces bit-identical output at window == 0.
    ///
    /// Why this exists: EVERY windowed FlashAttention stamp in flash_attn.cu is head_dim-256
    /// only (`fa_prefill_w_f32` == `fa_prefill_f32_body<256>`, and the quantized-view windowed
    /// twins likewise), while step35 is head_dim 128. Its SWA layers therefore have no windowed
    /// FA path and take this f32 floor in v0 — same cache bytes, same numeric class as the
    /// unwindowed quantized-view fallback, so the chunk-invariance contract holds on both.
    #[allow(clippy::too_many_arguments)]
    pub fn sdpa_naive_w_quantized_view(
        &self,
        q: &CudaSlice<f32>,
        k: &cudarc::driver::CudaView<u8>,
        v: &cudarc::driver::CudaView<u8>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        causal: bool,
        window: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let kv_dim = n_head_kv * head_dim;
        let mut kf = self.uninit(t_kv * kv_dim)?;
        let mut vf = self.uninit(t_kv * kv_dim)?;
        let f = self.func("fa_dequant_kv_ws_f32");
        let total = (2 * t_kv * kv_dim) as u64;
        let nblk = ((total + 255) / 256).min(65535 * 16) as u32;
        let cfg = LaunchConfig {
            grid_dim: (nblk.max(1), 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (kv_dim_i, t_kv_i) = (kv_dim as i32, t_kv as i32);
        let (k_tok_bytes_i, v_tok_bytes_i) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(k)
            .arg(v)
            .arg(&mut kf)
            .arg(&mut vf)
            .arg(&kv_dim_i)
            .arg(&kv_dim_i)
            .arg(&t_kv_i)
            .arg(&k_tok_bytes_i)
            .arg(&v_tok_bytes_i);
        unsafe { b.launch(cfg)? };
        self.sdpa_naive_w(
            q, &kf, &vf, o, head_dim, n_head, n_head_kv, t, t_kv, scale, causal, window,
        )
    }

    /// Hand-written FlashAttention prefill (sm_120, FA-2 online softmax on validated mma.sync,
    /// head_dim 256 or 128 (template-stamped twins), GQA, causal). Replaces sdpa_naive for T>1.
    /// Q/K/V/O [head_dim, n_head(_kv), T].
    pub fn fa_prefill(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        causal: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if portable_mma_gated() {
            return self.sdpa_naive(
                q, k, v, o, head_dim, n_head, n_head_kv, t, t_kv, scale, causal,
            );
        }
        // FA3 v10 arm (task #20, OPT-IN MEMRA_FA3=1 — harness-proven 883us vs the shipped
        // kernel's 993us at T=2048): TMA-swizzled wgmma FA, fresh causal hd256 only.
        // NEW NUMERIC CONFIG (GDN-mma precedent): online softmax / bf16-P class — the
        // run-gen argmax + greedy-stream batteries arbitrate; not bit-paired.
        // PROMOTED default-ON hopper (2026-07-27): 3-seed 2048-prime -> 128-decode
        // streams MATCH vs mma, full battery green, lane interleaved 5/5 (+2.4%).
        // MEMRA_FA3=0 reverts; kernel-check pins the mma config regardless.
        let fa3_on = head_dim == 256
            && causal
            && t == t_kv
            && match std::env::var("MEMRA_FA3").as_deref() {
                Ok("0") => false,
                // The force arm consults the arch now: the bf16 stage below calls
                // f32_to_bf16_into -> func("f32_to_bf16_bulk"), which cu/hybrid.cu:1623 omits on
                // a portable build. Refuse at the switch, not at the lookup.
                Ok("1") => {
                    refuse_portable_force("MEMRA_FA3=1", "the sm_90a fa3/bf16 kernels");
                    true
                }
                _ => cfg!(memra_hopper_mma),
            };
        if fa3_on {
            let n = t * n_head * head_dim;
            let nkv = t * n_head_kv * head_dim;
            let mut q16 = self.alloc_u8_uninit(n * 2)?;
            let mut k16 = self.alloc_u8_uninit(nkv * 2)?;
            let mut v16 = self.alloc_u8_uninit(nkv * 2)?;
            self.f32_to_bf16_into(q, &mut q16, n)?;
            self.f32_to_bf16_into(k, &mut k16, nkv)?;
            self.f32_to_bf16_into(v, &mut v16, nkv)?;
            let rc = {
                use cudarc::driver::{DevicePtr, DevicePtrMut};
                let stream = self.gpu.stream();
                let (qp, _g1) = q16.device_ptr(&stream);
                let (kp, _g2) = k16.device_ptr(&stream);
                let (vp, _g3) = v16.device_ptr(&stream);
                let (op, _g4) = o.device_ptr_mut(&stream);
                unsafe {
                    memra_fa3_prefill(
                        qp as *const core::ffi::c_void,
                        kp as *const core::ffi::c_void,
                        vp as *const core::ffi::c_void,
                        op as *mut f32,
                        t as i32,
                        n_head as i32,
                        n_head_kv as i32,
                        head_dim as i32,
                        scale,
                        stream.cu_stream() as *mut core::ffi::c_void,
                    )
                }
            };
            if rc != 0 {
                return Err(format!("memra_fa3_prefill rc={rc}").into());
            }
            return Ok(());
        }
        // FLOOR PORT (P2+P0a+P0b+P1): 4 warps/CTA, BLOCK_Q=64 query rows, BK=32 KV tile,
        // Q-in-reg + register-O, grid.y=n_head_kv (4 Q-heads share staged K/V).
        // P1 plain arm (MEMRA_FA_P1=1 opt-in until the qwen battery): the engine-study body
        // (FA2 schedule + boundary split + swizzle) on the non-windowed lane. bf16 pre-convert.
        static FA_P1: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let fa_p1 = *FA_P1.get_or_init(|| std::env::var("MEMRA_FA_P1").as_deref() == Ok("1"));
        if fa_p1 && head_dim == 256 && !std::env::var("MEMRA_FA_FLOOR").is_ok() {
            const BLOCK_Q: usize = 64;
            const BKX: usize = 32;
            let f = self.func("fa_prefill_bf16_p1");
            let shmem = (2 * (2 * BKX * head_dim + BLOCK_Q * BKX)
                + 4 * (BLOCK_Q * BKX + 2 * BLOCK_Q)) as u32;
            use cudarc::driver::sys::CUfunction_attribute_enum as A;
            f.set_attribute(
                A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                shmem as i32,
            )?;
            let cfg = LaunchConfig {
                grid_dim: (
                    (t as u32 + BLOCK_Q as u32 - 1) / BLOCK_Q as u32,
                    n_head as u32,
                    1,
                ),
                block_dim: (32, 4, 1),
                shared_mem_bytes: shmem,
            };
            let (hd, nh, nhkv, ti, tkvi, cz) = (
                head_dim as i32,
                n_head as i32,
                n_head_kv as i32,
                t as i32,
                t_kv as i32,
                causal as i32,
            );
            let qb = self.f32_to_bf16(q, t * n_head * head_dim)?;
            let kb = self.f32_to_bf16(k, t_kv * n_head_kv * head_dim)?;
            let vb = self.f32_to_bf16(v, t_kv * n_head_kv * head_dim)?;
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(&qb)
                .arg(&kb)
                .arg(&vb)
                .arg(o)
                .arg(&hd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(&ti)
                .arg(&tkvi)
                .arg(&scale)
                .arg(&cz);
            unsafe {
                b.launch(cfg)?;
            }
            return Ok(());
        }
        // Edge 5a (DEFAULT): fa_prefill_f32_pp — register-resident softmax (no sSw smem
        // round-trip), the FA3 softmax-GEMM overlap variant. ncu (pp512): short_scoreboard
        // 4.32->3.47, wait 1.99->1.45, per-call ~577us->~440us (1.31x) at flat 12.1% warps /
        // 255 regs / 2 CTAs (occupancy preserved). Bit-safe: 9B+27B argmax MATCH, rel 2.55e-3
        // vs floor 3.03e-3. MEMRA_FA_FLOOR reverts to the serialized-softmax floor kernel.
        const BK: usize = 32;
        // W2 lane (MEMRA_FA_PP_W2=1, ncu 2026-07-26): 2-warp/32-row CTA tile doubles grid.x —
        // bit-identical per-row math, pure coverage trade for the 6.25%-occupancy starvation.
        let w2 = std::env::var("MEMRA_FA_PP_W2").as_deref() == Ok("1");
        let (block_q, warps, w2_sfx): (usize, u32, &str) =
            if w2 { (32, 2, "_w2") } else { (64, 4, "") };
        // hd128 twins (2026-07-07): the prefill kernels are template-stamped at 256 (original
        // names, dispatch unchanged) and 128 (`_hd128`, the MiniMax-M3 class). Callers gate
        // other head_dims to sdpa_naive before reaching here.
        let hd_sfx = fa_hd_suffix(head_dim)?;
        let floor = std::env::var("MEMRA_FA_FLOOR").is_ok();
        // BF16-KV staging lane (2026-07-26, default ON): the kernel converts K/V to bf16
        // during staging anyway — pre-converting to bf16 mirrors is BIT-IDENTICAL (same
        // __float2bfloat16 values into the same mma) and turns the 67%-of-stalls scalar
        // staging into int4 vector copies. MEMRA_FA_BF16KV=0 reverts.
        let bf16kv = !floor && !w2 && std::env::var("MEMRA_FA_BF16KV").as_deref() != Ok("0");
        let (kb16, vb16) = if bf16kv {
            let n = t_kv * n_head_kv * head_dim;
            let mut kb = self.alloc_u8_uninit(n * 2)?;
            let mut vb = self.alloc_u8_uninit(n * 2)?;
            let fcv = self.func("f32_to_bf16_bulk");
            let ni = n as i64;
            let cfgc = LaunchConfig::for_num_elems((n as u32).div_ceil(4));
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&fcv);
            b.arg(k).arg(&mut kb).arg(&ni);
            unsafe {
                b.launch(cfgc)?;
            }
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&fcv);
            b.arg(v).arg(&mut vb).arg(&ni);
            unsafe {
                b.launch(cfgc)?;
            }
            (Some(kb), Some(vb))
        } else {
            (None, None)
        };
        let f = self.func(&if bf16kv {
            format!("fa_prefill_bf16kv_pp{hd_sfx}")
        } else {
            format!(
                "fa_prefill_f32{}{}{hd_sfx}",
                if floor { "" } else { "_pp" },
                if floor { "" } else { w2_sfx }
            )
        });
        // persistent smem: bf16*(KV_STAGES*(sK + sV) + sP) + f32*(sS + sM + sL);
        // the bf16kv ring doubles the K/V stages (KV_STAGES=2).
        let kv_stages = if bf16kv { 2 } else { 1 };
        let shmem = (2 * (kv_stages * 2 * BK * head_dim + block_q * BK)
            + 4 * (block_q * BK + 2 * block_q)) as u32;
        use cudarc::driver::sys::CUfunction_attribute_enum as A;
        f.set_attribute(
            A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
            shmem as i32,
        )?;
        let cfg = LaunchConfig {
            grid_dim: (
                (t as u32 + block_q as u32 - 1) / block_q as u32,
                n_head as u32,
                1,
            ),
            block_dim: (32, warps, 1),
            shared_mem_bytes: shmem,
        };
        let (hd, nh, nhkv, ti, tkvi, cz) = (
            head_dim as i32,
            n_head as i32,
            n_head_kv as i32,
            t as i32,
            t_kv as i32,
            causal as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q);
        match (&kb16, &vb16) {
            (Some(kb), Some(vb)) => {
                b.arg(kb).arg(vb);
            }
            _ => {
                b.arg(k).arg(v);
            }
        }
        b.arg(o)
            .arg(&hd)
            .arg(&nh)
            .arg(&nhkv)
            .arg(&ti)
            .arg(&tkvi)
            .arg(&scale)
            .arg(&cz);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Windowed FA prefill (gemma4 SWA layers past the sliding window, hd256): fa_prefill's
    /// exact dispatch (pp default, MEMRA_FA_FLOOR seam) with the sliding-window mask + tile
    /// skip in-kernel. Replaces the O(T*T_kv) scalar sdpa_naive_w on the prime path.
    #[allow(clippy::too_many_arguments)]
    pub fn fa_prefill_w(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        causal: bool,
        window: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        // sm_90a rides the mma twins (portable_mma_gated, 2026-07-31 — the raw
        // portable_cuda gate was stale-conservative on Hopper; fa_prefill already flipped).
        if portable_mma_gated() {
            return self.sdpa_naive_w(
                q, k, v, o, head_dim, n_head, n_head_kv, t, t_kv, scale, causal, window,
            );
        }
        // Default: bf16-prestaged twin (same treatment as hd512 — Q/K/V pre-converted once,
        // int4 stage copies; bit-identical, kernel_check-gated). MEMRA_FAW_STAGE=f32 reverts;
        // MEMRA_FA_FLOOR keeps the f32 floor stamp untouched.
        static FAW_F32: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let faw_f32 =
            *FAW_F32.get_or_init(|| std::env::var("MEMRA_FAW_STAGE").as_deref() == Ok("f32"));
        let floor = std::env::var("MEMRA_FA_FLOOR").is_ok();
        self.fa_prefill_w_arm(
            q,
            k,
            v,
            o,
            head_dim,
            n_head,
            n_head_kv,
            t,
            t_kv,
            scale,
            causal,
            window,
            floor || faw_f32,
            floor,
        )
    }

    /// Windowed FA prefill with PRE-CONVERTED bf16 operands (producer-emitted; 31B glue lane).
    /// Launches the P1 stamp directly — callers guarantee qb/kb/vb hold the exact bf16 of q/k/v.
    #[allow(clippy::too_many_arguments)]
    pub fn fa_prefill_w_pre(
        &self,
        qb: &CudaSlice<u8>,
        kb: &CudaSlice<u8>,
        vb: &CudaSlice<u8>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        causal: bool,
        window: usize,
        v_f16: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        const BLOCK_Q: usize = 64;
        const BK: usize = 32;
        debug_assert_eq!(head_dim, 256);
        let hp = fa_f16pv_on() && faw_hp_on() && n_head % 2 == 0 && (n_head / n_head_kv) % 2 == 0;
        debug_assert!(!v_f16 || hp, "f16 V emitted but the SWA hp arm is off");
        if hp {
            const BLOCK_QH: usize = 32;
            // V bytes must be f16 for the h2 stamp; producer normally emits f16 (v_f16),
            // else re-encode through the pooled scratch (stream-ordered reuse).
            let mut vguard = self.fa_vf16_scratch.lock().unwrap();
            let vh: &CudaSlice<u8> = if v_f16 {
                vb
            } else {
                let n = t_kv * n_head_kv * head_dim;
                if vguard.as_ref().map(|b| b.len() < n * 2).unwrap_or(true) {
                    *vguard = Some(self.alloc_uninit::<u8>(n * 2)?);
                }
                self.bf16_to_f16_into(vb, n, vguard.as_mut().unwrap())?;
                vguard.as_ref().unwrap()
            };
            let f = self.func("fa_prefill_w_bf16_p1h2");
            let shmem = (2 * (2 * BK * head_dim + 2 * BLOCK_QH * BK) + 4 * (2 * BLOCK_QH)) as u32;
            use cudarc::driver::sys::CUfunction_attribute_enum as A;
            f.set_attribute(
                A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                shmem as i32,
            )?;
            let cfg = LaunchConfig {
                grid_dim: ((t as u32).div_ceil(BLOCK_QH as u32), (n_head / 2) as u32, 1),
                block_dim: (32, 4, 1),
                shared_mem_bytes: shmem,
            };
            let (hd, nh, nhkv, ti, tkvi, cz, wi) = (
                head_dim as i32,
                n_head as i32,
                n_head_kv as i32,
                t as i32,
                t_kv as i32,
                causal as i32,
                window as i32,
            );
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(qb)
                .arg(kb)
                .arg(vh)
                .arg(o)
                .arg(&hd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(&ti)
                .arg(&tkvi)
                .arg(&scale)
                .arg(&cz)
                .arg(&wi);
            unsafe {
                b.launch(cfg)?;
            }
            return Ok(());
        }
        let f = self.func("fa_prefill_w_bf16_p1");
        let shmem =
            (2 * (2 * BK * head_dim + BLOCK_Q * BK) + 4 * (BLOCK_Q * BK + 2 * BLOCK_Q)) as u32;
        use cudarc::driver::sys::CUfunction_attribute_enum as A;
        f.set_attribute(
            A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
            shmem as i32,
        )?;
        let cfg = LaunchConfig {
            grid_dim: (
                (t as u32 + BLOCK_Q as u32 - 1) / BLOCK_Q as u32,
                n_head as u32,
                1,
            ),
            block_dim: (32, 4, 1),
            shared_mem_bytes: shmem,
        };
        let (hd, nh, nhkv, ti, tkvi, cz, wi) = (
            head_dim as i32,
            n_head as i32,
            n_head_kv as i32,
            t as i32,
            t_kv as i32,
            causal as i32,
            window as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(qb)
            .arg(kb)
            .arg(vb)
            .arg(o)
            .arg(&hd)
            .arg(&nh)
            .arg(&nhkv)
            .arg(&ti)
            .arg(&tkvi)
            .arg(&scale)
            .arg(&cz)
            .arg(&wi);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Windowed FA prefill with the stage arm FORCED — the kernel_check bit-identity entry.
    #[allow(clippy::too_many_arguments)]
    pub fn fa_prefill_w_arm(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        causal: bool,
        window: usize,
        f32_stage: bool,
        floor: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        const BLOCK_Q: usize = 64;
        const BK: usize = 32;
        debug_assert_eq!(head_dim, 256, "fa_prefill_w is stamped hd256 only");
        // P1 (2026-07-22 engine study): per-head Br=64 stamp with the FA2 schedule (V-copy
        // over GEMM0, next-K over softmax+GEMM1) + boundary/interior mask split. FP order
        // preserved -> bit-identical (gated). MEMRA_FAW_P1=0 reverts to the g4/o2 arms.
        static P1_ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let p1 = !floor
            && !f32_stage
            && *P1_ON.get_or_init(|| {
                std::env::var("MEMRA_FAW_P1")
                    .map(|v| v != "0")
                    .unwrap_or(true)
            });
        let hp =
            p1 && fa_f16pv_on() && faw_hp_on() && n_head % 2 == 0 && (n_head / n_head_kv) % 2 == 0;
        if hp {
            const BLOCK_QH: usize = 32;
            let f = self.func("fa_prefill_w_bf16_p1h2");
            let shmem = (2 * (2 * BK * head_dim + 2 * BLOCK_QH * BK) + 4 * (2 * BLOCK_QH)) as u32;
            use cudarc::driver::sys::CUfunction_attribute_enum as A;
            f.set_attribute(
                A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                shmem as i32,
            )?;
            let cfg = LaunchConfig {
                grid_dim: ((t as u32).div_ceil(BLOCK_QH as u32), (n_head / 2) as u32, 1),
                block_dim: (32, 4, 1),
                shared_mem_bytes: shmem,
            };
            let (hd, nh, nhkv, ti, tkvi, cz, wi) = (
                head_dim as i32,
                n_head as i32,
                n_head_kv as i32,
                t as i32,
                t_kv as i32,
                causal as i32,
                window as i32,
            );
            let qb = self.f32_to_bf16(q, t * n_head * head_dim)?;
            let kb = self.f32_to_bf16(k, t_kv * n_head_kv * head_dim)?;
            let vh = self.f32_to_f16(v, t_kv * n_head_kv * head_dim)?;
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(&qb)
                .arg(&kb)
                .arg(&vh)
                .arg(o)
                .arg(&hd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(&ti)
                .arg(&tkvi)
                .arg(&scale)
                .arg(&cz)
                .arg(&wi);
            unsafe {
                b.launch(cfg)?;
            }
            return Ok(());
        }
        if p1 {
            let f = self.func("fa_prefill_w_bf16_p1");
            let shmem =
                (2 * (2 * BK * head_dim + BLOCK_Q * BK) + 4 * (BLOCK_Q * BK + 2 * BLOCK_Q)) as u32;
            use cudarc::driver::sys::CUfunction_attribute_enum as A;
            f.set_attribute(
                A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                shmem as i32,
            )?;
            let cfg = LaunchConfig {
                grid_dim: (
                    (t as u32 + BLOCK_Q as u32 - 1) / BLOCK_Q as u32,
                    n_head as u32,
                    1,
                ),
                block_dim: (32, 4, 1),
                shared_mem_bytes: shmem,
            };
            let (hd, nh, nhkv, ti, tkvi, cz, wi) = (
                head_dim as i32,
                n_head as i32,
                n_head_kv as i32,
                t as i32,
                t_kv as i32,
                causal as i32,
                window as i32,
            );
            let qb = self.f32_to_bf16(q, t * n_head * head_dim)?;
            let kb = self.f32_to_bf16(k, t_kv * n_head_kv * head_dim)?;
            let vb = self.f32_to_bf16(v, t_kv * n_head_kv * head_dim)?;
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(&qb)
                .arg(&kb)
                .arg(&vb)
                .arg(o)
                .arg(&hd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(&ti)
                .arg(&tkvi)
                .arg(&scale)
                .arg(&cz)
                .arg(&wi);
            unsafe {
                b.launch(cfg)?;
            }
            return Ok(());
        }
        // MQA head-grouping (MEMRA_FAW_G4=0 reverts): 4 heads/CTA share the staged K/V —
        // per-(head,row) FP chain identical to the per-head stamp -> bit-identical (gated).
        static G4_ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let g4 = !floor
            && !f32_stage
            && n_head_kv == 1
            && n_head % 4 == 0
            && *G4_ON.get_or_init(|| {
                std::env::var("MEMRA_FAW_G4")
                    .map(|v| v != "0")
                    .unwrap_or(true)
            });
        if g4 {
            const SP_M: usize = 16;
            // Occupancy-2 twin (MEMRA_FAW_O2=0 reverts): one shared K/V buffer inside the dead
            // Q-stage region -> ~36.5KB smem, 2 CTA/SM (the llama hd256 mechanism). Bit-identical.
            static O2_ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
            let o2 = *O2_ON.get_or_init(|| {
                std::env::var("MEMRA_FAW_O2")
                    .map(|v| v != "0")
                    .unwrap_or(true)
            });
            let f = self.func(if o2 {
                "fa_prefill_w_bf16_g4o2"
            } else {
                "fa_prefill_w_bf16_g4"
            });
            let shmem = if o2 {
                (2 * (4 * SP_M * head_dim + 4 * SP_M * BK) + 4 * (4 * SP_M)) as u32
            } else {
                (2 * (2 * BK * head_dim + 4 * SP_M * head_dim + 4 * SP_M * BK) + 4 * (4 * SP_M))
                    as u32
            };
            use cudarc::driver::sys::CUfunction_attribute_enum as A;
            f.set_attribute(
                A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                shmem as i32,
            )?;
            let cfg = LaunchConfig {
                grid_dim: ((t as u32).div_ceil(SP_M as u32), (n_head / 4) as u32, 1),
                block_dim: (32, 4, 1),
                shared_mem_bytes: shmem,
            };
            let (hd, nh, nhkv, ti, tkvi, cz, wi) = (
                head_dim as i32,
                n_head as i32,
                n_head_kv as i32,
                t as i32,
                t_kv as i32,
                causal as i32,
                window as i32,
            );
            let qb = self.f32_to_bf16(q, t * n_head * head_dim)?;
            let kb = self.f32_to_bf16(k, t_kv * n_head_kv * head_dim)?;
            let vb = self.f32_to_bf16(v, t_kv * n_head_kv * head_dim)?;
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(&qb)
                .arg(&kb)
                .arg(&vb)
                .arg(o)
                .arg(&hd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(&ti)
                .arg(&tkvi)
                .arg(&scale)
                .arg(&cz)
                .arg(&wi);
            unsafe {
                b.launch(cfg)?;
            }
            return Ok(());
        }
        let f = self.func(if floor {
            "fa_prefill_w_f32"
        } else if f32_stage {
            "fa_prefill_w_f32_pp"
        } else {
            "fa_prefill_w_bf16_pp"
        });
        let shmem =
            (2 * (2 * BK * head_dim + BLOCK_Q * BK) + 4 * (BLOCK_Q * BK + 2 * BLOCK_Q)) as u32;
        use cudarc::driver::sys::CUfunction_attribute_enum as A;
        f.set_attribute(
            A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
            shmem as i32,
        )?;
        let cfg = LaunchConfig {
            grid_dim: (
                (t as u32 + BLOCK_Q as u32 - 1) / BLOCK_Q as u32,
                n_head as u32,
                1,
            ),
            block_dim: (32, 4, 1),
            shared_mem_bytes: shmem,
        };
        let (hd, nh, nhkv, ti, tkvi, cz, wi) = (
            head_dim as i32,
            n_head as i32,
            n_head_kv as i32,
            t as i32,
            t_kv as i32,
            causal as i32,
            window as i32,
        );
        if f32_stage {
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(q)
                .arg(k)
                .arg(v)
                .arg(o)
                .arg(&hd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(&ti)
                .arg(&tkvi)
                .arg(&scale)
                .arg(&cz)
                .arg(&wi);
            unsafe {
                b.launch(cfg)?;
            }
        } else {
            let qb = self.f32_to_bf16(q, t * n_head * head_dim)?;
            let kb = self.f32_to_bf16(k, t_kv * n_head_kv * head_dim)?;
            let vb = self.f32_to_bf16(v, t_kv * n_head_kv * head_dim)?;
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(&qb)
                .arg(&kb)
                .arg(&vb)
                .arg(o)
                .arg(&hd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(&ti)
                .arg(&tkvi)
                .arg(&scale)
                .arg(&cz)
                .arg(&wi);
            unsafe {
                b.launch(cfg)?;
            }
        }
        Ok(())
    }

    /// hd512 FA prefill (gemma4 GLOBAL layers): BLOCK_Q=32 x 2 warps, Q staged in smem,
    /// grid.z = 2 O-halves (each CTA computes the full 512-dim scores, accumulates half the
    /// V dims). Replaces the scalar sdpa_naive on the prime path's globals.
    #[allow(clippy::too_many_arguments)]
    pub fn fa_prefill_hd512(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        causal: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        // sm_90a rides the mma twins (portable_mma_gated, 2026-07-31 — same flip as _w).
        if portable_mma_gated() {
            return self.sdpa_naive(
                q, k, v, o, head_dim, n_head, n_head_kv, t, t_kv, scale, causal,
            );
        }
        // Default: pre-convert Q/K/V to bf16 once and stage int4 (8 bf16/copy) — at 1 CTA/SM the
        // synchronous stage serializes with compute and MQA re-stages the same K/V per head CTA;
        // pre-converting halves staged bytes and cuts stage instructions 8x. BIT-IDENTICAL to the
        // f32-staged kernel (the converter applies the same __float2bfloat16 the stage applied;
        // kernel_check gates the identity). MEMRA_FA512_STAGE=f32 = rollback to the f32 kernel.
        static F32_STAGE: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let f32_stage =
            *F32_STAGE.get_or_init(|| std::env::var("MEMRA_FA512_STAGE").as_deref() == Ok("f32"));
        // Single-pass arm (MEMRA_FA512_SP=0 reverts to the z=2 bf16 kernel): GEMM0 split-K across
        // the 2 warps instead of recomputed per O-half CTA — the 2026-07-22 kernel-diff excess.
        // Own numeric config (partial-sum order) — battery-gated.
        static SP_ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let sp = !f32_stage
            && *SP_ON.get_or_init(|| {
                std::env::var("MEMRA_FA512_SP")
                    .map(|v| v != "0")
                    .unwrap_or(true)
            });
        self.fa_prefill_hd512_arm(
            q,
            k,
            v,
            o,
            head_dim,
            n_head,
            n_head_kv,
            t,
            t_kv,
            scale,
            causal,
            f32_stage,
            sp,
            sp && fa_f16pv_on(),
        )
    }

    /// hd512 single-pass FA with PRE-CONVERTED bf16 operands (producer-emitted).
    #[allow(clippy::too_many_arguments)]
    pub fn fa_prefill_hd512_pre(
        &self,
        qb: &CudaSlice<u8>,
        kb: &CudaSlice<u8>,
        vb: &CudaSlice<u8>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        causal: bool,
        v_f16: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        debug_assert_eq!(head_dim, 512);
        const SP_M: usize = 16;
        const BKS: usize = 32;
        // f16-P/V door (MEMRA_FA_F16PV=1): P and the P@V accumulation in f16 (llama's fa=1 VKQ
        // class); KQ/softmax/rescale-band/final-normalize stay f32. Own numeric config —
        // battery-gated. V bytes must be f16 for the sp16 kernel (stage/ldmatrix are typeless).
        let f16pv = fa_f16pv_on();
        let nw = if f16pv { fa512_wide_warps() } else { 2 };
        let hp = f16pv && fa512_hp_on() && n_head % 2 == 0 && (n_head / n_head_kv) % 2 == 0;
        debug_assert!(!v_f16 || f16pv, "f16 V emitted without the door on");
        let mut vguard = self.fa_vf16_scratch.lock().unwrap();
        let vref: &CudaSlice<u8> = if f16pv && !v_f16 {
            // Fallback re-encode (producer emitted bf16); the emit lane normally hands f16.
            let n = t_kv * n_head_kv * head_dim;
            let need = n * 2;
            if vguard.as_ref().map(|b| b.len() < need).unwrap_or(true) {
                *vguard = Some(self.alloc_uninit::<u8>(need)?);
            }
            let dst = vguard.as_mut().unwrap();
            self.bf16_to_f16_into(vb, n, dst)?;
            vguard.as_ref().unwrap()
        } else {
            vb
        };
        let f = self.func(if hp {
            "fa_prefill_bf16_hd512_sp16h2"
        } else {
            match (f16pv, nw) {
                (true, 4) => "fa_prefill_bf16_hd512_sp16w4",
                (true, _) => "fa_prefill_bf16_hd512_sp16",
                _ => "fa_prefill_bf16_hd512_sp",
            }
        });
        let (nwarp, npart) = if hp {
            (4usize, 4usize)
        } else if nw > 2 {
            (nw, nw)
        } else {
            (2, 1)
        };
        // h2 drops sQ (Q register-resident) and doubles sP/sS/sL for the head pair.
        let shmem = if hp {
            (2 * (2 * BKS * head_dim + 2 * SP_M * BKS) + 4 * (2 * npart * SP_M * BKS + 2 * SP_M))
                as u32
        } else {
            (2 * (SP_M * head_dim + 2 * BKS * head_dim + SP_M * BKS)
                + 4 * (npart * SP_M * BKS + SP_M)) as u32
        };
        use cudarc::driver::sys::CUfunction_attribute_enum as A;
        f.set_attribute(
            A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
            shmem as i32,
        )?;
        let grid_y = if hp {
            (n_head / 2) as u32
        } else {
            n_head as u32
        };
        let cfg = LaunchConfig {
            grid_dim: ((t as u32).div_ceil(SP_M as u32), grid_y, 1),
            block_dim: (32, nwarp as u32, 1),
            shared_mem_bytes: shmem,
        };
        let (hd, nh, nhkv, ti, tkvi, cz) = (
            head_dim as i32,
            n_head as i32,
            n_head_kv as i32,
            t as i32,
            t_kv as i32,
            causal as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(qb)
            .arg(kb)
            .arg(vref)
            .arg(o)
            .arg(&hd)
            .arg(&nh)
            .arg(&nhkv)
            .arg(&ti)
            .arg(&tkvi)
            .arg(&scale)
            .arg(&cz);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// hd512 FA prefill with the stage/sp arms FORCED — the kernel_check gate entry
    /// (`fa_prefill_hd512` picks the arms from MEMRA_FA512_STAGE / MEMRA_FA512_SP).
    #[allow(clippy::too_many_arguments)]
    pub fn fa_prefill_hd512_arm(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        causal: bool,
        f32_stage: bool,
        sp: bool,
        f16pv: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        debug_assert_eq!(head_dim, 512, "fa_prefill_hd512 is hd512 only");
        if sp && !f32_stage {
            // Single-pass: 16 q-rows/CTA, 2 warps, grid (ceil(T/16), n_head, 1).
            // smem: sQ[16][512] + sK[32][512] + sV[32][512] + sP[16][32] (bf16) + sS[16][32]+sL f32.
            // f16pv: sp16 kernel — f16 P + f16 P@V accum, V operand encoded f16.
            const SP_M: usize = 16;
            const BKS: usize = 32;
            let nw = if f16pv { fa512_wide_warps() } else { 2 };
            let hp = f16pv && fa512_hp_on() && n_head % 2 == 0 && (n_head / n_head_kv) % 2 == 0;
            let f = self.func(if hp {
                "fa_prefill_bf16_hd512_sp16h2"
            } else {
                match (f16pv, nw) {
                    (true, 4) => "fa_prefill_bf16_hd512_sp16w4",
                    (true, _) => "fa_prefill_bf16_hd512_sp16",
                    _ => "fa_prefill_bf16_hd512_sp",
                }
            });
            let (nwarp, npart) = if hp {
                (4usize, 4usize)
            } else if nw > 2 {
                (nw, nw)
            } else {
                (2, 1)
            };
            let shmem = if hp {
                (2 * (2 * BKS * head_dim + 2 * SP_M * BKS)
                    + 4 * (2 * npart * SP_M * BKS + 2 * SP_M)) as u32
            } else {
                (2 * (SP_M * head_dim + 2 * BKS * head_dim + SP_M * BKS)
                    + 4 * (npart * SP_M * BKS + SP_M)) as u32
            };
            use cudarc::driver::sys::CUfunction_attribute_enum as A;
            f.set_attribute(
                A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                shmem as i32,
            )?;
            let grid_y = if hp {
                (n_head / 2) as u32
            } else {
                n_head as u32
            };
            let cfg = LaunchConfig {
                grid_dim: ((t as u32).div_ceil(SP_M as u32), grid_y, 1),
                block_dim: (32, nwarp as u32, 1),
                shared_mem_bytes: shmem,
            };
            let (hd, nh, nhkv, ti, tkvi, cz) = (
                head_dim as i32,
                n_head as i32,
                n_head_kv as i32,
                t as i32,
                t_kv as i32,
                causal as i32,
            );
            let qb = self.f32_to_bf16(q, t * n_head * head_dim)?;
            let kb = self.f32_to_bf16(k, t_kv * n_head_kv * head_dim)?;
            let vb = if f16pv {
                self.f32_to_f16(v, t_kv * n_head_kv * head_dim)?
            } else {
                self.f32_to_bf16(v, t_kv * n_head_kv * head_dim)?
            };
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(&qb)
                .arg(&kb)
                .arg(&vb)
                .arg(o)
                .arg(&hd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(&ti)
                .arg(&tkvi)
                .arg(&scale)
                .arg(&cz);
            unsafe {
                b.launch(cfg)?;
            }
            return Ok(());
        }
        const BLOCK_Q: usize = 32;
        const BK: usize = 32;
        const HALF: usize = 256;
        let f = self.func(if f32_stage {
            "fa_prefill_f32_hd512"
        } else {
            "fa_prefill_bf16_hd512"
        });
        // sQ[32][512] + sK[BK][512] + sV[BK][256] + sP[32][BK] (bf16) + sL[32] f32
        let shmem = (2 * (BLOCK_Q * head_dim + BK * head_dim + BK * HALF + BLOCK_Q * BK)
            + 4 * BLOCK_Q) as u32;
        use cudarc::driver::sys::CUfunction_attribute_enum as A;
        f.set_attribute(
            A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
            shmem as i32,
        )?;
        let cfg = LaunchConfig {
            grid_dim: (
                (t as u32 + BLOCK_Q as u32 - 1) / BLOCK_Q as u32,
                n_head as u32,
                2,
            ),
            block_dim: (32, 2, 1),
            shared_mem_bytes: shmem,
        };
        let (hd, nh, nhkv, ti, tkvi, cz) = (
            head_dim as i32,
            n_head as i32,
            n_head_kv as i32,
            t as i32,
            t_kv as i32,
            causal as i32,
        );
        if f32_stage {
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(q)
                .arg(k)
                .arg(v)
                .arg(o)
                .arg(&hd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(&ti)
                .arg(&tkvi)
                .arg(&scale)
                .arg(&cz);
            unsafe {
                b.launch(cfg)?;
            }
        } else {
            let qb = self.f32_to_bf16(q, t * n_head * head_dim)?;
            let kb = self.f32_to_bf16(k, t_kv * n_head_kv * head_dim)?;
            let vb = self.f32_to_bf16(v, t_kv * n_head_kv * head_dim)?;
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(&qb)
                .arg(&kb)
                .arg(&vb)
                .arg(o)
                .arg(&hd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(&ti)
                .arg(&tkvi)
                .arg(&scale)
                .arg(&cz);
            unsafe {
                b.launch(cfg)?;
            }
        }
        Ok(())
    }

    /// rope_neox2 with bf16 EMIT (31B glue lane): identical rope math/stores plus the post-rope
    /// values written as bf16 — the FA q/k operands come from this launch (bit-identical to the
    /// separate f32_to_bf16 the FA entries would run).
    #[allow(clippy::too_many_arguments)]
    pub fn rope_neox2_bf16e(
        &self,
        q: &mut CudaSlice<f32>,
        k: &mut CudaSlice<f32>,
        qb: &mut CudaSlice<u8>,
        kb: &mut CudaSlice<u8>,
        pos: &CudaSlice<i32>,
        head_dim: usize,
        n_dims: usize,
        nh_q: usize,
        nh_k: usize,
        n_tokens: usize,
        base: f32,
        freq_scale: f32,
        ff: Option<&CudaSlice<f32>>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("rope_neox2_bf16e_f32");
        let rows = ((nh_q + nh_k) * n_tokens) as u32;
        let cfg = LaunchConfig {
            grid_dim: (rows, 1, 1),
            block_dim: ((head_dim / 2) as u32, 1, 1),
            shared_mem_bytes: 0,
        };
        let theta_scale = base.powf(-2.0 / n_dims as f32);
        let (hd, nd, nhq, nhk, nt) = (
            head_dim as i32,
            n_dims as i32,
            nh_q as i32,
            nh_k as i32,
            n_tokens as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        match ff {
            Some(t) => {
                b.arg(&mut *q)
                    .arg(&mut *k)
                    .arg(&mut *qb)
                    .arg(&mut *kb)
                    .arg(pos)
                    .arg(&hd)
                    .arg(&nd)
                    .arg(&nhq)
                    .arg(&nhk)
                    .arg(&nt)
                    .arg(&theta_scale)
                    .arg(&freq_scale)
                    .arg(t);
                unsafe {
                    b.launch(cfg)?;
                }
            }
            None => {
                let null: u64 = 0;
                b.arg(&mut *q)
                    .arg(&mut *k)
                    .arg(&mut *qb)
                    .arg(&mut *kb)
                    .arg(pos)
                    .arg(&hd)
                    .arg(&nd)
                    .arg(&nhq)
                    .arg(&nhk)
                    .arg(&nt)
                    .arg(&theta_scale)
                    .arg(&freq_scale)
                    .arg(&null);
                unsafe {
                    b.launch(cfg)?;
                }
            }
        }
        Ok(())
    }

    /// Flat f32 -> bf16 conversion into a fresh scratch buffer (2 bytes/elem). `n % 4 == 0`
    /// (float4 in, 4x bf16 out). Feeds the bf16-staged hd512 FA prefill.
    pub fn f32_to_bf16(
        &self,
        x: &CudaSlice<f32>,
        n: usize,
    ) -> Result<CudaSlice<u8>, Box<dyn std::error::Error>> {
        assert!(n % 4 == 0, "f32_to_bf16 requires n % 4 == 0, got {n}");
        let mut y = self.alloc_uninit::<u8>(n * 2)?;
        let f = self.func("f32_to_bf16_flat");
        let n_i = n as i64;
        let cfg = LaunchConfig {
            grid_dim: (((n / 4) as u32).div_ceil(256), 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(&mut y).arg(&n_i);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(y)
    }

    pub fn f32_to_f16(
        &self,
        x: &CudaSlice<f32>,
        n: usize,
    ) -> Result<CudaSlice<u8>, Box<dyn std::error::Error>> {
        assert!(n % 4 == 0, "f32_to_f16 requires n % 4 == 0, got {n}");
        let mut y = self.alloc_uninit::<u8>(n * 2)?;
        let f = self.func("f32_to_f16_flat");
        let n_i = n as i64;
        let cfg = LaunchConfig {
            grid_dim: (((n / 4) as u32).div_ceil(256), 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(&mut y).arg(&n_i);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(y)
    }

    /// bf16 bytes -> f16 bytes, n elements (the f16-P/V door's V re-encode on the emit lane).
    pub fn bf16_to_f16(
        &self,
        xb: &CudaSlice<u8>,
        n: usize,
    ) -> Result<CudaSlice<u8>, Box<dyn std::error::Error>> {
        let mut y = self.alloc_uninit::<u8>(n * 2)?;
        self.bf16_to_f16_into(xb, n, &mut y)?;
        Ok(y)
    }

    /// Same conversion into a caller-owned (pooled) buffer; `y.len() >= n*2`.
    pub fn bf16_to_f16_into(
        &self,
        xb: &CudaSlice<u8>,
        n: usize,
        y: &mut CudaSlice<u8>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        assert!(n % 2 == 0, "bf16_to_f16 requires n % 2 == 0, got {n}");
        assert!(y.len() >= n * 2);
        let f = self.func("bf16_to_f16_flat");
        let n2 = (n / 2) as i64;
        let cfg = LaunchConfig {
            grid_dim: (((n / 2) as u32).div_ceil(256), 1, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(xb).arg(y).arg(&n2);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// task #18 (attn side): varlen FA — bf16 K/V mirrors (2 launches) + ONE
    /// fa_prefill_bf16kv launch for every fresh sequence. Same per-block math as the
    /// per-seq path (bit-gateable). Caller guarantees: fresh causal (T_kv == T),
    /// head_dim in {256, 128}, bf16kv lane on.
    #[allow(clippy::too_many_arguments)]
    pub fn fa_prefill_vl8(
        &self,
        seqs: &[FaSeqVl],
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        scale: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        const BK: usize = 32;
        let b = seqs.len();
        assert!(b >= 1 && b <= 8);
        let mut packed = [FaSeqVl::default(); 8];
        packed[..b].copy_from_slice(seqs);
        let v = FaVl8(packed);
        let max_t = seqs.iter().map(|s| s.t).max().unwrap() as u32;
        let ept = (n_head_kv * head_dim) as i32;
        {
            let f = self.func("fa_mirror_vl");
            let max_n = (max_t as i64) * ept as i64;
            let blocks = ((max_n as u32).div_ceil(4)).div_ceil(256);
            for which in 0..2i32 {
                let cfg = LaunchConfig {
                    grid_dim: (blocks, 1, b as u32),
                    block_dim: (256, 1, 1),
                    shared_mem_bytes: 0,
                };
                let __s_lb = self.gpu.stream();
                let mut lb = __s_lb.launch_builder(&f);
                lb.arg(&v).arg(&ept).arg(&which);
                unsafe {
                    lb.launch(cfg)?;
                }
            }
        }
        let hd_sfx = fa_hd_suffix(head_dim)?;
        let f = self.func(&format!("fa_prefill_bf16kv_vl{hd_sfx}"));
        let block_q = 64usize;
        let kv_stages = 2usize;
        let shmem = (2 * (kv_stages * 2 * BK * head_dim + block_q * BK)
            + 4 * (block_q * BK + 2 * block_q)) as u32;
        use cudarc::driver::sys::CUfunction_attribute_enum as A;
        f.set_attribute(
            A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
            shmem as i32,
        )?;
        let cfg = LaunchConfig {
            grid_dim: (max_t.div_ceil(block_q as u32), n_head as u32, b as u32),
            block_dim: (32, 4, 1),
            shared_mem_bytes: shmem,
        };
        let (hd, nh, nhkv) = (head_dim as i32, n_head as i32, n_head_kv as i32);
        let __s_lb = self.gpu.stream();
        let mut lb = __s_lb.launch_builder(&f);
        lb.arg(&v).arg(&hd).arg(&nh).arg(&nhkv).arg(&scale);
        unsafe {
            lb.launch(cfg)?;
        }
        Ok(())
    }

    /// task #18 (attn pre-FA): varlen split + QK-norm + RoPE + KV-append — FOUR launches
    /// for every fresh sequence (was 6 x B, plus the q/k/v split copies which the view
    /// inputs remove entirely). Fresh-only (append at t0=0, RoPE pos = token index).
    #[allow(clippy::too_many_arguments)]
    pub fn attn_pre_vl8(
        &self,
        seqs: &[AttnPreVl],
        wq: &CudaSlice<f32>,
        wk: &CudaSlice<f32>,
        head_dim: usize,
        rope_dims: usize,
        n_head: usize,
        n_head_kv: usize,
        eps: f32,
        freq_base: f32,
        freq_scale: f32,
        kv_dim_k: usize,
        kv_dim_v: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let b = seqs.len();
        assert!(b >= 1 && b <= 8);
        let mut packed = [AttnPreVl::default(); 8];
        packed[..b].copy_from_slice(seqs);
        let v = AttnPreVl8(packed);
        let max_t = seqs.iter().map(|s| s.t).max().unwrap() as u32;
        let (hd, nh, nhkv) = (head_dim as i32, n_head as i32, n_head_kv as i32);
        {
            let f = self.func("q_gate_split_vl");
            let n = max_t * (n_head * head_dim) as u32;
            let cfg = LaunchConfig {
                grid_dim: (n.div_ceil(256), 1, b as u32),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v).arg(&hd).arg(&nh);
            unsafe {
                lb.launch(cfg)?;
            }
        }
        {
            let f = self.func("attn_rms_vl");
            let cfg = LaunchConfig {
                grid_dim: (max_t * n_head as u32, 2, b as u32),
                block_dim: (rms_block(), 1, 1),
                shared_mem_bytes: 0,
            };
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v)
                .arg(wq)
                .arg(wk)
                .arg(&hd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(&eps);
            unsafe {
                lb.launch(cfg)?;
            }
        }
        {
            let f = self.func("attn_rope_vl");
            let theta_scale = freq_base.powf(-2.0 / rope_dims as f32);
            let nd = rope_dims as i32;
            let cfg = LaunchConfig {
                grid_dim: (max_t * n_head as u32, 2, b as u32),
                block_dim: ((head_dim / 2) as u32, 1, 1),
                shared_mem_bytes: 0,
            };
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v)
                .arg(&hd)
                .arg(&nd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(&theta_scale)
                .arg(&freq_scale);
            unsafe {
                lb.launch(cfg)?;
            }
        }
        {
            let f = self.func("append_kv_vl");
            let nblk = (kv_dim_k.max(kv_dim_v) / 32) as u32;
            let cfg = LaunchConfig {
                grid_dim: (nblk, max_t, b as u32),
                block_dim: (32, 1, 1),
                shared_mem_bytes: 0,
            };
            let (kdk, kdv) = (kv_dim_k as i32, kv_dim_v as i32);
            let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v).arg(&kdk).arg(&kdv).arg(&ktb).arg(&vtb);
            unsafe {
                lb.launch(cfg)?;
            }
        }
        Ok(())
    }

    /// FA prefill where K/V are QUANTIZED CudaViews into the resident byte KV cache (the T=K verify
    /// path, MTP-PLAN §D.3). Uses `fa_prefill_q` (inline-dequant during stage-to-smem). The view's
    /// base+offset pointer is honored; the kernel reads [0..t_kv*tok_bytes). Q is the T fresh query
    /// rows; t = T, t_kv = cache len. k_tok_bytes/v_tok_bytes are the per-token byte strides.
    pub fn fa_prefill_view(
        &self,
        q: &CudaSlice<f32>,
        k: &cudarc::driver::CudaView<u8>,
        v: &cudarc::driver::CudaView<u8>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        causal: bool,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        g: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if portable_mma_gated() {
            return self.sdpa_naive_quantized_view(
                q,
                k,
                v,
                o,
                head_dim,
                n_head,
                n_head_kv,
                t,
                t_kv,
                scale,
                causal,
                k_tok_bytes,
                v_tok_bytes,
            );
        }
        const BLOCK_Q: usize = 64;
        const BK: usize = 32;
        // g = e4m3 cache: the kernel parses via DQ_K_ELEM/DQ_V_ELEM (format macros) — the
        // kf8vf8-module stamp reads fp8 with the identical MMA/softmax/PV body.
        let name = format!("fa_prefill_q{}", fa_hd_suffix(head_dim)?);
        let f = if g {
            self.func_g(&name)
        } else {
            self.func(&name)
        };
        let shmem =
            (2 * (2 * BK * head_dim + BLOCK_Q * BK) + 4 * (BLOCK_Q * BK + 2 * BLOCK_Q)) as u32;
        use cudarc::driver::sys::CUfunction_attribute_enum as A;
        f.set_attribute(
            A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
            shmem as i32,
        )?;
        let cfg = LaunchConfig {
            grid_dim: (
                (t as u32 + BLOCK_Q as u32 - 1) / BLOCK_Q as u32,
                n_head as u32,
                1,
            ),
            block_dim: (32, 4, 1),
            shared_mem_bytes: shmem,
        };
        let (hd, nh, nhkv, ti, tkvi, cz) = (
            head_dim as i32,
            n_head as i32,
            n_head_kv as i32,
            t as i32,
            t_kv as i32,
            causal as i32,
        );
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(v)
            .arg(o)
            .arg(&hd)
            .arg(&nh)
            .arg(&nhkv)
            .arg(&ti)
            .arg(&tkvi)
            .arg(&scale)
            .arg(&cz)
            .arg(&ktb)
            .arg(&vtb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// ARC B (2026-07-05): dequant-once chunk-prime FA. Same contract as `fa_prefill_view`, but
    /// instead of every (q-block, head) CTA re-dequanting the whole quantized KV stream inline
    /// (T/64 x n_head redundant at chunk prime — 30.5% of the 32k prime wall), dequant the full
    /// [t_kv, kv_dim] K and V ONCE into a resident bf16 workspace (fa_dequant_kv_ws_bf16), then
    /// run `fa_prefill_qw` (the bf16-workspace twin) over it. EXACT: the workspace holds the same
    /// __float2bfloat16(dq_*_elem(...)) values fa_prefill_q stages to smem, and the twin's MMA/
    /// softmax/PV code is byte-identical -> bit-identical O (kernel_check pins bitdiff=0).
    /// The workspace allocation is REUSED across layers/chunks (grown to the largest shape);
    /// contents are rewritten per call. MEMRA_PRIME_DEQW=0 falls back to fa_prefill_view (callers gate).
    #[allow(clippy::too_many_arguments)]
    pub fn fa_prefill_view_ws(
        &self,
        q: &CudaSlice<f32>,
        k: &cudarc::driver::CudaView<u8>,
        v: &cudarc::driver::CudaView<u8>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        causal: bool,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        g: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if portable_mma_gated() {
            return self.sdpa_naive_quantized_view(
                q,
                k,
                v,
                o,
                head_dim,
                n_head,
                n_head_kv,
                t,
                t_kv,
                scale,
                causal,
                k_tok_bytes,
                v_tok_bytes,
            );
        }
        const BLOCK_Q: usize = 64;
        const BK: usize = 32;
        let kv_dim_k = n_head_kv * head_dim;
        let kv_dim_v = n_head_kv * head_dim;
        let k_ws_bytes = t_kv * kv_dim_k * 2; // bf16
        let v_ws_bytes = t_kv * kv_dim_v * 2;
        // Lock held across BOTH launches: enqueue-only (µs), all compute serializes on gpu.stream.
        let mut guard = self.prime_deqw_ws.lock().unwrap();
        let need_grow = match guard.as_ref() {
            Some((kw, vw)) => kw.len() < k_ws_bytes || vw.len() < v_ws_bytes,
            None => true,
        };
        if need_grow {
            let grow = |cur: usize, need: usize| if cur >= need { cur } else { need };
            let (ck, cv) = guard
                .as_ref()
                .map(|(a, b)| (a.len(), b.len()))
                .unwrap_or((0, 0));
            *guard = Some((
                self.alloc_u8(grow(ck, k_ws_bytes))?,
                self.alloc_u8(grow(cv, v_ws_bytes))?,
            ));
        }
        let (kw, vw) = guard.as_mut().unwrap();
        // pass 1: dequant K+V once into the bf16 workspace (grid-stride, 1 thread/elem)
        {
            // only THIS pass parses KV bytes — pass 2 reads the bf16 workspace (format-free).
            let f = if g {
                self.func_g("fa_dequant_kv_ws_bf16")
            } else {
                self.func("fa_dequant_kv_ws_bf16")
            };
            let total = (t_kv * (kv_dim_k + kv_dim_v)) as u64;
            let nblk = ((total + 255) / 256).min(65535 * 16) as u32;
            let cfg = LaunchConfig {
                grid_dim: (nblk.max(1), 1, 1),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let (kdk, kdv, tkvi) = (kv_dim_k as i32, kv_dim_v as i32, t_kv as i32);
            let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(k)
                .arg(v)
                .arg(&mut *kw)
                .arg(&mut *vw)
                .arg(&kdk)
                .arg(&kdv)
                .arg(&tkvi)
                .arg(&ktb)
                .arg(&vtb);
            unsafe {
                b.launch(cfg)?;
            }
        }
        // pass 2: the bf16-workspace prefill twin (same tile sizes/loop structure as fa_prefill_q).
        // DEFAULT: cp.async double-buffered staging twin (fa_prefill_qw_db, +32KB smem for the
        // second K/V tile pair, 1 CTA/SM): overlaps tile n+1's L2->smem copy with tile n's MMA.
        // Bit-identical output (staging is a pure byte copy; kernel_check pins bitdiff=0 under
        // both twins). A/B (27B rtx6000, N=3): 32k prime 17.10->16.51s, 16k 9.09->8.65s — the copy
        // latency hides behind the MMA pipe and beats the 2-CTA/SM occupancy of the sync twin.
        // MEMRA_PRIME_DEQW_DB=0 falls back to the single-buffer twin.
        let db = std::env::var("MEMRA_PRIME_DEQW_DB")
            .map(|v| v != "0")
            .unwrap_or(true);
        {
            let hd_sfx = fa_hd_suffix(head_dim)?;
            let f = self.func(&format!(
                "fa_prefill_qw{}{hd_sfx}",
                if db { "_db" } else { "" }
            ));
            let shmem = if db {
                // 4x KV tile buffers (bf16) + sP (bf16) + sL (f32)
                (2 * (4 * BK * head_dim + BLOCK_Q * BK) + 4 * BLOCK_Q) as u32
            } else {
                (2 * (2 * BK * head_dim + BLOCK_Q * BK) + 4 * (BLOCK_Q * BK + 2 * BLOCK_Q)) as u32
            };
            use cudarc::driver::sys::CUfunction_attribute_enum as A;
            f.set_attribute(
                A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                shmem as i32,
            )?;
            let cfg = LaunchConfig {
                grid_dim: (
                    (t as u32 + BLOCK_Q as u32 - 1) / BLOCK_Q as u32,
                    n_head as u32,
                    1,
                ),
                block_dim: (32, 4, 1),
                shared_mem_bytes: shmem,
            };
            let (hd, nh, nhkv, ti, tkvi, cz) = (
                head_dim as i32,
                n_head as i32,
                n_head_kv as i32,
                t as i32,
                t_kv as i32,
                causal as i32,
            );
            let (kdk, kdv) = (kv_dim_k as i32, kv_dim_v as i32);
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(q)
                .arg(&*kw)
                .arg(&*vw)
                .arg(o)
                .arg(&hd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(&ti)
                .arg(&tkvi)
                .arg(&scale)
                .arg(&cz)
                .arg(&kdk)
                .arg(&kdv);
            unsafe {
                b.launch(cfg)?;
            }
        }
        Ok(())
    }

    /// WINDOWED `fa_prefill_view_ws` twin at head_dim 128 (lane/pp-prefill 2026-08-07):
    /// step35's SWA prefill (win=512, 33 of 45 layers) previously had NO windowed FA prefill
    /// stamp — every windowed twin was hd256-only — and took `sdpa_naive_w_quantized_view`,
    /// the f32 floor, at 565 ms/layer on a pp4096 where the hd128 FA family does the harder
    /// causal-4096 in 3.3 ms (41% of the whole prime; research/pp-prefill-20260807 anatomy).
    /// Same two-pass shape as the unwindowed function: dequant K/V ONCE into the resident
    /// bf16 workspace, then the windowed qw kernel (`fa_prefill_qw_db_w_hd128`, cp.async
    /// double-buffered; MEMRA_PRIME_DEQW_DB=0 selects the single-buffer twin). The window
    /// mask is `fa_prefill_f32_body`'s exact predicate; `window == 0` is bit-identical to
    /// `fa_prefill_view_ws` by construction (default-arg body). NEW NUMERIC CLASS vs the
    /// f32 floor on SWA rows (bf16 MMA online-softmax vs f32 serial softmax) — adoption is
    /// gated by the full battery, and the class must change UNIFORMLY for a whole request
    /// (kernel selection keys on seq_end, never per chunk — the chunkfix law).
    /// hd128-only deliberately: the only windowed-prefill consumer at another head_dim is
    /// gemma4 (hd256), which already has `fa_prefill_w_f32`.
    #[allow(clippy::too_many_arguments)]
    pub fn fa_prefill_view_ws_w_hd128(
        &self,
        q: &CudaSlice<f32>,
        k: &cudarc::driver::CudaView<u8>,
        v: &cudarc::driver::CudaView<u8>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t: usize,
        t_kv: usize,
        scale: f32,
        causal: bool,
        window: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        assert_eq!(
            head_dim, 128,
            "fa_prefill_view_ws_w_hd128: only the hd128 twin is stamped"
        );
        if portable_mma_gated() {
            return self.sdpa_naive_w_quantized_view(
                q,
                k,
                v,
                o,
                head_dim,
                n_head,
                n_head_kv,
                t,
                t_kv,
                scale,
                causal,
                window,
                k_tok_bytes,
                v_tok_bytes,
            );
        }
        const BLOCK_Q: usize = 64;
        const BK: usize = 32;
        let kv_dim_k = n_head_kv * head_dim;
        let kv_dim_v = n_head_kv * head_dim;
        let k_ws_bytes = t_kv * kv_dim_k * 2; // bf16
        let v_ws_bytes = t_kv * kv_dim_v * 2;
        let mut guard = self.prime_deqw_ws.lock().unwrap();
        let need_grow = match guard.as_ref() {
            Some((kw, vw)) => kw.len() < k_ws_bytes || vw.len() < v_ws_bytes,
            None => true,
        };
        if need_grow {
            let grow = |cur: usize, need: usize| if cur >= need { cur } else { need };
            let (ck, cv) = guard
                .as_ref()
                .map(|(a, b)| (a.len(), b.len()))
                .unwrap_or((0, 0));
            *guard = Some((
                self.alloc_u8(grow(ck, k_ws_bytes))?,
                self.alloc_u8(grow(cv, v_ws_bytes))?,
            ));
        }
        let (kw, vw) = guard.as_mut().unwrap();
        // pass 1: dequant K+V once into the bf16 workspace (identical to fa_prefill_view_ws —
        // the workspace bytes are the SAME __float2bfloat16(dq(...)) values either way).
        {
            let f = self.func("fa_dequant_kv_ws_bf16");
            let total = (t_kv * (kv_dim_k + kv_dim_v)) as u64;
            let nblk = ((total + 255) / 256).min(65535 * 16) as u32;
            let cfg = LaunchConfig {
                grid_dim: (nblk.max(1), 1, 1),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let (kdk, kdv, tkvi) = (kv_dim_k as i32, kv_dim_v as i32, t_kv as i32);
            let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(k)
                .arg(v)
                .arg(&mut *kw)
                .arg(&mut *vw)
                .arg(&kdk)
                .arg(&kdv)
                .arg(&tkvi)
                .arg(&ktb)
                .arg(&vtb);
            unsafe {
                b.launch(cfg)?;
            }
        }
        // pass 2: the WINDOWED qw twin (db default, same as the unwindowed wrapper).
        let db = std::env::var("MEMRA_PRIME_DEQW_DB")
            .map(|v| v != "0")
            .unwrap_or(true);
        {
            let f = self.func(if db {
                "fa_prefill_qw_db_w_hd128"
            } else {
                "fa_prefill_qw_w_hd128"
            });
            let shmem = if db {
                (2 * (4 * BK * head_dim + BLOCK_Q * BK) + 4 * BLOCK_Q) as u32
            } else {
                (2 * (2 * BK * head_dim + BLOCK_Q * BK) + 4 * (BLOCK_Q * BK + 2 * BLOCK_Q)) as u32
            };
            use cudarc::driver::sys::CUfunction_attribute_enum as A;
            f.set_attribute(
                A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                shmem as i32,
            )?;
            let cfg = LaunchConfig {
                grid_dim: (
                    (t as u32 + BLOCK_Q as u32 - 1) / BLOCK_Q as u32,
                    n_head as u32,
                    1,
                ),
                block_dim: (32, 4, 1),
                shared_mem_bytes: shmem,
            };
            let (hd, nh, nhkv, ti, tkvi, cz) = (
                head_dim as i32,
                n_head as i32,
                n_head_kv as i32,
                t as i32,
                t_kv as i32,
                causal as i32,
            );
            let (kdk, kdv, wnd) = (kv_dim_k as i32, kv_dim_v as i32, window as i32);
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(q)
                .arg(&*kw)
                .arg(&*vw)
                .arg(o)
                .arg(&hd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(&ti)
                .arg(&tkvi)
                .arg(&scale)
                .arg(&cz)
                .arg(&kdk)
                .arg(&kdv)
                .arg(&wnd);
            unsafe {
                b.launch(cfg)?;
            }
        }
        Ok(())
    }

    /// FA decode (T=1 split-K) over the resident QUANTIZED KV cache (q8_0 K / q5_1 V) as u8 views.
    /// Replaces sdpa_naive_view for decode; inline-dequants per element. k_tok_bytes/v_tok_bytes are
    /// the per-token byte strides (differ: q8_0=34*nblk, q5_1=24*nblk per token).
    pub fn fa_decode(
        &self,
        q: &CudaSlice<f32>,
        k: &cudarc::driver::CudaView<u8>,
        v: &cudarc::driver::CudaView<u8>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t_kv: usize,
        scale: f32,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        self.fa_decode_kvmod(
            q,
            k,
            v,
            o,
            head_dim,
            n_head,
            n_head_kv,
            t_kv,
            scale,
            k_tok_bytes,
            v_tok_bytes,
            false,
        )
    }

    /// `fa_decode` with an explicit fp8-module flag (`g`): gemma windowed layers under
    /// MEMRA_GEMMA_WKV read an e4m3 cache — every kernel must come from the kf8vf8 module
    /// and the v4 lane (q8_0-hardcoded staging) is excluded.
    #[allow(clippy::too_many_arguments)]
    /// UNIFIED scalar decode launch (fa_decode_f32, nullable-ctr): ONE symbol for host-len
    /// (kvmod eager) and device-len (graph/stream) callers — the textually-identical f32_dc
    /// twin compiled apart and its ULP drift flipped 31B verify argmaxes (2026-07-12).
    #[allow(clippy::too_many_arguments)]
    #[allow(clippy::too_many_arguments)]
    fn fa_decode_scalar_unified(
        &self,
        q: &cudarc::driver::CudaView<f32>,
        k: &cudarc::driver::CudaView<u8>,
        v: &cudarc::driver::CudaView<u8>,
        o: &mut cudarc::driver::CudaViewMut<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t_kv_host: usize,
        t_kv_dev: Option<&CudaSlice<i32>>,
        scale: f32,
        n_splits: usize,
        split_keys: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        g: bool,
        part_o: &mut CudaSlice<f32>,
        part_m: &mut CudaSlice<f32>,
        part_l: &mut CudaSlice<f32>,
        q8_out: Option<(&mut CudaSlice<i8>, &mut CudaSlice<f32>)>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = if g {
            self.func_g("fa_decode_f32")
        } else {
            self.fa_func("fa_decode_f32", head_dim)
        };
        let cfg = LaunchConfig {
            grid_dim: (n_head as u32, n_splits as u32, 1),
            block_dim: (head_dim as u32, 1, 1),
            shared_mem_bytes: (4 * (head_dim + 32)) as u32,
        };
        let (hd, nh, nhkv, nsp) = (
            head_dim as i32,
            n_head as i32,
            n_head_kv as i32,
            n_splits as i32,
        );
        let (ktb, vtb, tkvi, ski) = (
            k_tok_bytes as i64,
            v_tok_bytes as i64,
            t_kv_host as i32,
            split_keys as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        match t_kv_dev {
            Some(d) => {
                b.arg(q)
                    .arg(k)
                    .arg(v)
                    .arg(&mut *part_o)
                    .arg(&mut *part_m)
                    .arg(&mut *part_l)
                    .arg(&hd)
                    .arg(&nh)
                    .arg(&nhkv)
                    .arg(&tkvi)
                    .arg(d)
                    .arg(&scale)
                    .arg(&nsp)
                    .arg(&ski)
                    .arg(&ktb)
                    .arg(&vtb);
                unsafe {
                    b.launch(cfg)?;
                }
            }
            None => {
                let null: u64 = 0;
                b.arg(q)
                    .arg(k)
                    .arg(v)
                    .arg(&mut *part_o)
                    .arg(&mut *part_m)
                    .arg(&mut *part_l)
                    .arg(&hd)
                    .arg(&nh)
                    .arg(&nhkv)
                    .arg(&tkvi)
                    .arg(&null)
                    .arg(&scale)
                    .arg(&nsp)
                    .arg(&ski)
                    .arg(&ktb)
                    .arg(&vtb);
                unsafe {
                    b.launch(cfg)?;
                }
            }
        }
        let cfg2 = LaunchConfig {
            grid_dim: (n_head as u32, 1, 1),
            block_dim: (head_dim as u32, 1, 1),
            shared_mem_bytes: 0,
        };
        if let Some((oq, od)) = q8_out {
            // wave-5b: q8-emitting combine — the wo matmul_pre consumes the pair directly.
            let fc = if g {
                self.func_g("fa_decode_combine_q8_1")
            } else {
                self.fa_func("fa_decode_combine_q8_1", head_dim)
            };
            let __s_b2 = self.gpu.stream();
            let mut b2 = __s_b2.launch_builder(&fc);
            b2.arg(&*part_o)
                .arg(&*part_m)
                .arg(&*part_l)
                .arg(oq)
                .arg(od)
                .arg(&hd)
                .arg(&nh)
                .arg(&nsp);
            unsafe {
                b2.launch(cfg2)?;
            }
            return Ok(());
        }
        let fc = if g {
            self.func_g("fa_decode_combine_f32")
        } else {
            self.fa_func("fa_decode_combine_f32", head_dim)
        };
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&fc);
        b2.arg(&*part_o)
            .arg(&*part_m)
            .arg(&*part_l)
            .arg(o)
            .arg(&hd)
            .arg(&nh)
            .arg(&nsp);
        unsafe {
            b2.launch(cfg2)?;
        }
        Ok(())
    }

    pub fn fa_decode_kvmod(
        &self,
        q: &CudaSlice<f32>,
        k: &cudarc::driver::CudaView<u8>,
        v: &cudarc::driver::CudaView<u8>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t_kv: usize,
        scale: f32,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        g: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let q_view = q.as_view();
        let mut o_view = o.as_view_mut();
        self.fa_decode_kvmod_view(
            &q_view,
            k,
            v,
            &mut o_view,
            head_dim,
            n_head,
            n_head_kv,
            t_kv,
            scale,
            k_tok_bytes,
            v_tok_bytes,
            g,
        )
    }

    /// Row-view entry into `fa_decode_kvmod`. The kernel sees the selected Q/output rows as its
    /// base pointers, so the launch geometry and arithmetic are identical to the owned-slice entry.
    /// Batched fallback callers use this to avoid materializing rows around an otherwise unchanged
    /// per-session KV view and FA launch.
    #[allow(clippy::too_many_arguments)]
    pub fn fa_decode_kvmod_view(
        &self,
        q: &cudarc::driver::CudaView<f32>,
        k: &cudarc::driver::CudaView<u8>,
        v: &cudarc::driver::CudaView<u8>,
        o: &mut cudarc::driver::CudaViewMut<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t_kv: usize,
        scale: f32,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        g: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        // PERF-4: the warp-per-token vec path replaces the scalar element-per-thread fa_decode_f32 —
        // warp-per-token fa_decode_vec_q (grid=(n_head_kv,n_splits), block=(32,gqa_ratio)).
        // The block dequants each KV tile ONCE into smem (bf16) and broadcasts to all gqa Q-head
        // warps -> each KV byte leaves HBM/L2 ~1x/group (vs 4x). ARGS identical; func/grid/block/
        // smem/n_splits differ. fa_decode_f32 stays the bit-reference fallback. Combine is shared.
        //
        // SPLIT-K: the scalar path has grid.x=n_head (32) blocks; the vec path only has
        // grid.x=n_head_kv (8). To avoid starving the GPU at mid ctx, the vec path splits MORE
        // aggressively (64 keys/split vs 256) so grid.y rises and 8*n_splits fills the SMs.
        // At VERY short ctx (t_kv<96) even 1 split can't fill the GPU from 8 KV heads, so the
        // broadcast can't beat the scalar path's 4x-more-blocks latency hiding — fall back to
        // scalar there (measured crossover: vec 0.68x at t_kv=64, 1.23x at t_kv=96, 2.2x at 256).
        // DEFAULT-ON (2026-06-28): clean clock-locked sweep proved vec beats scalar at every
        // t_kv>=96 and the gain WIDENS with ctx (graph decode: +9.5% @128, +11.6% @512, +11.8%
        // @2048) — the KV-byte-broadcast (4x fewer HBM reads/group) compounds as attention grows.
        // MEMRA_NO_FA_VEC forces the scalar bit-reference. Below FA_VEC_MIN_TKV the scalar path's
        // 4x-more-blocks (grid.x=n_head=32 vs n_head_kv=8) hides latency better, so keep scalar there.
        // g + no-v4: the g-module REGISTER twin mis-decodes the gemma windowed shape
        // (root-cause open, jsonl) — only reachable by forcing v4 off (MEMRA_FA_V4_MAX);
        // fall to the exact scalar there instead of the broken register arm.
        let mut fa_vec = std::env::var("MEMRA_NO_FA_VEC").is_err() && t_kv >= fa_vec_min_tkv();
        // hd256 under g: v4-or-scalar. hd128/other under g ride the REGISTER g-lane (the
        // dq_K_lane/dq_V_lane macros are format-aware; v3 is format-gated off, v2/smem
        // arms are excluded under g). Mirrored in kvmod / fa_decode_dc / fa_geom_eager.
        if g && head_dim == 256 && !fa_v4_at(t_kv) {
            fa_vec = false;
        }
        let sp = fa_split_keys(t_kv, n_head_kv);
        let n_splits = if fa_vec {
            ((t_kv + sp - 1) / sp).max(1)
        } else {
            ((t_kv + 255) / 256).max(1)
        };
        let o_len = n_head * n_splits * head_dim;
        let ml_len = n_head * n_splits;
        let mut part_guard = self.fa_part_pool.lock().unwrap();
        if part_guard
            .as_ref()
            .map(|pp| pp.0.len() < o_len || pp.1.len() < ml_len)
            .unwrap_or(true)
        {
            // RETIRE-ON-GROW, never free (#68 root cause, 2026-08-04): captured graphs (the
            // per-session persistent draft graph, the decode/prime graph doors) BAKE these pool
            // buffer addresses. Dropping the old buffers on grow returns them to the async pool,
            // later live allocations land at those addresses, and the next graph REPLAY writes
            // its fa partials over them — the ST serve-spec corruption (acceptance collapse +
            // output corruption began the burst after the trunk's t_kv growth first realloc'd
            // this pool past the draft-capture size; research/fp8ship-20260804). Retiring keeps
            // the baked addresses alive (single-stream: eager writes the new buffers, replays
            // touch only the old — never concurrently). Doubling growth bounds retired VRAM
            // (total retired < final size).
            let old = part_guard.take();
            let (co, cm) = old
                .as_ref()
                .map(|pp| (pp.0.len(), pp.1.len()))
                .unwrap_or((0, 0));
            if let Some(old) = old {
                self.fa_part_retired.lock().unwrap().push(old);
            }
            if std::env::var("MEMRA_DEBUG_FAPOOL").is_ok() {
                eprintln!(
                    "[fa-pool] REALLOC o {} -> {} ml {} -> {} (old retired)",
                    co, o_len, cm, ml_len
                );
            }
            *part_guard = Some((
                self.alloc_uninit::<f32>(o_len.max(2 * co))?,
                self.alloc_uninit::<f32>(ml_len.max(2 * cm))?,
                self.alloc_uninit::<f32>(ml_len.max(2 * cm))?,
            ));
        }
        let pg = part_guard.as_mut().unwrap();
        self.gpu
            .stream()
            .memset_zeros(&mut pg.0.slice_mut(0..o_len))?;
        self.gpu
            .stream()
            .memset_zeros(&mut pg.1.slice_mut(0..ml_len))?;
        self.gpu
            .stream()
            .memset_zeros(&mut pg.2.slice_mut(0..ml_len))?;
        let (part_o, part_m, part_l) = (&mut pg.0, &mut pg.1, &mut pg.2);
        let (part_o, part_m, part_l) = (&mut *part_o, &mut *part_m, &mut *part_l);
        let (hd, nh, nhkv, tkvi, nsp) = (
            head_dim as i32,
            n_head as i32,
            n_head_kv as i32,
            t_kv as i32,
            n_splits as i32,
        );
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        // The vec kernel holds head_dim/32 register accumulators (FA_DEC_MAX_DPL=8 -> head_dim<=256).
        // All shipped models use head_dim=256; fall back to scalar for anything wider rather than
        // silently truncating the accumulator.
        let fa_vec = fa_vec && head_dim <= 512 && head_dim % 32 == 0;
        // hd-512 vec crossover (MEMRA_FA512_MIN, default 512): the DPL16 twin wins at depth
        // (82.5 -> vec at 1736) but the scalar's more-blocks latency hiding wins at tiny t_kv
        // (the same scalar-floor physics as hd256's old 96 floor; short-ctx plain regressed
        // 178.4 -> 173.7 when 512 rode vec unconditionally).
        let fa512_min = fa512_min_tkv();
        // FA-DEEP pick (bit-identical twins, see fa_deep_at): default module only — the
        // g-module keeps the v4 pick (its class is not the depth-decay class).
        let deep = fa_vec
            && head_dim == 256
            && fa_v4_at(t_kv)
            && !g
            && fa_deep_at(t_kv)
            && !matches!(fa_v4_mode(), "noB3" | "stage");
        let (f, cfg) = if fa_vec && head_dim == 512 && t_kv >= fa512_min {
            // gemma4 globals (hd 512): the DPL16 register twin (fa_decode_vec_q body with a
            // 16-slot accumulator ceiling). Scalar fallback measured 82.5us/layer at 1736 ctx.
            let gqa = (n_head / n_head_kv).max(1) as u32;
            let fv = self.fa_func("fa_decode_vec_q_dpl16", head_dim);
            (
                fv,
                LaunchConfig {
                    grid_dim: (n_head_kv as u32, n_splits as u32, 1),
                    block_dim: (32, gqa, 1),
                    shared_mem_bytes: 0,
                },
            )
        } else if fa_vec && head_dim <= 256 {
            let gqa = (n_head / n_head_kv).max(1) as u32;
            // DEEP-CTX smem twin (2026-07-05): the register-dequant path's GQA reuse rides L2,
            // which holds to ~8k ctx but dies at 40k (layer KV ~37MB) — the 4 GQA warps then
            // re-read every KV byte from DRAM (4x traffic). Above MEMRA_FA_SMEM_TKV (default
            // 1024 — the 2026-07-05 crossover re-sweep on real prompts: p3 spec 73.8->79.2 at
            // 2048, flat down to 512, p2 +5%, p1/9B unchanged; the ARC-A probe's synthetic
            // 2.1x smem-at-all-depths pointed here; 0=never) dispatch the smem-broadcast twin:
            // dequant each tile ONCE per block.
            // Bit-identical per (token,split): same bf16 round-trip, same accumulation order,
            // same partial layout -> same combine. Short/mid ctx keeps the register path (it won
            // there by 12x — latency, not bandwidth, rules small KV).
            static SMEM_TKV: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
            let smem_tkv = *SMEM_TKV.get_or_init(|| {
                std::env::var("MEMRA_FA_SMEM_TKV")
                    .ok()
                    .and_then(|v| v.parse().ok())
                    .unwrap_or_else(|| {
                        FA_SMEM_TKV_DEFAULT.load(std::sync::atomic::Ordering::Relaxed)
                    })
            });
            if fa_v4_at(t_kv) && head_dim == 256 {
                // FA v4 lane (2026-07-10): key-per-lane score phase, zero shuffles per key.
                // NEW NUMERIC CONFIG (chunk-serial per-key dot) — battery-arbitrated.
                // g (fp8-windowed): the v4 staging is format-aware (2026-07-12) — kf8vf8 module.
                let v4name = match fa_v4_mode() {
                    "noB3" => "fa_decode_vec_q_v4_noB3", // phase probe (WRONG OUTPUT)
                    "stage" => "fa_decode_vec_q_v4_stage", // phase probe (WRONG OUTPUT)
                    _ if deep => "fa_decode_vec_q_v4_deep",
                    _ => "fa_decode_vec_q_v4",
                };
                let fv = if g {
                    self.func_g(v4name)
                } else {
                    self.func(v4name)
                };
                // fa_v4_smem (deep: fa_v4_deep_smem, +640B row pads) + sV (g: raw e4m3 sV
                // tile = 1B/elem — half the smem, 3->5 blocks/SM)
                let shmem = (if deep { 12160 } else { 11520 }
                    + 32 * head_dim * if g { 1 } else { 2 }) as u32;
                use cudarc::driver::sys::CUfunction_attribute_enum as A;
                fv.set_attribute(
                    A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                    shmem as i32,
                )?;
                (
                    fv,
                    LaunchConfig {
                        grid_dim: (n_head_kv as u32, n_splits as u32, 1),
                        block_dim: (32, gqa, 1),
                        shared_mem_bytes: shmem,
                    },
                )
            } else if fa_v3_active(head_dim) {
                // FA v3 lane: dp4a-K hybrid (register-quantized Q, raw q8_0 K, staged-V kept).
                // smem = sV only (half of v2's).
                let fv = if g {
                    self.func_g("fa_decode_vec_q_v3")
                } else {
                    self.func("fa_decode_vec_q_v3")
                };
                let shmem = (32 * head_dim * 2) as u32; // sV bf16 [FA_DEC_TILE=32][hd]
                (
                    fv,
                    LaunchConfig {
                        grid_dim: (n_head_kv as u32, n_splits as u32, 1),
                        block_dim: (32, gqa, 1),
                        shared_mem_bytes: shmem,
                    },
                )
            } else if fa_v2_on() {
                // FAVENDOR lane: llama fattn-vec tile-batched softmax + wide-load staging on
                // OUR smem KV broadcast. Replaces BOTH per-key twins when on; same grid/block/
                // partials; same 32KB sK+sV tile as the smem twin.
                let fv = if g {
                    self.func_g("fa_decode_vec_q_v2")
                } else {
                    self.func("fa_decode_vec_q_v2")
                };
                let shmem = (2 * 32 * head_dim * 2) as u32; // sK+sV bf16 [FA_DEC_TILE=32][hd]
                (
                    fv,
                    LaunchConfig {
                        grid_dim: (n_head_kv as u32, n_splits as u32, 1),
                        block_dim: (32, gqa, 1),
                        shared_mem_bytes: shmem,
                    },
                )
            } else if smem_tkv > 0 && t_kv >= smem_tkv && !g && !(head_dim == 512 && Self::gkv_on())
            {
                // (fp8 exclusions: the smem twin's V-stage is q5_1-hardcoded — neither the wkv
                // windowed layers (g) nor the gkv globals (hd512) may be forced onto it via
                // MEMRA_FA_SMEM_TKV; they fall through to the format-clean register/scalar arms.)
                let fv = if g {
                    self.func_g("fa_decode_vec_q_smem")
                } else {
                    self.func("fa_decode_vec_q_smem")
                };
                let shmem = (2 * 32 * head_dim * 2) as u32; // sK+sV bf16 [FA_DEC_TILE=32][hd]
                use cudarc::driver::sys::CUfunction_attribute_enum as A;
                fv.set_attribute(
                    A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                    shmem as i32,
                )?;
                (
                    fv,
                    LaunchConfig {
                        grid_dim: (n_head_kv as u32, n_splits as u32, 1),
                        block_dim: (32, gqa, 1),
                        shared_mem_bytes: shmem,
                    },
                )
            } else {
                // REGISTER-DEQUANT kernel (2026-07-03): per-warp direct q8_0/q5_1 register
                // dequant, zero dynamic shared memory.
                let fv = if g {
                    self.func_g("fa_decode_vec_q")
                } else {
                    self.func("fa_decode_vec_q")
                };
                (
                    fv,
                    LaunchConfig {
                        grid_dim: (n_head_kv as u32, n_splits as u32, 1),
                        block_dim: (32, gqa, 1),
                        shared_mem_bytes: 0,
                    },
                )
            }
        } else {
            // UNIFIED scalar (nullable-ctr symbol shared with graph/stream callers). The
            // split ladder value rides along so ns_eff reproduces THIS n_splits in-kernel.
            return self.fa_decode_scalar_unified(
                q,
                k,
                v,
                o,
                head_dim,
                n_head,
                n_head_kv,
                t_kv,
                None,
                scale,
                n_splits,
                if fa_vec { sp } else { 256 },
                k_tok_bytes,
                v_tok_bytes,
                g,
                part_o,
                part_m,
                part_l,
                None,
            );
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(v)
            .arg(&mut *part_o)
            .arg(&mut *part_m)
            .arg(&mut *part_l)
            .arg(&hd)
            .arg(&nh)
            .arg(&nhkv)
            .arg(&tkvi)
            .arg(&scale)
            .arg(&nsp)
            .arg(&ktb)
            .arg(&vtb);
        unsafe {
            b.launch(cfg)?;
        }
        // (combine re-tile refuted in the fa-deep lane — flat/worse both shapes; the v4
        // combine stays for all arms. Receipts research/fa-decode-deep-20260802/.)
        let (fc, cfg2) = (
            if g {
                self.func_g("fa_decode_combine_f32")
            } else {
                self.fa_func("fa_decode_combine_f32", head_dim)
            },
            LaunchConfig {
                grid_dim: (n_head as u32, 1, 1),
                block_dim: (head_dim as u32, 1, 1),
                shared_mem_bytes: 0,
            },
        );
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&fc);
        b2.arg(&*part_o)
            .arg(&*part_m)
            .arg(&*part_l)
            .arg(o)
            .arg(&hd)
            .arg(&nh)
            .arg(&nsp);
        unsafe {
            b2.launch(cfg2)?;
        }
        Ok(())
    }

    /// BATCHED-TICK increment 2: ONE fa_decode launch covering ALL B sequences of the
    /// batched decode step (blockIdx.z = sequence). Per-seq K/V cache bases ride a device
    /// pointer table (`kv_ptrs`, [2B] interleaved k0,v0,...); per-seq key bounds ride the
    /// tick's position table (`pos_seq`, T_kv = pos+1). v4-lane only: the CALLER
    /// (decode_batch) gates every row through `fa_seqs_eligible` AND one `fa_split_keys`
    /// rung (`split_keys`), so each sequence's split partition, key walk and combine order
    /// reproduce its per-seq eager v4 program exactly (kernel-check pins seqs-vs-loop bit
    /// identity; decode-batch-gate strict pins the whole tick vs decode_step_h).
    /// q is the stacked [B, n_head, head_dim] tick buffer read in place (no per-seq q
    /// copies); o is written [B, n_head, head_dim] in place (no per-seq a copies).
    #[allow(clippy::too_many_arguments)]
    pub fn fa_decode_batch_seqs_v4(
        &self,
        q: &CudaSlice<f32>,
        kv_ptrs: &cudarc::driver::CudaView<u64>,
        pos_seq: &CudaSlice<i32>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        b_n: usize,
        t_kv_max: usize,
        scale: f32,
        split_keys: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        debug_assert!(head_dim == 256, "seqs twin is v4-stamped (hd256 only)");
        let n_splits_max = (t_kv_max + split_keys - 1) / split_keys;
        let o_len = b_n * n_head * n_splits_max * head_dim;
        let ml_len = b_n * n_head * n_splits_max;
        let mut part_guard = self.fa_part_pool.lock().unwrap();
        if part_guard
            .as_ref()
            .map(|pp| pp.0.len() < o_len || pp.1.len() < ml_len)
            .unwrap_or(true)
        {
            // RETIRE-ON-GROW, never free (#68 root cause, 2026-08-04): captured graphs (the
            // per-session persistent draft graph, the decode/prime graph doors) BAKE these pool
            // buffer addresses. Dropping the old buffers on grow returns them to the async pool,
            // later live allocations land at those addresses, and the next graph REPLAY writes
            // its fa partials over them — the ST serve-spec corruption (acceptance collapse +
            // output corruption began the burst after the trunk's t_kv growth first realloc'd
            // this pool past the draft-capture size; research/fp8ship-20260804). Retiring keeps
            // the baked addresses alive (single-stream: eager writes the new buffers, replays
            // touch only the old — never concurrently). Doubling growth bounds retired VRAM
            // (total retired < final size).
            let old = part_guard.take();
            let (co, cm) = old
                .as_ref()
                .map(|pp| (pp.0.len(), pp.1.len()))
                .unwrap_or((0, 0));
            if let Some(old) = old {
                self.fa_part_retired.lock().unwrap().push(old);
            }
            if std::env::var("MEMRA_DEBUG_FAPOOL").is_ok() {
                eprintln!(
                    "[fa-pool] REALLOC o {} -> {} ml {} -> {} (old retired)",
                    co, o_len, cm, ml_len
                );
            }
            *part_guard = Some((
                self.alloc_uninit::<f32>(o_len.max(2 * co))?,
                self.alloc_uninit::<f32>(ml_len.max(2 * cm))?,
                self.alloc_uninit::<f32>(ml_len.max(2 * cm))?,
            ));
        }
        let pg = part_guard.as_mut().unwrap();
        self.gpu
            .stream()
            .memset_zeros(&mut pg.0.slice_mut(0..o_len))?;
        self.gpu
            .stream()
            .memset_zeros(&mut pg.1.slice_mut(0..ml_len))?;
        self.gpu
            .stream()
            .memset_zeros(&mut pg.2.slice_mut(0..ml_len))?;
        let (part_o, part_m, part_l) = (&mut pg.0, &mut pg.1, &mut pg.2);
        let (hd, nh, nhkv) = (head_dim as i32, n_head as i32, n_head_kv as i32);
        let (nspm, spk) = (n_splits_max as i32, split_keys as i32);
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let gqa = (n_head / n_head_kv).max(1) as u32;
        let f = self.func("fa_decode_vec_q_seqs_v4");
        // fa_v4_smem (11520B) + sV bf16 tile — the v4 eager arm's sizing on the default module.
        let shmem = (11520 + 32 * head_dim * 2) as u32;
        use cudarc::driver::sys::CUfunction_attribute_enum as A;
        f.set_attribute(
            A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
            shmem as i32,
        )?;
        let cfg = LaunchConfig {
            grid_dim: (n_head_kv as u32, n_splits_max as u32, b_n as u32),
            block_dim: (32, gqa, 1),
            shared_mem_bytes: shmem,
        };
        {
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(q)
                .arg(kv_ptrs)
                .arg(pos_seq)
                .arg(&mut *part_o)
                .arg(&mut *part_m)
                .arg(&mut *part_l)
                .arg(&hd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(&scale)
                .arg(&nspm)
                .arg(&spk)
                .arg(&ktb)
                .arg(&vtb);
            unsafe {
                b.launch(cfg)?;
            }
        }
        let fc = self.func("fa_decode_combine_seqs");
        let cfg2 = LaunchConfig {
            grid_dim: (n_head as u32, b_n as u32, 1),
            block_dim: (head_dim as u32, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&fc);
        b2.arg(&*part_o)
            .arg(&*part_m)
            .arg(&*part_l)
            .arg(o)
            .arg(&hd)
            .arg(&nh)
            .arg(pos_seq)
            .arg(&nspm)
            .arg(&spk);
        unsafe {
            b2.launch(cfg2)?;
        }
        Ok(())
    }

    /// BATCHED-TICK increment 2: z-batched decode KV append — one launch appends this
    /// step's B rows, each into ITS OWN sequence cache at slot pos_seq[z], through the same
    /// [2B] interleaved pointer table the seqs FA reads. Each (block, z) warp executes the
    /// per-token appender's exact warp program on row z of the stacked [B, kv_dim] k/v —
    /// written cache bytes are BIT-IDENTICAL to the B per-seq calls it replaces
    /// (kernel-check pins the bytes). Default flash module only (callers exclude fp8-KV).
    #[allow(clippy::too_many_arguments)]
    pub fn append_kv_quantized_seqs(
        &self,
        k_rows: &CudaSlice<f32>,
        v_rows: &CudaSlice<f32>,
        kv_ptrs: &cudarc::driver::CudaView<u64>,
        pos_seq: &CudaSlice<i32>,
        b_n: usize,
        kv_dim_k: usize,
        kv_dim_v: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("append_quantize_kv_q8_0_q5_1_seqs");
        let nblk = (kv_dim_k.max(kv_dim_v) / 32) as u32;
        let cfg = LaunchConfig {
            grid_dim: (nblk, b_n as u32, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (kdk, kdv) = (kv_dim_k as i32, kv_dim_v as i32);
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(k_rows)
            .arg(v_rows)
            .arg(kv_ptrs)
            .arg(pos_seq)
            .arg(&kdk)
            .arg(&kdv)
            .arg(&ktb)
            .arg(&vtb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// True iff the MULTI-ROW verify FA (`fa_decode_rows`) is usable for a verify batch whose
    /// FIRST row attends `base_len + 1` keys: every row must take the SAME kernel eager decode
    /// would (the vec path) — mirrors fa_decode's gate exactly (MEMRA_NO_FA_VEC + FA_VEC_MIN_TKV +
    /// head_dim), evaluated at the MINIMUM row bound so no row could have picked scalar.
    /// MEMRA_FA_ROWS_OFF=1 is the A/B + fallback seam (per-row loop).
    pub fn fa_rows_eligible(&self, base_len: usize, head_dim: usize) -> bool {
        std::env::var("MEMRA_NO_FA_VEC").is_err()
            && std::env::var("MEMRA_FA_ROWS_OFF").is_err()
            && base_len + 1 >= fa_vec_min_tkv()
            && head_dim <= 256
            && head_dim % 32 == 0
    }

    /// MULTI-ROW verify FA: run fa_decode_vec_q's EXACT per-row program for T causal query rows
    /// (row r attends keys [0..base_len+r+1)) in ONE kernel launch with grid.z = row, plus ONE
    /// row-batched combine. Replaces the T separate (fa_decode + combine) launches of the spec
    /// verify — same per-row split partition (n_splits_r = ceil(t_kv_r/split_keys), the
    /// fa_split_keys formula), same key-walk order, same reduce shapes => bit-identical outputs
    /// per row (kernel-check pins rows-vs-loop byte identity; run-spec is the end gate).
    /// Caller must have checked `fa_rows_eligible(base_len, head_dim)`.
    /// q is the verify's token-major [T, n_head, head_dim] stack; o is written [T, n_head, head_dim].
    #[allow(clippy::too_many_arguments)]
    pub fn fa_decode_rows(
        &self,
        q: &CudaSlice<f32>,
        k: &cudarc::driver::CudaView<u8>,
        v: &cudarc::driver::CudaView<u8>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        base_len: usize,
        t: usize,
        scale: f32,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        // hd512 dpl16 twin is DEVICE-LEN (graph arc): base_dev/plus feed the
        // kernel; host base_len keeps sizing the splits/partials. hd256 twins
        // keep the host arg. None is a bug for hd512 (asserted below).
        base_dev: Option<(&CudaSlice<i32>, i32)>,
        // K and V planes hold the same values (gemma globals, wv:=wk): pick
        // the _kv twin — V plane never read, value rides the q8_0 key dq.
        kv_shared: bool,
        // this layer's cache is e4m3 (gemma windowed under wkv): resolve the
        // hd256 rows kernel from the kf8vf8 module. PER-CALL — a global env
        // check here hijacked qwen/kernel-check hd256 rows (8 FAILs, 230ebbe).
        g: bool,
        // t=1 decode arm only: emit (int8, per-32 scales) from the dc combine
        // (hd512 path) — the standalone quantize launch folds away.
        mut q8_out: Option<(&mut CudaSlice<i8>, &mut CudaSlice<f32>)>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        debug_assert!(base_len + 1 >= fa_vec_min_tkv() && head_dim <= 512 && head_dim % 32 == 0);
        let t_kv_max = base_len + t; // LAST row's key bound
        let mut sp = fa_split_keys(t_kv_max, n_head_kv); // env/default — same value every row
        // hd512 split override (MEMRA_FA_SP512, 2026-07-11): gemma globals have n_head_kv=2 so
        // the grid is (2 x n_splits) — at depth ~29 splits = 58 blocks on 82 SMs (half idle,
        // rows_dpl16 8x off its byte floor). EVERY gemma hd512 caller shares THIS wrapper
        // (parity law), so the partition is freely tunable — verify and decode move together.
        if head_dim == 512 {
            static SP512: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
            // default 16 (2026-07-11 depth sweep, N=2: plain 155.4->156.5, depth spec
            // 236.9->250.4; 12/24/32 all worse). hd512 exists only on gemma globals.
            let v = *SP512.get_or_init(|| {
                std::env::var("MEMRA_FA_SP512")
                    .ok()
                    .and_then(|x| x.parse().ok())
                    .unwrap_or(0)
            });
            sp = if v >= 8 {
                v
            } else {
                FA_SP512_DEFAULT.load(std::sync::atomic::Ordering::Relaxed)
            };
        }
        let (hd, nh, nhkv) = (head_dim as i32, n_head as i32, n_head_kv as i32);
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let gqa = (n_head / n_head_kv).max(1) as u32;
        // LADDER-RUNG STRADDLE FIX (issue #10, 2026-07-13, rtx6000-proven): one sp for every row
        // diverges from eager decode when a split-ladder rung falls INSIDE the batch — row r's
        // eager twin used fa_split_keys(t_kv_r), the batch used fa_split_keys(t_kv_max), and
        // the different partition changes the combine's FP order (greedy tie flips at depth;
        // MEMRA_FA_SPLIT=64 pin -> PASS on the exact rtx6000 failing config). Fix: group
        // consecutive rows by their OWN ladder value and launch once per group — each row then
        // executes the exact per-row program eager ran. Rungs land once per doubling, so this
        // is 1 launch in the common case and 2 on a crossing round. hd512 keeps one group (its
        // sp override is t_kv-independent by construction).
        let mut groups: Vec<(usize, usize, usize)> = Vec::new(); // (row0, t_g, sp_g)
        if head_dim == 512 || fa_split_keys(base_len + 1, n_head_kv) == sp {
            groups.push((0, t, sp));
        } else {
            let mut r0 = 0usize;
            while r0 < t {
                let sp_g = fa_split_keys(base_len + r0 + 1, n_head_kv);
                let mut r1 = r0 + 1;
                while r1 < t && fa_split_keys(base_len + r1 + 1, n_head_kv) == sp_g {
                    r1 += 1;
                }
                groups.push((r0, r1 - r0, sp_g));
                r0 = r1;
            }
        }
        // Deep-ctx smem twin for the VERIFY rows (2026-07-05): same threshold + rationale as
        // fa_decode's dispatch — at 40k the register path's GQA L2-reuse premise is dead and the
        // verify multiplies the 4x DRAM re-read by T rows. Bit-identical per (row,token,split).
        static SMEM_TKV_R: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
        let smem_tkv = *SMEM_TKV_R.get_or_init(|| {
            std::env::var("MEMRA_FA_SMEM_TKV")
                .ok()
                .and_then(|v| v.parse().ok())
                .unwrap_or_else(|| FA_SMEM_TKV_DEFAULT.load(std::sync::atomic::Ordering::Relaxed))
        });
        let v4 = fa_v4_at(base_len + t) && head_dim == 256;
        let v3 = fa_v3_active(head_dim);
        let smem_rows =
            head_dim <= 256 && !v3 && !fa_v2_on() && smem_tkv > 0 && t_kv_max >= smem_tkv;
        // kv_shared twin RETIRED (2026-07-11 depth run-gen gate): the wv:=wk premise fails
        // POST-cache — cached K is k-normed+roped, cached V is not; the twin fed roped keys
        // in as values. Verify/decode/stream gates were blind (both sides shared the wrong
        // symbol — the parity law's blind spot); only prefill-vs-decode at depth caught it.
        let _ = kv_shared;
        // i2 twin: 2-key interleaved walk (MEMRA_FA_I2=0 reverts). i4 probed NEGATIVE
        // (157.3 vs 161.2 depth plain — register pressure past i2's sweet spot; jsonl).
        let i2 = head_dim == 512 && std::env::var("MEMRA_FA_I2").as_deref() != Ok("0");
        // v4-hd512 (MEMRA_FA_V512=1 opt-in, 2026-07-14): the v4 key-per-lane recipe on the
        // globals lane (depth profile: i2 ~4.6x off its byte floor — the v3-class
        // reduce-per-key latency signature). NEW NUMERIC CONFIG shared by every hd512
        // caller (decode+verify flip together); run-gen argmax + acceptance arbitrate.
        // T-BATCHED hd512 (DEFAULT ON 2026-07-14, MEMRA_FA_TB512=0 seam): one block per
        // (kv_head, split) stages its tile once and loops the rows over it — kills the
        // x t DRAM re-read of the full-ctx globals (depth cell +1.4%, plain flat, N=3
        // interleaved). FIXED absolute partition = NEW NUMERIC for the combine order,
        // shared by every hd512 caller through this wrapper (decode+verify flip together;
        // depth stream identical, acceptance unshifted, spec 256/256 x3 models).
        // Requires sp <= 32 (single staged tile; acc reused per row). The z-form v4_512
        // sibling (in-kernel dp4a port alone) probed FLAT — hd512 was DRAM-re-read-bound,
        // not unpack-bound; jsonl 2026-07-14.
        static TB512: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        // gqa <= 16 = fa_v4_smem_512's q-array capacity; past it fall to the register twins.
        let tb512 = head_dim == 512
            && sp <= 32
            && n_head / n_head_kv.max(1) <= 16
            && *TB512.get_or_init(|| std::env::var("MEMRA_FA_TB512").as_deref() != Ok("0"));
        let fname = if tb512 {
            "fa_decode_vec_q_rows_v4_512_tb"
        } else if i2 {
            "fa_decode_vec_q_rows_dpl16_i2"
        } else if head_dim == 512 {
            "fa_decode_vec_q_rows_dpl16"
        }
        // gemma globals (parity law)
        else if v4 {
            "fa_decode_vec_q_rows_v4"
        } else if v3 {
            "fa_decode_vec_q_rows_v3"
        } else if fa_v2_on() {
            "fa_decode_vec_q_rows_v2"
        } else if smem_rows {
            "fa_decode_vec_q_rows_smem"
        } else {
            "fa_decode_vec_q_rows"
        };
        let f = if head_dim == 512 {
            self.fa_func(fname, head_dim)
        } else if g {
            // FP8-WINDOWED: hd256 rows over an e4m3 cache — kf8vf8 module, SAME symbol
            // choice as decode's kvmod dispatch (parity law: excluding v4 here paired
            // g-module rows against decode's g-module v4 — different programs, short-VG
            // maxdiff 2.0 / spec stream 0/128, 2026-07-12). rows_v4 is format-aware
            // since fda9790; only the smem twin stays excluded (V-stage q5_1-only).
            // hd128 (qwen fp8-KV) lands on the base/register rows via fname — the
            // dq macros are format-aware.
            self.func_g(if smem_rows {
                "fa_decode_vec_q_rows"
            } else {
                fname
            })
        } else {
            self.func(fname)
        };
        let shmem = if tb512 {
            // fa_v4_smem_512 (q 9KB gqa<=16 + k tile 18KB) + sV 32*512 (e4m3 module halves it)
            let gk = Self::gkv_on();
            let sh =
                (8192 + 1024 + 32 * 512 + 32 * 64 + 32 * head_dim * if gk { 1 } else { 2 }) as u32;
            use cudarc::driver::sys::CUfunction_attribute_enum as A;
            f.set_attribute(
                A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                sh as i32,
            )?;
            sh
        } else if v4 || v3 || smem_rows || fa_v2_on() {
            // v4: fa_v4_smem (11.5KB) + sV; v3 stages sV only; v2/smem twins stage sK+sV.
            let sh = (if v4 {
                11520 + 32 * head_dim * if g { 1 } else { 2 }
            } else if v3 {
                32 * head_dim * 2
            } else {
                2 * 32 * head_dim * 2
            }) as u32;
            use cudarc::driver::sys::CUfunction_attribute_enum as A;
            f.set_attribute(
                A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                sh as i32,
            )?;
            sh
        } else {
            0
        };
        // Per-GROUP launches (single group in the common case — identical to the pre-fix
        // single launch there): each group gets its own partials (the rows kernel indexes
        // partials by its LOCAL grid.z row) and q/o row-offset views.
        for &(r0, t_g, sp_g) in &groups {
            let n_splits_g = (base_len + r0 + t_g).div_ceil(sp_g);
            let (nspm, spk) = (n_splits_g as i32, sp_g as i32);
            let base_i = (base_len + r0) as i32;
            let o_len = t_g * n_head * n_splits_g * head_dim;
            let ml_len = t_g * n_head * n_splits_g;
            let mut part_guard = self.fa_part_pool.lock().unwrap();
            if part_guard
                .as_ref()
                .map(|pp| pp.0.len() < o_len || pp.1.len() < ml_len)
                .unwrap_or(true)
            {
                // RETIRE-ON-GROW, never free (#68 root cause, 2026-08-04): captured graphs (the
                // per-session persistent draft graph, the decode/prime graph doors) BAKE these pool
                // buffer addresses. Dropping the old buffers on grow returns them to the async pool,
                // later live allocations land at those addresses, and the next graph REPLAY writes
                // its fa partials over them — the ST serve-spec corruption (acceptance collapse +
                // output corruption began the burst after the trunk's t_kv growth first realloc'd
                // this pool past the draft-capture size; research/fp8ship-20260804). Retiring keeps
                // the baked addresses alive (single-stream: eager writes the new buffers, replays
                // touch only the old — never concurrently). Doubling growth bounds retired VRAM
                // (total retired < final size).
                let old = part_guard.take();
                let (co, cm) = old
                    .as_ref()
                    .map(|pp| (pp.0.len(), pp.1.len()))
                    .unwrap_or((0, 0));
                if let Some(old) = old {
                    self.fa_part_retired.lock().unwrap().push(old);
                }
                if std::env::var("MEMRA_DEBUG_FAPOOL").is_ok() {
                    eprintln!(
                        "[fa-pool] REALLOC o {} -> {} ml {} -> {} (old retired)",
                        co, o_len, cm, ml_len
                    );
                }
                *part_guard = Some((
                    self.alloc_uninit::<f32>(o_len.max(2 * co))?,
                    self.alloc_uninit::<f32>(ml_len.max(2 * cm))?,
                    self.alloc_uninit::<f32>(ml_len.max(2 * cm))?,
                ));
            }
            let pg = part_guard.as_mut().unwrap();
            self.gpu
                .stream()
                .memset_zeros(&mut pg.0.slice_mut(0..o_len))?;
            self.gpu
                .stream()
                .memset_zeros(&mut pg.1.slice_mut(0..ml_len))?;
            self.gpu
                .stream()
                .memset_zeros(&mut pg.2.slice_mut(0..ml_len))?;
            let (part_o, part_m, part_l) = (&mut pg.0, &mut pg.1, &mut pg.2);
            let (part_o, part_m, part_l) = (&mut *part_o, &mut *part_m, &mut *part_l);
            let qv = self.view(q, t * n_head * head_dim);
            let q_g = qv.slice(r0 * n_head * head_dim..(r0 + t_g) * n_head * head_dim);
            let cfg = LaunchConfig {
                grid_dim: (n_head_kv as u32, n_splits_g as u32, t_g as u32),
                block_dim: (32, gqa, 1),
                shared_mem_bytes: shmem,
            };
            {
                let __s_b = self.gpu.stream();
                let mut b = __s_b.launch_builder(&f);
                if tb512 {
                    // rows-inner launch: grid.z dropped, the kernel loops n_rows itself.
                    let (bd, plus) =
                        base_dev.expect("hd512 rows twin requires a device base counter");
                    let plus_g = plus + r0 as i32;
                    let nr = t_g as i32;
                    if Self::pdl_on() && Self::pdl_wb_on() {
                        // wave-B2b: flavor mirrors fa_func(fname, 512) = gkv.
                        use cudarc::driver::{DevicePtr, DevicePtrMut};
                        let s = &self.gpu.stream();
                        let (pq, _b0) = q_g.device_ptr(s);
                        let (pk, _b1) = k.device_ptr(s);
                        let (pv, _b2) = v.device_ptr(s);
                        let (po, _b3) = part_o.device_ptr_mut(s);
                        let (pm, _b4) = part_m.device_ptr_mut(s);
                        let (pl, _b5) = part_l.device_ptr_mut(s);
                        let (pb, _b6) = bd.device_ptr(s);
                        let mut ps = [
                            &pq as *const _ as *mut std::ffi::c_void,
                            &pk as *const _ as *mut _,
                            &pv as *const _ as *mut _,
                            &po as *const _ as *mut _,
                            &pm as *const _ as *mut _,
                            &pl as *const _ as *mut _,
                            &hd as *const _ as *mut _,
                            &nh as *const _ as *mut _,
                            &nhkv as *const _ as *mut _,
                            &pb as *const _ as *mut _,
                            &plus_g as *const _ as *mut _,
                            &scale as *const _ as *mut _,
                            &nspm as *const _ as *mut _,
                            &spk as *const _ as *mut _,
                            &ktb as *const _ as *mut _,
                            &vtb as *const _ as *mut _,
                            &nr as *const _ as *mut _,
                        ];
                        unsafe {
                            self.launch_pdl_flash(
                                Self::gkv_on(),
                                "fa_decode_vec_q_rows_v4_512_tb",
                                (n_head_kv as u32, n_splits_g as u32, 1),
                                (32, gqa, 1),
                                shmem,
                                &mut ps,
                            )?;
                        }
                    } else {
                        let cfg_tb = LaunchConfig {
                            grid_dim: (n_head_kv as u32, n_splits_g as u32, 1),
                            block_dim: (32, gqa, 1),
                            shared_mem_bytes: shmem,
                        };
                        b.arg(&q_g)
                            .arg(k)
                            .arg(v)
                            .arg(&mut *part_o)
                            .arg(&mut *part_m)
                            .arg(&mut *part_l)
                            .arg(&hd)
                            .arg(&nh)
                            .arg(&nhkv)
                            .arg(bd)
                            .arg(&plus_g)
                            .arg(&scale)
                            .arg(&nspm)
                            .arg(&spk)
                            .arg(&ktb)
                            .arg(&vtb)
                            .arg(&nr);
                        unsafe {
                            b.launch(cfg_tb)?;
                        }
                    }
                } else if head_dim == 512 {
                    let (bd, plus) =
                        base_dev.expect("hd512 rows twin requires a device base counter");
                    let plus_g = plus + r0 as i32;
                    b.arg(&q_g)
                        .arg(k)
                        .arg(v)
                        .arg(&mut *part_o)
                        .arg(&mut *part_m)
                        .arg(&mut *part_l)
                        .arg(&hd)
                        .arg(&nh)
                        .arg(&nhkv)
                        .arg(bd)
                        .arg(&plus_g)
                        .arg(&scale)
                        .arg(&nspm)
                        .arg(&spk)
                        .arg(&ktb)
                        .arg(&vtb);
                    unsafe {
                        b.launch(cfg)?;
                    }
                } else {
                    b.arg(&q_g)
                        .arg(k)
                        .arg(v)
                        .arg(&mut *part_o)
                        .arg(&mut *part_m)
                        .arg(&mut *part_l)
                        .arg(&hd)
                        .arg(&nh)
                        .arg(&nhkv)
                        .arg(&base_i)
                        .arg(&scale)
                        .arg(&nspm)
                        .arg(&spk)
                        .arg(&ktb)
                        .arg(&vtb);
                    unsafe {
                        b.launch(cfg)?;
                    }
                }
            }
            let cfg2 = LaunchConfig {
                grid_dim: (n_head as u32, t_g as u32, 1),
                block_dim: (head_dim as u32, 1, 1),
                shared_mem_bytes: 0,
            };
            let mut o_g = o.slice_mut(r0 * n_head * head_dim..(r0 + t_g) * n_head * head_dim);
            if head_dim == 512 {
                // device-len combine (shared by verify/eager/graph — parity by symbol): the
                // per-row n_splits derives from the SAME counter the rows kernel read.
                let (bd, plus) = base_dev.unwrap();
                let plus_g = plus + r0 as i32;
                if let Some((oq, od)) = q8_out.as_mut() {
                    // wave-5b port (2026-07-23, t=1 decode only): q8-emitting dc combine.
                    debug_assert!(t == 1, "rows q8 emit is a t=1 decode arm");
                    if Self::pdl_on() && Self::pdl_wb_on() {
                        // wave-B2: flavor mirrors fa_func (hd512 + gkv → kf8vf8).
                        use cudarc::driver::{DevicePtr, DevicePtrMut};
                        let s = &self.gpu.stream();
                        let (po, _g0) = part_o.device_ptr(s);
                        let (pm, _g1) = part_m.device_ptr(s);
                        let (pl, _g2) = part_l.device_ptr(s);
                        let (pq, _g3) = oq.device_ptr_mut(s);
                        let (pd, _g4) = od.device_ptr_mut(s);
                        let (pb, _g5) = bd.device_ptr(s);
                        let mut ps = [
                            &po as *const _ as *mut std::ffi::c_void,
                            &pm as *const _ as *mut _,
                            &pl as *const _ as *mut _,
                            &pq as *const _ as *mut _,
                            &pd as *const _ as *mut _,
                            &hd as *const _ as *mut _,
                            &nh as *const _ as *mut _,
                            &pb as *const _ as *mut _,
                            &plus_g as *const _ as *mut _,
                            &nspm as *const _ as *mut _,
                            &spk as *const _ as *mut _,
                        ];
                        unsafe {
                            self.launch_pdl_flash(
                                Self::gkv_on(),
                                "fa_decode_combine_rows_dc_q8_1",
                                cfg2.grid_dim,
                                cfg2.block_dim,
                                0,
                                &mut ps,
                            )?;
                        }
                        continue;
                    }
                    let fc = self.fa_func("fa_decode_combine_rows_dc_q8_1", head_dim);
                    let __s_b2 = self.gpu.stream();
                    let mut b2 = __s_b2.launch_builder(&fc);
                    b2.arg(&*part_o)
                        .arg(&*part_m)
                        .arg(&*part_l)
                        .arg(&mut **oq)
                        .arg(&mut **od)
                        .arg(&hd)
                        .arg(&nh)
                        .arg(bd)
                        .arg(&plus_g)
                        .arg(&nspm)
                        .arg(&spk);
                    unsafe {
                        b2.launch(cfg2)?;
                    }
                    continue;
                }
                let fc = self.fa_func("fa_decode_combine_rows_dc", head_dim);
                let __s_b2 = self.gpu.stream();
                let mut b2 = __s_b2.launch_builder(&fc);
                b2.arg(&*part_o)
                    .arg(&*part_m)
                    .arg(&*part_l)
                    .arg(&mut o_g)
                    .arg(&hd)
                    .arg(&nh)
                    .arg(bd)
                    .arg(&plus_g)
                    .arg(&nspm)
                    .arg(&spk);
                unsafe {
                    b2.launch(cfg2)?;
                }
            } else {
                // q8 emit is wired for the hd512 dc-combine arm only — a Some here would
                // leave the caller's pair unwritten (consumer would read garbage).
                assert!(
                    q8_out.is_none(),
                    "rows q8 emit requires the hd512 dc combine"
                );
                let fc = self.func("fa_decode_combine_rows");
                let __s_b2 = self.gpu.stream();
                let mut b2 = __s_b2.launch_builder(&fc);
                b2.arg(&*part_o)
                    .arg(&*part_m)
                    .arg(&*part_l)
                    .arg(&mut o_g)
                    .arg(&hd)
                    .arg(&nh)
                    .arg(&base_i)
                    .arg(&nspm)
                    .arg(&spk);
                unsafe {
                    b2.launch(cfg2)?;
                }
            }
        }
        Ok(())
    }

    /// WINDOWED verify rows (gemma R6 deep-ctx): every row attends exactly `window` keys —
    /// bit-identical per row to the T=1 decode's fa_decode over the window VIEW. Caller gates
    /// base_len + 1 >= window (no under-window rows) and head_dim == 256 (v4 stamp).
    #[allow(clippy::too_many_arguments)]
    pub fn fa_decode_rows_w(
        &self,
        q: &CudaSlice<f32>,
        k: &cudarc::driver::CudaView<u8>,
        v: &cudarc::driver::CudaView<u8>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        base_dev: &CudaSlice<i32>,
        base_plus: i32,
        t: usize,
        scale: f32,
        window: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        q8_out: Option<(&mut CudaSlice<i8>, &mut CudaSlice<f32>)>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        // DEVICE-LEN (graph arc step 1, 2026-07-11): the causal base rides an i32 counter
        // (kernel T_kv = dev[0] + base_plus + r + 1) so depth graphs can replay with len
        // advancing on-device. dc paths pass kvl.len_d with plus=-1; verify/eager sync the
        // counter with one async set_i32_one first. Partials/splits size from `window` (host).
        debug_assert!(head_dim == 256);
        // windowed split (MEMRA_FA_SPW, default 32 — re-swept 2026-07-12 under the raw-e4m3 sV
        // occupancy ceiling (4 blocks/SM): t=1 decode is GRID-limited (win/sp splits x nkv
        // blocks), so smaller splits fill the ceiling — 1.7k 174.4/174.0 vs 48's 170.7/170.3,
        // 4.9k 159.8 vs 157.4 (N=2 interleaved, stable window). Spec serving prefers 64
        // (verify t=K+1 fills the grid via grid.z=t; depth K=7 281.3 vs 249.3 at 32) — set
        // MEMRA_FA_SPW=64 there, same config law as MEMRA_GEMMA_GKV=0. MUST be one value for
        // ALL widths: a t-keyed probe broke decode-vs-verify combine order (stream 9/128).
        let sp = {
            static SPW: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
            let v = *SPW.get_or_init(|| {
                std::env::var("MEMRA_FA_SPW")
                    .ok()
                    .and_then(|x| x.parse().ok())
                    .unwrap_or(0)
            });
            if v >= 8 {
                v
            } else {
                FA_SPW_DEFAULT.load(std::sync::atomic::Ordering::Relaxed)
            }
        };
        let n_splits_max = (window + sp - 1) / sp;
        let (hd, nh, nhkv) = (head_dim as i32, n_head as i32, n_head_kv as i32);
        let (nspm, spk, wini) = (n_splits_max as i32, sp as i32, window as i32);
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let gqa = (n_head / n_head_kv).max(1) as u32;
        let o_len = t * n_head * n_splits_max * head_dim;
        let ml_len = t * n_head * n_splits_max;
        let mut part_guard = self.fa_part_pool.lock().unwrap();
        if part_guard
            .as_ref()
            .map(|pp| pp.0.len() < o_len || pp.1.len() < ml_len)
            .unwrap_or(true)
        {
            // RETIRE-ON-GROW, never free (#68 root cause, 2026-08-04): captured graphs (the
            // per-session persistent draft graph, the decode/prime graph doors) BAKE these pool
            // buffer addresses. Dropping the old buffers on grow returns them to the async pool,
            // later live allocations land at those addresses, and the next graph REPLAY writes
            // its fa partials over them — the ST serve-spec corruption (acceptance collapse +
            // output corruption began the burst after the trunk's t_kv growth first realloc'd
            // this pool past the draft-capture size; research/fp8ship-20260804). Retiring keeps
            // the baked addresses alive (single-stream: eager writes the new buffers, replays
            // touch only the old — never concurrently). Doubling growth bounds retired VRAM
            // (total retired < final size).
            let old = part_guard.take();
            let (co, cm) = old
                .as_ref()
                .map(|pp| (pp.0.len(), pp.1.len()))
                .unwrap_or((0, 0));
            if let Some(old) = old {
                self.fa_part_retired.lock().unwrap().push(old);
            }
            if std::env::var("MEMRA_DEBUG_FAPOOL").is_ok() {
                eprintln!(
                    "[fa-pool] REALLOC o {} -> {} ml {} -> {} (old retired)",
                    co, o_len, cm, ml_len
                );
            }
            *part_guard = Some((
                self.alloc_uninit::<f32>(o_len.max(2 * co))?,
                self.alloc_uninit::<f32>(ml_len.max(2 * cm))?,
                self.alloc_uninit::<f32>(ml_len.max(2 * cm))?,
            ));
        }
        let pg = part_guard.as_mut().unwrap();
        self.gpu
            .stream()
            .memset_zeros(&mut pg.0.slice_mut(0..o_len))?;
        self.gpu
            .stream()
            .memset_zeros(&mut pg.1.slice_mut(0..ml_len))?;
        self.gpu
            .stream()
            .memset_zeros(&mut pg.2.slice_mut(0..ml_len))?;
        let (part_o, part_m, part_l) = (&mut pg.0, &mut pg.1, &mut pg.2);
        // Lane pick: decode AND verify both land here in the windowed regime (parity law —
        // hybrid_forward verify_attn), so the pick only needs internal consistency, not
        // clone-of-decode bit fidelity (SASS-proven impossible for textually identical
        // kernels, jsonl 2026-07-10). v4 under the threshold; smem twin at/above the smem
        // floor (deep-ctx broadcast win); register twin between.
        static SMEM_TKV_W: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
        let smem_tkv = *SMEM_TKV_W.get_or_init(|| {
            std::env::var("MEMRA_FA_SMEM_TKV")
                .ok()
                .and_then(|v| v.parse().ok())
                .unwrap_or_else(|| FA_SMEM_TKV_DEFAULT.load(std::sync::atomic::Ordering::Relaxed))
        });
        // MULTI-ROW v4: resurrected 2026-07-14 (the '33 tok/s collapse' was a paired-map
        // partial-write bug, not the mechanism) and falsified HONESTLY at gqa 2: bit-exact
        // but −1.7% on the 31B depth cell — the sp helper warp already hides staging
        // in-block, and mr trades L2-cheap redundant bytes for serialized per-warp gqa
        // score/B3 chains. Arm deleted; jsonl row 2026-07-14 is the record.
        use cudarc::driver::sys::CUfunction_attribute_enum as A;
        // FP8-WINDOWED (wkv): the v4 family is format-aware (2026-07-12 KFMT/VFMT staging
        // arms) — wkv rides the SAME lane logic, resolved from the kf8vf8 module. One symbol
        // per (lane, format-module) keeps parity structural; the old register-i2 detour
        // (-33%) is retired.
        let wg = Self::wkv_on();
        // STAGING-PARALLEL v4 (MEMRA_FA_SPW2, default ON at gqa==1): warp 1 = staging helper
        // (v4 is 61% staging); score phases identical to v4_w. Same symbol all t.
        let sp2 =
            gqa <= 4 && fa_v4_at(window) && std::env::var("MEMRA_FA_SPW2").as_deref() != Ok("0");
        if sp2 {
            let sh = (11520 + 32 * head_dim * if wg { 1 } else { 2 }) as u32;
            if Self::pdl_on() && Self::pdl_wb_on() {
                // wave-B2b: flavor mirrors wg.
                use cudarc::driver::{DevicePtr, DevicePtrMut};
                let s = &self.gpu.stream();
                let (pq, _b0) = q.device_ptr(s);
                let (pk, _b1) = k.device_ptr(s);
                let (pv, _b2) = v.device_ptr(s);
                let (po, _b3) = part_o.device_ptr_mut(s);
                let (pm, _b4) = part_m.device_ptr_mut(s);
                let (pl, _b5) = part_l.device_ptr_mut(s);
                let (pb, _b6) = base_dev.device_ptr(s);
                let mut ps = [
                    &pq as *const _ as *mut std::ffi::c_void,
                    &pk as *const _ as *mut _,
                    &pv as *const _ as *mut _,
                    &po as *const _ as *mut _,
                    &pm as *const _ as *mut _,
                    &pl as *const _ as *mut _,
                    &hd as *const _ as *mut _,
                    &nh as *const _ as *mut _,
                    &nhkv as *const _ as *mut _,
                    &pb as *const _ as *mut _,
                    &base_plus as *const _ as *mut _,
                    &scale as *const _ as *mut _,
                    &nspm as *const _ as *mut _,
                    &spk as *const _ as *mut _,
                    &ktb as *const _ as *mut _,
                    &vtb as *const _ as *mut _,
                    &wini as *const _ as *mut _,
                ];
                unsafe {
                    self.launch_pdl_flash(
                        wg,
                        "fa_decode_vec_q_rows_v4_w_sp",
                        (n_head_kv as u32, n_splits_max as u32, t as u32),
                        (32, gqa + 1, 1),
                        sh,
                        &mut ps,
                    )?;
                }
            } else {
                let f = if wg {
                    self.func_g("fa_decode_vec_q_rows_v4_w_sp")
                } else {
                    self.func("fa_decode_vec_q_rows_v4_w_sp")
                };
                f.set_attribute(
                    A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                    sh as i32,
                )?;
                let cfg = LaunchConfig {
                    grid_dim: (n_head_kv as u32, n_splits_max as u32, t as u32),
                    block_dim: (32, gqa + 1, 1),
                    shared_mem_bytes: sh,
                };
                let __s_b = self.gpu.stream();
                let mut b = __s_b.launch_builder(&f);
                b.arg(q)
                    .arg(k)
                    .arg(v)
                    .arg(&mut *part_o)
                    .arg(&mut *part_m)
                    .arg(&mut *part_l)
                    .arg(&hd)
                    .arg(&nh)
                    .arg(&nhkv)
                    .arg(base_dev)
                    .arg(&base_plus)
                    .arg(&scale)
                    .arg(&nspm)
                    .arg(&spk)
                    .arg(&ktb)
                    .arg(&vtb)
                    .arg(&wini);
                unsafe {
                    b.launch(cfg)?;
                }
            }
        } else {
            if fa_v4_at(window) && Self::pdl_on() && Self::pdl_wb_on() {
                // wave-B2b: the v4_w pick only (smem/reg twins stay builder-launched).
                let sh = (11520 + 32 * head_dim * if wg { 1 } else { 2 }) as u32;
                use cudarc::driver::{DevicePtr, DevicePtrMut};
                let s = &self.gpu.stream();
                let (pq, _b0) = q.device_ptr(s);
                let (pk, _b1) = k.device_ptr(s);
                let (pv, _b2) = v.device_ptr(s);
                let (po, _b3) = part_o.device_ptr_mut(s);
                let (pm, _b4) = part_m.device_ptr_mut(s);
                let (pl, _b5) = part_l.device_ptr_mut(s);
                let (pb, _b6) = base_dev.device_ptr(s);
                let mut ps = [
                    &pq as *const _ as *mut std::ffi::c_void,
                    &pk as *const _ as *mut _,
                    &pv as *const _ as *mut _,
                    &po as *const _ as *mut _,
                    &pm as *const _ as *mut _,
                    &pl as *const _ as *mut _,
                    &hd as *const _ as *mut _,
                    &nh as *const _ as *mut _,
                    &nhkv as *const _ as *mut _,
                    &pb as *const _ as *mut _,
                    &base_plus as *const _ as *mut _,
                    &scale as *const _ as *mut _,
                    &nspm as *const _ as *mut _,
                    &spk as *const _ as *mut _,
                    &ktb as *const _ as *mut _,
                    &vtb as *const _ as *mut _,
                    &wini as *const _ as *mut _,
                ];
                unsafe {
                    self.launch_pdl_flash(
                        wg,
                        "fa_decode_vec_q_rows_v4_w",
                        (n_head_kv as u32, n_splits_max as u32, t as u32),
                        (32, gqa, 1),
                        sh,
                        &mut ps,
                    )?;
                }
            } else {
                let pick = |name: &str| {
                    if wg {
                        self.func_g(name)
                    } else {
                        self.func(name)
                    }
                };
                let (f, sh) = if fa_v4_at(window) {
                    let f = pick("fa_decode_vec_q_rows_v4_w");
                    (f, (11520 + 32 * head_dim * if wg { 1 } else { 2 }) as u32)
                } else if smem_tkv > 0 && window >= smem_tkv {
                    // NOTE: the smem twin's V-stage is still q5_1-hardcoded — unreachable under wkv
                    // at the gemma window (v4 covers it); revisit if the smem floor ever drops.
                    (
                        pick("fa_decode_vec_q_rows_smem_w"),
                        (2 * 32 * head_dim * 2) as u32,
                    )
                } else {
                    (pick("fa_decode_vec_q_rows_reg_w"), 0u32)
                };
                f.set_attribute(
                    A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                    sh as i32,
                )?;
                let cfg = LaunchConfig {
                    grid_dim: (n_head_kv as u32, n_splits_max as u32, t as u32),
                    block_dim: (32, gqa, 1),
                    shared_mem_bytes: sh,
                };
                let __s_b = self.gpu.stream();
                let mut b = __s_b.launch_builder(&f);
                b.arg(q)
                    .arg(k)
                    .arg(v)
                    .arg(&mut *part_o)
                    .arg(&mut *part_m)
                    .arg(&mut *part_l)
                    .arg(&hd)
                    .arg(&nh)
                    .arg(&nhkv)
                    .arg(base_dev)
                    .arg(&base_plus)
                    .arg(&scale)
                    .arg(&nspm)
                    .arg(&spk)
                    .arg(&ktb)
                    .arg(&vtb)
                    .arg(&wini);
                unsafe {
                    b.launch(cfg)?;
                }
            }
        }
        let cfg2 = LaunchConfig {
            grid_dim: (n_head as u32, t as u32, 1),
            block_dim: (head_dim as u32, 1, 1),
            shared_mem_bytes: 0,
        };
        if let Some((oq, od)) = q8_out {
            // wave-5b port (2026-07-23): q8-emitting combine — the t=1 decode's wo matvec
            // consumes the pair directly; the standalone quantize launch folds away.
            if Self::pdl_on() && Self::pdl_wb_on() {
                // wave-B2: flavor mirrors the builder's wg choice.
                use cudarc::driver::{DevicePtr, DevicePtrMut};
                let s = &self.gpu.stream();
                let (po, _g0) = part_o.device_ptr(s);
                let (pm, _g1) = part_m.device_ptr(s);
                let (pl, _g2) = part_l.device_ptr(s);
                let (pq, _g3) = oq.device_ptr_mut(s);
                let (pd, _g4) = od.device_ptr_mut(s);
                let mut ps = [
                    &po as *const _ as *mut std::ffi::c_void,
                    &pm as *const _ as *mut _,
                    &pl as *const _ as *mut _,
                    &pq as *const _ as *mut _,
                    &pd as *const _ as *mut _,
                    &hd as *const _ as *mut _,
                    &nh as *const _ as *mut _,
                    &nspm as *const _ as *mut _,
                    &spk as *const _ as *mut _,
                    &wini as *const _ as *mut _,
                ];
                unsafe {
                    self.launch_pdl_flash(
                        wg,
                        "fa_decode_combine_rows_w_q8_1",
                        cfg2.grid_dim,
                        cfg2.block_dim,
                        0,
                        &mut ps,
                    )?;
                }
                return Ok(());
            }
            let fc = if wg {
                self.func_g("fa_decode_combine_rows_w_q8_1")
            } else {
                self.func("fa_decode_combine_rows_w_q8_1")
            };
            let __s_b2 = self.gpu.stream();
            let mut b2 = __s_b2.launch_builder(&fc);
            b2.arg(&*part_o)
                .arg(&*part_m)
                .arg(&*part_l)
                .arg(oq)
                .arg(od)
                .arg(&hd)
                .arg(&nh)
                .arg(&nspm)
                .arg(&spk)
                .arg(&wini);
            unsafe {
                b2.launch(cfg2)?;
            }
            return Ok(());
        }
        let fc = if wg {
            self.func_g("fa_decode_combine_rows_w")
        } else {
            self.func("fa_decode_combine_rows_w")
        };
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&fc);
        b2.arg(&*part_o)
            .arg(&*part_m)
            .arg(&*part_l)
            .arg(o)
            .arg(&hd)
            .arg(&nh)
            .arg(&nspm)
            .arg(&spk)
            .arg(&wini);
        unsafe {
            b2.launch(cfg2)?;
        }
        Ok(())
    }

    /// ROUND-STREAM stage (c): fa rows with the causal base from a device counter. Two lanes:
    /// v3 (qwen stream, fa_v3_active) and v4 (gemma hd256 burst — rows_v4_dc, g-module aware);
    /// `t_kv_upper` sizes splits/partials — the same one-sp-for-all-rows approximation class
    /// the host rows path already uses (battery-arbitrated); actual per-row bounds derive
    /// in-kernel from the counter (+ base_plus, v4 lane only — v3's kernel has no plus arg).
    #[allow(clippy::too_many_arguments)]
    pub fn fa_decode_rows_dc(
        &self,
        q: &CudaSlice<f32>,
        k: &cudarc::driver::CudaView<u8>,
        v: &cudarc::driver::CudaView<u8>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        base_dev: &CudaSlice<i32>,
        t_kv_upper: usize,
        t: usize,
        scale: f32,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        base_plus: i32,
        g: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let v4 = head_dim == 256 && fa_v4_at(t_kv_upper);
        assert!(
            v4 || fa_v3_active(head_dim),
            "stream fa rows requires the v3 or v4 lane"
        );
        assert!(v4 || base_plus == 0, "v3_dc kernel takes no plus arg");
        if v4 {
            let sp = fa_split_keys(t_kv_upper, n_head_kv);
            let n_splits_max = (t_kv_upper + sp - 1) / sp;
            let (hd, nh, nhkv) = (head_dim as i32, n_head as i32, n_head_kv as i32);
            let (nspm, spk) = (n_splits_max as i32, sp as i32);
            let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
            let gqa = (n_head / n_head_kv).max(1) as u32;
            let o_len = t * n_head * n_splits_max * head_dim;
            let ml_len = t * n_head * n_splits_max;
            let mut part_guard = self.fa_part_pool.lock().unwrap();
            if part_guard
                .as_ref()
                .map(|pp| pp.0.len() < o_len || pp.1.len() < ml_len)
                .unwrap_or(true)
            {
                // RETIRE-ON-GROW, never free (#68 root cause, 2026-08-04): captured graphs (the
                // per-session persistent draft graph, the decode/prime graph doors) BAKE these pool
                // buffer addresses. Dropping the old buffers on grow returns them to the async pool,
                // later live allocations land at those addresses, and the next graph REPLAY writes
                // its fa partials over them — the ST serve-spec corruption (acceptance collapse +
                // output corruption began the burst after the trunk's t_kv growth first realloc'd
                // this pool past the draft-capture size; research/fp8ship-20260804). Retiring keeps
                // the baked addresses alive (single-stream: eager writes the new buffers, replays
                // touch only the old — never concurrently). Doubling growth bounds retired VRAM
                // (total retired < final size).
                let old = part_guard.take();
                let (co, cm) = old
                    .as_ref()
                    .map(|pp| (pp.0.len(), pp.1.len()))
                    .unwrap_or((0, 0));
                if let Some(old) = old {
                    self.fa_part_retired.lock().unwrap().push(old);
                }
                if std::env::var("MEMRA_DEBUG_FAPOOL").is_ok() {
                    eprintln!(
                        "[fa-pool] REALLOC o {} -> {} ml {} -> {} (old retired)",
                        co, o_len, cm, ml_len
                    );
                }
                *part_guard = Some((
                    self.alloc_uninit::<f32>(o_len.max(2 * co))?,
                    self.alloc_uninit::<f32>(ml_len.max(2 * cm))?,
                    self.alloc_uninit::<f32>(ml_len.max(2 * cm))?,
                ));
            }
            let pg = part_guard.as_mut().unwrap();
            self.gpu
                .stream()
                .memset_zeros(&mut pg.0.slice_mut(0..o_len))?;
            self.gpu
                .stream()
                .memset_zeros(&mut pg.1.slice_mut(0..ml_len))?;
            self.gpu
                .stream()
                .memset_zeros(&mut pg.2.slice_mut(0..ml_len))?;
            let (part_o, part_m, part_l) = (&mut pg.0, &mut pg.1, &mut pg.2);
            let f = if g {
                self.func_g("fa_decode_vec_q_rows_v4_dc")
            } else {
                self.func("fa_decode_vec_q_rows_v4_dc")
            };
            let sh = (11520 + 32 * head_dim * if g { 1 } else { 2 }) as u32;
            use cudarc::driver::sys::CUfunction_attribute_enum as A;
            f.set_attribute(
                A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                sh as i32,
            )?;
            let cfg = LaunchConfig {
                grid_dim: (n_head_kv as u32, n_splits_max as u32, t as u32),
                block_dim: (32, gqa, 1),
                shared_mem_bytes: sh,
            };
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(q)
                .arg(k)
                .arg(v)
                .arg(&mut *part_o)
                .arg(&mut *part_m)
                .arg(&mut *part_l)
                .arg(&hd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(base_dev)
                .arg(&base_plus)
                .arg(&scale)
                .arg(&nspm)
                .arg(&spk)
                .arg(&ktb)
                .arg(&vtb);
            unsafe {
                b.launch(cfg)?;
            }
            let fc = self.func("fa_decode_combine_rows_dc");
            let cfg2 = LaunchConfig {
                grid_dim: (n_head as u32, t as u32, 1),
                block_dim: (head_dim as u32, 1, 1),
                shared_mem_bytes: 0,
            };
            let __s_b2 = self.gpu.stream();
            let mut b2 = __s_b2.launch_builder(&fc);
            b2.arg(&*part_o)
                .arg(&*part_m)
                .arg(&*part_l)
                .arg(o)
                .arg(&hd)
                .arg(&nh)
                .arg(base_dev)
                .arg(&base_plus)
                .arg(&nspm)
                .arg(&spk);
            unsafe {
                b2.launch(cfg2)?;
            }
            return Ok(());
        }
        let sp = fa_split_keys(t_kv_upper, n_head_kv);
        let n_splits_max = (t_kv_upper + sp - 1) / sp;
        let (hd, nh, nhkv) = (head_dim as i32, n_head as i32, n_head_kv as i32);
        let (nspm, spk) = (n_splits_max as i32, sp as i32);
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let gqa = (n_head / n_head_kv).max(1) as u32;
        let o_len = t * n_head * n_splits_max * head_dim;
        let ml_len = t * n_head * n_splits_max;
        let mut part_guard = self.fa_part_pool.lock().unwrap();
        if part_guard
            .as_ref()
            .map(|pp| pp.0.len() < o_len || pp.1.len() < ml_len)
            .unwrap_or(true)
        {
            // RETIRE-ON-GROW, never free (#68 root cause, 2026-08-04): captured graphs (the
            // per-session persistent draft graph, the decode/prime graph doors) BAKE these pool
            // buffer addresses. Dropping the old buffers on grow returns them to the async pool,
            // later live allocations land at those addresses, and the next graph REPLAY writes
            // its fa partials over them — the ST serve-spec corruption (acceptance collapse +
            // output corruption began the burst after the trunk's t_kv growth first realloc'd
            // this pool past the draft-capture size; research/fp8ship-20260804). Retiring keeps
            // the baked addresses alive (single-stream: eager writes the new buffers, replays
            // touch only the old — never concurrently). Doubling growth bounds retired VRAM
            // (total retired < final size).
            let old = part_guard.take();
            let (co, cm) = old
                .as_ref()
                .map(|pp| (pp.0.len(), pp.1.len()))
                .unwrap_or((0, 0));
            if let Some(old) = old {
                self.fa_part_retired.lock().unwrap().push(old);
            }
            if std::env::var("MEMRA_DEBUG_FAPOOL").is_ok() {
                eprintln!(
                    "[fa-pool] REALLOC o {} -> {} ml {} -> {} (old retired)",
                    co, o_len, cm, ml_len
                );
            }
            *part_guard = Some((
                self.alloc_uninit::<f32>(o_len.max(2 * co))?,
                self.alloc_uninit::<f32>(ml_len.max(2 * cm))?,
                self.alloc_uninit::<f32>(ml_len.max(2 * cm))?,
            ));
        }
        let pg = part_guard.as_mut().unwrap();
        self.gpu
            .stream()
            .memset_zeros(&mut pg.0.slice_mut(0..o_len))?;
        self.gpu
            .stream()
            .memset_zeros(&mut pg.1.slice_mut(0..ml_len))?;
        self.gpu
            .stream()
            .memset_zeros(&mut pg.2.slice_mut(0..ml_len))?;
        let (part_o, part_m, part_l) = (&mut pg.0, &mut pg.1, &mut pg.2);
        let f = self.func("fa_decode_vec_q_rows_v3_dc");
        let sh = (32 * head_dim * 2) as u32;
        use cudarc::driver::sys::CUfunction_attribute_enum as A;
        f.set_attribute(
            A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
            sh as i32,
        )?;
        let cfg = LaunchConfig {
            grid_dim: (n_head_kv as u32, n_splits_max as u32, t as u32),
            block_dim: (32, gqa, 1),
            shared_mem_bytes: sh,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(v)
            .arg(&mut *part_o)
            .arg(&mut *part_m)
            .arg(&mut *part_l)
            .arg(&hd)
            .arg(&nh)
            .arg(&nhkv)
            .arg(base_dev)
            .arg(&scale)
            .arg(&nspm)
            .arg(&spk)
            .arg(&ktb)
            .arg(&vtb);
        unsafe {
            b.launch(cfg)?;
        }
        let fc = self.func("fa_decode_combine_rows_dc");
        let cfg2 = LaunchConfig {
            grid_dim: (n_head as u32, t as u32, 1),
            block_dim: (head_dim as u32, 1, 1),
            shared_mem_bytes: 0,
        };
        let plus0 = 0i32;
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&fc);
        b2.arg(&*part_o)
            .arg(&*part_m)
            .arg(&*part_l)
            .arg(o)
            .arg(&hd)
            .arg(&nh)
            .arg(base_dev)
            .arg(&plus0)
            .arg(&nspm)
            .arg(&spk);
        unsafe {
            b2.launch(cfg2)?;
        }
        Ok(())
    }

    /// Device-counter variant of `fa_decode` (CUDA-GRAPH-PLAN Phase 2). The sequence length is read
    /// from `t_kv_dev[0]` (resident device i32[1]) for the attention loop bound + per-split key range;
    /// the GRID `n_splits` is sized for `bucket_max` (the bucket's max t_kv — baked at capture time).
    /// Empty splits (key range beyond the actual t_kv) write an empty partial (m=NEG_INF) so the
    /// shared combine skips them -> bit-correct for ANY actual t_kv <= bucket_max.
    ///
    /// BIT-IDENTITY (the gate): pass `bucket_max == actual_t_kv` and this reproduces `fa_decode`
    /// EXACTLY (same n_splits, same per, same split boundaries, same combine) while reading t_kv from
    /// device. Bucketing (bucket_max > t_kv) is for the future captured path and changes split
    /// grouping (different but mathematically-equal log-sum-exp merge).
    pub fn fa_decode_dc(
        &self,
        q: &CudaSlice<f32>,
        k: &cudarc::driver::CudaView<u8>,
        v: &cudarc::driver::CudaView<u8>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t_kv_dev: &CudaSlice<i32>,
        bucket_max: usize,
        scale: f32,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        g: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        self.fa_decode_dc_q8(
            q,
            k,
            v,
            o,
            head_dim,
            n_head,
            n_head_kv,
            t_kv_dev,
            bucket_max,
            scale,
            k_tok_bytes,
            v_tok_bytes,
            g,
            None,
        )
    }

    /// `fa_decode_dc` with an optional q8_1 sink (wave 5b): when `q8_out` is given the
    /// combine emits (int8, per-32 scales) for the wo matmul_pre and skips the f32 O write.
    #[allow(clippy::too_many_arguments)]
    pub fn fa_decode_dc_q8(
        &self,
        q: &CudaSlice<f32>,
        k: &cudarc::driver::CudaView<u8>,
        v: &cudarc::driver::CudaView<u8>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        t_kv_dev: &CudaSlice<i32>,
        bucket_max: usize,
        scale: f32,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        g: bool,
        q8_out: Option<(&mut CudaSlice<i8>, &mut CudaSlice<f32>)>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        // The fa_vec gate + n_splits are sized from bucket_max (host, fixed at capture). The kernel
        // reads the ACTUAL t_kv from t_kv_dev for the per-split bound. DEFAULT-ON to MATCH the eager
        // `fa_decode` gate above — graph capture must mirror eager's kernel choice or the graph-vs-eager
        // bit-identity gate breaks. MEMRA_NO_FA_VEC forces scalar on BOTH paths in lockstep.
        // `g` = this layer's cache is e4m3 (gemma windowed under wkv) — every pick below must
        // mirror fa_decode_kvmod's g-routing or the graph diverges from eager (short/mid 1/96,
        // 2026-07-12).
        let mut fa_vec =
            std::env::var("MEMRA_NO_FA_VEC").is_err() && bucket_max >= fa_vec_min_tkv();
        if g && head_dim == 256 && !fa_v4_at(bucket_max) {
            fa_vec = false;
        } // mirror kvmod/geom
        let sp = fa_split_keys(bucket_max, n_head_kv);
        let n_splits = if fa_vec {
            ((bucket_max + sp - 1) / sp).max(1)
        } else {
            ((bucket_max + 255) / 256).max(1)
        };
        let o_len = n_head * n_splits * head_dim;
        let ml_len = n_head * n_splits;
        let mut part_guard = self.fa_part_pool.lock().unwrap();
        if part_guard
            .as_ref()
            .map(|pp| pp.0.len() < o_len || pp.1.len() < ml_len)
            .unwrap_or(true)
        {
            // RETIRE-ON-GROW, never free (#68 root cause, 2026-08-04): captured graphs (the
            // per-session persistent draft graph, the decode/prime graph doors) BAKE these pool
            // buffer addresses. Dropping the old buffers on grow returns them to the async pool,
            // later live allocations land at those addresses, and the next graph REPLAY writes
            // its fa partials over them — the ST serve-spec corruption (acceptance collapse +
            // output corruption began the burst after the trunk's t_kv growth first realloc'd
            // this pool past the draft-capture size; research/fp8ship-20260804). Retiring keeps
            // the baked addresses alive (single-stream: eager writes the new buffers, replays
            // touch only the old — never concurrently). Doubling growth bounds retired VRAM
            // (total retired < final size).
            let old = part_guard.take();
            let (co, cm) = old
                .as_ref()
                .map(|pp| (pp.0.len(), pp.1.len()))
                .unwrap_or((0, 0));
            if let Some(old) = old {
                self.fa_part_retired.lock().unwrap().push(old);
            }
            if std::env::var("MEMRA_DEBUG_FAPOOL").is_ok() {
                eprintln!(
                    "[fa-pool] REALLOC o {} -> {} ml {} -> {} (old retired)",
                    co, o_len, cm, ml_len
                );
            }
            *part_guard = Some((
                self.alloc_uninit::<f32>(o_len.max(2 * co))?,
                self.alloc_uninit::<f32>(ml_len.max(2 * cm))?,
                self.alloc_uninit::<f32>(ml_len.max(2 * cm))?,
            ));
        }
        let pg = part_guard.as_mut().unwrap();
        self.gpu
            .stream()
            .memset_zeros(&mut pg.0.slice_mut(0..o_len))?;
        self.gpu
            .stream()
            .memset_zeros(&mut pg.1.slice_mut(0..ml_len))?;
        self.gpu
            .stream()
            .memset_zeros(&mut pg.2.slice_mut(0..ml_len))?;
        let (part_o, part_m, part_l) = (&mut pg.0, &mut pg.1, &mut pg.2);
        let (hd, nh, nhkv, nsp) = (
            head_dim as i32,
            n_head as i32,
            n_head_kv as i32,
            n_splits as i32,
        );
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let fa_vec = fa_vec && head_dim <= 512 && head_dim % 32 == 0;
        // FA-DEEP pick keyed on bucket_max (the fa_v4_at precedent) — bit-identical twins,
        // so a threshold falling between t_kv and bucket_max cannot diverge eager-vs-graph.
        let deep = fa_vec
            && head_dim == 256
            && fa_v4_at(bucket_max)
            && !g
            && fa_deep_at(bucket_max)
            && !matches!(fa_v4_mode(), "noB3" | "stage");
        let (f, cfg) = if fa_vec
            && head_dim == 512
            && bucket_max >= {
                static FA512_MIN_DC: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
                *FA512_MIN_DC.get_or_init(|| {
                    std::env::var("MEMRA_FA512_MIN")
                        .ok()
                        .and_then(|v| v.parse().ok())
                        .unwrap_or(512)
                })
            } {
            // gemma globals dc twin (mirror the eager dpl16 pick incl the crossover floor).
            let gqa = (n_head / n_head_kv).max(1) as u32;
            (
                self.fa_func("fa_decode_vec_q_dpl16_dc", head_dim),
                LaunchConfig {
                    grid_dim: (n_head_kv as u32, n_splits as u32, 1),
                    block_dim: (32, gqa, 1),
                    shared_mem_bytes: 0,
                },
            )
        } else if fa_vec && head_dim == 512 {
            // under the 512 floor eager runs scalar — the SAME unified symbol, ctr non-null;
            // ns_eff in-kernel reproduces eager's ceil(t_kv/sp) partition for the LIVE len.
            let q_view = q.as_view();
            let mut o_view = o.as_view_mut();
            return self.fa_decode_scalar_unified(
                &q_view,
                k,
                v,
                &mut o_view,
                head_dim,
                n_head,
                n_head_kv,
                0,
                Some(t_kv_dev),
                scale,
                n_splits,
                sp,
                k_tok_bytes,
                v_tok_bytes,
                g,
                &mut *part_o,
                &mut *part_m,
                &mut *part_l,
                q8_out,
            );
        } else if fa_vec && head_dim == 256 && fa_v4_at(bucket_max) {
            // gemma/qwen v4 dc twin (eager default lane) — capture must mirror eager's pick,
            // incl the g-module route + raw-e4m3 sV sizing.
            let gqa = (n_head / n_head_kv).max(1) as u32;
            let fv = if g {
                self.func_g("fa_decode_vec_q_v4_dc")
            } else if deep {
                self.func("fa_decode_vec_q_v4_deep_dc")
            } else {
                self.func("fa_decode_vec_q_v4_dc")
            };
            let shmem =
                (if deep { 12160 } else { 11520 } + 32 * head_dim * if g { 1 } else { 2 }) as u32;
            use cudarc::driver::sys::CUfunction_attribute_enum as A;
            fv.set_attribute(
                A::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                shmem as i32,
            )?;
            (
                fv,
                LaunchConfig {
                    grid_dim: (n_head_kv as u32, n_splits as u32, 1),
                    block_dim: (32, gqa, 1),
                    shared_mem_bytes: shmem,
                },
            )
        } else if fa_vec && fa_v3_active(head_dim) {
            // FA v3 lane _dc twin: the captured graph must run the SAME walk body as eager
            // under MEMRA_FA_V3=1 (eager, rows-verify and graph switch together).
            let gqa = (n_head / n_head_kv).max(1) as u32;
            let fv = if g {
                self.func_g("fa_decode_vec_q_v3_dc")
            } else {
                self.func("fa_decode_vec_q_v3_dc")
            };
            let shmem = (32 * head_dim * 2) as u32; // sV bf16 [FA_DEC_TILE=32][hd]
            (
                fv,
                LaunchConfig {
                    grid_dim: (n_head_kv as u32, n_splits as u32, 1),
                    block_dim: (32, gqa, 1),
                    shared_mem_bytes: shmem,
                },
            )
        } else if fa_vec && fa_v2_on() {
            // FAVENDOR lane: v2 _dc twin — the captured graph must run the SAME walk body as
            // eager under MEMRA_FA_V2=1 or graph_decode_gate's bit-identity breaks (the flag is
            // a numeric config; eager, rows-verify and graph all switch together).
            let gqa = (n_head / n_head_kv).max(1) as u32;
            let fv = if g {
                self.func_g("fa_decode_vec_q_v2_dc")
            } else {
                self.func("fa_decode_vec_q_v2_dc")
            };
            let shmem = (2 * 32 * head_dim * 2) as u32; // sK+sV bf16 [FA_DEC_TILE=32][hd]
            (
                fv,
                LaunchConfig {
                    grid_dim: (n_head_kv as u32, n_splits as u32, 1),
                    block_dim: (32, gqa, 1),
                    shared_mem_bytes: shmem,
                },
            )
        } else if fa_vec {
            let gqa = (n_head / n_head_kv).max(1) as u32;
            // REGISTER-DEQUANT twin: zero dynamic smem (see fa_decode above).
            let fv = if g {
                self.func_g("fa_decode_vec_q_dc")
            } else {
                self.func("fa_decode_vec_q_dc")
            };
            (
                fv,
                LaunchConfig {
                    grid_dim: (n_head_kv as u32, n_splits as u32, 1),
                    block_dim: (32, gqa, 1),
                    shared_mem_bytes: 0,
                },
            )
        } else {
            let q_view = q.as_view();
            let mut o_view = o.as_view_mut();
            return self.fa_decode_scalar_unified(
                &q_view,
                k,
                v,
                &mut o_view,
                head_dim,
                n_head,
                n_head_kv,
                0,
                Some(t_kv_dev),
                scale,
                n_splits,
                if fa_vec { sp } else { 256 },
                k_tok_bytes,
                v_tok_bytes,
                g,
                &mut *part_o,
                &mut *part_m,
                &mut *part_l,
                q8_out,
            );
        };
        let ski = sp as i32; // one-partition law: the twins derive ns_eff from (T_kv, ski)
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(v)
            .arg(&mut *part_o)
            .arg(&mut *part_m)
            .arg(&mut *part_l)
            .arg(&hd)
            .arg(&nh)
            .arg(&nhkv)
            .arg(t_kv_dev)
            .arg(&scale)
            .arg(&nsp)
            .arg(&ski)
            .arg(&ktb)
            .arg(&vtb);
        unsafe {
            b.launch(cfg)?;
        }
        let cfg2 = LaunchConfig {
            grid_dim: (n_head as u32, 1, 1),
            block_dim: (head_dim as u32, 1, 1),
            shared_mem_bytes: 0,
        };
        if let Some((oq, od)) = q8_out {
            let fc = if g {
                self.func_g("fa_decode_combine_q8_1")
            } else {
                self.fa_func("fa_decode_combine_q8_1", head_dim)
            };
            let __s_b2 = self.gpu.stream();
            let mut b2 = __s_b2.launch_builder(&fc);
            b2.arg(&*part_o)
                .arg(&*part_m)
                .arg(&*part_l)
                .arg(oq)
                .arg(od)
                .arg(&hd)
                .arg(&nh)
                .arg(&nsp);
            unsafe {
                b2.launch(cfg2)?;
            }
            return Ok(());
        }
        let fc = if g {
            self.func_g("fa_decode_combine_f32")
        } else {
            self.fa_func("fa_decode_combine_f32", head_dim)
        };
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&fc);
        b2.arg(&*part_o)
            .arg(&*part_m)
            .arg(&*part_l)
            .arg(o)
            .arg(&hd)
            .arg(&nh)
            .arg(&nsp);
        unsafe {
            b2.launch(cfg2)?;
        }
        Ok(())
    }

    /// _dcw append (t=1): physical write row = len_dev[0] - base_dev[0] in-kernel; follow
    /// with `inc_i32(len_dev)` on the same stream. Bit-identical bytes to the host-row append
    /// at equal rows.
    #[allow(clippy::too_many_arguments)]
    pub fn append_kv_quantized_dcw(
        &self,
        k_row: &CudaSlice<f32>,
        v_row: &CudaSlice<f32>,
        kc: &mut CudaSlice<u8>,
        vc: &mut CudaSlice<u8>,
        len_dev: &CudaSlice<i32>,
        base_dev: Option<&CudaSlice<i32>>,
        kv_dim_k: usize,
        kv_dim_v: usize,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("append_quantize_kv_q8_0_q5_1_dcw");
        let nblk = (kv_dim_k.max(kv_dim_v) / 32) as u32;
        let cfg = LaunchConfig {
            grid_dim: (nblk, 1, 1),
            block_dim: (32, 1, 1),
            shared_mem_bytes: 0,
        };
        let (kdk, kdv) = (kv_dim_k as i32, kv_dim_v as i32);
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let null: u64 = 0;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(k_row).arg(v_row).arg(kc).arg(vc).arg(len_dev);
        match base_dev {
            Some(base) => {
                b.arg(base);
            }
            None => {
                b.arg(&null);
            }
        }
        b.arg(&kdk).arg(&kdv).arg(&ktb).arg(&vtb);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Increment a device i32 counter (graph-capturable; the `inc_i32` kernel).
    pub fn inc_i32(&self, counter: &mut CudaSlice<i32>) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("inc_i32");
        let cfg = LaunchConfig {
            grid_dim: (1, 1, 1),
            block_dim: (1, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(counter);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Windowed device-counter fa decode (step TP graph increment A): the KV view derives
    /// entirely from device state — `len_dev` (staged length), `base_dev` (physical row of
    /// logical 0 after the last ring rebase; None reads as 0), and `window` (0 = global) — so
    /// a captured child replays with ZERO per-token node updates. v3-vec only (the default
    /// kernel class on this lane); callers keep eager below the vec floor and for any other
    /// class. Scratch comes from the engine's fa partial pool sized at `bucket_max` (for SWA
    /// layers pass min(bucket, window)); the pool's retire-on-grow keeps captured addresses
    /// alive across bucket growth.
    #[allow(clippy::too_many_arguments)]
    /// Retire-on-grow ensure for the fa partial pool (see the #68 comment on the eager
    /// twin). Split out so graph capture can pre-run it OUTSIDE the capture region — an
    /// alloc inside a captured section becomes a mem node, and child graphs reject those.
    fn fa_part_pool_grow(
        &self,
        part_guard: &mut Option<(CudaSlice<f32>, CudaSlice<f32>, CudaSlice<f32>)>,
        o_len: usize,
        ml_len: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if part_guard
            .as_ref()
            .map(|pp| pp.0.len() < o_len || pp.1.len() < ml_len)
            .unwrap_or(true)
        {
            let old = part_guard.take();
            let (co, cm) = old
                .as_ref()
                .map(|pp| (pp.0.len(), pp.1.len()))
                .unwrap_or((0, 0));
            if let Some(old) = old {
                self.fa_part_retired.lock().unwrap().push(old);
            }
            *part_guard = Some((
                self.alloc_uninit::<f32>(o_len.max(2 * co))?,
                self.alloc_uninit::<f32>(ml_len.max(2 * cm))?,
                self.alloc_uninit::<f32>(ml_len.max(2 * cm))?,
            ));
        }
        Ok(())
    }

    /// Pre-grow the fa partial pool for a dcw call at (n_head, bucket_max) geometry, from
    /// OUTSIDE any capture region. Idempotent and cheap when already big enough.
    pub fn fa_dcw_pool_ensure(
        &self,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        bucket_max: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let sp = fa_split_keys(bucket_max, n_head_kv);
        let n_splits = ((bucket_max + sp - 1) / sp).max(1);
        let o_len = n_head * n_splits * head_dim;
        let ml_len = n_head * n_splits;
        let mut part_guard = self.fa_part_pool.lock().unwrap();
        self.fa_part_pool_grow(&mut part_guard, o_len, ml_len)
    }

    /// T=2 dcw decode attention (MEMRA_SPEC_FA2): both verify columns' rows are ALREADY
    /// appended; one launch walks the KV stream once with two query rows (per-row causal
    /// bounds len-1 / len) and the per-row combine consumes each half of the partials.
    /// BIT-IDENTICAL per row to that row's own per-column launch under the equal-partition
    /// guard the caller enforces (ns_eff/per equal for both bounds; boundary rounds fall
    /// back per column). `q2` = [2, n_head, head_dim]; `o2` = [2, n_head*head_dim] gated
    /// outputs (the head gate fuses into the combine as in the t=1 path).
    #[allow(clippy::too_many_arguments)]
    pub fn fa_decode_dcw2(
        &self,
        q2: &CudaSlice<f32>,
        k_ring: &cudarc::driver::CudaView<u8>,
        v_ring: &cudarc::driver::CudaView<u8>,
        o2: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        len_dev: &CudaSlice<i32>,
        base_dev: Option<&CudaSlice<i32>>,
        window: usize,
        bucket_max: usize,
        scale: f32,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        gate2: &CudaSlice<f32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let fa_vec = std::env::var("MEMRA_NO_FA_VEC").is_err() && bucket_max >= fa_vec_min_tkv();
        if !fa_vec || head_dim > 256 || head_dim % 32 != 0 || !fa_v3_on() {
            return Err("fa_decode_dcw2 supports the default v3-vec class only".into());
        }
        let sp = fa_split_keys(bucket_max, n_head_kv);
        let n_splits = ((bucket_max + sp - 1) / sp).max(1);
        // Partials for BOTH rows: row-major halves.
        let o_len = 2 * n_head * n_splits * head_dim;
        let ml_len = 2 * n_head * n_splits;
        let mut part_guard = self.fa_part_pool.lock().unwrap();
        Self::fa_part_pool_grow(self, &mut part_guard, o_len, ml_len)?;
        let pg = part_guard.as_mut().unwrap();
        let (part_o, part_m, part_l) = (&mut pg.0, &mut pg.1, &mut pg.2);
        let (hd, nh, nhkv, nsp) = (
            head_dim as i32,
            n_head as i32,
            n_head_kv as i32,
            n_splits as i32,
        );
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let (ski, win) = (sp as i32, window as i32);
        let gqa = (n_head / n_head_kv).max(1) as u32;
        let smem = (32 * head_dim * 2) as u32;
        let f = self.func("fa_decode_vec_q_v3_dcw2");
        let cfg = LaunchConfig {
            grid_dim: (n_head_kv as u32, n_splits as u32, 1),
            block_dim: (32, gqa, 1),
            shared_mem_bytes: smem,
        };
        let null: u64 = 0;
        {
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(q2)
                .arg(k_ring)
                .arg(v_ring)
                .arg(&mut *part_o)
                .arg(&mut *part_m)
                .arg(&mut *part_l)
                .arg(&hd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(len_dev);
            match base_dev {
                Some(base) => {
                    b.arg(base);
                }
                None => {
                    b.arg(&null);
                }
            }
            b.arg(&win)
                .arg(&scale)
                .arg(&nsp)
                .arg(&ski)
                .arg(&ktb)
                .arg(&vtb);
            unsafe {
                b.launch(cfg)?;
            }
        }
        // Per-row combine+gate: the t=1 combine kernel over each half (its `head` axis spans
        // 2*n_head rows laid out row-major, and the gate rows are stacked the same way), so
        // one launch covers both rows with the exact t=1 program per (row, head).
        let fc = {
            static CS: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
            if *CS.get_or_init(|| std::env::var("MEMRA_FA_COMBINE_S").as_deref() == Ok("1")) {
                self.func("fa_decode_combine_gate_f32_s")
            } else {
                self.func("fa_decode_combine_gate_f32")
            }
        };
        let combine_shared = std::env::var("MEMRA_FA_COMBINE_S").as_deref() == Ok("1");
        let nh2 = (2 * n_head) as i32;
        let cfg2 = LaunchConfig {
            grid_dim: ((2 * n_head) as u32, 1, 1),
            block_dim: (head_dim as u32, 1, 1),
            shared_mem_bytes: if combine_shared {
                (2 * n_splits * 4) as u32
            } else {
                0
            },
        };
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&fc);
        b2.arg(&*part_o)
            .arg(&*part_m)
            .arg(&*part_l)
            .arg(gate2)
            .arg(o2)
            .arg(&hd)
            .arg(&nh2)
            .arg(&nsp);
        unsafe {
            b2.launch(cfg2)?;
        }
        Ok(())
    }

    /// T-ROW dcw decode attention over a per-row session table (the per-session
    /// distributed-KV primitive). `tab` = t entries of five u64 words {k_ring, v_ring,
    /// len_ptr, base_ptr, len_back}; every (row, head, split) block runs the t=1 dcw
    /// program verbatim with that row's ring/len/base and its own split geometry, so each
    /// row is bit-identical to its own per-row launch. The kernel embeds the big-rig
    /// split ladder, so this refuses when the ladder env overrides are armed or the rig
    /// is not the >=128-SM class. `q_rows` = [t, n_head, head_dim]; `o_rows` = [t,
    /// n_head*head_dim] gated; `gate_rows` = [t, n_head].
    #[allow(clippy::too_many_arguments)]
    pub fn fa_decode_dcw_rows(
        &self,
        q_rows: &CudaSlice<f32>,
        tab: &CudaSlice<u64>,
        o_rows: &mut CudaSlice<f32>,
        t: usize,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        window: usize,
        max_ns: usize,
        scale: f32,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        gate_rows: &CudaSlice<f32>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if std::env::var("MEMRA_NO_FA_VEC").is_ok()
            || head_dim > 256
            || head_dim % 32 != 0
            || !fa_v3_on()
        {
            return Err("fa_decode_dcw_rows supports the default v3-vec class only".into());
        }
        if fa_sm_count() < 128
            || std::env::var("MEMRA_FA_SPLIT").is_ok()
            || std::env::var("MEMRA_FA_SP_SHORT").is_ok()
            || std::env::var("MEMRA_FA_SP16").is_ok()
        {
            return Err(
                "fa_decode_dcw_rows embeds the big-rig split ladder; env split overrides \
                 (or a <128-SM rig) keep the per-row path"
                    .into(),
            );
        }
        if t == 0 || t > 32 || max_ns == 0 || tab.len() < t * 6 {
            return Err("fa_decode_dcw_rows geometry".into());
        }
        let o_len = t * n_head * max_ns * head_dim;
        let ml_len = t * n_head * max_ns;
        let mut part_guard = self.fa_part_pool.lock().unwrap();
        Self::fa_part_pool_grow(self, &mut part_guard, o_len, ml_len)?;
        let pg = part_guard.as_mut().unwrap();
        let (part_o, part_m, part_l) = (&mut pg.0, &mut pg.1, &mut pg.2);
        let (hd, nh, nhkv) = (head_dim as i32, n_head as i32, n_head_kv as i32);
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let (win, mns) = (window as i32, max_ns as i32);
        let gqa = (n_head / n_head_kv).max(1) as u32;
        let smem = (32 * head_dim * 2) as u32;
        let f = self.func("fa_decode_vec_q_v3_dcw_rows");
        let cfg = LaunchConfig {
            grid_dim: (n_head_kv as u32, max_ns as u32, t as u32),
            block_dim: (32, gqa, 1),
            shared_mem_bytes: smem,
        };
        {
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(q_rows)
                .arg(tab)
                .arg(&mut *part_o)
                .arg(&mut *part_m)
                .arg(&mut *part_l)
                .arg(&hd)
                .arg(&nh)
                .arg(&nhkv)
                .arg(&win)
                .arg(&scale)
                .arg(&mns)
                .arg(&ktb)
                .arg(&vtb);
            unsafe {
                b.launch(cfg)?;
            }
        }
        // Per-(row, head) combine+gate: the t=1 combine over t*n_head stacked heads —
        // row r head h reads its own partial bank; splits past a row's ns_eff carry
        // (-inf, 0) partials the NEG_INF guard no-ops bit-exactly.
        let fc = {
            static CS: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
            if *CS.get_or_init(|| std::env::var("MEMRA_FA_COMBINE_S").as_deref() == Ok("1")) {
                self.func("fa_decode_combine_gate_f32_s")
            } else {
                self.func("fa_decode_combine_gate_f32")
            }
        };
        let combine_shared = std::env::var("MEMRA_FA_COMBINE_S").as_deref() == Ok("1");
        let nht = (t * n_head) as i32;
        let cfg2 = LaunchConfig {
            grid_dim: ((t * n_head) as u32, 1, 1),
            block_dim: (head_dim as u32, 1, 1),
            shared_mem_bytes: if combine_shared {
                (2 * max_ns * 4) as u32
            } else {
                0
            },
        };
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&fc);
        b2.arg(&*part_o)
            .arg(&*part_m)
            .arg(&*part_l)
            .arg(gate_rows)
            .arg(o_rows)
            .arg(&hd)
            .arg(&nht)
            .arg(&mns);
        unsafe {
            b2.launch(cfg2)?;
        }
        Ok(())
    }

    pub fn fa_decode_dcw(
        &self,
        q: &CudaSlice<f32>,
        k_ring: &cudarc::driver::CudaView<u8>,
        v_ring: &cudarc::driver::CudaView<u8>,
        o: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        n_head_kv: usize,
        len_dev: &CudaSlice<i32>,
        base_dev: Option<&CudaSlice<i32>>,
        window: usize,
        bucket_max: usize,
        scale: f32,
        k_tok_bytes: usize,
        v_tok_bytes: usize,
        // FUSION #2d: Some(gate_row) fuses the head gate into the combine (bit-identical,
        // one launch saved); `o` then receives the GATED output and the caller skips its
        // attn_head_gate call.
        fused_gate: Option<&CudaSlice<f32>>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let fa_vec = std::env::var("MEMRA_NO_FA_VEC").is_err() && bucket_max >= fa_vec_min_tkv();
        if !fa_vec || head_dim > 256 || head_dim % 32 != 0 || !fa_v3_on() {
            return Err("fa_decode_dcw supports the default v3-vec class only                         (bucket >= vec floor, head_dim <= 256, MEMRA_FA_V3 on);                         keep eager outside it"
                .into());
        }
        let sp = fa_split_keys(bucket_max, n_head_kv);
        let n_splits = ((bucket_max + sp - 1) / sp).max(1);
        let o_len = n_head * n_splits * head_dim;
        let ml_len = n_head * n_splits;
        let mut part_guard = self.fa_part_pool.lock().unwrap();
        Self::fa_part_pool_grow(self, &mut part_guard, o_len, ml_len)?;
        let pg = part_guard.as_mut().unwrap();
        // MEMRA_FA_DCW_MEMSET=0: skip the partial-pool zeroing — every (head, split) in
        // [0, nsp) writes its partial before the combine reads it (per = ceil(len/nsp), so
        // split s starts at s*per < len for all s < nsp), making the zeros dead stores.
        // Door-gated pending the identity battery; =0 saves 3 memset launches/rank/layer.
        static MEMSET_ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        // Token-graph capture ALWAYS keeps the memsets: the retarget path (increment C)
        // finds the attention children BY their three-memset signature and updates the
        // memset widths per bucket — capturing without them silently kills retargeting
        // (battery-v8 token drift, 2026-08-21).
        let memset_on = *MEMSET_ON
            .get_or_init(|| std::env::var("MEMRA_FA_DCW_MEMSET").as_deref() != Ok("0"))
            || crate::tp::token_graph_building();
        if memset_on {
            self.gpu
                .stream()
                .memset_zeros(&mut pg.0.slice_mut(0..o_len))?;
            self.gpu
                .stream()
                .memset_zeros(&mut pg.1.slice_mut(0..ml_len))?;
            self.gpu
                .stream()
                .memset_zeros(&mut pg.2.slice_mut(0..ml_len))?;
        }
        let (part_o, part_m, part_l) = (&mut pg.0, &mut pg.1, &mut pg.2);
        let (hd, nh, nhkv, nsp) = (
            head_dim as i32,
            n_head as i32,
            n_head_kv as i32,
            n_splits as i32,
        );
        let (ktb, vtb) = (k_tok_bytes as i64, v_tok_bytes as i64);
        let (ski, win) = (sp as i32, window as i32);
        let gqa = (n_head / n_head_kv).max(1) as u32;
        let smem = (32 * head_dim * 2) as u32; // sV bf16 [FA_DEC_TILE=32][hd] (v3 uses sV only)
        // MEMRA_FA_UNROLL=8: the B1-unroll-8 twin (deeper K load pipeline, bit-identical —
        // see fa_dec_v3_walk_u). Same launch geometry.
        static U8: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        static HOIST: std::sync::OnceLock<u8> = std::sync::OnceLock::new();
        let hoist = *HOIST.get_or_init(|| match std::env::var("MEMRA_FA_HOIST").as_deref() {
            Ok("2") => 2,
            Ok("1") => 1,
            _ => 0,
        });
        // MEMRA_FA_PROF=1: clock64() phase profile of the decode-attention walk. ncu is
        // permission-blocked in this container and the module params are not exposed, so this
        // is how the ~1.18us/key gets localised. Diagnostic only (extra atomics per block);
        // prints cumulative cycle shares every 430 launches.
        static FPROF: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let fprof = *FPROF.get_or_init(|| std::env::var("MEMRA_FA_PROF").as_deref() == Ok("1"));
        static PROF_BUF: std::sync::Mutex<Option<(usize, CudaSlice<u64>)>> =
            std::sync::Mutex::new(None);
        // MEMRA_FA_HSPLIT=2: split each kv_head's gqa warp group across TWO blocks (2x grid,
        // duplicated Phase A staging) — bit-identical per (head, split). Tests whether B1's
        // 59-63% cycle share is occupancy-starved latency (grid is only n_head_kv x n_splits).
        static HS: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        let hs2 = *HS.get_or_init(|| std::env::var("MEMRA_FA_HSPLIT").as_deref() == Ok("2"))
            && (n_head / n_head_kv) % 2 == 0
            && (n_head / n_head_kv) >= 2;
        let f = if fprof {
            self.func("fa_decode_vec_q_v3_dcw_prof")
        } else if hs2 {
            self.func("fa_decode_vec_q_v3_dcw_hs2")
        } else if hoist == 2 {
            // + typed 4-byte K loads (memcpy from uint8_t* can lower to byte loads).
            self.func("fa_decode_vec_q_v3_dcw_hc")
        } else if hoist == 1 {
            // Loop-invariant K alignment class hoisted out of B1 (bit-identical).
            self.func("fa_decode_vec_q_v3_dcw_h")
        } else if *U8.get_or_init(|| std::env::var("MEMRA_FA_UNROLL").as_deref() == Ok("8")) {
            self.func("fa_decode_vec_q_v3_dcw_u8")
        } else {
            self.func("fa_decode_vec_q_v3_dcw")
        };
        let cfg = LaunchConfig {
            grid_dim: if hs2 {
                ((2 * n_head_kv) as u32, n_splits as u32, 1)
            } else {
                (n_head_kv as u32, n_splits as u32, 1)
            },
            block_dim: if hs2 { (32, gqa / 2, 1) } else { (32, gqa, 1) },
            shared_mem_bytes: smem,
        };
        let null: u64 = 0;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k_ring)
            .arg(v_ring)
            .arg(&mut *part_o)
            .arg(&mut *part_m)
            .arg(&mut *part_l)
            .arg(&hd)
            .arg(&nh)
            .arg(&nhkv)
            .arg(len_dev);
        match base_dev {
            Some(base) => {
                b.arg(base);
            }
            None => {
                b.arg(&null);
            }
        }
        b.arg(&win)
            .arg(&scale)
            .arg(&nsp)
            .arg(&ski)
            .arg(&ktb)
            .arg(&vtb);
        if fprof {
            let mut guard = PROF_BUF.lock().map_err(|_| "fa prof buffer lock")?;
            if guard
                .as_ref()
                .is_none_or(|(d, _)| *d != self.ctx().ordinal())
            {
                *guard = Some((self.ctx().ordinal(), self.htod_u64(&vec![0u64; 8])?));
            }
            let (_, buf) = guard.as_mut().expect("armed above");
            b.arg(&*buf);
            unsafe {
                b.launch(cfg)?;
            }
            static CALLS: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(0);
            let n = CALLS.fetch_add(1, std::sync::atomic::Ordering::Relaxed) + 1;
            if n % 430 == 0 {
                self.stream().synchronize()?;
                let h = self.dtoh_u64(buf)?;
                let phases = ["setup", "stageV", "b1_klo", "b2_soft", "sync", "b3_vacc"];
                let tot: u64 = h[..6].iter().sum();
                let mut line = format!("[fa-prof] calls={n} keys={} cycles={tot}", h[6]);
                for (i, name) in phases.iter().enumerate() {
                    let pct = if tot > 0 {
                        h[i] as f64 / tot as f64 * 100.0
                    } else {
                        0.0
                    };
                    line.push_str(&format!(" {name}={pct:.1}%"));
                }
                if h[6] > 0 {
                    line.push_str(&format!(" cyc/key={:.0}", tot as f64 / h[6] as f64));
                }
                eprintln!("{line}");
            }
        } else {
            unsafe {
                b.launch(cfg)?;
            }
        }
        let mut combine_shared = false;
        let fc = if fused_gate.is_some() {
            // MEMRA_FA_COMBINE_S=1: shared-staged split metadata (bit-identical; kills the
            // n_splits-deep dependent global load chain every thread used to walk twice).
            static CS: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
            if *CS.get_or_init(|| std::env::var("MEMRA_FA_COMBINE_S").as_deref() == Ok("1")) {
                combine_shared = true;
                self.func("fa_decode_combine_gate_f32_s")
            } else {
                self.func("fa_decode_combine_gate_f32")
            }
        } else {
            self.fa_func("fa_decode_combine_f32", head_dim)
        };
        let cfg2 = LaunchConfig {
            grid_dim: (n_head as u32, 1, 1),
            block_dim: (head_dim as u32, 1, 1),
            shared_mem_bytes: if combine_shared {
                (2 * n_splits * 4) as u32
            } else {
                0
            },
        };
        let __s_b2 = self.gpu.stream();
        let mut b2 = __s_b2.launch_builder(&fc);
        b2.arg(&*part_o).arg(&*part_m).arg(&*part_l);
        if let Some(gate_row) = fused_gate {
            b2.arg(gate_row);
        }
        b2.arg(o).arg(&hd).arg(&nh).arg(&nsp);
        unsafe {
            b2.launch(cfg2)?;
        }
        Ok(())
    }

    /// EAGER fa_decode geometry for a given actual `t_kv` (CUDA-GRAPH-PLAN §3.3 bucketing). Returns
    /// `(fa_vec, n_splits)` EXACTLY as `fa_decode` computes them so the graph-capture path can key its
    /// bucket on the same `(kernel, n_splits)` pair and pass a `bucket_max` that reproduces eager's
    /// n_splits bit-for-bit. (Per = ceil(t_kv/n_splits) is then recomputed from the DEVICE t_kv inside
    /// the kernel and matches eager when n_splits matches — the bit-identity contract.)
    pub fn fa_geom_eager(
        &self,
        t_kv: usize,
        head_dim: usize,
        n_head_kv: usize,
        g: bool,
    ) -> (bool, usize) {
        // MUST mirror `fa_decode` / `fa_decode_dc` (default-ON 2026-06-28). This is the bucket-key
        // source: if it disagrees with the actual kernel pick, the graph captures the wrong path and
        // replay diverges from eager. All three sites read MEMRA_NO_FA_VEC in lockstep.
        let fa_ok = std::env::var("MEMRA_NO_FA_VEC").is_err() && t_kv >= fa_vec_min_tkv();
        // hd512 dpl16 vec lane (gemma globals, 2026-07-11 graph-arc fix): the original key
        // hardcoded vec = hd<=256, so for hd512 it bucketed by the SCALAR 256-key splits while
        // the dpl16/rows_dpl16 kernels split by the ladder — n_splits changed WITHIN a bucket
        // (mid-ctx graph mismatch at pos 19 + partials OOB at longer runs). Mirror the real
        // fa_decode dispatch: vec512 above the fa512 floor, vec256 as before.
        let vec512 = fa_ok && head_dim == 512 && t_kv >= fa512_min_tkv();
        let mut fa_vec = vec512 || (fa_ok && head_dim <= 256 && head_dim % 32 == 0);
        // g (fp8-windowed): mirror kvmod's clamp — only the v4 lane parses e4m3 in the vec
        // family; everything else falls to the g-module scalar.
        // hd256 under g: v4-or-scalar. hd128/other under g ride the REGISTER g-lane (the
        // dq_K_lane/dq_V_lane macros are format-aware; v3 is format-gated off, v2/smem
        // arms are excluded under g). Mirrored in kvmod / fa_decode_dc / fa_geom_eager.
        if g && head_dim == 256 && !fa_v4_at(t_kv) {
            fa_vec = false;
        }
        let sp = fa_split_keys(t_kv, n_head_kv);
        let n_splits = if fa_vec {
            ((t_kv + sp - 1) / sp).max(1)
        } else {
            ((t_kv + 255) / 256).max(1)
        };
        (fa_vec, n_splits)
    }

    /// `bucket_max` (host t_kv to feed `fa_decode_dc` / `full_attn_decode_dc`) that makes the _dc
    /// kernel pick the SAME (fa_vec, n_splits) as eager would for actual `t_kv`. Because the dc
    /// launcher derives both from `bucket_max` via the same formulas, we just hand it `t_kv` itself:
    /// the n_splits is then identical, and the per-split boundaries (computed from the DEVICE t_kv in
    /// the kernel) match eager exactly. The bucket KEY (for the graph HashMap) is `(fa_vec, n_splits)`.
    pub fn fa_bucket_key(
        &self,
        t_kv: usize,
        head_dim: usize,
        n_head_kv: usize,
        g: bool,
    ) -> (bool, usize) {
        self.fa_geom_eager(t_kv, head_dim, n_head_kv, g)
    }

    /// CUDA-graph capture wrapper (CUDA-GRAPH-PLAN §3.2, llama.cpp warmup pattern). Runs `step`
    /// inline TWICE (warmup — lets the caching allocator settle to stable pointers and any one-time
    /// kernel attribute/JIT happen outside capture), then captures a THIRD invocation on the Engine's
    /// decode stream (RELAXED mode) and instantiates it into a replayable `CudaGraph`. The closure
    /// must enqueue ONLY device work on `e.stream()` (no dtoh / no synchronize / no host branch on
    /// device data) — every per-step varying scalar must come from a device counter. Returns the
    /// instantiated graph; `CudaGraph::launch()` replays the whole step in one dispatch.
    /// `capture_graph` with CAPTURE-RETAIN: every Engine allocation made during the warmups
    /// and the capture is kept alive in the returned keeper — hold it as long as the graph
    /// replays (transients returning to the pool get reused by unrelated work and corrupt
    /// replays; the draft-graph root cause). Model-generic, next capture reuses it.
    pub fn capture_graph_retained<F>(
        &self,
        step: F,
    ) -> Result<
        (
            cudarc::driver::CudaGraph,
            Vec<Box<dyn std::any::Any + Send>>,
        ),
        Box<dyn std::error::Error>,
    >
    where
        F: FnMut(&Engine) -> Result<(), Box<dyn std::error::Error>>,
    {
        use cudarc::driver::sys::CUgraphInstantiate_flags;
        self.capture_graph_retained_flags(
            CUgraphInstantiate_flags::CUDA_GRAPH_INSTANTIATE_FLAG_AUTO_FREE_ON_LAUNCH,
            step,
        )
    }

    /// Retained capture with an explicit instantiate flag. ALLOC-FREE captured graphs
    /// (zero mem nodes — the gemma slotted door) should pass UPLOAD instead of
    /// AUTO_FREE_ON_LAUNCH: the auto-free flag's launch-time mem-pool scan was measured at
    /// ~0.25us/node (205us on the 826-node step) even with nothing to free.
    pub fn capture_graph_retained_flags<F>(
        &self,
        flags: cudarc::driver::sys::CUgraphInstantiate_flags,
        mut step: F,
    ) -> Result<
        (
            cudarc::driver::CudaGraph,
            Vec<Box<dyn std::any::Any + Send>>,
        ),
        Box<dyn std::error::Error>,
    >
    where
        F: FnMut(&Engine) -> Result<(), Box<dyn std::error::Error>>,
    {
        use cudarc::driver::sys::CUstreamCaptureMode;
        // KEEP scope = WARMUPS ONLY (2026-07-13): keep_if_capturing retains via
        // CudaSlice::clone, which is a device ALLOC + D2D COPY on the stream — clones made
        // while the capture region is open become dead copy NODES replayed every launch
        // (E4B: 1440 copies = 0.74ms/token, the whole graph-vs-eager regression). The
        // warmup runs allocate the same transient sequence at the same pool addresses, so
        // retaining the warmup clones preserves the draft-graph fix without polluting the
        // captured graph.
        self.capture_keep.lock().unwrap().clear();
        let was_tracking = self.gpu.ctx.is_event_tracking();
        if was_tracking {
            unsafe {
                self.gpu.ctx.disable_event_tracking();
            }
        }
        let mut run = || -> Result<cudarc::driver::CudaGraph, Box<dyn std::error::Error>> {
            self.capture_keep_on
                .store(true, std::sync::atomic::Ordering::Relaxed);
            let w = (|| {
                step(self)?;
                step(self)
            })();
            self.capture_keep_on
                .store(false, std::sync::atomic::Ordering::Relaxed);
            w?;
            self.gpu.stream().synchronize()?;
            self.gpu
                .stream()
                .begin_capture(CUstreamCaptureMode::CU_STREAM_CAPTURE_MODE_RELAXED)?;
            let r = step(self);
            let g = self.gpu.stream().end_capture(flags);
            r?;
            let graph = g?.ok_or("capture produced no graph (stream was not capturing)")?;
            graph.upload()?;
            Ok(graph)
        };
        let result = run();
        self.capture_keep_on
            .store(false, std::sync::atomic::Ordering::Relaxed);
        if was_tracking {
            unsafe {
                self.gpu.ctx.enable_event_tracking();
            }
        }
        let keeper = std::mem::take(&mut *self.capture_keep.lock().unwrap());
        Ok((result?, keeper))
    }

    /// Retained capture WITHOUT the two warmup executions. The warmups exist for transient
    /// pool-address stability (draft-graph lanes); the step TP token-graph sections are
    /// alloc-free with persistent operands, and their bodies carry device side effects
    /// (dcw KV appends + counter incs) that a warmup would REALLY EXECUTE — measured as a
    /// +2/rank len_d drift per bucket build that marched appends past the ring planes.
    pub fn capture_graph_retained_nowarm<F>(
        &self,
        mut step: F,
    ) -> Result<
        (
            cudarc::driver::CudaGraph,
            Vec<Box<dyn std::any::Any + Send>>,
        ),
        Box<dyn std::error::Error>,
    >
    where
        F: FnMut(&Engine) -> Result<(), Box<dyn std::error::Error>>,
    {
        use cudarc::driver::sys::{CUgraphInstantiate_flags, CUstreamCaptureMode};
        let was_tracking = self.gpu.ctx.is_event_tracking();
        if was_tracking {
            unsafe {
                self.gpu.ctx.disable_event_tracking();
            }
        }
        let mut run = || -> Result<cudarc::driver::CudaGraph, Box<dyn std::error::Error>> {
            self.gpu.stream().synchronize()?;
            self.gpu
                .stream()
                .begin_capture(CUstreamCaptureMode::CU_STREAM_CAPTURE_MODE_RELAXED)?;
            let r = step(self);
            let g = self.gpu.stream().end_capture(
                CUgraphInstantiate_flags::CUDA_GRAPH_INSTANTIATE_FLAG_AUTO_FREE_ON_LAUNCH,
            );
            r?;
            let graph = g?.ok_or("capture produced no graph (stream was not capturing)")?;
            graph.upload()?;
            Ok(graph)
        };
        let result = run();
        if was_tracking {
            unsafe {
                self.gpu.ctx.enable_event_tracking();
            }
        }
        Ok((result?, Vec::new()))
    }

    pub fn capture_graph<F>(
        &self,
        mut step: F,
    ) -> Result<cudarc::driver::CudaGraph, Box<dyn std::error::Error>>
    where
        F: FnMut(&Engine) -> Result<(), Box<dyn std::error::Error>>,
    {
        use cudarc::driver::sys::{CUgraphInstantiate_flags, CUstreamCaptureMode};
        // EVENT TRACKING OFF for capture. The Engine creates a 2nd stream (copy_stream) so cudarc is in
        // multi-stream mode and, by default, records a CudaEvent per CudaSlice alloc/use to serialize
        // cross-stream access. Those per-buffer event waits issue stream ops that are NOT permitted
        // inside a capture region (CUDA_ERROR_STREAM_CAPTURE_UNSUPPORTED). The captured decode step is
        // strictly SINGLE-STREAM (every kernel on gpu.stream), so this synchronization is unnecessary
        // here — disable it for the whole warmup+capture, re-enable after. SAFETY: the decode-dc path
        // touches only gpu.stream; no buffer crosses to copy_stream during capture.
        let was_tracking = self.gpu.ctx.is_event_tracking();
        if was_tracking {
            unsafe {
                self.gpu.ctx.disable_event_tracking();
            }
        }
        // Q1 PROBE (MEMRA_GRAPH_IFLAG): the generic capture body's cuMemAllocAsync nodes are
        // EXACTLY BALANCED by in-graph free nodes (measured census q27: 1589 ALLOC / 1589
        // FREE), so AUTO_FREE_ON_LAUNCH has nothing to reclaim at launch — it only pays its
        // per-node launch-time mem-pool scan. `upload` / `none` select the alternatives to
        // measure that scan's real cost on the generic path. Diagnostic door only; the
        // default stays AUTO_FREE until a measured A/B justifies moving it.
        let iflag = {
            static F: std::sync::OnceLock<CUgraphInstantiate_flags> = std::sync::OnceLock::new();
            *F.get_or_init(|| match std::env::var("MEMRA_GRAPH_IFLAG").as_deref() {
                // UPLOAD = the gemma slotted door's zero-mem-node choice; PRIORITY = the flag
                // hybrid_forward.rs:5935 actually ships (both drop the auto-free launch scan).
                Ok("upload") => CUgraphInstantiate_flags::CUDA_GRAPH_INSTANTIATE_FLAG_UPLOAD,
                Ok("priority") => {
                    CUgraphInstantiate_flags::CUDA_GRAPH_INSTANTIATE_FLAG_USE_NODE_PRIORITY
                }
                _ => CUgraphInstantiate_flags::CUDA_GRAPH_INSTANTIATE_FLAG_AUTO_FREE_ON_LAUNCH,
            })
        };
        // MEMRA_GRAPH_CAPTIME=1 (Q1 lane): phase-resolved capture cost. Recapture is paid at
        // every kernel-class crossing, so it — not steady-state decode — is the quantity a
        // mem-node reduction could plausibly shrink. Only `instantiate` (cuStreamEndCapture +
        // cuGraphInstantiateWithFlags) and `upload` scale with node count; the warmups are
        // eager step executions and are node-count-invariant. Printing the split bounds the
        // refactor's ceiling instead of assuming it.
        let ct = {
            static T: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
            *T.get_or_init(|| std::env::var("MEMRA_GRAPH_CAPTIME").as_deref() == Ok("1"))
        };
        // MEMRA_GRAPH_WARMUPS (Q1 lane; DEFAULT 1 since lane/graph-warmups 2026-08-05): the
        // phase split showed the eager warmups are 80% of recapture cost (q27 27.4 of 34.4 ms
        // pod / 42% of 52.6 ms 5090) — 3x larger than the ENTIRE mem-node ceiling the audit
        // chased, and node-count-invariant, so no capture-body refactor could touch it.
        // Warmup 2's theorized job was async-pool ADDRESS STABILITY: warmup 1's allocs may
        // grow/map the pool, warmup 2 re-walks the same sequence over the freed blocks so the
        // captured third run bakes settled addresses. That hazard is the #68 stale-baked-
        // address class — which the engine now guards STRUCTURALLY rather than by re-walking:
        // in-body transients are captured as BALANCED in-graph alloc/free node pairs (census
        // 1589/1589 — replays allocate for themselves; no baked transient pointers), every
        // externally-referenced buffer is stable-pointer by design (fa_part_pool retires-on-
        // grow and never frees, resident counters/scratch, cache set in place), and the
        // draft-graph path additionally rides capture_graph_retained (capture_keep holds all
        // warmup+capture allocs alive). One warmup therefore suffices for kernel-attr
        // settling and pool mapping. Arbitrated adversarially, not by taste:
        // graph-warmup-stress (pool-growth cycles large<->small x10, overlap arm, forced
        // recaptures over freed blocks — bit-identity vs eager + canary teeth) is GREEN at
        // warmups=1 on the deployment rig, plus graph-decode-gate 256-step bit-identity,
        // graph-session-gate, run-spec K=1..8 (receipts research/graph-warmups-5090-20260805/
        // + the pod's research/graph-allocfree-20260805/). Measured: recapture -38..-42% q27 /
        // -41% q9, decode +~1%, capture+prime -13ms. MEMRA_GRAPH_WARMUPS=2 = the rollback
        // seam; tools/graph-warmup-stress-gate.sh = the gate any regression re-runs.
        let warmups = {
            static W: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
            *W.get_or_init(|| {
                std::env::var("MEMRA_GRAPH_WARMUPS")
                    .ok()
                    .and_then(|v| v.parse().ok())
                    .filter(|n| *n >= 1)
                    .unwrap_or(1)
            })
        };
        let mut run = || -> Result<cudarc::driver::CudaGraph, Box<dyn std::error::Error>> {
            let t_w = std::time::Instant::now();
            // warmup: inline runs (no capture) so allocator pointers + kernel attrs are stable.
            for _ in 0..warmups {
                step(self)?;
            }
            self.gpu.stream().synchronize()?;
            let ms_warm = t_w.elapsed().as_secs_f64() * 1e3;
            // capture the third run.
            let t_c = std::time::Instant::now();
            self.gpu
                .stream()
                .begin_capture(CUstreamCaptureMode::CU_STREAM_CAPTURE_MODE_RELAXED)?;
            // If the body errors mid-capture, end the capture before propagating so the stream isn't
            // left in a capturing state.
            let r = step(self);
            let ms_body = t_c.elapsed().as_secs_f64() * 1e3;
            let t_i = std::time::Instant::now();
            let g = self.gpu.stream().end_capture(iflag);
            let ms_inst = t_i.elapsed().as_secs_f64() * 1e3;
            r?;
            let graph = g?.ok_or("capture produced no graph (stream was not capturing)")?;
            let t_u = std::time::Instant::now();
            graph.upload()?;
            if ct {
                println!(
                    "[graph-captime] warmup2x {ms_warm:.2} ms  capture-body {ms_body:.2} ms  \
                          instantiate {ms_inst:.2} ms  upload {:.2} ms",
                    t_u.elapsed().as_secs_f64() * 1e3
                );
            }
            Ok(graph)
        };
        let result = run();
        if was_tracking {
            unsafe {
                self.gpu.ctx.enable_event_tracking();
            }
        }
        result
    }

    /// gdn_scan variant where state_in/out are CudaViews (resident SSM state, in-place per step).
    pub fn gdn_scan_s128_view(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        g: &CudaSlice<f32>,
        beta: &CudaSlice<f32>,
        state_in: &cudarc::driver::CudaView<f32>,
        state_out: &mut cudarc::driver::CudaViewMut<f32>,
        o: &mut CudaSlice<f32>,
        n_head: usize,
        t: usize,
        scale: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gdn_scan_s128");
        const S_V: u32 = 128;
        const WARP: u32 = 32;
        const COLS: u32 = 4;
        let cfg = LaunchConfig {
            grid_dim: (n_head as u32, 1, S_V / COLS),
            block_dim: (WARP, COLS, 1),
            shared_mem_bytes: 0,
        };
        let (h, ti) = (n_head as i32, t as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(v)
            .arg(g)
            .arg(beta)
            .arg(state_in)
            .arg(state_out)
            .arg(o)
            .arg(&h)
            .arg(&ti)
            .arg(&scale);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// conv1d where the input is a CudaView (resident conv state assembled in place).
    pub fn ssm_conv1d_view(
        &self,
        x: &cudarc::driver::CudaView<f32>,
        w: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        conv_dim: usize,
        t: usize,
        d_conv: usize,
        silu: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("ssm_conv1d_silu_f32");
        // grid.x = channel, grid.y = T-tiles (block 256 strides over T) — parallel over both axes.
        let cfg = LaunchConfig {
            grid_dim: (conv_dim as u32, ((t as u32 + 255) / 256).max(1), 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (cd, ti, dc, s) = (conv_dim as i32, t as i32, d_conv as i32, silu as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(w).arg(y).arg(&cd).arg(&ti).arg(&dc).arg(&s);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Depthwise causal conv1d + optional SiLU.
    /// x:[conv_dim, T+d_conv-1] channel-major (first d_conv-1 cols = carried state),
    /// w:[d_conv, conv_dim] kernel-major, y:[conv_dim, T] channel-major.
    /// FUSED prefill conv (token-major input, zero left-state): replaces
    /// transpose + zeros + conv_left_pad + ssm_conv1d with ONE launch reading the matmul output
    /// directly. Output channel-major [conv_dim, T], SiLU applied. BIT-IDENTICAL accumulation.
    pub fn ssm_conv1d_tm(
        &self,
        qkv_tm: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        conv_dim: usize,
        t: usize,
        d_conv: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("ssm_conv1d_tm_f32");
        let cfg = LaunchConfig {
            grid_dim: (((conv_dim + 255) / 256) as u32, t as u32, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (cd, ti, dc) = (conv_dim as i32, t as i32, d_conv as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(qkv_tm).arg(w).arg(y).arg(&cd).arg(&ti).arg(&dc);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// BATCHED verify conv (T>1, carried state): window reads the resident conv ring for
    /// negative rows; separate ring-update launch afterwards. BIT-IDENTICAL per value to the
    /// T=1 chain. T >= pad rides the pure input-column ring update (unchanged legacy path);
    /// T < pad (the MEMRA_SPEC_M2 t=2 verify arm) needs old-ring sources for the roll — the
    /// update kernel would race reading the ring it rewrites, so that arm clones the ring
    /// (dtod) and rolls via ssm_conv_ring_rebuild (PURE COPIES: the ring stores raw input
    /// columns; the final ring == what T sequential decode ring rolls leave).
    pub fn ssm_conv1d_tm_state(
        &self,
        qkv_tm: &CudaSlice<f32>,
        conv_state: &mut CudaSlice<f32>,
        w: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        conv_dim: usize,
        t: usize,
        d_conv: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        self.ssm_conv1d_tm_state_pad(qkv_tm, conv_state, w, y, conv_dim, t, d_conv, None)
    }

    /// task #14: `pad_len` = device true length for PADDED prime graphs — the ring update
    /// reads rows [len-pad, len) instead of the pad tail. None = the classic host-T path.
    #[allow(clippy::too_many_arguments)]
    pub fn ssm_conv1d_tm_state_pad(
        &self,
        qkv_tm: &CudaSlice<f32>,
        conv_state: &mut CudaSlice<f32>,
        w: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        conv_dim: usize,
        t: usize,
        d_conv: usize,
        pad_len: Option<&CudaSlice<i32>>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        assert!(t >= 1, "ssm_conv1d_tm_state requires T >= 1");
        // clone BEFORE the window kernel is issued is not required (stream-ordered: the dtod and
        // the window kernel both read the pre-roll ring; the roll launches after both) — but
        // cloning first keeps the ordering trivially correct under any future stream split.
        let ring_old = if t < d_conv - 1 {
            Some(self.clone_dtod(conv_state)?)
        } else {
            None
        };
        {
            let f = self.func("ssm_conv1d_tm_state_f32");
            let cfg = LaunchConfig {
                grid_dim: (((conv_dim + 255) / 256) as u32, t as u32, 1),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let (cd, ti, dc) = (conv_dim as i32, t as i32, d_conv as i32);
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(qkv_tm)
                .arg(&*conv_state)
                .arg(w)
                .arg(y)
                .arg(&cd)
                .arg(&ti)
                .arg(&dc);
            unsafe {
                b.launch(cfg)?;
            }
        }
        match (ring_old, pad_len) {
            (None, Some(len_d)) => {
                let f = self.func("ssm_conv_ring_update_dev_f32");
                let n = conv_dim * (d_conv - 1);
                let cfg = LaunchConfig::for_num_elems(n as u32);
                let (cd, dc) = (conv_dim as i32, d_conv as i32);
                let __s_b = self.gpu.stream();
                let mut b = __s_b.launch_builder(&f);
                b.arg(qkv_tm).arg(conv_state).arg(len_d).arg(&cd).arg(&dc);
                unsafe {
                    b.launch(cfg)?;
                }
            }
            (None, None) => {
                let f = self.func("ssm_conv_ring_update_f32");
                let n = conv_dim * (d_conv - 1);
                let cfg = LaunchConfig::for_num_elems(n as u32);
                let (cd, ti, dc) = (conv_dim as i32, t as i32, d_conv as i32);
                let __s_b = self.gpu.stream();
                let mut b = __s_b.launch_builder(&f);
                b.arg(qkv_tm).arg(conv_state).arg(&cd).arg(&ti).arg(&dc);
                unsafe {
                    b.launch(cfg)?;
                }
            }
            (Some(old), _) => {
                self.ssm_conv_ring_rebuild(qkv_tm, &old, conv_state, conv_dim, t, d_conv)?
            }
        }
        Ok(())
    }

    /// qkv-view twin (task #16): batched prime reads the concat GEMM output directly.
    pub fn ssm_conv1d_tm_state_pad_v(
        &self,
        qkv_tm: &cudarc::driver::CudaView<f32>,
        conv_state: &mut CudaSlice<f32>,
        w: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        conv_dim: usize,
        t: usize,
        d_conv: usize,
        pad_len: Option<&CudaSlice<i32>>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        assert!(t >= 1, "ssm_conv1d_tm_state requires T >= 1");
        // clone BEFORE the window kernel is issued is not required (stream-ordered: the dtod and
        // the window kernel both read the pre-roll ring; the roll launches after both) — but
        // cloning first keeps the ordering trivially correct under any future stream split.
        let ring_old = if t < d_conv - 1 {
            Some(self.clone_dtod(conv_state)?)
        } else {
            None
        };
        {
            let f = self.func("ssm_conv1d_tm_state_f32");
            let cfg = LaunchConfig {
                grid_dim: (((conv_dim + 255) / 256) as u32, t as u32, 1),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let (cd, ti, dc) = (conv_dim as i32, t as i32, d_conv as i32);
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(qkv_tm)
                .arg(&*conv_state)
                .arg(w)
                .arg(y)
                .arg(&cd)
                .arg(&ti)
                .arg(&dc);
            unsafe {
                b.launch(cfg)?;
            }
        }
        match (ring_old, pad_len) {
            (None, Some(len_d)) => {
                let f = self.func("ssm_conv_ring_update_dev_f32");
                let n = conv_dim * (d_conv - 1);
                let cfg = LaunchConfig::for_num_elems(n as u32);
                let (cd, dc) = (conv_dim as i32, d_conv as i32);
                let __s_b = self.gpu.stream();
                let mut b = __s_b.launch_builder(&f);
                b.arg(qkv_tm).arg(conv_state).arg(len_d).arg(&cd).arg(&dc);
                unsafe {
                    b.launch(cfg)?;
                }
            }
            (None, None) => {
                let f = self.func("ssm_conv_ring_update_f32");
                let n = conv_dim * (d_conv - 1);
                let cfg = LaunchConfig::for_num_elems(n as u32);
                let (cd, ti, dc) = (conv_dim as i32, t as i32, d_conv as i32);
                let __s_b = self.gpu.stream();
                let mut b = __s_b.launch_builder(&f);
                b.arg(qkv_tm).arg(conv_state).arg(&cd).arg(&ti).arg(&dc);
                unsafe {
                    b.launch(cfg)?;
                }
            }
            (Some(_), _) => unreachable!(
                "ssm_conv1d_tm_state_pad_v: T < d_conv-1 has no view path (PRIME_MIN_T gates it)"
            ),
        }
        Ok(())
    }

    /// PREFIX conv-ring rebuild (spec REPLAY-FREE partial accept): overwrite the resident ring
    /// with the state a T=1 chain holds after only the FIRST `tc` columns of `qkv_tm` — the last
    /// `pad` entries of [ring_old | cols 0..tc-1]. PURE COPIES (the ring stores raw inputs; no
    /// arithmetic, cannot perturb FP order). `ring_old` = the pre-round snapshot ring.
    pub fn ssm_conv_ring_rebuild(
        &self,
        qkv_tm: &CudaSlice<f32>,
        ring_old: &CudaSlice<f32>,
        conv_state: &mut CudaSlice<f32>,
        conv_dim: usize,
        tc: usize,
        d_conv: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("ssm_conv_ring_rebuild_f32");
        let n = conv_dim * (d_conv - 1);
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let (cd, ti, dc) = (conv_dim as i32, tc as i32, d_conv as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(qkv_tm)
            .arg(ring_old)
            .arg(conv_state)
            .arg(&cd)
            .arg(&ti)
            .arg(&dc);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// FUSED decode GDN prep (T=1): repack + q/k L2-norm + beta sigmoid + g_log in one launch.
    /// Replaces 5 tiny serialized kernels on the decode critical path. L2 reduce runs as a 32-lane
    /// warp tree (vs l2_norm_f32's 256-thread two-level tree) — same math, different FP sum order;
    /// the argmax + run-spec gates are the authority.
    #[allow(clippy::too_many_arguments)]
    pub fn gdn_prep_decode(
        &self,
        conv_out: &CudaSlice<f32>,
        beta_raw: &CudaSlice<f32>,
        alpha: &CudaSlice<f32>,
        dt_bias: &CudaSlice<f32>,
        a: &CudaSlice<f32>,
        q_l2: &mut CudaSlice<f32>,
        k_l2: &mut CudaSlice<f32>,
        v_g: &mut CudaSlice<f32>,
        beta: &mut CudaSlice<f32>,
        g_log: &mut CudaSlice<f32>,
        d_state: usize,
        num_v: usize,
        num_k: usize,
        key_dim: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gdn_prep_decode_f32");
        let cfg = LaunchConfig {
            grid_dim: (num_v as u32, 1, 1),
            block_dim: (32, 4, 1),
            shared_mem_bytes: 0,
        };
        let (ds, nv, nk, kd) = (d_state as i32, num_v as i32, num_k as i32, key_dim as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(conv_out)
            .arg(beta_raw)
            .arg(alpha)
            .arg(dt_bias)
            .arg(a)
            .arg(q_l2)
            .arg(k_l2)
            .arg(v_g)
            .arg(beta)
            .arg(g_log)
            .arg(&ds)
            .arg(&nv)
            .arg(&nk)
            .arg(&kd)
            .arg(&eps);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// FUSED prefill conv + GDN repack: token-major qkv -> q_g/k_g/v_g in ONE launch (no conv_out
    /// materialization, no qkv_to_gdn_repack pass). BIT-IDENTICAL values; scatter matches
    /// qkv_to_gdn_repack's modulo head-repeat mapping exactly.
    #[allow(clippy::too_many_arguments)]
    pub fn ssm_conv1d_gdn(
        &self,
        qkv_tm: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        q_g: &mut CudaSlice<f32>,
        k_g: &mut CudaSlice<f32>,
        v_g: &mut CudaSlice<f32>,
        conv_dim: usize,
        t: usize,
        d_conv: usize,
        d_state: usize,
        num_v: usize,
        num_k: usize,
        key_dim: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("ssm_conv1d_gdn_f32");
        let cfg = LaunchConfig {
            grid_dim: (((conv_dim + 255) / 256) as u32, t as u32, 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (cd, ti, dc) = (conv_dim as i32, t as i32, d_conv as i32);
        let (ds, nv, nk, kd) = (d_state as i32, num_v as i32, num_k as i32, key_dim as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(qkv_tm)
            .arg(w)
            .arg(q_g)
            .arg(k_g)
            .arg(v_g)
            .arg(&cd)
            .arg(&ti)
            .arg(&dc)
            .arg(&ds)
            .arg(&nv)
            .arg(&nk)
            .arg(&kd);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn ssm_conv1d(
        &self,
        x: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        conv_dim: usize,
        t: usize,
        d_conv: usize,
        silu: bool,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("ssm_conv1d_silu_f32");
        let cfg = LaunchConfig {
            grid_dim: (conv_dim as u32, ((t as u32 + 255) / 256).max(1), 1),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (cd, ti, dc, s) = (conv_dim as i32, t as i32, d_conv as i32, silu as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(w).arg(y).arg(&cd).arg(&ti).arg(&dc).arg(&s);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Gated DeltaNet scan, S_v=128. q,k,v:[128,H,T]; g,beta:[H,T]; state:[128,128,H] transposed;
    /// o:[128,H,T]. Single sequence.
    pub fn gdn_scan_s128(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        g: &CudaSlice<f32>,
        beta: &CudaSlice<f32>,
        state_in: &CudaSlice<f32>,
        state_out: &mut CudaSlice<f32>,
        o: &mut CudaSlice<f32>,
        n_head: usize,
        t: usize,
        scale: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gdn_scan_s128");
        const S_V: u32 = 128;
        const WARP: u32 = 32;
        const COLS_PER_BLOCK: u32 = 4;
        let cfg = LaunchConfig {
            grid_dim: (n_head as u32, 1, S_V / COLS_PER_BLOCK),
            block_dim: (WARP, COLS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let (h, ti) = (n_head as i32, t as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(v)
            .arg(g)
            .arg(beta)
            .arg(state_in)
            .arg(state_out)
            .arg(o)
            .arg(&h)
            .arg(&ti)
            .arg(&scale);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    // ==== B2' batched decode state ops (decode_batch.rs) ====
    // Per-seq state pointers ride device u64 arrays (views into the per-step pointer table).
    // Bodies are the single-seq kernels per sequence — bit-identical per row.

    #[allow(clippy::too_many_arguments)]
    pub fn ssm_conv1d_fused_decode_b(
        &self,
        qkv_cols: &CudaSlice<f32>,
        conv_state_ptrs: &cudarc::driver::CudaView<u64>,
        w: &CudaSlice<f32>,
        conv_outs: &mut CudaSlice<f32>,
        conv_dim: usize,
        d_conv: usize,
        b_n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("ssm_conv1d_fused_decode_b_f32");
        let cfg = LaunchConfig {
            grid_dim: (((conv_dim + 255) / 256) as u32, 1, b_n as u32),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (cd, dc) = (conv_dim as i32, d_conv as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(qkv_cols)
            .arg(conv_state_ptrs)
            .arg(w)
            .arg(conv_outs)
            .arg(&cd)
            .arg(&dc);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    #[allow(clippy::too_many_arguments)]
    pub fn gdn_prep_decode_b(
        &self,
        conv_outs: &CudaSlice<f32>,
        beta_raws: &CudaSlice<f32>,
        alphas: &CudaSlice<f32>,
        dt_bias: &CudaSlice<f32>,
        a: &CudaSlice<f32>,
        q_l2: &mut CudaSlice<f32>,
        k_l2: &mut CudaSlice<f32>,
        v_g: &mut CudaSlice<f32>,
        beta: &mut CudaSlice<f32>,
        g_log: &mut CudaSlice<f32>,
        d_state: usize,
        num_v: usize,
        num_k: usize,
        key_dim: usize,
        eps: f32,
        conv_dim: usize,
        b_n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gdn_prep_decode_b_f32");
        let cfg = LaunchConfig {
            grid_dim: (num_v as u32, 1, b_n as u32),
            block_dim: (32, 4, 1),
            shared_mem_bytes: 0,
        };
        let (ds, nv, nk, kd, cd) = (
            d_state as i32,
            num_v as i32,
            num_k as i32,
            key_dim as i32,
            conv_dim as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(conv_outs)
            .arg(beta_raws)
            .arg(alphas)
            .arg(dt_bias)
            .arg(a)
            .arg(q_l2)
            .arg(k_l2)
            .arg(v_g)
            .arg(beta)
            .arg(g_log)
            .arg(&ds)
            .arg(&nv)
            .arg(&nk)
            .arg(&kd)
            .arg(&eps)
            .arg(&cd);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    #[allow(clippy::too_many_arguments)]
    pub fn gdn_scan_s128_batched(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        g: &CudaSlice<f32>,
        beta: &CudaSlice<f32>,
        state_in_ptrs: &cudarc::driver::CudaView<u64>,
        state_out_ptrs: &cudarc::driver::CudaView<u64>,
        o: &mut CudaSlice<f32>,
        n_head: usize,
        b_n: usize,
        scale: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gdn_scan_s128_b");
        const S_V: u32 = 128;
        const WARP: u32 = 32;
        const COLS_PER_BLOCK: u32 = 4;
        let cfg = LaunchConfig {
            grid_dim: (n_head as u32, b_n as u32, S_V / COLS_PER_BLOCK),
            block_dim: (WARP, COLS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let h = n_head as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(v)
            .arg(g)
            .arg(beta)
            .arg(state_in_ptrs)
            .arg(state_out_ptrs)
            .arg(o)
            .arg(&h)
            .arg(&scale);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// VIEW twins of the three GDN decode state wrappers (lane/verify-launchslim): identical
    /// launches, row args as views into the caller's packed [T, ...] buffers — the t-parallel
    /// verify's per-row loop passes slices instead of paying an arithmetic-free dtod per row
    /// per kernel (48 layers x T rows x 4 copies/round on the money path). Same kernels, same
    /// numeric class; only the pointer arithmetic moved host-side.
    #[allow(clippy::too_many_arguments)]
    pub fn ssm_conv1d_fused_decode_b_view(
        &self,
        qkv_cols: &cudarc::driver::CudaView<f32>,
        conv_state_ptrs: &cudarc::driver::CudaView<u64>,
        w: &CudaSlice<f32>,
        conv_outs: &mut CudaSlice<f32>,
        conv_dim: usize,
        d_conv: usize,
        b_n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("ssm_conv1d_fused_decode_b_f32");
        let cfg = LaunchConfig {
            grid_dim: (((conv_dim + 255) / 256) as u32, 1, b_n as u32),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let (cd, dc) = (conv_dim as i32, d_conv as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(qkv_cols)
            .arg(conv_state_ptrs)
            .arg(w)
            .arg(conv_outs)
            .arg(&cd)
            .arg(&dc);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    #[allow(clippy::too_many_arguments)]
    pub fn gdn_prep_decode_b_view(
        &self,
        conv_outs: &CudaSlice<f32>,
        beta_raws: &cudarc::driver::CudaView<f32>,
        alphas: &cudarc::driver::CudaView<f32>,
        dt_bias: &CudaSlice<f32>,
        a: &CudaSlice<f32>,
        q_l2: &mut CudaSlice<f32>,
        k_l2: &mut CudaSlice<f32>,
        v_g: &mut CudaSlice<f32>,
        beta: &mut CudaSlice<f32>,
        g_log: &mut CudaSlice<f32>,
        d_state: usize,
        num_v: usize,
        num_k: usize,
        key_dim: usize,
        eps: f32,
        conv_dim: usize,
        b_n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gdn_prep_decode_b_f32");
        let cfg = LaunchConfig {
            grid_dim: (num_v as u32, 1, b_n as u32),
            block_dim: (32, 4, 1),
            shared_mem_bytes: 0,
        };
        let (ds, nv, nk, kd, cd) = (
            d_state as i32,
            num_v as i32,
            num_k as i32,
            key_dim as i32,
            conv_dim as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(conv_outs)
            .arg(beta_raws)
            .arg(alphas)
            .arg(dt_bias)
            .arg(a)
            .arg(q_l2)
            .arg(k_l2)
            .arg(v_g)
            .arg(beta)
            .arg(g_log)
            .arg(&ds)
            .arg(&nv)
            .arg(&nk)
            .arg(&kd)
            .arg(&eps)
            .arg(&cd);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    #[allow(clippy::too_many_arguments)]
    pub fn gdn_scan_s128_batched_view(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        g: &CudaSlice<f32>,
        beta: &CudaSlice<f32>,
        state_in_ptrs: &cudarc::driver::CudaView<u64>,
        state_out_ptrs: &cudarc::driver::CudaView<u64>,
        o: &mut cudarc::driver::CudaViewMut<f32>,
        n_head: usize,
        b_n: usize,
        scale: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gdn_scan_s128_b");
        const S_V: u32 = 128;
        const WARP: u32 = 32;
        const COLS_PER_BLOCK: u32 = 4;
        let cfg = LaunchConfig {
            grid_dim: (n_head as u32, b_n as u32, S_V / COLS_PER_BLOCK),
            block_dim: (WARP, COLS_PER_BLOCK, 1),
            shared_mem_bytes: 0,
        };
        let h = n_head as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(q)
            .arg(k)
            .arg(v)
            .arg(g)
            .arg(beta)
            .arg(state_in_ptrs)
            .arg(state_out_ptrs)
            .arg(o)
            .arg(&h)
            .arg(&scale);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// A4 seam: chunked WY GDN prefill. DEFAULT ON (`MEMRA_GDN_CHUNKED=0` = rollback to the
    /// sequential scan). Flipped 2026-07-04 with the full battery green: kernel-check ALL
    /// GREEN x {9B, 27B} incl the f64-truth chunk gates; run-gen argmax 82==82 both models
    /// on AND off (24/24 sweep runs); run-spec K={1,2,3,4,6,8} PASS x {9B synth, 9B text,
    /// 27B p2, 27B p3}; e2e first-16-token agreement 6/6 (full-256 drifts at index 47-125
    /// on 5/6 prompts — accepted cache-state-FP class, batched-prime precedent).
    /// PREFILL-ONLY: decode + spec verify never route here (decode==verify dispatch
    /// identity law); prime_cache/forward/forward_last are the only callers.
    pub fn gdn_chunked_enabled() -> bool {
        static E: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *E.get_or_init(|| {
            std::env::var("MEMRA_GDN_CHUNKED")
                .map(|v| v != "0")
                .unwrap_or(true)
        })
    }

    /// A4 chunk size (MEMRA_GDN_CHUNK, default 32 — the sweep winner: the O(T*C) chunk
    /// matrices grow with C while the sequential state pass is C-flat, so smaller chunks
    /// win; C=32/64 also get the register-history solve template). Clamped to multiples
    /// of 32 in [32, 128] (kernel row mappings require it).
    pub fn gdn_chunk_size() -> usize {
        static C: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
        *C.get_or_init(|| {
            let c: usize = std::env::var("MEMRA_GDN_CHUNK")
                .ok()
                .and_then(|v| v.parse().ok())
                .unwrap_or(32);
            c.clamp(32, 128) / 32 * 32
        })
    }

    /// A4: chunked WY / blockwise-inverse GDN prefill (see cu/hybrid.cu K1-K5 header for the
    /// math). Same contract as `gdn_scan_s128` (layouts, state ping-pong) but chunk-parallel:
    /// NOT bit-identical to the sequential scan (chunked FP accumulation order); run-gen
    /// argmax + run-spec batteries are the accuracy authority. PREFILL callers only.
    #[allow(clippy::too_many_arguments)]
    /// task #18: K1-K3 of the chunked WY scan (shared by the per-seq path and the
    /// batched-prime varlen path). Returns (gcum, P, U, W); `A` is K3-internal.
    #[allow(clippy::too_many_arguments, clippy::type_complexity)]
    #[allow(clippy::too_many_arguments)]
    pub fn gdn_chunk_k123(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        g: &CudaSlice<f32>,
        beta: &CudaSlice<f32>,
        wb16: Option<&mut CudaSlice<u8>>,
        n_head: usize,
        t: usize,
        c: usize,
        hk: usize,
        k2w: Option<(&CudaSlice<u8>, &CudaSlice<u8>, &mut CudaSlice<u8>)>,
    ) -> Result<
        (
            CudaSlice<f32>,
            CudaSlice<f32>,
            CudaSlice<f32>,
            CudaSlice<f32>,
        ),
        Box<dyn std::error::Error>,
    > {
        const D: usize = 128;
        let h = n_head;
        let nc = (t + c - 1) / c;
        let (hi, ti, ci) = (h as i32, t as i32, c as i32);
        let mut gcum = self.uninit(t * h)?;
        let mut a = self.uninit(nc * h * c * c)?;
        let mut p = self.uninit(nc * h * c * c)?;
        let mut u = self.uninit(nc * h * c * D)?;
        let mut w = self.uninit(nc * h * c * D)?;
        {
            // K1
            let f = self.func("gdn_chunk_cumgate_f32");
            let cfg = LaunchConfig {
                grid_dim: (nc as u32, h as u32, 1),
                block_dim: (32, 1, 1),
                shared_mem_bytes: 0,
            };
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(g).arg(&mut gcum).arg(&hi).arg(&ti).arg(&ci);
            unsafe {
                b.launch(cfg)?;
            }
        }
        if let Some((qb, kb, pb)) = k2w {
            // K2-wgmma (MEMRA_GDN_WGMMA path, c==32): A + pre-masked Pb16 in one kernel;
            // the P f32 buffer stays UNWRITTEN (its only wgmma-path consumer is Pb16).
            assert!(c == 32, "gdn_k2_wgmma is a C==32 tile");
            let f = self.func("gdn_k2_wgmma");
            let cfg = LaunchConfig {
                grid_dim: (nc as u32, h as u32, 1),
                block_dim: (128, 1, 1),
                shared_mem_bytes: 0,
            };
            let hki = hk as i32;
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(qb)
                .arg(kb)
                .arg(&gcum)
                .arg(beta)
                .arg(&mut a)
                .arg(&mut *pb)
                .arg(&hi)
                .arg(&ti)
                .arg(&ci)
                .arg(&hki);
            unsafe {
                b.launch(cfg)?;
            }
        } else if c <= 64 && !portable_mma_gated() {
            // K2 register-tiled (2x2 outputs/thread, whole-chunk smem k tile)
            let f = self.func("gdn_chunk_attn_f32");
            f.set_attribute(
                CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                GDN_K2_DYNAMIC_SHARED_BYTES as i32,
            )?;
            let jt = ((c + 31) / 32) as u32;
            let cfg = LaunchConfig {
                grid_dim: (nc as u32, h as u32, jt),
                block_dim: (256, 1, 1),
                shared_mem_bytes: GDN_K2_DYNAMIC_SHARED_BYTES,
            };
            let hki = hk as i32;
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(q)
                .arg(k)
                .arg(&gcum)
                .arg(beta)
                .arg(&mut a)
                .arg(&mut p)
                .arg(&hi)
                .arg(&ti)
                .arg(&ci)
                .arg(&hki);
            unsafe {
                b.launch(cfg)?;
            }
        } else {
            // K2 generic (C = 128, or the portable target's low-smem fallback)
            assert!(
                hk == h,
                "generic K2 is broadcast-only (de-broadcast rides C==32)"
            );
            let f = self.func("gdn_chunk_attn_g_f32");
            let cfg = LaunchConfig {
                grid_dim: (nc as u32, h as u32, 1),
                block_dim: (32, 8, 1),
                shared_mem_bytes: 0,
            };
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(q)
                .arg(k)
                .arg(&gcum)
                .arg(beta)
                .arg(&mut a)
                .arg(&mut p)
                .arg(&hi)
                .arg(&ti)
                .arg(&ci);
            unsafe {
                b.launch(cfg)?;
            }
        }
        {
            // K3 (register-history templates for C=32/64; local-memory generic otherwise)
            let cfg = LaunchConfig {
                grid_dim: (nc as u32, h as u32, 1),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            match c {
                32 | 64 => {
                    let f = self.func(if c == 32 {
                        "gdn_chunk_solve32_f32"
                    } else {
                        "gdn_chunk_solve64_f32"
                    });
                    // mirror-fold: W's bf16 twin emitted on store (0 = skip)
                    let wb: u64 = match wb16 {
                        Some(d) => self.addr_u8(d),
                        None => 0,
                    };
                    let hki = hk as i32;
                    let __s_b = self.gpu.stream();
                    let mut b = __s_b.launch_builder(&f);
                    b.arg(v)
                        .arg(k)
                        .arg(&a)
                        .arg(&gcum)
                        .arg(&mut u)
                        .arg(&mut w)
                        .arg(&wb)
                        .arg(&hi)
                        .arg(&ti)
                        .arg(&hki);
                    unsafe {
                        b.launch(cfg)?;
                    }
                }
                _ => {
                    assert!(hk == h, "generic K3 is broadcast-only");
                    let f = self.func("gdn_chunk_solve_f32");
                    let __s_b = self.gpu.stream();
                    let mut b = __s_b.launch_builder(&f);
                    b.arg(v)
                        .arg(k)
                        .arg(&a)
                        .arg(&gcum)
                        .arg(&mut u)
                        .arg(&mut w)
                        .arg(&hi)
                        .arg(&ti)
                        .arg(&ci);
                    unsafe {
                        b.launch(cfg)?;
                    }
                }
            }
        }
        Ok((gcum, p, u, w))
    }

    /// task #21 de-broadcast seam: q/k stored at num_k distinct GQA heads instead of
    /// the num_v broadcast. MEMRA_GDN_DB=0 reverts. Only the chunked prefill path
    /// consumes the compact layout (hk plumbed; hk == H reproduces broadcast exactly).
    pub fn gdn_db_on() -> bool {
        std::env::var("MEMRA_GDN_DB").as_deref() != Ok("0")
    }

    /// Whether the K4/K5 mma pair serves at chunk size `c` (mirrors gdn_scan_chunked's
    /// seam read — env re-read per call ON PURPOSE, kernel-check pins both configs).
    /// DEFAULT ON for sm_120a builds too (lane/moeprime-nvfp4-direct, 2026-08-21): the pair
    /// was qualified on 90a only and left env-opt-in elsewhere; measured on Blackwell it
    /// wins on BOTH rigs — one RTX PRO 6000 (ornith15 pp14715 12,036 -> 12,751/12,957,
    /// +6-8%, both orders) and the local 5090 (q38-27b pp6435 1,397/1,429 -> 1,427/1,446,
    /// both orders) — with kernel-check/run-gen/margin-gate/run-spec green under the flag.
    /// bf16 HMMA (m16n8k16) is sm_80-class PTX; only the wgmma nest stays Hopper-gated.
    pub fn gdn_mma_enabled(&self, c: usize) -> bool {
        !portable_mma_gated()
            && c == 32
            && match std::env::var("MEMRA_GDN_MMA").as_deref() {
                Ok("1") => true,
                Ok("0") => false,
                _ => gdn_mma_default_on(),
            }
    }

    /// task #22: whether the fused K4+K5 (+K2) wgmma path serves (nested inside the
    /// mma config; same per-call env read discipline). HARD-gated to the Hopper build:
    /// the wgmma asm bodies exist only at __CUDA_ARCH__ == 900 (MEMRA_K45_REAL,
    /// wgmma_common.cuh) — on every other arch the kernel compiles EMPTY, so an env
    /// force would silently produce garbage. Required since the sm_120a mma default
    /// flip made MEMRA_GDN_WGMMA=1 alone reach this branch there.
    pub fn gdn_wgmma_on(&self, c: usize) -> bool {
        cfg!(memra_hopper_mma)
            && self.gdn_mma_enabled(c)
            && std::env::var("MEMRA_GDN_WGMMA").as_deref() != Ok("0")
    }

    /// task #18 conv-fuse: carried-ring conv + SiLU + GDN repack in ONE pass (the
    /// conv_out intermediate and its transposed re-read disappear — 11.8ms of the
    /// T=2048 prime). Ring update stays the separate follow-up launch (pad-aware).
    /// BIT-IDENTICAL values to ssm_conv1d_tm_state_pad + qkv_to_gdn_repack.
    #[allow(clippy::too_many_arguments)]
    pub fn ssm_conv1d_gdn_state_pad(
        &self,
        qkv_tm: &cudarc::driver::CudaView<f32>,
        conv_state: &mut CudaSlice<f32>,
        w: &CudaSlice<f32>,
        q_g: &mut CudaSlice<f32>,
        k_g: &mut CudaSlice<f32>,
        v_g: &mut CudaSlice<f32>,
        conv_dim: usize,
        t: usize,
        d_conv: usize,
        d_state: usize,
        num_v: usize,
        num_k: usize,
        key_dim: usize,
        hk: usize,
        pad_len: Option<&CudaSlice<i32>>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        assert!(
            t >= d_conv - 1,
            "fused state conv requires T >= pad (PRIME_MIN_T gates)"
        );
        {
            let f = self.func("ssm_conv1d_gdn_state_f32");
            let cfg = LaunchConfig {
                grid_dim: (((conv_dim + 255) / 256) as u32, t as u32, 1),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let (cd, ti, dc) = (conv_dim as i32, t as i32, d_conv as i32);
            let (ds, nv, nk, kd, hki) = (
                d_state as i32,
                num_v as i32,
                num_k as i32,
                key_dim as i32,
                hk as i32,
            );
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(qkv_tm)
                .arg(&*conv_state)
                .arg(w)
                .arg(q_g)
                .arg(k_g)
                .arg(v_g)
                .arg(&cd)
                .arg(&ti)
                .arg(&dc)
                .arg(&ds)
                .arg(&nv)
                .arg(&nk)
                .arg(&kd)
                .arg(&hki);
            unsafe {
                b.launch(cfg)?;
            }
        }
        match pad_len {
            Some(len_d) => {
                let f = self.func("ssm_conv_ring_update_dev_f32");
                let n = conv_dim * (d_conv - 1);
                let cfg = LaunchConfig::for_num_elems(n as u32);
                let (cd, dc) = (conv_dim as i32, d_conv as i32);
                let __s_b = self.gpu.stream();
                let mut b = __s_b.launch_builder(&f);
                b.arg(qkv_tm).arg(conv_state).arg(len_d).arg(&cd).arg(&dc);
                unsafe {
                    b.launch(cfg)?;
                }
            }
            None => {
                let f = self.func("ssm_conv_ring_update_f32");
                let n = conv_dim * (d_conv - 1);
                let cfg = LaunchConfig::for_num_elems(n as u32);
                let (cd, ti, dc) = (conv_dim as i32, t as i32, d_conv as i32);
                let __s_b = self.gpu.stream();
                let mut b = __s_b.launch_builder(&f);
                b.arg(qkv_tm).arg(conv_state).arg(&cd).arg(&ti).arg(&dc);
                unsafe {
                    b.launch(cfg)?;
                }
            }
        }
        Ok(())
    }

    /// task #18 increment 2: allocate ONE sequence's chunk buffers (no launches) —
    /// K1-K5 all run varlen afterwards. `a`/`w` become struct members so the varlen
    /// K2/K3 can write them.
    pub fn gdn_chunk_alloc(
        &self,
        n_head: usize,
        t: usize,
        c: usize,
        hk: usize,
    ) -> Result<GdnChunkBufs, Box<dyn std::error::Error>> {
        const D: usize = 128;
        assert!(
            c == 32,
            "gdn_chunk_alloc: varlen chain is the C==32 mma pair"
        );
        let h = n_head;
        let nc = (t + c - 1) / c;
        Ok(GdnChunkBufs {
            gcum: self.uninit(t * h)?,
            a: self.uninit(nc * h * c * c)?,
            p: self.uninit(nc * h * c * c)?,
            u: self.uninit(nc * h * c * D)?,
            w: self.uninit(nc * h * c * D)?,
            kb16: self.alloc_u8_uninit(t * hk * D * 2)?,
            wb16: self.alloc_u8_uninit(nc * h * c * D * 2)?,
            y16: self.alloc_u8_uninit(nc * h * c * D * 2)?,
            ssnap16: self.alloc_u8_uninit(nc * h * D * D * 2)?,
            qb16: self.alloc_u8_uninit(t * hk * D * 2)?,
            pb16: self.alloc_u8_uninit(nc * h * c * c * 2)?,
            o: self.uninit(D * h * t)?,
            t,
            nc,
        })
    }

    /// view-source twin of f32_to_bf16 (the batched FA3 v mirror reads a concat view).
    pub fn f32_to_bf16_v(
        &self,
        x: &cudarc::driver::CudaView<f32>,
        dst: &mut CudaSlice<u8>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("f32_to_bf16_bulk");
        let ni = n as i64;
        let cfg = LaunchConfig::for_num_elems((n as u32).div_ceil(4));
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(dst).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// f32 -> bf16 bulk mirror into a caller buffer (the K4/K5 operand mirrors).
    pub fn f32_to_bf16_into(
        &self,
        x: &CudaSlice<f32>,
        dst: &mut CudaSlice<u8>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("f32_to_bf16_bulk");
        let ni = n as i64;
        let cfg = LaunchConfig::for_num_elems((n as u32).div_ceil(4));
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(dst).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// task #18 increment 2: varlen K1+K2+K3 — three launches run every sequence's
    /// cumgate/attn/solve (per-block math identical to the per-seq kernels).
    pub fn gdn_chunk_k123_vl8(
        &self,
        seqs: &[GdnSeqVl],
        n_head: usize,
        hk: usize,
        wq: Option<&GdnWVl8>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let b = seqs.len();
        assert!(b >= 1 && b <= 8, "gdn_chunk_k123_vl8: 1..=8 sequences");
        let mut packed = [GdnSeqVl::default(); 8];
        packed[..b].copy_from_slice(seqs);
        let v = GdnVl8(packed);
        let (hi, ci) = (n_head as i32, 32i32);
        let max_nc = seqs.iter().map(|a| a.nc).max().unwrap() as u32;
        {
            let f = self.func("gdn_chunk_cumgate_vl");
            let cfg = LaunchConfig {
                grid_dim: (max_nc, n_head as u32, b as u32),
                block_dim: (32, 1, 1),
                shared_mem_bytes: 0,
            };
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v).arg(&hi).arg(&ci);
            unsafe {
                lb.launch(cfg)?;
            }
        }
        let hki = hk as i32;
        if let Some(w) = wq {
            // K2-wgmma vl twin (writes A + pre-masked Pb16)
            let f = self.func("gdn_k2_wgmma_vl");
            let cfg = LaunchConfig {
                grid_dim: (max_nc, n_head as u32, b as u32),
                block_dim: (128, 1, 1),
                shared_mem_bytes: 0,
            };
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v).arg(w).arg(&hi).arg(&ci).arg(&hki);
            unsafe {
                lb.launch(cfg)?;
            }
        } else {
            let f = self.func("gdn_chunk_attn_vl");
            f.set_attribute(
                CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES,
                GDN_K2_DYNAMIC_SHARED_BYTES as i32,
            )?;
            let cfg = LaunchConfig {
                grid_dim: (max_nc, n_head as u32, b as u32),
                block_dim: (256, 1, 1),
                shared_mem_bytes: GDN_K2_DYNAMIC_SHARED_BYTES,
            };
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v).arg(&hi).arg(&ci).arg(&hki);
            unsafe {
                lb.launch(cfg)?;
            }
        }
        {
            let f = self.func("gdn_chunk_solve32_vl");
            let cfg = LaunchConfig {
                grid_dim: (max_nc, n_head as u32, b as u32),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v).arg(&hi).arg(&ci).arg(&hki);
            unsafe {
                lb.launch(cfg)?;
            }
        }
        Ok(())
    }

    /// task #18 increment 3: varlen PREP chain — conv(+ring) / repack / fused-l2 /
    /// fused gate-prep, 5 launches for every sequence (per-element math identical
    /// to the per-seq kernels; l2/gate fusions write disjoint outputs).
    #[allow(clippy::too_many_arguments)]
    pub fn gdn_prep_vl8(
        &self,
        seqs: &[GdnPrepVl],
        conv_w: &CudaSlice<f32>,
        dt_bias: &CudaSlice<f32>,
        a: &CudaSlice<f32>,
        conv_dim: usize,
        d_conv: usize,
        d_state: usize,
        num_v: usize,
        num_k: usize,
        key_dim: usize,
        hk: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let b = seqs.len();
        assert!(b >= 1 && b <= 8);
        let mut packed = [GdnPrepVl::default(); 8];
        packed[..b].copy_from_slice(seqs);
        let v = GdnPrepVl8(packed);
        let max_t = seqs.iter().map(|s| s.t).max().unwrap() as u32;
        let (cdi, dci) = (conv_dim as i32, d_conv as i32);
        let conv_fuse = std::env::var("MEMRA_CONV_FUSE").as_deref() != Ok("0");
        assert!(
            conv_fuse || hk == num_v,
            "de-broadcast requires the fused conv"
        );
        if conv_fuse {
            let f = self.func("ssm_conv1d_gdn_state_vl");
            let cfg = LaunchConfig {
                grid_dim: ((conv_dim as u32).div_ceil(256), max_t, b as u32),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let (dsi, nvi, nki, kdi, hki) = (
                d_state as i32,
                num_v as i32,
                num_k as i32,
                key_dim as i32,
                hk as i32,
            );
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v)
                .arg(conv_w)
                .arg(&cdi)
                .arg(&dci)
                .arg(&dsi)
                .arg(&nvi)
                .arg(&nki)
                .arg(&kdi)
                .arg(&hki);
            unsafe {
                lb.launch(cfg)?;
            }
        } else {
            let f = self.func("ssm_conv1d_tm_state_vl");
            let cfg = LaunchConfig {
                grid_dim: ((conv_dim as u32).div_ceil(256), max_t, b as u32),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v).arg(conv_w).arg(&cdi).arg(&dci);
            unsafe {
                lb.launch(cfg)?;
            }
        }
        {
            let f = self.func("ssm_conv_ring_update_vl");
            let n = (conv_dim * (d_conv - 1)) as u32;
            let cfg = LaunchConfig {
                grid_dim: (n.div_ceil(256), 1, b as u32),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v).arg(&cdi).arg(&dci);
            unsafe {
                lb.launch(cfg)?;
            }
        }
        if !conv_fuse {
            let f = self.func("qkv_to_gdn_repack_vl");
            let n = max_t * (num_v * d_state) as u32;
            let cfg = LaunchConfig {
                grid_dim: (n.div_ceil(256), 1, b as u32),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let (dsi, nvi, nki, kdi) = (d_state as i32, num_v as i32, num_k as i32, key_dim as i32);
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v).arg(&dsi).arg(&nvi).arg(&nki).arg(&kdi);
            unsafe {
                lb.launch(cfg)?;
            }
        }
        if Self::l2_v2_on(d_state) {
            let f = self.func("gdn_l2_v2_vl");
            let cfg = LaunchConfig {
                grid_dim: ((max_t * hk as u32).div_ceil(8), 2, b as u32),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let (dsi, nvi) = (d_state as i32, hk as i32);
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v).arg(&dsi).arg(&nvi).arg(&eps);
            unsafe {
                lb.launch(cfg)?;
            }
        } else {
            let f = self.func("gdn_l2_vl");
            let cfg = LaunchConfig {
                grid_dim: (max_t * hk as u32, 2, b as u32),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let (dsi, nvi) = (d_state as i32, hk as i32);
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v).arg(&dsi).arg(&nvi).arg(&eps);
            unsafe {
                lb.launch(cfg)?;
            }
        }
        {
            let f = self.func("gdn_gate_prep_vl");
            let n = max_t * num_v as u32;
            let cfg = LaunchConfig {
                grid_dim: (n.div_ceil(256), 1, b as u32),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let nvi = num_v as i32;
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v).arg(dt_bias).arg(a).arg(&nvi);
            unsafe {
                lb.launch(cfg)?;
            }
        }
        Ok(())
    }

    /// varlen bf16 mirrors over the gdnseq_t table (which: 0 = k_l2 -> kb16, 1 = w -> wb16).
    pub fn gdn_mirror_vl8(
        &self,
        seqs: &[GdnSeqVl],
        n_head: usize,
        which: i32,
        hk: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let b = seqs.len();
        assert!(b >= 1 && b <= 8);
        let mut packed = [GdnSeqVl::default(); 8];
        packed[..b].copy_from_slice(seqs);
        let v = GdnVl8(packed);
        let ept = (if which == 0 { hk } else { n_head } * 128) as i32;
        let max_n = seqs
            .iter()
            .map(|s| {
                if which == 0 {
                    s.t as i64 * ept as i64
                } else {
                    s.nc as i64 * ept as i64 * 32
                }
            })
            .max()
            .unwrap();
        let f = self.func("gdn_mirror_vl");
        let blocks = ((max_n as u32).div_ceil(4)).div_ceil(256);
        let cfg = LaunchConfig {
            grid_dim: (blocks, 1, b as u32),
            block_dim: (256, 1, 1),
            shared_mem_bytes: 0,
        };
        let __s_lb = self.gpu.stream();
        let mut lb = __s_lb.launch_builder(&f);
        lb.arg(&v).arg(&ept).arg(&which);
        unsafe {
            lb.launch(cfg)?;
        }
        Ok(())
    }

    /// varlen gated-norm tail (+f16out) — one launch replaces B gated_rmsnorm calls.
    pub fn gdn_tail_vl8(
        &self,
        seqs: &[GdnPrepVl],
        norm_w: &CudaSlice<f32>,
        d_state: usize,
        num_v: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let b = seqs.len();
        assert!(b >= 1 && b <= 8);
        let mut packed = [GdnPrepVl::default(); 8];
        packed[..b].copy_from_slice(seqs);
        let v = GdnPrepVl8(packed);
        let max_t = seqs.iter().map(|s| s.t).max().unwrap() as u32;
        let f = self.func("gated_rmsnorm_f16out_vl");
        // block_dim MUST match gated_rmsnorm's (128): the reduction tree order pins the scale
        let cfg = LaunchConfig {
            grid_dim: (max_t * num_v as u32, 1, b as u32),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (dsi, nvi) = (d_state as i32, num_v as i32);
        let __s_lb = self.gpu.stream();
        let mut lb = __s_lb.launch_builder(&f);
        lb.arg(&v).arg(norm_w).arg(&dsi).arg(&nvi).arg(&eps);
        unsafe {
            lb.launch(cfg)?;
        }
        Ok(())
    }

    /// Raw device address helpers for the varlen by-value arg struct (single-stream
    /// launches; every buffer outlives the call — the f16 FFI discipline).
    pub fn addr_f32(&self, x: &CudaSlice<f32>) -> u64 {
        use cudarc::driver::DevicePtr;
        let s = self.gpu.stream();
        let (p, _g) = x.device_ptr(&s);
        p as u64
    }
    pub fn addr_f32_mut(&self, x: &mut CudaSlice<f32>) -> u64 {
        use cudarc::driver::DevicePtrMut;
        let s = self.gpu.stream();
        let (p, _g) = x.device_ptr_mut(&s);
        p as u64
    }
    pub fn addr_f32v(&self, x: &cudarc::driver::CudaView<f32>) -> u64 {
        use cudarc::driver::DevicePtr;
        let s = self.gpu.stream();
        let (p, _g) = x.device_ptr(&s);
        p as u64
    }
    pub fn addr_u8(&self, x: &CudaSlice<u8>) -> u64 {
        use cudarc::driver::DevicePtr;
        let s = self.gpu.stream();
        let (p, _g) = x.device_ptr(&s);
        p as u64
    }

    /// task #18: the varlen K4+K5 pair — TWO launches run every sequence's state pass
    /// and output pass (grid gains a seq dim; per-block math identical to the per-seq
    /// launches, so this is strictly bit-gateable against them).
    pub fn gdn_chunk_vl8(
        &self,
        seqs: &[GdnSeqVl],
        n_head: usize,
        scale: f32,
        hk: usize,
        wq: Option<&GdnWVl8>,
    ) -> Result<(), Box<dyn std::error::Error>> {
        const NSPLIT: u32 = 4;
        let b = seqs.len();
        assert!(b >= 1 && b <= 8, "gdn_chunk_vl8: 1..=8 sequences");
        let mut packed = [GdnSeqVl::default(); 8];
        packed[..b].copy_from_slice(seqs);
        let v = GdnVl8(packed);
        let (hi, ci) = (n_head as i32, 32i32);
        let max_nc = seqs.iter().map(|a| a.nc).max().unwrap() as u32;
        let hki = hk as i32;
        if let Some(w) = wq {
            // K4+K5 fused wgmma vl twin: one launch, Y/Ssnap never materialized.
            let f = self.func("gdn_k45_wgmma_vl");
            let cfg = LaunchConfig {
                grid_dim: (n_head as u32, NSPLIT, b as u32),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v).arg(w).arg(&scale).arg(&hi).arg(&ci).arg(&hki);
            unsafe {
                lb.launch(cfg)?;
            }
            let _ = max_nc;
            return Ok(());
        }
        {
            let f = self.func("gdn_chunk_state_mma_vl");
            let cfg = LaunchConfig {
                grid_dim: (n_head as u32, NSPLIT, b as u32),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v).arg(&hi).arg(&ci).arg(&hki);
            unsafe {
                lb.launch(cfg)?;
            }
        }
        {
            let f = self.func("gdn_chunk_output_mma_vl");
            let cfg = LaunchConfig {
                grid_dim: (max_nc, n_head as u32, b as u32),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let __s_lb = self.gpu.stream();
            let mut lb = __s_lb.launch_builder(&f);
            lb.arg(&v).arg(&hi).arg(&ci).arg(&scale).arg(&hki);
            unsafe {
                lb.launch(cfg)?;
            }
        }
        Ok(())
    }
    pub fn gdn_scan_chunked(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        g: &CudaSlice<f32>,
        beta: &CudaSlice<f32>,
        kb16_pre: Option<&CudaSlice<u8>>,
        qb16_pre: Option<&CudaSlice<u8>>,
        state_in: &CudaSlice<f32>,
        state_out: &mut CudaSlice<f32>,
        o: &mut CudaSlice<f32>,
        n_head: usize,
        t: usize,
        scale: f32,
        c: usize,
        hk: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        const D: usize = 128;
        const NSPLIT: u32 = 4;
        assert!(c >= 1 && c <= 128, "gdn_scan_chunked: C must be in 1..=128");
        let h = n_head;
        let nc = (t + c - 1) / c;
        let (hi, ti, ci) = (h as i32, t as i32, c as i32);
        // mirror-fold (round 27): on the mma path W's bf16 twin is emitted by K3's store
        // (wb16 pre-allocated and threaded through k123) and k's by the producer l2 when
        // the caller hands `kb16_pre` — both standalone mirror passes disappear.
        // default mirrors gdn_mma_enabled (incl. the sm_120a-build ON, lane/moeprime-nvfp4-direct)
        let gdn_mma_pre = !portable_mma_gated()
            && c == 32
            && match std::env::var("MEMRA_GDN_MMA").as_deref() {
                Ok("1") => true,
                Ok("0") => false,
                _ => gdn_mma_default_on(),
            };
        let mut wb16_pre: Option<CudaSlice<u8>> = if gdn_mma_pre {
            Some(self.alloc_u8_uninit(nc * h * c * D * 2)?)
        } else {
            None
        };
        // K2-wgmma pre-work (MEMRA_GDN_WGMMA): the kb16/qb16 mirrors hoist ABOVE K123 so
        // K2 rides them via cp.async; K2 writes the pre-masked Pb16 directly (the
        // gdn_p_bf16_masked pass and the in-branch mirror builds disappear).
        // mirrors gdn_wgmma_on: hard Hopper-build gate (empty wgmma bodies elsewhere)
        let gdn_wgmma_pre = cfg!(memra_hopper_mma)
            && gdn_mma_pre
            && std::env::var("MEMRA_GDN_WGMMA").as_deref() != Ok("0");
        let nk = t * hk * D;
        let mut kb16_local: Option<CudaSlice<u8>> = None;
        if gdn_mma_pre && kb16_pre.is_none() {
            let mut kb = self.alloc_u8_uninit(nk * 2)?;
            let f = self.func("f32_to_bf16_bulk");
            let n2 = nk as i64;
            let cfg2 = LaunchConfig::for_num_elems((nk as u32).div_ceil(4));
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(k).arg(&mut kb).arg(&n2);
            unsafe {
                b.launch(cfg2)?;
            }
            kb16_local = Some(kb);
        }
        let kb16_ref0: Option<&CudaSlice<u8>> = kb16_local.as_ref().or(kb16_pre);
        if let Some(kb) = kb16_pre {
            assert!(kb.len() >= nk * 2, "kb16_pre too small");
        }
        let mut qb16: Option<CudaSlice<u8>> = None;
        let mut pb16: Option<CudaSlice<u8>> = None;
        if gdn_wgmma_pre {
            // mirror-fold (round 35): prep's l2 v2 emits qb16 in-epilogue (kb16 pattern);
            // the standalone bulk cvt only serves callers without the prep mirror.
            if qb16_pre.is_none() {
                let mut qb = self.alloc_u8_uninit(nk * 2)?;
                let f = self.func("f32_to_bf16_bulk");
                let n2 = nk as i64;
                let cfg2 = LaunchConfig::for_num_elems((nk as u32).div_ceil(4));
                let __s_b = self.gpu.stream();
                let mut b = __s_b.launch_builder(&f);
                b.arg(q).arg(&mut qb).arg(&n2);
                unsafe {
                    b.launch(cfg2)?;
                }
                qb16 = Some(qb);
            } else if let Some(qb) = qb16_pre {
                assert!(qb.len() >= nk * 2, "qb16_pre too small");
            }
            pb16 = Some(self.alloc_u8_uninit(nc * h * c * c * 2)?);
        }
        let qb16_ref0: Option<&CudaSlice<u8>> = qb16.as_ref().or(qb16_pre);
        let k2w = if gdn_wgmma_pre {
            Some((
                *qb16_ref0.as_ref().unwrap(),
                *kb16_ref0.as_ref().unwrap(),
                pb16.as_mut().unwrap(),
            ))
        } else {
            None
        };
        let (gcum, p, u, w) =
            self.gdn_chunk_k123(q, k, v, g, beta, wb16_pre.as_mut(), n_head, t, c, hk, k2w)?;
        let _ = &w;
        let mut y = self.uninit(nc * h * c * D)?;
        let mut ssnap = self.uninit(nc * h * D * D)?; // chunk-start state snapshots (K5 phase 1)
        // K4-MMA seam (MEMRA_GDN_MMA; harness verdict 1.75x — tools/bench_gdn_k4.cu, ledger
        // 2026-07-26): M in mma accumulator fragments, bf16 W/k mirrors through a cp.async
        // ring. C==32 only (the kernel's tile). PROMOTED default-ON on the Hopper lane
        // after the STATE-CARRY battery (2026-07-26): 2048-token prime (64 in-kernel state
        // carries) -> 256 greedy decode tokens IDENTICAL to f32 on 3 seeds, AND chunked-
        // continuation prime (MEMRA_PRIME_CHUNK=512, 4 cross-call carries via cache.recur)
        // IDENTICAL on 2 seeds; plus argmax MATCH, pp512 +3.5% (17286), oracle out
        // mean_rel ~1e-4. kernel-check pins BOTH configs (f32 tight band forced =0; mma
        // band 8e-2/8e-1 vs f64 truth). =0 reverts; portable stays f32. NOT read via
        // OnceLock ON PURPOSE: kernel-check toggles the env per call to pin both forms.
        // default mirrors gdn_mma_enabled (incl. the sm_120a-build ON) — all three read
        // sites must agree or the pre-work arms while the scan takes the scalar route.
        let gdn_mma = !portable_mma_gated()
            && c == 32
            && match std::env::var("MEMRA_GDN_MMA").as_deref() {
                Ok("1") => true,
                Ok("0") => false,
                _ => gdn_mma_default_on(),
            };
        if gdn_mma {
            let wb16 = wb16_pre
                .take()
                .expect("mma path pre-allocates wb16 (K3 store fold)");
            let kb16_ref: &CudaSlice<u8> = kb16_ref0.expect("mma path pre-builds kb16 above K123");
            // K4+K5 FUSED wgmma seam (MEMRA_GDN_WGMMA, task #22; harness verdict
            // tools/bench_gdn_wgmma.cu v5, ledger 1f08b997: in-band Y 1.07e-2 / state
            // 1.03e-2 / O 1.08e-2, 91.3us vs 70.4 K4-only at H=32 T=512). K5's output
            // pass runs inside the persistent-M kernel; Y and Ssnap are never
            // materialized. New numeric class (gk folds into k^T instead of ys) —
            // explicit opt-in until the state-carry battery promotes it. Env read per
            // call (kernel-check pins configs by toggling env, GDN_MMA precedent).
            // PROMOTED default-ON hopper (2026-07-27): full battery green — harness
            // in-band, argmax gate PASS, 3-seed greedy IDENTICAL after ~2k prime,
            // chunked-continuation IDENTICAL, kernel-check + decode-batch gates green,
            // official prefill lane +0.74% interleaved x5 (5/5 rounds). =0 reverts.
            if gdn_wgmma_pre {
                // qb16/pb16 pre-built above K123 (K2-wgmma wrote the masked Pb16).
                let qb16 = qb16_ref0.unwrap();
                let pb16 = pb16.as_ref().unwrap();
                {
                    let f = self.func("gdn_k45_wgmma");
                    let cfg = LaunchConfig {
                        grid_dim: (h as u32, 4, 1),
                        block_dim: (256, 1, 1),
                        shared_mem_bytes: 0,
                    };
                    let hki = hk as i32;
                    let __s_b = self.gpu.stream();
                    let mut b = __s_b.launch_builder(&f);
                    b.arg(kb16_ref)
                        .arg(&gcum)
                        .arg(beta)
                        .arg(&u)
                        .arg(&wb16)
                        .arg(qb16)
                        .arg(pb16)
                        .arg(o)
                        .arg(&scale)
                        .arg(state_in)
                        .arg(&mut *state_out)
                        .arg(&hi)
                        .arg(&ti)
                        .arg(&ci)
                        .arg(&hki);
                    unsafe {
                        b.launch(cfg)?;
                    }
                }
                return Ok(());
            }
            // COUPLED PAIR: K4-mma writes Y and Ssnap as bf16 (their only consumer is
            // K5-mma, which rounds to bf16 regardless — identical numerics, half the
            // traffic; harness K5 63.0 -> 35.3us). Fresh bf16 buffers replace the f32 ones.
            let mut y16 = self.alloc_u8_uninit(nc * h * c * D * 2)?;
            let mut ssnap16 = self.alloc_u8_uninit(nc * h * D * D * 2)?;
            {
                let f = self.func("gdn_chunk_state_mma");
                let cfg = LaunchConfig {
                    grid_dim: (h as u32, NSPLIT, 1),
                    block_dim: (256, 1, 1),
                    shared_mem_bytes: 0,
                };
                let hki = hk as i32;
                let __s_b = self.gpu.stream();
                let mut b = __s_b.launch_builder(&f);
                b.arg(kb16_ref)
                    .arg(&gcum)
                    .arg(beta)
                    .arg(&u)
                    .arg(&wb16)
                    .arg(&mut y16)
                    .arg(&mut ssnap16)
                    .arg(state_in)
                    .arg(&mut *state_out)
                    .arg(&hi)
                    .arg(&ti)
                    .arg(&ci)
                    .arg(&hki);
                unsafe {
                    b.launch(cfg)?;
                }
            }
            {
                // K5-mma (bf16 St/Y consumers)
                let f = self.func("gdn_chunk_output_mma");
                let jt = ((c + 31) / 32) as u32;
                let cfg = LaunchConfig {
                    grid_dim: (nc as u32, h as u32, jt),
                    block_dim: (256, 1, 1),
                    shared_mem_bytes: 0,
                };
                let hki = hk as i32;
                let __s_b = self.gpu.stream();
                let mut b = __s_b.launch_builder(&f);
                b.arg(q)
                    .arg(&gcum)
                    .arg(&p)
                    .arg(&y16)
                    .arg(&ssnap16)
                    .arg(o)
                    .arg(&hi)
                    .arg(&ti)
                    .arg(&ci)
                    .arg(&scale)
                    .arg(&hki);
                unsafe {
                    b.launch(cfg)?;
                }
            }
            return Ok(());
        }
        {
            // K4 (sequential over chunks inside; blocks col-partition the state)
            let f = self.func("gdn_chunk_state_f32");
            let cfg = LaunchConfig {
                grid_dim: (h as u32, NSPLIT, 1),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(k)
                .arg(&gcum)
                .arg(beta)
                .arg(&u)
                .arg(&w)
                .arg(&mut y)
                .arg(&mut ssnap)
                .arg(state_in)
                .arg(&mut *state_out)
                .arg(&hi)
                .arg(&ti)
                .arg(&ci);
            unsafe {
                b.launch(cfg)?;
            }
        }
        {
            // K5 (j-blocked: grid.z = 32-row output blocks per chunk; writes o fully)
            let f = self.func("gdn_chunk_output_f32");
            let jt = ((c + 31) / 32) as u32;
            let cfg = LaunchConfig {
                grid_dim: (nc as u32, h as u32, jt),
                block_dim: (256, 1, 1),
                shared_mem_bytes: 0,
            };
            let __s_b = self.gpu.stream();
            let mut b = __s_b.launch_builder(&f);
            b.arg(q)
                .arg(&gcum)
                .arg(&p)
                .arg(&y)
                .arg(&ssnap)
                .arg(o)
                .arg(&hi)
                .arg(&ti)
                .arg(&ci)
                .arg(&scale);
            unsafe {
                b.launch(cfg)?;
            }
        }
        Ok(())
    }

    /// PREFILL GDN scan dispatch (the A4 seam): chunked WY form when enabled and T is in the
    /// batched-prefill regime, else the sequential scan. Callers: hybrid_forward::linear_attn
    /// (forward/forward_last) + linear_attn_prime (prime_cache). Decode (T=1) and the spec
    /// verify call `gdn_scan_s128` DIRECTLY — the decode==verify dispatch identity is untouched.
    ///
    /// MEMRA_GDN_DIFF=1: numerical-oracle mode — runs BOTH forms on the same inputs, prints the
    /// per-call (== per-layer, in call order) output/state error distribution, and keeps the
    /// SEQUENTIAL results so the run stays on the shipped path (stage-1 prototype evidence).
    #[allow(clippy::too_many_arguments)]
    #[allow(clippy::too_many_arguments)]
    pub fn gdn_scan_prefill(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        g: &CudaSlice<f32>,
        beta: &CudaSlice<f32>,
        kb16_pre: Option<&CudaSlice<u8>>,
        qb16_pre: Option<&CudaSlice<u8>>,
        state_in: &CudaSlice<f32>,
        state_out: &mut CudaSlice<f32>,
        o: &mut CudaSlice<f32>,
        n_head: usize,
        t: usize,
        scale: f32,
        hk: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if std::env::var("MEMRA_GDN_DIFF").is_ok() && t >= 16 {
            assert!(hk == n_head, "GDN_DIFF oracle is broadcast-only");
            return self.gdn_scan_diff(q, k, v, g, beta, state_in, state_out, o, n_head, t, scale);
        }
        if Self::gdn_chunked_enabled() && t >= 16 {
            self.gdn_scan_chunked(
                q,
                k,
                v,
                g,
                beta,
                kb16_pre,
                qb16_pre,
                state_in,
                state_out,
                o,
                n_head,
                t,
                scale,
                Self::gdn_chunk_size(),
                hk,
            )
        } else {
            assert!(
                hk == n_head,
                "s128 scan is broadcast-only (prep guarantees by predicate)"
            );
            self.gdn_scan_s128(q, k, v, g, beta, state_in, state_out, o, n_head, t, scale)
        }
    }

    /// Stage-1 oracle: run sequential AND chunked, report per-call error stats, keep sequential.
    #[allow(clippy::too_many_arguments)]
    fn gdn_scan_diff(
        &self,
        q: &CudaSlice<f32>,
        k: &CudaSlice<f32>,
        v: &CudaSlice<f32>,
        g: &CudaSlice<f32>,
        beta: &CudaSlice<f32>,
        state_in: &CudaSlice<f32>,
        state_out: &mut CudaSlice<f32>,
        o: &mut CudaSlice<f32>,
        n_head: usize,
        t: usize,
        scale: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        static CALL: std::sync::atomic::AtomicUsize = std::sync::atomic::AtomicUsize::new(0);
        let call = CALL.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
        let mut o_c = self.uninit(o.len())?;
        let mut st_c = self.uninit(state_out.len())?;
        self.gdn_scan_chunked(
            q,
            k,
            v,
            g,
            beta,
            None,
            None,
            state_in,
            &mut st_c,
            &mut o_c,
            n_head,
            t,
            scale,
            Self::gdn_chunk_size(),
            n_head,
        )?;
        self.gdn_scan_s128(q, k, v, g, beta, state_in, state_out, o, n_head, t, scale)?;
        let (oh_s, oh_c) = (self.dtoh(o)?, self.dtoh(&o_c)?);
        let (sh_s, sh_c) = (self.dtoh(state_out)?, self.dtoh(&st_c)?);
        let stats = |a: &[f32], b: &[f32]| -> (f32, f32, f64) {
            let mut max_abs = 0f32;
            let mut max_rel = 0f32;
            let mut sum_rel = 0f64;
            for (x, y) in a.iter().zip(b) {
                let ad = (x - y).abs();
                let rel = ad / x.abs().max(y.abs()).max(1e-3);
                if ad > max_abs {
                    max_abs = ad;
                }
                if rel > max_rel {
                    max_rel = rel;
                }
                sum_rel += rel as f64;
            }
            (max_abs, max_rel, sum_rel / a.len() as f64)
        };
        let (o_ma, o_mr, o_mean) = stats(&oh_s, &oh_c);
        let (s_ma, s_mr, s_mean) = stats(&sh_s, &sh_c);
        println!(
            "[gdn-diff call {call:3} T={t} C={}] out: max_abs={o_ma:.3e} max_rel={o_mr:.3e} mean_rel={o_mean:.3e} | \
                  state: max_abs={s_ma:.3e} max_rel={s_mr:.3e} mean_rel={s_mean:.3e}",
            Self::gdn_chunk_size()
        );
        Ok(())
    }

    /// softplus-based g_log: g_log[h,t] = a[h] * softplus(alpha[h,t] + dt_bias[h]). a pre-negated.
    pub fn gdn_glog(
        &self,
        alpha: &CudaSlice<f32>,
        dt_bias: &CudaSlice<f32>,
        a: &CudaSlice<f32>,
        g_log: &mut CudaSlice<f32>,
        n_head: usize,
        t: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gdn_glog_f32");
        let cfg = LaunchConfig::for_num_elems((n_head * t) as u32);
        let (h, ti) = (n_head as i32, t as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(alpha).arg(dt_bias).arg(a).arg(g_log).arg(&h).arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// view twins (task #16): the batched prime's GDN core reads the CONCAT projection
    /// buffers at row offsets (CudaView) — same kernels, same values, no split copies.
    pub fn sigmoid_v(
        &self,
        x: &cudarc::driver::CudaView<f32>,
        y: &mut CudaSlice<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("sigmoid_f32");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let ni = n as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(y).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn gdn_glog_v(
        &self,
        alpha: &cudarc::driver::CudaView<f32>,
        dt_bias: &CudaSlice<f32>,
        a: &CudaSlice<f32>,
        g_log: &mut CudaSlice<f32>,
        n_head: usize,
        t: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gdn_glog_f32");
        let cfg = LaunchConfig::for_num_elems((n_head * t) as u32);
        let (h, ti) = (n_head as i32, t as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(alpha).arg(dt_bias).arg(a).arg(g_log).arg(&h).arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn sigmoid(
        &self,
        x: &CudaSlice<f32>,
        y: &mut CudaSlice<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("sigmoid_f32");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let ni = n as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(x).arg(y).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// attn out-gate fused epilogue (task #17): dst = a * sigmoid(g) + fp16 twin, one launch
    /// (replaces sigmoid + mul + convert). Bit-identical class.
    pub fn sig_mul_f16out(
        &self,
        a: &CudaSlice<f32>,
        g: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        dst16: &mut CudaSlice<u8>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("sig_mul_f16out_f32");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let ni = n as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(a).arg(g).arg(dst).arg(dst16).arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// step35 (Step-3.7-Flash) SEPARATE head-wise attention gate: one scalar per query head,
    /// broadcast over head_dim. `dst = a * sigmoid(g)` where `a`/`dst` are `[head_dim, n_head, T]`
    /// (the `q_gate_split` layout) and `g` is the PRE-sigmoid `attn_gate` projection output in
    /// token-major `[T, n_head]`. `dst16` is the optional fp16 operand for wo (None -> skipped).
    ///
    /// NOT interchangeable with `sig_mul_f16out`, which gates FULL WIDTH (qwen35 packs one gate
    /// value per (head, dim) element inside wq). Using this for that, or that for this, silently
    /// applies the wrong number of distinct gate values.
    #[allow(clippy::too_many_arguments)]
    pub fn attn_head_gate(
        &self,
        a: &CudaSlice<f32>,
        g: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        dst16: Option<&mut CudaSlice<u8>>,
        head_dim: usize,
        n_head: usize,
        t: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("attn_head_gate_f32");
        let cfg = LaunchConfig::for_num_elems((head_dim * n_head * t) as u32);
        let (hd, nh, ti) = (head_dim as i32, n_head as i32, t as i32);
        // nullable device pointer by value (0 = skip), same convention as `l2_norm_pp`.
        let d16: u64 = match dst16 {
            Some(d) => self.addr_u8(d),
            None => 0,
        };
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(a)
            .arg(g)
            .arg(dst)
            .arg(&d16)
            .arg(&hd)
            .arg(&nh)
            .arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// step35 CLAMPED SwiGLU: `dst = min(silu(gate*gs), limit) * clamp(up*us, +-limit)`.
    /// Verbatim from llama.cpp `llama-graph.cpp:2146-2165` (routed, `swiglu_clamp_exp`) and
    /// `:1751-1770` (shared, `swiglu_clamp_shexp`), non-DEEPSEEK4 branch.
    ///
    /// This is NOT `swigluoai_mul_scaled`: that one clamps the gate BEFORE swish and multiplies by
    /// `(1 + clamp(up))`. Caller MUST check `limit > 1e-6` (upstream's eps gate) and use the plain
    /// `silu_mul_scaled` path otherwise — at limit=0 this kernel would clamp every positive
    /// activation to zero. On Step-3.7-Flash only layers 43 (7.0) and 44 (16.0) have a live limit.
    #[allow(clippy::too_many_arguments)]
    pub fn swiglu_clamped_mul_scaled(
        &self,
        gate: &CudaSlice<f32>,
        up: &CudaSlice<f32>,
        gs: f32,
        us: f32,
        limit: f32,
        dst: &mut CudaSlice<f32>,
        n: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        debug_assert!(
            limit > 1e-6,
            "swiglu_clamped needs a live limit; use silu_mul_scaled"
        );
        let f = self.func("swiglu_clamped_mul_scaled_f32");
        let cfg = LaunchConfig::for_num_elems(n as u32);
        let ni = n as i32;
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(gate)
            .arg(up)
            .arg(&gs)
            .arg(&us)
            .arg(&limit)
            .arg(dst)
            .arg(&ni);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// gated RMSNorm: dst = RMSNorm(o, w[ncols]) * silu(z), per row of ncols. nrows blocks.
    pub fn gated_rmsnorm(
        &self,
        o: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        z: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gated_rmsnorm_f32");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, e) = (ncols as i32, eps);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(o).arg(w).arg(z).arg(dst).arg(&nc).arg(&e);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// f16out twin of `gated_rmsnorm` (task #17): epilogue also emits the fp16 operand for
    /// the ssm_out GEMM. Bit-identical class (same floats + the cvt kernel's __float2half).
    pub fn gated_rmsnorm_f16out(
        &self,
        o: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        z: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        dst16: &mut CudaSlice<u8>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gated_rmsnorm_f16out_f32");
        // block_dim MUST match gated_rmsnorm's (128): the reduction tree order pins the scale
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, e) = (ncols as i32, eps);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(o).arg(w).arg(z).arg(dst).arg(dst16).arg(&nc).arg(&e);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// add+RMSNorm emitting the f32 normed row AND its q8_1 quantization in one launch (the MoE
    /// layer input: z feeds the router matmul as f32, the expert dp4a as q8_1). BIT-IDENTICAL to
    /// add_rms_norm + quantize_q8_1. Returns (q, d) alongside the caller-provided res/z buffers.
    #[allow(clippy::too_many_arguments)]
    pub fn add_rms_norm_zq8(
        &self,
        a: &CudaSlice<f32>,
        b_in: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        res: &mut CudaSlice<f32>,
        z: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(CudaSlice<i8>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        assert!(ncols % 32 == 0);
        let mut q = self.alloc_uninit::<i8>(nrows * ncols)?;
        let mut d = self.alloc_uninit::<f32>(nrows * (ncols / 32))?;
        let f = self.func("add_rms_norm_zq8");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (1024, 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, ep) = (ncols as i32, eps);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(a)
            .arg(b_in)
            .arg(w)
            .arg(res)
            .arg(z)
            .arg(&mut q)
            .arg(&mut d)
            .arg(&nc)
            .arg(&ep);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((q, d))
    }

    /// gated RMSNorm emitting q8_1 directly (fused quantize epilogue) — the ssm_out matvec input.
    /// BIT-IDENTICAL bytes to gated_rmsnorm + quantize_q8_1 (ncols % 32 == 0; blocks never straddle
    /// rows). Saves one launch per linear-attn layer (36/token on the 9B).
    /// z-view twins of gated_rmsnorm(+f16out) — task #16 batched-prime split removal.
    pub fn gated_rmsnorm_zv(
        &self,
        o: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        z: &cudarc::driver::CudaView<f32>,
        dst: &mut CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gated_rmsnorm_f32");
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, e) = (ncols as i32, eps);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(o).arg(w).arg(z).arg(dst).arg(&nc).arg(&e);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn gated_rmsnorm_f16out_zv(
        &self,
        o: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        z: &cudarc::driver::CudaView<f32>,
        dst: &mut CudaSlice<f32>,
        dst16: &mut CudaSlice<u8>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("gated_rmsnorm_f16out_f32");
        // block_dim MUST match gated_rmsnorm's (128): the reduction tree order pins the scale
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, e) = (ncols as i32, eps);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(o).arg(w).arg(z).arg(dst).arg(dst16).arg(&nc).arg(&e);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    pub fn gated_rmsnorm_q8_1(
        &self,
        o: &CudaSlice<f32>,
        w: &CudaSlice<f32>,
        z: &CudaSlice<f32>,
        ncols: usize,
        nrows: usize,
        eps: f32,
    ) -> Result<(CudaSlice<i8>, CudaSlice<f32>), Box<dyn std::error::Error>> {
        assert!(ncols % 32 == 0);
        let f = self.func("gated_rmsnorm_q8_1");
        let mut out_q = self.alloc_uninit::<i8>(nrows * ncols)?;
        let mut out_d = self.alloc_uninit::<f32>(nrows * (ncols / 32))?;
        let cfg = LaunchConfig {
            grid_dim: (nrows as u32, 1, 1),
            block_dim: (128, 1, 1),
            shared_mem_bytes: 0,
        };
        let (nc, ep) = (ncols as i32, eps);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(o)
            .arg(w)
            .arg(z)
            .arg(&mut out_q)
            .arg(&mut out_d)
            .arg(&nc)
            .arg(&ep);
        unsafe {
            b.launch(cfg)?;
        }
        Ok((out_q, out_d))
    }

    /// transpose [rows,cols] row-major -> [cols,rows] row-major.
    pub fn transpose(
        &self,
        inp: &CudaSlice<f32>,
        rows: usize,
        cols: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let f = self.func("transpose_f32");
        let mut out = self.zeros(rows * cols)?;
        let cfg = LaunchConfig::for_num_elems((rows * cols) as u32);
        let (r, c) = (rows as i32, cols as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(inp).arg(&mut out).arg(&r).arg(&c);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(out)
    }

    /// repeat-interleave heads: in[head_dim,n_in,T] -> out[head_dim,n_out,T].
    pub fn repeat_heads(
        &self,
        inp: &CudaSlice<f32>,
        out: &mut CudaSlice<f32>,
        head_dim: usize,
        n_in: usize,
        n_out: usize,
        t: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("repeat_heads_f32");
        let cfg = LaunchConfig::for_num_elems((head_dim * n_out * t) as u32);
        let (hd, ni, no, ti) = (head_dim as i32, n_in as i32, n_out as i32, t as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(inp).arg(out).arg(&hd).arg(&ni).arg(&no).arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// q|gate split (on-device). qf:[T, n_head*2*head_dim] -> q_out,gate_out:[head_dim,n_head,T].
    /// Replaces the dtoh->host-double-loop->htod in full_attn / full_attn_decode.
    ///
    /// Bounds-checked against `qf` before the launch: the kernel reads `2*head_dim*n_head*t`
    /// floats, and running it on a `wq` output that carries no fused gate reads 2x off the end
    /// (silently, on the device). A layout mismatch is a typed `FusedQGateExtent` here instead.
    pub fn q_gate_split(
        &self,
        qf: &CudaSlice<f32>,
        q_out: &mut CudaSlice<f32>,
        gate_out: &mut CudaSlice<f32>,
        head_dim: usize,
        n_head: usize,
        t: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        memra_gguf::config::check_fused_q_gate_extent(qf.len(), head_dim, n_head, t)?;
        let out_need = head_dim * n_head * t;
        if q_out.len() < out_need || gate_out.len() < out_need {
            return Err(format!(
                "q_gate_split destinations too small: need {out_need} each, have q={} gate={}",
                q_out.len(),
                gate_out.len()
            )
            .into());
        }
        let f = self.func("q_gate_split_f32");
        let cfg = LaunchConfig::for_num_elems((head_dim * n_head * t) as u32);
        let (hd, nh, ti) = (head_dim as i32, n_head as i32, t as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(qf)
            .arg(q_out)
            .arg(gate_out)
            .arg(&hd)
            .arg(&nh)
            .arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// qkv->GDN repack (on-device). conv_out:[conv_dim,T] channel-major ->
    /// q_g/k_g/v_g:[d_state,num_v,T] with q/k head-repeat kh = vh % num_k (validated modulo mapping).
    /// Replaces the dtoh->host-q/k/v-repack->3x-htod in linear_attn / linear_attn_decode.
    pub fn qkv_to_gdn_repack(
        &self,
        conv_out: &CudaSlice<f32>,
        q_g: &mut CudaSlice<f32>,
        k_g: &mut CudaSlice<f32>,
        v_g: &mut CudaSlice<f32>,
        d_state: usize,
        num_v: usize,
        num_k: usize,
        key_dim: usize,
        t: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("qkv_to_gdn_repack_f32");
        let cfg = LaunchConfig::for_num_elems((d_state * num_v * t) as u32);
        let (ds, nv, nk, kd, ti) = (
            d_state as i32,
            num_v as i32,
            num_k as i32,
            key_dim as i32,
            t as i32,
        );
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(conv_out)
            .arg(q_g)
            .arg(k_g)
            .arg(v_g)
            .arg(&ds)
            .arg(&nv)
            .arg(&nk)
            .arg(&kd)
            .arg(&ti);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// conv left zero-pad (prefill from zero state). src:[conv_dim,T] -> dst:[conv_dim,T+pad],
    /// cols 0..pad = 0, cols pad..pad+T = src. `dst` MUST be pre-zeroed. No dtoh/host-loop/htod.
    pub fn conv_left_pad(
        &self,
        src: &CudaSlice<f32>,
        dst: &mut CudaSlice<f32>,
        conv_dim: usize,
        t: usize,
        pad: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("conv_left_pad_f32");
        let cfg = LaunchConfig::for_num_elems((conv_dim * t) as u32);
        let (cd, ti, p) = (conv_dim as i32, t as i32, pad as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(src).arg(dst).arg(&cd).arg(&ti).arg(&p);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// conv-state assemble + ring roll (decode T=1). conv_state:[conv_dim,pad] (resident),
    /// qkv_col:[conv_dim] -> conv_in:[conv_dim,pad+1]; AND rolls conv_state (keep last pad cols).
    /// Replaces the dtoh->host-conv-ring-assemble->ring-update->htod in linear_attn_decode.
    pub fn conv_assemble_and_roll(
        &self,
        qkv_col: &CudaSlice<f32>,
        conv_state: &mut CudaSlice<f32>,
        conv_in: &mut CudaSlice<f32>,
        conv_dim: usize,
        pad: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("conv_assemble_and_roll_f32");
        let cfg = LaunchConfig::for_num_elems(conv_dim as u32);
        let (cd, p) = (conv_dim as i32, pad as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(qkv_col).arg(conv_state).arg(conv_in).arg(&cd).arg(&p);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// RANK3 LEVER (conv fuse, T=1 DECODE): fused conv_assemble_and_roll + ssm_conv1d_silu in ONE
    /// launch. Assembles the conv window [conv_state | qkv_col] in registers, computes the depthwise
    /// causal conv + SiLU into `conv_out`, and rolls the ring — never materializing conv_in to HBM.
    /// Replaces e.conv_assemble_and_roll(...) + e.ssm_conv1d(...). BIT-IDENTICAL to that two-kernel
    /// sequence (same 8-wide accumulation order, same SiLU). `conv_out` is [conv_dim] (T=1).
    pub fn ssm_conv1d_fused_decode(
        &self,
        qkv_col: &CudaSlice<f32>,
        conv_state: &mut CudaSlice<f32>,
        w: &CudaSlice<f32>,
        conv_out: &mut CudaSlice<f32>,
        conv_dim: usize,
        d_conv: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let f = self.func("ssm_conv1d_fused_decode_f32");
        let cfg = LaunchConfig::for_num_elems(conv_dim as u32);
        let (cd, dc) = (conv_dim as i32, d_conv as i32);
        let __s_b = self.gpu.stream();
        let mut b = __s_b.launch_builder(&f);
        b.arg(qkv_col)
            .arg(conv_state)
            .arg(w)
            .arg(conv_out)
            .arg(&cd)
            .arg(&dc);
        unsafe {
            b.launch(cfg)?;
        }
        Ok(())
    }

    /// Copy a contiguous range [start, start+len) out of src into a fresh slice (device→device via host).
    /// Used for qkv split views. Small/rare; not perf-critical in Stage 1.
    pub fn slice_range(
        &self,
        src: &CudaSlice<f32>,
        start: usize,
        len: usize,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        let host = self.gpu.stream().clone_dtoh(src)?;
        self.gpu.stream().synchronize()?;
        Ok(self.htod(&host[start..start + len])?)
    }
}

#[cfg(test)]
mod target_dispatch_tests {
    use super::legacy_quant_gemm_allowed;

    #[test]
    fn legacy_quant_gemm_arch_policy_honors_the_escape_hatch() {
        // sm_120a native lane
        assert!(legacy_quant_gemm_allowed(false, false, false));
        assert!(!legacy_quant_gemm_allowed(false, false, true));
        // pure portable lane (sm_89): gated
        assert!(!legacy_quant_gemm_allowed(true, false, false));
        assert!(!legacy_quant_gemm_allowed(true, false, true));
        // Hopper-MMA lane (sm_90a): portable build, int8-MMA GEMM re-admitted
        assert!(legacy_quant_gemm_allowed(true, true, false));
        assert!(!legacy_quant_gemm_allowed(true, true, true));
    }

    #[cfg(all(memra_portable_cuda, not(memra_hopper_mma)))]
    #[test]
    fn portable_build_disables_legacy_quant_gemm_without_an_env_override() {
        assert!(!legacy_quant_gemm_allowed(
            cfg!(memra_portable_cuda),
            cfg!(memra_hopper_mma),
            false
        ));
    }

    #[cfg(memra_hopper_mma)]
    #[test]
    fn hopper_mma_build_re_admits_legacy_quant_gemm() {
        assert!(legacy_quant_gemm_allowed(
            cfg!(memra_portable_cuda),
            cfg!(memra_hopper_mma),
            false
        ));
        assert!(super::portable_mma_gated() == false);
    }
}

/// The memra-kv device seam (Phase D): the cache's 7 ops delegate to the engine's
/// inherent methods (inherent methods win name resolution, so no recursion).
impl memra_kv::KvDev for Engine {
    fn zeros(&self, n: usize) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        Engine::zeros(self, n)
    }
    fn uninit(&self, n: usize) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        Engine::uninit(self, n)
    }
    fn alloc_u8(&self, n: usize) -> Result<CudaSlice<u8>, Box<dyn std::error::Error>> {
        Engine::alloc_u8(self, n)
    }
    fn htod_i32(&self, v: &[i32]) -> Result<CudaSlice<i32>, Box<dyn std::error::Error>> {
        Engine::htod_i32(self, v)
    }
    fn clone_dtod(
        &self,
        src: &CudaSlice<f32>,
    ) -> Result<CudaSlice<f32>, Box<dyn std::error::Error>> {
        Engine::clone_dtod(self, src)
    }
    fn copy_into(
        &self,
        dst: &mut CudaSlice<f32>,
        off: usize,
        src: &CudaSlice<f32>,
        len: usize,
    ) -> Result<(), Box<dyn std::error::Error>> {
        Engine::copy_into(self, dst, off, src, len)
    }
    fn set_i32_one(
        &self,
        d: &mut CudaSlice<i32>,
        v: i32,
    ) -> Result<(), Box<dyn std::error::Error>> {
        Engine::set_i32_one(self, d, v)
    }
}

#[cfg(test)]
mod fused_gate_bounds_tests {
    use super::*;

    /// The fused `[q|gate]` split's read-site guard, on the device.
    ///
    /// `q_gate_split_f32` reads `2*head_dim*n_head*T` floats out of `qf`. A checkpoint whose gate
    /// is a SEPARATE tensor produces a `wq` output of exactly half that, so before 2026-08-19 the
    /// kernel launched and read 2x past the end of the allocation — an out-of-bounds DEVICE read:
    /// no panic, no error, just whatever memory follows. The guard turns it into a typed
    /// `FusedQGateExtent` before the launch.
    ///
    /// Catch demonstration for this test (guard temporarily removed, then restored):
    /// `compute-sanitizer --tool memcheck` on the half-width case reported invalid `__global__`
    /// reads of size 4 in `q_gate_split_f32`; with the guard in place the same run is clean and
    /// the call returns `Err`. Receipt in the lane report.
    #[test]
    #[ignore = "requires a CUDA GPU"]
    fn q_gate_split_refuses_a_separate_gate_wq_instead_of_reading_past_it() {
        let e = Engine::new(0).unwrap();
        let (head_dim, n_head, t) = (8usize, 4usize, 2usize);
        let fused = 2 * head_dim * n_head * t;
        let out_n = head_dim * n_head * t;

        // half-width `qf` = the separate-gate / ungated layout. MUST be refused.
        let narrow = e.htod(&vec![1.0f32; out_n]).unwrap();
        let mut q = e.uninit(out_n).unwrap();
        let mut gate = e.uninit(out_n).unwrap();
        let err = e
            .q_gate_split(&narrow, &mut q, &mut gate, head_dim, n_head, t)
            .expect_err("half-width wq must be refused, not read past")
            .to_string();
        assert!(err.contains("NO fused gate"), "{err}");
        assert!(err.contains(&format!("{fused}")), "{err}");

        // full-width `qf` = a real qwen3.5 fused layout. MUST still run, and split correctly:
        // per head hh the block is [q(head_dim) | gate(head_dim)] at stride 2*head_dim.
        let host: Vec<f32> = (0..fused).map(|i| i as f32).collect();
        let wide = e.htod(&host).unwrap();
        e.q_gate_split(&wide, &mut q, &mut gate, head_dim, n_head, t)
            .expect("full-width wq splits");
        let (qh, gh) = (e.dtoh(&q).unwrap(), e.dtoh(&gate).unwrap());
        for tok in 0..t {
            for hh in 0..n_head {
                for d in 0..head_dim {
                    let base = tok * (n_head * 2 * head_dim) + hh * (2 * head_dim);
                    let idx = tok * (n_head * head_dim) + hh * head_dim + d;
                    assert_eq!(qh[idx], host[base + d], "q t{tok} h{hh} d{d}");
                    assert_eq!(gh[idx], host[base + head_dim + d], "gate t{tok} h{hh} d{d}");
                }
            }
        }

        // undersized destinations are refused too (the other half of the extent contract)
        let mut small = e.uninit(out_n - 1).unwrap();
        assert!(
            e.q_gate_split(&wide, &mut small, &mut gate, head_dim, n_head, t)
                .is_err()
        );
    }
}

/// FULL-WIDTH-ROPE CONTRACT on the fused rms_norm+qkv+rope kernels
/// (lane/graph-s-key-exactness-20260819, probe O-10 follow-up). CPU-only: the guard runs before
/// any launch, so the refusal is testable without a device.
#[cfg(test)]
mod fused_rope_width_tests {
    use super::Engine;

    /// gemma-4: rotary width == head width on both classes (GGUF 256/256 and 512/512, and the
    /// safetensors route derives the same), which is why the fusion is legal there today.
    #[test]
    fn full_width_is_accepted() {
        assert!(Engine::full_width_rope_only("rms_norm_qkv_rope", 256, 256).is_ok());
        assert!(Engine::full_width_rope_only("rms_norm_qkv_rope_cat", 512, 512).is_ok());
        assert!(Engine::full_width_rope_only("rms_norm_qkv_rope_append", 128, 128).is_ok());
    }

    /// The widths the gemma-4 31B OFFICIAL artifact declares, read from its own GGUF header
    /// (`gemma-4-31B-it-official-Q8_0-MTP.gguf`, box3, 2026-08-19):
    ///
    /// ```text
    /// attention.key_length     512   rope.dimension_count     512   (global class)
    /// attention.key_length_swa 256   rope.dimension_count_swa 256   (SWA class)
    /// ```
    ///
    /// Both classes satisfy `n_rot == head_dim`, which is why the fusion is legal for gemma and
    /// why `HybridModel::gemma4_rope_dims` can feed this guard without refusing what we serve.
    /// An artifact that ever declares otherwise gets a loud refusal at the first fused launch
    /// instead of a silently over-rotated head.
    #[test]
    fn gemma4_official_artifact_widths_pass() {
        assert!(Engine::full_width_rope_only("rms_norm_qkv_rope", 512, 512).is_ok());
        assert!(Engine::full_width_rope_only("rms_norm_qkv_rope_append_dc", 256, 256).is_ok());
    }

    /// The bug this guard exists to make impossible: a partial-rotary arch fused onto a kernel
    /// with no `n_dims`, silently rotating the pass-through band.
    #[test]
    fn partial_rotary_is_refused_with_the_geometry_named() {
        // qwen3.5: n_rot 64 of head_dim 256 (the shape probe O-10 pinned in the split path).
        let err = Engine::full_width_rope_only("rms_norm_qkv_rope", 64, 256)
            .expect_err("partial rotary must refuse");
        let msg = err.to_string();
        assert!(msg.contains("PARTIAL ROTARY REFUSED"), "{msg}");
        assert!(msg.contains("n_rot 64"), "{msg}");
        assert!(msg.contains("head_dim 256"), "{msg}");
        assert!(
            msg.contains("64..256"),
            "names the band it would corrupt: {msg}"
        );
        // step35 full-attn: 64 of 128 (upstream halves n_rot_full).
        assert!(Engine::full_width_rope_only("rms_norm_qkv_rope_append_dc", 64, 128).is_err());
        // and the reverse mismatch (a wider rope than the head) is not "close enough" either.
        assert!(Engine::full_width_rope_only("rms_norm_qkv_rope", 256, 128).is_err());
    }
}