mlx_native/

kernel_registry.rs

//! [`KernelRegistry`] — lazy compilation and caching of Metal compute pipelines.
//!
//! MSL shader source is embedded at compile time via `include_str!`.  On first
//! access, the source is compiled into a Metal library, the named function is
//! extracted, and a `ComputePipelineState` is created and cached.  Subsequent
//! calls return the cached pipeline.
//!
//! ## Precompiled `.metallib` fast path
//!
//! `build.rs` runs `xcrun metal -O3` on every `.metal` file under
//! `src/shaders/` and links the results into a single `default.metallib`
//! placed in `OUT_DIR`.  We embed the bytes via `include_bytes!`.
//!
//! When `MLX_PRECOMPILED_METALLIB=1` is set, `get_pipeline` and
//! `get_pipeline_with_constants` first try to resolve the kernel function
//! against this precompiled library; if found, build the pipeline from it
//! (saves Apple's runtime source-compile pass).  On any failure (function
//! missing, empty embedded blob, load error) the code transparently falls
//! back to the original source-compile path — byte-identical behavior.
//!
//! Default-ON; precompiled gives ~+6% on gemma4 Q-sliding decode (M5 Max).

use std::collections::HashMap;
use std::sync::OnceLock;

use metal::{ComputePipelineDescriptor, ComputePipelineState, FunctionConstantValues, MTLDataType};

use crate::error::{MlxError, Result};

/// Bytes of the precompiled `default.metallib` produced by `build.rs` from
/// every `src/shaders/*.metal` file.  Empty when `MLX_NATIVE_SKIP_METALLIB`
/// was set at build time or xcrun was unavailable.
const EMBEDDED_METALLIB: &[u8] = include_bytes!(concat!(env!("OUT_DIR"), "/default.metallib"));

/// Returns `true` when the precompiled `.metallib` fast path is enabled
/// (default-ON).  Set `MLX_PRECOMPILED_METALLIB=0` (or `false`, `off`)
/// to opt out — useful for diagnosing kernel-compile regressions or
/// A/B benching.
fn precompiled_enabled() -> bool {
    static FLAG: OnceLock<bool> = OnceLock::new();
    *FLAG.get_or_init(|| {
        match std::env::var("MLX_PRECOMPILED_METALLIB").as_deref() {
            Ok("0") | Ok("false") | Ok("off") => false,
            _ => true,
        }
    })
}

/// Returns `true` when the precompiled `.metallib` is consulted for
/// `get_pipeline_with_constants` (FCV-specialized) kernels.  Inherits
/// the master gate [`precompiled_enabled`]; both must be ON for the
/// FCV path to use precompiled.  Default-ON.
fn precompiled_fcv_enabled() -> bool {
    static FLAG: OnceLock<bool> = OnceLock::new();
    *FLAG.get_or_init(|| {
        match std::env::var("MLX_PRECOMPILED_METALLIB_FCV").as_deref() {
            Ok("0") | Ok("false") | Ok("off") => false,
            _ => true,
        }
    })
}

// MTLDataType numeric values (from metal-rs argument.rs, confirmed in Apple Metal spec):
//   Int  = 29
//   Bool = 53
// These are used when calling set_constant_value_at_index so the Metal runtime
// knows how wide each constant value is.

/// Registry that lazily compiles and caches Metal compute pipelines from
/// embedded MSL source.
///
/// # Usage
///
/// ```ignore
/// let mut registry = KernelRegistry::new();
/// let pipeline = registry.get_pipeline("elementwise_add", device.metal_device())?;
/// encoder.encode(&pipeline, &buffers, grid, tg);
/// ```
///
/// # Thread Safety
///
/// `KernelRegistry` is **not** `Sync` by default (it uses `&mut self` for
/// `get_pipeline` to allow mutable cache insertion).  If you need concurrent
/// access, wrap it in a `Mutex` or use one registry per thread.
pub struct KernelRegistry {
    /// Cached pipelines keyed by kernel function name.
    cache: HashMap<String, ComputePipelineState>,
    /// MSL source text keyed by kernel function name.
    ///
    /// Populated at construction time with all embedded shader sources.
    sources: HashMap<String, &'static str>,
    /// Precompiled `default.metallib` (lazy).
    ///
    /// `None` initially.  On first `get_pipeline*` call under
    /// `MLX_PRECOMPILED_METALLIB=1`, populated via
    /// `device.new_library_with_data(EMBEDDED_METALLIB)` — or set to a
    /// sentinel "load-failed" marker (still `None`) so we don't retry.
    /// Set to `Some(library)` on success.
    ///
    /// Lazily filled because we need a `&metal::DeviceRef` to load it
    /// and `KernelRegistry::new()` does not have one.
    precompiled_lib: Option<metal::Library>,
    /// Whether we've already attempted to load the precompiled library.
    /// Prevents repeated load attempts on failure.
    precompiled_load_attempted: bool,
}

impl KernelRegistry {
    /// Create a new registry with all embedded shader sources pre-registered.
    ///
    /// No compilation happens here — shaders are compiled lazily on first use.
    pub fn new() -> Self {
        let mut sources = HashMap::new();

        // Register embedded shader sources.
        sources.insert(
            "placeholder".into(),
            include_str!("shaders/placeholder.metal"),
        );
        sources.insert(
            "quantized_matmul".into(),
            include_str!("shaders/quantized_matmul.metal"),
        );
        sources.insert(
            "quantized_matmul_simd".into(),
            include_str!("shaders/quantized_matmul.metal"),
        );
        sources.insert(
            "quantized_matmul_simd_bf16".into(),
            include_str!("shaders/quantized_matmul.metal"),
        );
        sources.insert(
            "quantized_matmul_simd_bf16_expert".into(),
            include_str!("shaders/quantized_matmul.metal"),
        );

        // GGML block-format quantized mat-vec kernels (ADR-006 Phase 3)
        let ggml_src: &'static str =
            include_str!("shaders/quantized_matmul_ggml.metal");
        sources.insert("kernel_mul_mv_q4_0_f32".into(), ggml_src);
        sources.insert("kernel_mul_mv_q8_0_f32".into(), ggml_src);
        // ADR-028 iter-368: peer-style NSG=4 NR=2 variant (128 threads/TG).
        sources.insert("kernel_mul_mv_q8_0_f32_nr2".into(), ggml_src);
        sources.insert("kernel_mul_mv_q6_K_f32".into(), ggml_src);
        // ADR-028 iter-309 — q6_K mat-vec with nr0=2 + cached yl[16]
        // (peer-pattern port of llama.cpp's `kernel_mul_mv_q6_K_f32_impl`
        // with N_R0_Q6_K=2; 4 rows/TG vs baseline's 2).  Env-gated via
        // `HF2Q_Q6K_MV_NR2=1` in the dispatcher.
        sources.insert("kernel_mul_mv_q6_K_f32_nr2".into(), ggml_src);
        // ADR-022 Phase 1 — Q5_1 / IQ4_NL dense mat-vec.
        sources.insert("kernel_mul_mv_q5_1_f32".into(), ggml_src);
        sources.insert("kernel_mul_mv_iq4_nl_f32".into(), ggml_src);
        // ADR-013 P7 — Q4_K dense decode mat-vec (port of llama.cpp's
        // kernel_mul_mv_q4_K_f32 at ggml-metal.metal:7715-7821).
        sources.insert("kernel_mul_mv_q4_K_f32".into(), ggml_src);
        // ADR-022 Phase 2 — Q5_K dense mv kernel.
        sources.insert("kernel_mul_mv_q5_K_f32".into(), ggml_src);

        // GGML block-format quantized matrix-matrix kernels
        // (ADR-011 Phase 3 Wave P3a: port of llama.cpp's kernel_mul_mm_<q>_f32).
        // Used at prefill m > 8 to reuse each weight tile across a 32-row
        // block via threadgroup-staged simdgroup MMA, instead of re-reading
        // every block per prompt-token as the mv kernel does.
        let ggml_mm_src: &'static str =
            include_str!("shaders/quantized_matmul_mm.metal");
        sources.insert("kernel_mul_mm_q4_0_f32".into(), ggml_mm_src);
        sources.insert("kernel_mul_mm_q8_0_f32".into(), ggml_mm_src);
        sources.insert("kernel_mul_mm_q6_K_f32".into(), ggml_mm_src);
        // ADR-022 Phase 1 — dense Q5_1 / IQ4_NL mm.
        sources.insert("kernel_mul_mm_q5_1_f32".into(), ggml_mm_src);
        sources.insert("kernel_mul_mm_iq4_nl_f32".into(), ggml_mm_src);
        // ADR-022 Phase 2 — dense Q5_K mm.
        sources.insert("kernel_mul_mm_q5_K_f32".into(), ggml_mm_src);
        // ADR-022 Phase 3 — dense Q4_K mm.
        sources.insert("kernel_mul_mm_q4_K_f32".into(), ggml_mm_src);

        // GGML block-format quantized matrix-matrix kernels — tensor API
        // variant (ADR-011 Phase 3 Wave P3b-tensor: port of llama.cpp's
        // kernel_mul_mm_impl `#ifdef GGML_METAL_HAS_TENSOR` branch).
        // Uses Apple's MetalPerformancePrimitives `tensor_ops::matmul2d`
        // primitive which on M3+ dispatches to hardware tensor cores for
        // 2-3x the effective FLOP throughput vs the simdgroup MMA path.
        // Only compiled on devices where the tensor API is available; the
        // kernel_registry's runtime-probe (see MlxDevice::has_tensor) gates
        // compilation so non-tensor devices transparently fall back to the
        // non-tensor `kernel_mul_mm_<q>_f32` kernels.
        let ggml_mm_tensor_src: &'static str =
            include_str!("shaders/quantized_matmul_mm_tensor.metal");
        sources.insert("kernel_mul_mm_q4_0_tensor_f32".into(), ggml_mm_tensor_src);
        sources.insert("kernel_mul_mm_q4_0_tensor_bf16_perm021".into(), ggml_mm_tensor_src);
        sources.insert("kernel_mul_mm_q6_K_tensor_bf16_perm021".into(), ggml_mm_tensor_src);
        sources.insert("kernel_mul_mm_q8_0_tensor_f32".into(), ggml_mm_tensor_src);
        sources.insert("kernel_mul_mm_q6_K_tensor_f32".into(), ggml_mm_tensor_src);
        // ADR-022 Phase 1 — Q5_1 / IQ4_NL tensor mm.
        sources.insert("kernel_mul_mm_q5_1_tensor_f32".into(), ggml_mm_tensor_src);
        sources.insert("kernel_mul_mm_iq4_nl_tensor_f32".into(), ggml_mm_tensor_src);
        // ADR-022 Phase 2 — Q5_K tensor mm.
        sources.insert("kernel_mul_mm_q5_K_tensor_f32".into(), ggml_mm_tensor_src);
        // ADR-022 Phase 3 — Q4_K tensor mm + Q8_0 perm021.
        sources.insert("kernel_mul_mm_q4_K_tensor_f32".into(), ggml_mm_tensor_src);
        sources.insert("kernel_mul_mm_q8_0_tensor_bf16_perm021".into(), ggml_mm_tensor_src);
        // ADR-029 iter-30 H29-speed — F16-weight V2 large-tile mm.
        // Same source file as the V2 quantized variants; reads F16 weight
        // directly from device memory (no per-call dequant).  Used when
        // MlxQWeight.f16_shadow is populated and m > MM_ROUTING_THRESHOLD.
        sources.insert("hf2q_mul_mm_tensor_v2_f16".into(), ggml_mm_tensor_src);
        // ADR-029 iter-36 H28-D — F16-weight perm021 mm for O-projection.
        // Same source file; reads F16 weight from MlxQWeight.f16_shadow when
        // populated, bypassing the per-call quantized dequant.  B-stage
        // (bfloat permuted [n_heads, seq_len, head_dim] input) is byte-
        // identical to the quantized variant.
        sources.insert("kernel_mul_mm_f16_tensor_bf16_perm021".into(), ggml_mm_tensor_src);
        // ADR-029 iter-23 H28-A — V2 large-tile tensor mm (NRA=64 M, NRB=128 N).
        // Same source file as V1 tensor mm; distinct kernel host names so the
        // dispatcher can pick V1 vs V2 at runtime via HF2Q_LARGE_TILE_MM.
        sources.insert("kernel_mul_mm_q4_0_tensor_v2_f32".into(), ggml_mm_tensor_src);
        sources.insert("kernel_mul_mm_q8_0_tensor_v2_f32".into(), ggml_mm_tensor_src);
        sources.insert("kernel_mul_mm_q6_K_tensor_v2_f32".into(), ggml_mm_tensor_src);
        sources.insert("kernel_mul_mm_q5_1_tensor_v2_f32".into(), ggml_mm_tensor_src);
        sources.insert("kernel_mul_mm_iq4_nl_tensor_v2_f32".into(), ggml_mm_tensor_src);
        sources.insert("kernel_mul_mm_q5_K_tensor_v2_f32".into(), ggml_mm_tensor_src);
        sources.insert("kernel_mul_mm_q4_K_tensor_v2_f32".into(), ggml_mm_tensor_src);
        // ADR-029 iter-28 H29 — whole-tensor dequant from block_q → F16.
        // Used at model load to materialize an F16 shadow of attn/dense MLP
        // weights so the runtime dispatch can use kernel_mul_mm_f16_f32_*
        // (peer's gemma4 pattern).  Trades ~1 GB resident memory for 2-3×
        // faster per-call dense matmul at prefill.
        let dequant_to_f16_src: &'static str =
            include_str!("shaders/dequant_to_f16.metal");
        sources.insert("hf2q_dequant_q4_0_to_f16".into(), dequant_to_f16_src);
        sources.insert("hf2q_dequant_q8_0_to_f16".into(), dequant_to_f16_src);
        sources.insert("hf2q_dequant_q5_1_to_f16".into(), dequant_to_f16_src);
        sources.insert("hf2q_dequant_iq4_nl_to_f16".into(), dequant_to_f16_src);
        sources.insert("hf2q_dequant_q4_K_to_f16".into(), dequant_to_f16_src);
        sources.insert("hf2q_dequant_q5_K_to_f16".into(), dequant_to_f16_src);
        sources.insert("hf2q_dequant_q6_K_to_f16".into(), dequant_to_f16_src);

        // ADR-022 Phase 1 P1.7 — Q5_1 / IQ4_NL mul_mv_ext r1 family.
        // Eight instantiations (2 types × 4 r1ptg widths). Each PSO is
        // additionally specialized at PSO-compile time with FC_mul_mv_nsg
        // (function_constant 600) and FC_mul_mv_nxpsg (function_constant 601).
        let mul_mv_ext_src: &'static str = include_str!("shaders/mul_mv_ext.metal");
        sources.insert("kernel_mul_mv_ext_q5_1_f32_r1_2".into(), mul_mv_ext_src);
        sources.insert("kernel_mul_mv_ext_q5_1_f32_r1_3".into(), mul_mv_ext_src);
        sources.insert("kernel_mul_mv_ext_q5_1_f32_r1_4".into(), mul_mv_ext_src);
        sources.insert("kernel_mul_mv_ext_q5_1_f32_r1_5".into(), mul_mv_ext_src);
        sources.insert("kernel_mul_mv_ext_iq4_nl_f32_r1_2".into(), mul_mv_ext_src);
        sources.insert("kernel_mul_mv_ext_iq4_nl_f32_r1_3".into(), mul_mv_ext_src);
        sources.insert("kernel_mul_mv_ext_iq4_nl_f32_r1_4".into(), mul_mv_ext_src);
        sources.insert("kernel_mul_mv_ext_iq4_nl_f32_r1_5".into(), mul_mv_ext_src);
        // ADR-022 Phase 4 — Q4_0 / Q8_0 / Q4_K / Q5_K / Q6_K mv_ext.
        // 5 types × 4 r1ptg widths = 20 instantiations.
        for r1 in [2, 3, 4, 5].iter() {
            for ty in ["q4_0", "q8_0", "q4_K", "q5_K", "q6_K"].iter() {
                let name = format!("kernel_mul_mv_ext_{ty}_f32_r1_{r1}");
                sources.insert(name, mul_mv_ext_src);
            }
        }

        // Dense bf16×f32 → f32 tensor-API matmul (non-flash-attention
        // prefill Q@K^T and scores@V, modeled on llama.cpp's
        // kernel_mul_mm_bf16_f32 with the GGML_METAL_HAS_TENSOR branch
        // active).  Tile geometry and write-back identical to the
        // quantized tensor kernel; only the A-stage copy (bfloat →
        // bfloat, no dequantize) differs.
        let dense_mm_bf16_tensor_src: &'static str =
            include_str!("shaders/dense_mm_bf16_tensor.metal");
        sources.insert("hf2q_dense_mm_bf16_f32_tensor".into(), dense_mm_bf16_tensor_src);
        // ADR-029 iter-80 H60: V2 large-tile variant (NRA=64, NRB=128).
        // Same source file (`dense_mm_bf16_tensor.metal`) — second host_name
        // entry resolves to the V2 kernel appended at the bottom of that
        // file. Picked at dispatch time when HF2Q_LARGE_TILE_MM=1.
        sources.insert("hf2q_dense_mm_bf16_f32_tensor_v2".into(), dense_mm_bf16_tensor_src);

        // Dense f32×f32 → f32 tensor-API matmul (F32-everywhere
        // sibling of dense_mm_bf16_tensor).  Used by hf2q's ADR-005
        // iter-118 BF16-vs-F32 ViT attention A/B diagnostic to remove
        // the BF16 K-stage cast as a confounding variable.  Port of
        // llama.cpp's kernel_mul_mm_f32_f32 specialization
        // (ggml-metal.metal:10098) on the GGML_METAL_HAS_TENSOR
        // branch.  Same tile geometry (NR0=64 NR1=32 NK=32) but
        // float-everywhere shmem staging.
        let dense_mm_f32_f32_tensor_src: &'static str =
            include_str!("shaders/dense_mm_f32_f32.metal");
        sources.insert("hf2q_dense_mm_f32_f32_tensor".into(), dense_mm_f32_f32_tensor_src);

        // Dense f16×f32 → f32 tensor-API matmul (F16-staging sibling
        // of dense_mm_bf16_tensor).  Used by hf2q's ADR-005 Phase 2c
        // iter-128 gemma4v ViT precision-parity path: every mmproj
        // weight is stored as F16 in GGUF, peer's `kernel_mul_mm_f16_f32`
        // (`ggml-metal.metal:10099`) stages BOTH A and B as `half` in
        // shmem and computes on `simdgroup_half8x8`.  Matches peer
        // per-element rounding budget exactly (10-bit mantissa vs
        // BF16's 7-bit), closing the 1.16x/block cascade compound that
        // iter-127 numerically bisected to BF16 staging.  Same tile
        // geometry as the BF16 sibling (NR0=64 NR1=32 NK=32, 8 KB
        // shmem) — half and bfloat share 16-bit storage.
        let dense_mm_f16_tensor_src: &'static str =
            include_str!("shaders/dense_mm_f16_tensor.metal");
        sources.insert("hf2q_dense_mm_f16_f32_tensor".into(), dense_mm_f16_tensor_src);

        // Dense bf16×f32 → f32 GEMV (matrix-vector multiply) — optimized
        // for M=1 single-token decode.  Port of llama.cpp's
        // kernel_mul_mv_bf16_f32_4 (bfloat4-vectorized GEMV kernel).
        // Used in apply_linear_projection_f32 when seq_len=1 and the
        // weight matrix is BF16, replacing the MM kernel (~2× faster for
        // M=1 due to better memory bandwidth utilization per thread).
        let dense_gemv_bf16_src: &'static str =
            include_str!("shaders/dense_gemv_bf16.metal");
        sources.insert("hf2q_dense_gemv_bf16_f32_4".into(), dense_gemv_bf16_src);

        // Fused scale-mask-softmax for the non-flash-attention prefill
        // path.  One row-local threadgroup per (head, query) pair
        // replaces three separate dispatches (scale, mask-add, softmax);
        // reads a bf16 mask (-INF at masked positions, matching
        // flash_attn_prefill_mask.metal) that is shared across heads.
        let scale_mask_softmax_src: &'static str =
            include_str!("shaders/scale_mask_softmax.metal");
        sources.insert("scale_mask_softmax_f32".into(), scale_mask_softmax_src);
        // ADR-029 iter-93 H71: float4-vectorized variant for peer parity
        // with kernel_soft_max_f32_4. Same source file; v4 host_name resolves
        // to the second kernel appended at the bottom of scale_mask_softmax.metal.
        sources.insert("scale_mask_softmax_f32_v4".into(), scale_mask_softmax_src);

        // Expert-routed (MoE) quantized matmul kernel (Story 2.1)
        sources.insert(
            "quantized_matmul_id".into(),
            include_str!("shaders/quantized_matmul_id.metal"),
        );

        // Expert-routed (MoE) GGML block-format quantized matmul kernels
        let ggml_id_src: &'static str =
            include_str!("shaders/quantized_matmul_id_ggml.metal");
        sources.insert("kernel_mul_mv_id_q4_0_f32".into(), ggml_id_src);
        sources.insert("kernel_mul_mv_id_q8_0_f32".into(), ggml_id_src);
        // ADR-013 P7 — Q4_K MoE expert-routed mat-vec (port of
        // llama.cpp's kernel_mul_mv_id_q4_K_f32 at ggml-metal.metal:10349).
        sources.insert("kernel_mul_mv_id_q4_K_f32".into(), ggml_id_src);
        sources.insert("kernel_mul_mv_id_q5_K_f32".into(), ggml_id_src);
        sources.insert("kernel_mul_mv_id_q6_K_f32".into(), ggml_id_src);
        // ADR-028 iter-321 — q6_K _id with nr0=2 + cached yl[16]
        // (peer-pattern port mirroring iter-309's non-_id variant).
        // Env-gated via HF2Q_Q6K_ID_MV_NR2=1 in dispatch_id_mv.
        sources.insert("kernel_mul_mv_id_q6_K_f32_nr2".into(), ggml_id_src);
        // ADR-029 iter-6 — q8_0 _id with nr0=2 + nsg=4 cross-SG reduce
        // (peer-pattern port; peer N_R0_Q8_0=2 + N_SG_Q8_0=4 in
        //  /opt/llama.cpp/ggml/src/ggml-metal/ggml-metal-impl.h:27,40).
        // Env-gated via HF2Q_Q8_0_ID_MV_NR2=1 in dispatch_id_mv.
        sources.insert("kernel_mul_mv_id_q8_0_f32_nr2".into(), ggml_id_src);
        // ADR-022 Phase 1 — Q5_1 / IQ4_NL MoE expert-routed mat-vec.
        sources.insert("kernel_mul_mv_id_q5_1_f32".into(), ggml_id_src);
        sources.insert("kernel_mul_mv_id_iq4_nl_f32".into(), ggml_id_src);
        // Fused-SwiGLU mv_id variants (ADR-012 §Optimize / Task #15):
        // computes y[r][n] = sum_k(dequant(W[expert][n][k]) * silu(gate[r][k]) * up[r][k])
        // in one dispatch — replaces silu_mul + expert_down sequence.
        sources.insert("kernel_mul_mv_id_q4_0_f32_swiglu".into(), ggml_id_src);

        // Expert-routed (MoE) GGML block-format QUANTIZED MATRIX-MATRIX kernels
        // (ADR-011 Phase 3 Wave P3a: port of llama.cpp's
        // `kernel_mul_mm_id_map0_ne20_N` + `kernel_mul_mm_id_<q>_f32`).
        // Two-stage dispatch: map0 regroups the token-to-expert table into
        // per-expert routed-token lists, then mm_id stages a 64x32 expert
        // weight tile into threadgroup shmem and reuses it across a 32-row
        // block of that expert's routed tokens.
        let ggml_id_mm_src: &'static str =
            include_str!("shaders/quantized_matmul_id_mm.metal");
        sources.insert("kernel_mul_mm_id_map0_ne20_1".into(), ggml_id_mm_src);
        sources.insert("kernel_mul_mm_id_map0_ne20_8".into(), ggml_id_mm_src);
        sources.insert("kernel_mul_mm_id_q4_0_f32".into(), ggml_id_mm_src);
        sources.insert("kernel_mul_mm_id_q8_0_f32".into(), ggml_id_mm_src);
        sources.insert("kernel_mul_mm_id_q6_K_f32".into(), ggml_id_mm_src);
        // ADR-013 P16 — Q4_K mm_id (port of llama.cpp ggml-metal.metal:10169).
        sources.insert("kernel_mul_mm_id_q4_K_f32".into(), ggml_id_mm_src);
        // ADR-022 Phase 1 P1.6 — Q5_1 / IQ4_NL mm_id template instantiations.
        sources.insert("kernel_mul_mm_id_q5_1_f32".into(), ggml_id_mm_src);
        sources.insert("kernel_mul_mm_id_iq4_nl_f32".into(), ggml_id_mm_src);
        // ADR-022 Phase 2 — Q5_K mm_id template instantiation.
        sources.insert("kernel_mul_mm_id_q5_K_f32".into(), ggml_id_mm_src);

        // ADR-033 §Pi Task #20 / ADR-034 §93 — fused MoE gate+up+silu_mul
        // mm_id kernel for Q6_K. Replaces 3 dispatches (gate_mm_id, up_mm_id,
        // silu_mul_id) with 1 fused dispatch per MoE FFN per layer. Closes
        // hf2q-vs-llama.cpp prefill gap at production Qwen MoE shapes.
        let fused_q6_k_mm_id_src: &'static str =
            include_str!("shaders/fused_gate_up_silu_mm_id_q6_K.metal");
        sources.insert("kernel_fused_gate_up_silu_mm_id_q6_K_f32".into(), fused_q6_k_mm_id_src);

        // MoE-routed quantized matrix-matrix kernels — tensor API variant
        // (ADR-011 Phase 3 Wave P3b-tensor).  Uses the MPP tensor_ops
        // matmul2d primitive for hardware-tensor-core MMA on M3+.  Only
        // the mm_id kernel is ported — map0 is a short pre-pass (not
        // matmul) and continues to use the simdgroup version.
        let ggml_id_mm_tensor_src: &'static str =
            include_str!("shaders/quantized_matmul_id_mm_tensor.metal");
        sources.insert("kernel_mul_mm_id_q4_0_tensor_f32".into(), ggml_id_mm_tensor_src);
        sources.insert("kernel_mul_mm_id_q8_0_tensor_f32".into(), ggml_id_mm_tensor_src);
        sources.insert("kernel_mul_mm_id_q6_K_tensor_f32".into(), ggml_id_mm_tensor_src);
        // ADR-013 P16 — Q4_K tensor-API mm_id.
        sources.insert("kernel_mul_mm_id_q4_K_tensor_f32".into(), ggml_id_mm_tensor_src);
        // ADR-022 Phase 1 P1.6 — Q5_1 / IQ4_NL tensor-API mm_id.
        sources.insert("kernel_mul_mm_id_q5_1_tensor_f32".into(), ggml_id_mm_tensor_src);
        sources.insert("kernel_mul_mm_id_iq4_nl_tensor_f32".into(), ggml_id_mm_tensor_src);
        // ADR-022 Phase 2 — Q5_K tensor-API mm_id.
        sources.insert("kernel_mul_mm_id_q5_K_tensor_f32".into(), ggml_id_mm_tensor_src);

        // Embedding kernels (Story 1.5)
        let embedding_src: &'static str = include_str!("shaders/embedding.metal");
        sources.insert("embedding_gather_4bit".into(), embedding_src);
        sources.insert("embedding_gather_6bit".into(), embedding_src);

        // MoE gate kernel (Story 1.5)
        let moe_gate_src: &'static str = include_str!("shaders/moe_gate.metal");
        sources.insert("moe_gate".into(), moe_gate_src);

        // MoE dispatch kernels (Story 1.5)
        let moe_dispatch_src: &'static str = include_str!("shaders/moe_dispatch.metal");
        sources.insert("fused_gelu_mul".into(), moe_dispatch_src);
        sources.insert("moe_swiglu_fused".into(), moe_dispatch_src);
        sources.insert("moe_swiglu_batch".into(), moe_dispatch_src);
        sources.insert("moe_swiglu_seq".into(), moe_dispatch_src);
        sources.insert("moe_accumulate".into(), moe_dispatch_src);
        sources.insert("moe_weighted_sum".into(), moe_dispatch_src);
        sources.insert("moe_weighted_sum_seq".into(), moe_dispatch_src);
        sources.insert("zero_buffer".into(), moe_dispatch_src);
        sources.insert("naive_matvec_f32".into(), moe_dispatch_src);
        sources.insert("moe_gather_topk_weights".into(), moe_dispatch_src);
        // bf16 variants (Phase 2 bf16 activation path)
        sources.insert("fused_gelu_mul_bf16".into(), moe_dispatch_src);
        sources.insert("moe_swiglu_seq_bf16".into(), moe_dispatch_src);
        sources.insert("moe_weighted_sum_seq_bf16_input".into(), moe_dispatch_src);

        // ADR-033 §Pi next-iter arc — two-pass MoE mm_id (iter A: map0).
        // Pre-pass that sorts tokens by expert assignment before the main
        // mm_id kernel. Ported from llama.cpp's kernel_mul_mm_id_map0; one
        // template specialization per supported ne20 (n_expert_used).
        let moe_mm_id_map0_src: &'static str =
            include_str!("shaders/moe_mm_id_map0.metal");
        sources.insert("moe_mm_id_map0_ne20_1".into(),  moe_mm_id_map0_src);
        sources.insert("moe_mm_id_map0_ne20_2".into(),  moe_mm_id_map0_src);
        sources.insert("moe_mm_id_map0_ne20_4".into(),  moe_mm_id_map0_src);
        sources.insert("moe_mm_id_map0_ne20_5".into(),  moe_mm_id_map0_src);
        sources.insert("moe_mm_id_map0_ne20_6".into(),  moe_mm_id_map0_src);
        sources.insert("moe_mm_id_map0_ne20_8".into(),  moe_mm_id_map0_src);
        sources.insert("moe_mm_id_map0_ne20_10".into(), moe_mm_id_map0_src);
        sources.insert("moe_mm_id_map0_ne20_16".into(), moe_mm_id_map0_src);
        sources.insert("moe_mm_id_map0_ne20_22".into(), moe_mm_id_map0_src);

        // ADR-033 §Pi iter B-1 — main mm_id kernel skeleton (Q4_0). Body
        // pending iter B-2 (simdgroup matmul + Q4_0 dequant chain). NOT
        // registered: the shader currently writes zeros for routed tiles, so
        // registering it would expose a known-wrong pipeline (and prewarm_all
        // would compile it). The shader file is retained for iter B-2; re-add
        // the source insert once the real matmul body lands.
        //   let moe_mm_id_q4_0_src = include_str!("shaders/moe_mm_id_q4_0.metal");
        //   sources.insert("moe_mm_id_q4_0_f32_skeleton".into(), moe_mm_id_q4_0_src);
        // ADR-020 iter-11h-e3a: backward kernels for moe_weighted_sum_seq.
        sources.insert(
            "moe_weighted_sum_seq_backward_outputs_f32".into(),
            moe_dispatch_src,
        );
        sources.insert(
            "moe_weighted_sum_seq_backward_weights_f32".into(),
            moe_dispatch_src,
        );
        // ADR-020 iter-11h-e3b: fused backward kernel for moe_swiglu_seq.
        sources.insert(
            "moe_swiglu_seq_backward_f32".into(),
            moe_dispatch_src,
        );

        // Batched KV cache copy kernels
        let kv_cache_src: &'static str = include_str!("shaders/kv_cache_copy.metal");
        sources.insert("kv_cache_copy_batch_f32".into(), kv_cache_src);
        sources.insert("kv_cache_copy_batch_f32_to_f16".into(), kv_cache_src);
        sources.insert("kv_cache_copy_seq_f32".into(), kv_cache_src);
        sources.insert("kv_cache_copy_seq_f32_to_f16".into(), kv_cache_src);
        // Wave P4.11 — fused K+V copy variants
        sources.insert("kv_cache_copy_seq_f32_kv_dual".into(), kv_cache_src);
        sources.insert("kv_cache_copy_seq_f32_to_f16_kv_dual".into(), kv_cache_src);
        // ADR-028 iter-145 — fused single-position K+V copy variants (decode shape)
        sources.insert("kv_cache_copy_batch_f32_kv_dual".into(), kv_cache_src);
        sources.insert("kv_cache_copy_batch_f32_to_f16_kv_dual".into(), kv_cache_src);
        // bf16-source KV cache copy (Phase 2 bf16 activation path)
        sources.insert("kv_cache_copy_seq_bf16".into(), kv_cache_src);
        // ADR-030 iter-95 — bit-exact BF16→BF16 head-major cache copy for
        // Option A xlen verify (avoids F16 round-trip precision drift).
        sources.insert("kv_cache_copy_seq_bf16_to_bf16_head_major".into(), kv_cache_src);

        // Elementwise and transpose kernels (Story 1.5)
        let elementwise_src: &'static str = include_str!("shaders/elementwise.metal");
        sources.insert("elementwise_add_f32".into(), elementwise_src);
        sources.insert("elementwise_add_f16".into(), elementwise_src);
        sources.insert("elementwise_mul_f32".into(), elementwise_src);
        sources.insert("elementwise_mul_f16".into(), elementwise_src);
        sources.insert("elementwise_add_bf16".into(), elementwise_src);
        sources.insert("elementwise_mul_bf16".into(), elementwise_src);
        sources.insert("cast_f16_to_f32".into(), elementwise_src);
        sources.insert("cast_f32_to_f16".into(), elementwise_src);
        sources.insert("cast_bf16_to_f32".into(), elementwise_src);
        sources.insert("cast_f32_to_bf16".into(), elementwise_src);
        sources.insert("cast_bf16_to_f16".into(), elementwise_src);
        sources.insert("cast_f16_to_bf16".into(), elementwise_src);
        sources.insert("scalar_mul_bf16".into(), elementwise_src);
        sources.insert("scalar_mul_f32".into(), elementwise_src);
        sources.insert("embedding_gather_scale_f32".into(), elementwise_src);
        sources.insert("embedding_gather_scale_batch_f32".into(), elementwise_src);
        sources.insert("permute_021_bf16".into(), elementwise_src);
        sources.insert("transpose_last2_bf16".into(), elementwise_src);
        sources.insert("transpose_last2_f16".into(), elementwise_src);
        sources.insert("permute_021_f32".into(), elementwise_src);
        sources.insert("permute_021_bf16_to_f32".into(), elementwise_src);
        sources.insert("permute_021_f32_to_f16".into(), elementwise_src);
        sources.insert("transpose_2d_f32".into(), elementwise_src);
        sources.insert("transpose_2d_f16".into(), elementwise_src);

        // Attention kernels (Story 1.3)
        let sdpa_src: &'static str = include_str!("shaders/sdpa.metal");
        sources.insert("sdpa".into(), sdpa_src);
        sources.insert("sdpa_bf16".into(), sdpa_src);
        let sdpa_sliding_src: &'static str = include_str!("shaders/sdpa_sliding.metal");
        sources.insert("sdpa_sliding".into(), sdpa_sliding_src);
        sources.insert("sdpa_sliding_bf16".into(), sdpa_sliding_src);

        // Flash-attention tiled prefill kernel (ADR-011 Phase 1).
        // Ten entry points; all backed by the same shader source.
        // Pipelines are compiled with function constants via
        // `get_pipeline_with_bool_constants` — not `get_pipeline`.
        let flash_attn_prefill_src: &'static str =
            include_str!("shaders/flash_attn_prefill.metal");
        // D=256 variants (BQ=32, BK=16, WM=4, WN=1 — 128 threads/threadgroup)
        sources.insert(
            "steel_attention_float32_bq32_bk16_bd256_wm4_wn1_maskfloat32".into(),
            flash_attn_prefill_src,
        );
        sources.insert(
            "steel_attention_float32_bq32_bk16_bd256_wm4_wn1_maskbool_".into(),
            flash_attn_prefill_src,
        );
        sources.insert(
            "steel_attention_bfloat16_bq32_bk16_bd256_wm4_wn1_maskbfloat16".into(),
            flash_attn_prefill_src,
        );
        sources.insert(
            "steel_attention_bfloat16_bq32_bk16_bd256_wm4_wn1_maskbool_".into(),
            flash_attn_prefill_src,
        );
        sources.insert(
            "steel_attention_float16_bq32_bk16_bd256_wm4_wn1_maskfloat16".into(),
            flash_attn_prefill_src,
        );
        sources.insert(
            "steel_attention_float16_bq32_bk16_bd256_wm4_wn1_maskbool_".into(),
            flash_attn_prefill_src,
        );
        // D=512 variants (BQ=8, BK=8, WM=1, WN=1 — 32 threads/threadgroup)
        // NOTE: f32 at D=512 is NOT instantiated — threadgroup memory exceeds
        // the 32 KB Metal limit (candle sdpa.rs:86-94).
        sources.insert(
            "steel_attention_bfloat16_bq8_bk8_bd512_wm1_wn1_maskbfloat16".into(),
            flash_attn_prefill_src,
        );
        sources.insert(
            "steel_attention_bfloat16_bq8_bk8_bd512_wm1_wn1_maskbool_".into(),
            flash_attn_prefill_src,
        );
        sources.insert(
            "steel_attention_float16_bq8_bk8_bd512_wm1_wn1_maskfloat16".into(),
            flash_attn_prefill_src,
        );
        sources.insert(
            "steel_attention_float16_bq8_bk8_bd512_wm1_wn1_maskbool_".into(),
            flash_attn_prefill_src,
        );

        // Flash attention vector kernels — SIMD-vectorized decode-path SDPA
        // (ported from llama.cpp flash_attn_ext_vec)
        let flash_attn_vec_src: &'static str =
            include_str!("shaders/flash_attn_vec.metal");
        sources.insert("flash_attn_vec_dk256".into(), flash_attn_vec_src);
        sources.insert("flash_attn_vec_dk512".into(), flash_attn_vec_src);
        sources.insert("flash_attn_vec_reduce_dk128".into(), flash_attn_vec_src);
        sources.insert("flash_attn_vec_reduce_dk256".into(), flash_attn_vec_src);
        sources.insert("flash_attn_vec_reduce_dk512".into(), flash_attn_vec_src);
        // F16 KV variants (Phase 4a)
        sources.insert("flash_attn_vec_f16kv_dk256".into(), flash_attn_vec_src);
        sources.insert("flash_attn_vec_f16kv_dk512".into(), flash_attn_vec_src);

        // ADR-037 Phase E1.1 (2026-05-22) — tree-attention kernel for
        // EAGLE-3 + dynamic tree speculative decoding. Variant of
        // flash_attn_vec consuming an explicit per-(query, kv_pos) mask
        // buffer instead of implicit causal. Reduce pass reuses
        // flash_attn_vec_reduce_* (identical output layout).
        let tree_attention_src: &'static str =
            include_str!("shaders/tree_attention.metal");
        sources.insert("tree_attention_dk128".into(), tree_attention_src);
        sources.insert("tree_attention_dk256".into(), tree_attention_src);
        sources.insert("tree_attention_dk512".into(), tree_attention_src);
        sources.insert("tree_attention_f16kv_dk128".into(), tree_attention_src);
        sources.insert("tree_attention_f16kv_dk256".into(), tree_attention_src);
        sources.insert("tree_attention_f16kv_dk512".into(), tree_attention_src);

        // RoPE, normalization, activation kernels (Story 1.4)
        let rope_src: &'static str = include_str!("shaders/rope.metal");
        sources.insert("rope_f32".into(), rope_src);
        sources.insert("rope_f16".into(), rope_src);
        sources.insert("rope_bf16".into(), rope_src);
        sources.insert("rope_neox_bf16".into(), rope_src);
        sources.insert("rope_neox_f32".into(), rope_src);
        let rms_norm_src: &'static str = include_str!("shaders/rms_norm.metal");
        sources.insert("rms_norm_f32".into(), rms_norm_src);
        // ADR-028 iter-310 — float4 + simd_sum variants (peer-pattern,
        // ported from llama.cpp kernel_rms_norm_fuse_impl<float4, 1>).
        // Env-gated via HF2Q_RMS_NORM_V2=1 in the dispatchers.
        sources.insert("rms_norm_f32_v2".into(), rms_norm_src);
        sources.insert("rms_norm_no_scale_f32_v2".into(), rms_norm_src);
        sources.insert("rms_norm_f16".into(), rms_norm_src);
        sources.insert("rms_norm_bf16".into(), rms_norm_src);
        sources.insert("rms_norm_no_scale_bf16".into(), rms_norm_src);
        sources.insert("rms_norm_no_scale_f32".into(), rms_norm_src);
        sources.insert("rms_norm_no_scale_f32_dual".into(), rms_norm_src);
        sources.insert("rms_norm_f32_triple".into(), rms_norm_src);
        sources.insert("fused_post_attn_triple_norm_f32".into(), rms_norm_src);
        // ADR-028 iter-370: V2 (float4 + simd_sum) variant of triple_norm.
        sources.insert("fused_post_attn_triple_norm_f32_v2".into(), rms_norm_src);
        // ADR-028 iter-217: fused post-FF norm 2 + end-of-layer FINAL
        // (combines 2 sequential fused_norm_add dispatches into 1 kernel).
        sources.insert("fused_post_ff_norm2_endlayer_f32".into(), rms_norm_src);
        // ADR-028 iter-362: V2 (float4 + simd_sum) variant of the above.
        // Same math, 75% fewer barriers per dispatch (4 vs 16 at tg=256).
        sources.insert("fused_post_ff_norm2_endlayer_f32_v2".into(), rms_norm_src);
        // ADR-028 iter-367: V2 fusion of moe_weighted_sum INTO Path A end-of-layer.
        // Eliminates 1 dispatch + moe_accum round-trip from gemma4 decode default.
        sources.insert("fused_moe_wsum_post_ff_norm2_endlayer_f32_v2".into(), rms_norm_src);
        sources.insert("rms_norm_no_scale_f32_dual_perm".into(), rms_norm_src);
        // Fused RMS norm + elementwise multiply kernels (Phase 4e.2)
        sources.insert("rms_norm_mul_f32".into(), rms_norm_src);
        sources.insert("rms_norm_mul_f16".into(), rms_norm_src);
        sources.insert("rms_norm_mul_bf16".into(), rms_norm_src);
        // L2 norm kernels (ADR-013 Decision 3 — Gated DeltaNet Q/K norm)
        let l2_norm_src: &'static str = include_str!("shaders/l2_norm.metal");
        sources.insert("l2_norm_f32".into(), l2_norm_src);
        sources.insert("l2_norm_f16".into(), l2_norm_src);
        sources.insert("l2_norm_bf16".into(), l2_norm_src);
        // ADR-015 iter59a — fused L2 norm + scalar multiply (DN q-path).
        sources.insert("l2_norm_scale_f32".into(), l2_norm_src);
        // Cumulative-sum kernels (ADR-013 Decision 4 — DeltaNet decay-mask base)
        let cumsum_src: &'static str = include_str!("shaders/cumsum.metal");
        sources.insert("cumsum_f32".into(), cumsum_src);
        sources.insert("cumsum_bf16".into(), cumsum_src);
        // SSM conv kernels (ADR-013 Decision 7 — DeltaNet 1D causal conv + SiLU)
        let ssm_conv_src: &'static str = include_str!("shaders/ssm_conv.metal");
        sources.insert("ssm_conv_forward_f32".into(), ssm_conv_src);
        sources.insert("ssm_conv_forward_bf16".into(), ssm_conv_src);
        sources.insert("ssm_conv_state_update_f32".into(), ssm_conv_src);
        sources.insert("ssm_conv_state_update_bf16".into(), ssm_conv_src);
        // Tri-solve kernels (ADR-013 Decision 5 — chunked DeltaNet debug path)
        let tri_solve_src: &'static str = include_str!("shaders/tri_solve.metal");
        sources.insert("tri_solve_lower_unit_f32".into(), tri_solve_src);
        sources.insert("tri_solve_lower_unit_bf16".into(), tri_solve_src);
        // Rope-multi kernels (ADR-013 Decision 10 — IMROPE for Qwen3.5)
        let rope_multi_src: &'static str = include_str!("shaders/rope_multi.metal");
        sources.insert("rope_multi_f32".into(), rope_multi_src);
        sources.insert("rope_multi_bf16".into(), rope_multi_src);
        // Gated DeltaNet fused kernel (ADR-013 Decision 6 — centerpiece)
        let gdn_src: &'static str = include_str!("shaders/gated_delta_net.metal");
        sources.insert("gated_delta_net_f32".into(), gdn_src);
        // ADR-015 iter56 — decode-only `simd_sum` variant. Three NSG-templated
        // host names share the same source; selection is by D_k via
        // `dispatch_gated_delta_net_decode`. Drop-in for the fused kernel
        // above when n_tokens=1.
        let gdn_decode_src: &'static str =
            include_str!("shaders/gated_delta_net_decode.metal");
        sources.insert("gated_delta_net_decode_f32_1".into(), gdn_decode_src);
        sources.insert("gated_delta_net_decode_f32_2".into(), gdn_decode_src);
        sources.insert("gated_delta_net_decode_f32_4".into(), gdn_decode_src);
        // Wave 5b — chunk-parallel inter-chunk state-recurrence kernel
        // (the one new kernel in the chunk-parallel pipeline; spec source:
        // arXiv 2412.06464 §4 + FLA chunk_delta_h.py:43-298).
        let gdn_chunk_src: &'static str =
            include_str!("shaders/gated_delta_net_chunk.metal");
        sources.insert(
            "gated_delta_net_chunk_inter_state_bf16".into(),
            gdn_chunk_src,
        );
        // ADR-033 §Pi Task #25 iter 19 — K=256 native variant. Same algorithm
        // as gated_delta_net_chunk_inter_state_bf16 but with compile-time
        // 32-tile MMA loops (vs 16 for K=128). Required for Qwen3.6's
        // head_dim=256 chunk-scan path support. Per the K=128 kernel's
        // documented constraint at gated_delta_net_chunk.metal:441, runtime-K
        // bounds defeat MMA scheduling — this separate kernel keeps K=256
        // compile-time-known, avoiding the 3.15× regression.
        let gdn_chunk_k256_src: &'static str =
            include_str!("shaders/gated_delta_net_chunk_k256.metal");
        sources.insert(
            "gated_delta_net_chunk_inter_state_bf16_k256".into(),
            gdn_chunk_k256_src,
        );
        // ADR-033 §Pi Task #25 iter 20 — K=256 native chunk_o variant.
        // Sister kernel to iter 19's inter_state_k256. Bumped from K=128's
        // num_k_tiles=16 to num_k_tiles=32; bo_acc/bA_acc accumulators are
        // V/BT-indexed (not K-indexed) so they keep their original sizes.
        let gdn_chunk_o_k256_src: &'static str =
            include_str!("shaders/gated_delta_net_chunk_o_k256.metal");
        sources.insert(
            "gated_delta_net_chunk_o_bf16_k256".into(),
            gdn_chunk_o_k256_src,
        );
        // Wave 5b.1 iter 2 — chunk_scaled_dot_kkt kernel (input-side of
        // the chunk pipeline; spec source: FLA chunk_scaled_dot_kkt.py:36-99).
        let gdn_kkt_src: &'static str =
            include_str!("shaders/gated_delta_net_kkt.metal");
        sources.insert("gated_delta_net_kkt_bf16".into(), gdn_kkt_src);
        // Wave 5b.1 iter 2 — recompute_w_u_fwd kernel (applies post-solve A
        // to (β·v) and (β·k·exp(g)) to produce w and u; spec source: FLA
        // wy_fast.py:29-117).
        let gdn_recompute_wu_src: &'static str =
            include_str!("shaders/gated_delta_net_recompute_wu.metal");
        sources.insert(
            "gated_delta_net_recompute_wu_bf16".into(),
            gdn_recompute_wu_src,
        );
        // Wave 5b.1 iter 3 — chunk_fwd_o kernel (per-chunk output: closes
        // the chunk pipeline; spec source: FLA chunk_o.py:42-138).
        let gdn_chunk_o_src: &'static str =
            include_str!("shaders/gated_delta_net_chunk_o.metal");
        sources.insert("gated_delta_net_chunk_o_bf16".into(), gdn_chunk_o_src);
        // Wave 5b.1 iter 4 — orchestrator helper kernels:
        //   chunk_local_cumsum_g_f32      — per-chunk prefix sum on g [B, T, H]
        //   chunk_tri_solve_invert_f32    — per-chunk-block (I + A_strict)^-1
        //                                   on FLA's [B, T, H, BT] layout.
        let chunk_local_cumsum_g_src: &'static str =
            include_str!("shaders/chunk_local_cumsum_g.metal");
        sources.insert(
            "chunk_local_cumsum_g_f32".into(),
            chunk_local_cumsum_g_src,
        );
        let chunk_tri_solve_invert_src: &'static str =
            include_str!("shaders/chunk_gated_delta_rule_tri_solve_invert.metal");
        sources.insert(
            "chunk_tri_solve_invert_f32".into(),
            chunk_tri_solve_invert_src,
        );
        // Sigmoid-gated elementwise multiply (ADR-013 Decision 9 — full-attn output gate)
        let sigmoid_mul_src: &'static str = include_str!("shaders/sigmoid_mul.metal");
        sources.insert("sigmoid_mul_f32".into(), sigmoid_mul_src);
        sources.insert("sigmoid_mul_bf16".into(), sigmoid_mul_src);
        let silu_mul_src: &'static str = include_str!("shaders/silu_mul.metal");
        sources.insert("silu_mul_f32".into(), silu_mul_src);
        // ADR-033 §Pi Task #25 iter 16 — K-bank slice copy for K=256 → 2×K=128
        // bank-split chunk-scan path (Qwen3.6 head_dim=256 support).
        let bank_slice_bf16_src: &'static str =
            include_str!("shaders/bank_slice_bf16.metal");
        sources.insert("bank_slice_bf16".into(), bank_slice_bf16_src);
        // ADR-033 §Pi Task #25 iter 17 — F32 variants (for h0 input and
        // final_state output) + concat (inverse of slice, for assembling
        // the K=256 final_state from per-bank K=128 outputs). Same source
        // file — multiple kernels share the BankSliceParams struct.
        sources.insert("bank_slice_f32".into(), bank_slice_bf16_src);
        sources.insert("bank_concat_f32".into(), bank_slice_bf16_src);
        // ADR-034 task #93 — fused gate_proj + up_proj + silu_mul Q8_0.
        let fused_gate_up_silu_q8_0_src: &'static str =
            include_str!("shaders/fused_gate_up_silu_q8_0.metal");
        sources.insert(
            "kernel_fused_gate_up_silu_q8_0_f32".into(),
            fused_gate_up_silu_q8_0_src,
        );
        // ADR-034 task #94 — fused dual Q4_0 projection (FA Q/K/V/gate fuse).
        let fused_dual_proj_q4_0_src: &'static str =
            include_str!("shaders/fused_dual_proj_q4_0.metal");
        sources.insert(
            "kernel_fused_dual_proj_q4_0_f32".into(),
            fused_dual_proj_q4_0_src,
        );
        // ADR-034 task #93 cont. 24 — fused gate+up+silu_mul Q4_K (broader quant coverage).
        #[allow(non_snake_case)]
        let fused_gate_up_silu_q4_K_src: &'static str =
            include_str!("shaders/fused_gate_up_silu_q4_K.metal");
        sources.insert(
            "kernel_fused_gate_up_silu_q4_K_f32".into(),
            fused_gate_up_silu_q4_K_src,
        );
        // ADR-034 task #93 cont. 26 — fused gate+up+silu_mul IQ4_NL.
        let fused_gate_up_silu_iq4_nl_src: &'static str =
            include_str!("shaders/fused_gate_up_silu_iq4_nl.metal");
        sources.insert(
            "kernel_fused_gate_up_silu_iq4_nl_f32".into(),
            fused_gate_up_silu_iq4_nl_src,
        );
        // ADR-034 task #93 cont. 27 — fused gate+up+silu_mul Q5_K.
        #[allow(non_snake_case)]
        let fused_gate_up_silu_q5_K_src: &'static str =
            include_str!("shaders/fused_gate_up_silu_q5_K.metal");
        sources.insert(
            "kernel_fused_gate_up_silu_q5_K_f32".into(),
            fused_gate_up_silu_q5_K_src,
        );
        // ADR-034 task #93 cont. 28 — fused gate+up+silu_mul Q6_K.
        #[allow(non_snake_case)]
        let fused_gate_up_silu_q6_K_src: &'static str =
            include_str!("shaders/fused_gate_up_silu_q6_K.metal");
        sources.insert(
            "kernel_fused_gate_up_silu_q6_K_f32".into(),
            fused_gate_up_silu_q6_K_src,
        );
        let compute_g_beta_src: &'static str = include_str!("shaders/compute_g_beta.metal");
        sources.insert("compute_g_beta_f32".into(), compute_g_beta_src);
        let ssm_norm_gate_src: &'static str = include_str!("shaders/ssm_norm_gate.metal");
        sources.insert("ssm_norm_gate_f32".into(), ssm_norm_gate_src);
        let gelu_src: &'static str = include_str!("shaders/gelu.metal");
        sources.insert("gelu_f32".into(), gelu_src);
        sources.insert("gelu_f16".into(), gelu_src);
        sources.insert("gelu_bf16".into(), gelu_src);
        let softmax_src: &'static str = include_str!("shaders/softmax.metal");
        sources.insert("softmax_f32".into(), softmax_src);
        sources.insert("softmax_f16".into(), softmax_src);
        sources.insert("softmax_bf16".into(), softmax_src);
        let softmax_backward_src: &'static str =
            include_str!("shaders/softmax_backward.metal");
        sources.insert("softmax_backward_f32".into(), softmax_backward_src);
        let log_elementwise_src: &'static str =
            include_str!("shaders/log_elementwise.metal");
        sources.insert("log_f32".into(), log_elementwise_src);
        sources.insert("log_backward_f32".into(), log_elementwise_src);
        let row_sum_src: &'static str = include_str!("shaders/row_sum.metal");
        sources.insert("row_sum_f32".into(), row_sum_src);
        sources.insert("row_sum_backward_f32".into(), row_sum_src);
        // ADR-020 iter-10a: GGUF-legacy quantize-dequantize round-trip kernels
        // (Q4_0 + Q8_0).  Used by hf2q's dynamic_quant Track 1 to produce
        // W_low / W_high for the gradient-Taylor sensitivity formula.
        let qdq_legacy_src: &'static str = include_str!("shaders/qdq_legacy.metal");
        sources.insert("qdq_q4_0_f32".into(), qdq_legacy_src);
        sources.insert("qdq_q8_0_f32".into(), qdq_legacy_src);
        // ADR-020 iter-10b: RMSNorm reverse-mode autograd kernels.
        // r_inv helper is reused by both backward kernels; dx and dw cover
        // the full backward identity for `y = x * rsqrt(mean(x²) + eps) * w`.
        let rms_norm_backward_src: &'static str =
            include_str!("shaders/rms_norm_backward.metal");
        sources.insert(
            "rms_norm_compute_rms_inv_f32".into(),
            rms_norm_backward_src,
        );
        sources.insert("rms_norm_backward_dx_f32".into(), rms_norm_backward_src);
        sources.insert("rms_norm_backward_dw_f32".into(), rms_norm_backward_src);
        // ADR-020 iter-11a: 2-D row-major slice + concat-by-column kernels.
        // Used by hf2q's multi-head SDPA on GpuTape (slice Q/K/V into
        // per-head views, run per-head SDPA, concat per-head contexts
        // back to full attention output).
        let slice_concat_2d_src: &'static str =
            include_str!("shaders/slice_concat_2d.metal");
        sources.insert("slice_2d_cols_f32".into(), slice_concat_2d_src);
        sources.insert("copy_2d_cols_into_f32".into(), slice_concat_2d_src);
        // ADR-020 iter-11b: SiLU forward + backward kernels for GpuTape
        // SwiGLU FFN composition.
        let silu_backward_src: &'static str =
            include_str!("shaders/silu_backward.metal");
        sources.insert("silu_f32".into(), silu_backward_src);
        sources.insert("silu_backward_f32".into(), silu_backward_src);
        // ADR-020 iter-11d: FP32 embedding lookup + scatter-add backward.
        let embedding_autograd_src: &'static str =
            include_str!("shaders/embedding_autograd.metal");
        sources.insert("embedding_lookup_f32".into(), embedding_autograd_src);
        sources.insert(
            "embedding_scatter_add_f32".into(),
            embedding_autograd_src,
        );
        // ADR-020 iter-13a: Adam optimizer step kernel for Track 2
        // DWQ-proper training loop.
        let adam_update_src: &'static str =
            include_str!("shaders/adam_update.metal");
        sources.insert("adam_update_f32".into(), adam_update_src);
        // ADR-020 iter-13b: differentiable affine qdq kernels for the
        // DWQ-proper training loop.  Init + forward + backward (scales,
        // biases) — q_int is FROZEN, scales+biases learnable.
        let qdq_affine_src: &'static str =
            include_str!("shaders/qdq_affine.metal");
        sources.insert("qdq_affine_init_f32".into(), qdq_affine_src);
        sources.insert("qdq_affine_forward_f32".into(), qdq_affine_src);
        sources.insert(
            "qdq_affine_backward_scales_f32".into(),
            qdq_affine_src,
        );
        sources.insert(
            "qdq_affine_backward_biases_f32".into(),
            qdq_affine_src,
        );
        // ADR-020 iter-15: fused affine quantized matmul for DWQ inference.
        // Per-element kernel; one thread per (m, n) output element.
        // Tiled + simdgroup-MMA variant lands in iter-15b.
        let qmm_affine_src: &'static str =
            include_str!("shaders/qmm_affine.metal");
        sources.insert("qmm_affine_t_f32".into(), qmm_affine_src);
        // ADR-020 iter-15b: tiled variant — 16x16 thread block with
        // cooperative-load X/W tiles in threadgroup-shared memory for
        // 2-5x speedup over the per-element kernel.
        let qmm_affine_tiled_src: &'static str =
            include_str!("shaders/qmm_affine_tiled.metal");
        sources.insert(
            "qmm_affine_t_f32_tiled".into(),
            qmm_affine_tiled_src,
        );
        // ADR-020 iter-15c: simdgroup-MMA variant — uses Apple GPU
        // hardware `simdgroup_matrix<float, 8, 8>` MMA for the inner
        // reduction.  Per-tile algorithmic 8× over scalar tiled, lands
        // as ~3-4× wall after launch / load amortization.
        let qmm_affine_simd_src: &'static str =
            include_str!("shaders/qmm_affine_simd.metal");
        sources.insert(
            "qmm_affine_t_f32_simd".into(),
            qmm_affine_simd_src,
        );
        // ADR-020 iter-15c-2: 4-simdgroup-per-TG variant — 32×32
        // output tile, 4 simdgroups arranged as 2×2 grid each owning
        // a 16×16 sub-tile = 4 simdgroup_matrix accumulators.  Same
        // math as 15c-1, fuller warp-pool exploitation.
        let qmm_affine_simd4_src: &'static str =
            include_str!("shaders/qmm_affine_simd4.metal");
        sources.insert(
            "qmm_affine_t_f32_simd4".into(),
            qmm_affine_simd4_src,
        );
        // ADR-020 iter-15c-2b: gs=64 variant (mlx-lm dynamic_quant
        // canonical default).  Same 4-simdgroup geometry, BK=64
        // instead of 32 (= 8 sub-K-tiles per K-step instead of 4).
        let qmm_affine_simd4_gs64_src: &'static str =
            include_str!("shaders/qmm_affine_simd4_gs64.metal");
        sources.insert(
            "qmm_affine_t_f32_simd4_gs64".into(),
            qmm_affine_simd4_gs64_src,
        );
        // ADR-020 AC#5 Iter A: packed-U32 dense affine matmul (bits=4,
        // gs=32) — production decode/prefill kernel for serving DWQ
        // safetensors directly without a load-time unpack pass.
        let qmm_affine_t_packed_simd4_b4_src: &'static str =
            include_str!("shaders/qmm_affine_t_packed_simd4_b4.metal");
        sources.insert(
            "qmm_affine_t_packed_simd4_b4".into(),
            qmm_affine_t_packed_simd4_b4_src,
        );
        // ADR-020 iter-11h-b: training-mode causal depthwise 1D
        // convolution (forward + backward dx + backward dw).  Used by
        // GpuTape autograd for differentiable Qwen3.5MoE forward
        // (GatedDeltaNet's conv1d step).
        let conv1d_dwc_src: &'static str =
            include_str!("shaders/conv1d_depthwise_causal.metal");
        sources.insert(
            "conv1d_depthwise_causal_forward_f32".into(),
            conv1d_dwc_src,
        );
        sources.insert(
            "conv1d_depthwise_causal_backward_dx_f32".into(),
            conv1d_dwc_src,
        );
        sources.insert(
            "conv1d_depthwise_causal_backward_dw_f32".into(),
            conv1d_dwc_src,
        );
        // ADR-020 iter-11h-c1: elementwise exp forward + backward.
        // Building block for GatedDeltaNet's alpha = exp(-g) state-decay.
        let exp_src: &'static str =
            include_str!("shaders/exp_elementwise.metal");
        sources.insert("exp_f32".into(), exp_src);
        sources.insert("exp_backward_f32".into(), exp_src);
        // ADR-020 iter-11h-c2: vector outer product (forward + dlhs +
        // drhs).  Building block for gated_delta_update's
        // outer(delta, k) state-update term.
        let outer_src: &'static str =
            include_str!("shaders/outer_product.metal");
        sources.insert("outer_product_f32".into(), outer_src);
        sources.insert("outer_product_backward_lhs_f32".into(), outer_src);
        sources.insert("outer_product_backward_rhs_f32".into(), outer_src);
        // ADR-020 iter-11h-e1: take_along_axis (gather) + scatter-backward.
        // Building block for MoE router on GpuTape.
        let taa_src: &'static str =
            include_str!("shaders/take_along_axis.metal");
        sources.insert("take_along_axis_f32".into(), taa_src);
        sources.insert("take_along_axis_backward_f32".into(), taa_src);
        // ADR-020 iter-11h-misc-1: elementwise divide forward + backward.
        let div_src: &'static str =
            include_str!("shaders/divide_elementwise.metal");
        sources.insert("divide_f32".into(), div_src);
        sources.insert("divide_backward_f32".into(), div_src);
        // ADR-020 iter-11h-misc-3: elementwise sqrt forward + backward.
        let sqrt_src: &'static str =
            include_str!("shaders/sqrt_elementwise.metal");
        sources.insert("sqrt_f32".into(), sqrt_src);
        sources.insert("sqrt_backward_f32".into(), sqrt_src);
        let softcap_src: &'static str = include_str!("shaders/softcap.metal");
        sources.insert("softcap_f32".into(), softcap_src);
        sources.insert("softcap_f16".into(), softcap_src);
        sources.insert("softcap_bf16".into(), softcap_src);

        // Fused norm-add kernels — Gemma4 post-attention / post-FFN ordering:
        //   normed = rms_norm(input, weight, eps);  output = residual + normed
        let fused_norm_add_src: &'static str =
            include_str!("shaders/fused_norm_add_bf16.metal");
        sources.insert("fused_norm_add_bf16".into(), fused_norm_add_src);
        sources.insert("fused_norm_add_no_weight_bf16".into(), fused_norm_add_src);

        // Fused head-norm + RoPE f32 kernel — replaces separate rms_norm + rope_neox_f32
        let fused_hnr_f32_src: &'static str =
            include_str!("shaders/fused_head_norm_rope_f32.metal");
        sources.insert("fused_head_norm_rope_f32".into(), fused_hnr_f32_src);
        // ADR-028 iter-337 — float4 + simd_sum Phase 1 variant.  Phases
        // 2-4 byte-identical to v1; race-fix barrier preserved.  Env-gated
        // via HF2Q_FUSED_HEAD_NORM_ROPE_V2 (default ON, opt-out via =0).
        sources.insert("fused_head_norm_rope_f32_v2".into(), fused_hnr_f32_src);

        // Fused head-norm + RoPE bf16 kernels (single-token + batch prefill)
        // Both entry points live in the same .metal file.
        let fused_hnr_bf16_src: &'static str =
            include_str!("shaders/fused_head_norm_rope_bf16.metal");
        sources.insert("fused_head_norm_rope_bf16".into(), fused_hnr_bf16_src);
        sources.insert("fused_head_norm_rope_batch_bf16".into(), fused_hnr_bf16_src);

        // Fused norm-add f32 kernels — post-attention / post-FFN / end-of-layer
        let fused_norm_add_f32_src: &'static str =
            include_str!("shaders/fused_norm_add_f32.metal");
        sources.insert("fused_norm_add_f32".into(), fused_norm_add_f32_src);
        // ADR-028 iter-331 — float4 + simd_sum variant (peer-pattern,
        // ported from llama.cpp kernel_rms_norm_fuse_impl<float4, 3>).
        // Env-gated via HF2Q_FUSED_NORM_ADD_V2=1 in the dispatcher
        // (default ON since iter-331; opt-out via =0/false/off).
        sources.insert("fused_norm_add_f32_v2".into(), fused_norm_add_f32_src);
        sources.insert("fused_residual_norm_f32".into(), fused_norm_add_f32_src);
        sources.insert("fused_residual_norm_scalar_f32".into(), fused_norm_add_f32_src);
        sources.insert("fused_moe_routing_f32".into(), fused_norm_add_f32_src);
        // ADR-028 iter-363: V2 (simd_max + simd_sum) variant of MoE routing.
        sources.insert("fused_moe_routing_f32_v2".into(), fused_norm_add_f32_src);
        // ADR-029 iter-175 Step 1i: V3 = V2 + parallel SG-tournament top-K
        // (replaces V2's single-thread serial scan for k = 0..top_k).
        sources.insert("fused_moe_routing_f32_v3".into(), fused_norm_add_f32_src);
        sources.insert("fused_moe_routing_batch_f32".into(), fused_norm_add_f32_src);
        // ADR-029 iter-175 Step 1j: batched-prefill V3 (same parallel
        // SG-tournament top-K as fused_moe_routing_f32_v3, applied per-token
        // within each TG of the batched dispatch).
        sources.insert("fused_moe_routing_batch_f32_v3".into(), fused_norm_add_f32_src);
        sources.insert("fused_norm_add_scalar_f32".into(), fused_norm_add_f32_src);
        sources.insert("fused_moe_wsum_norm_add_f32".into(), fused_norm_add_f32_src);
        sources.insert("fused_moe_wsum_dnorm_add_f32".into(), fused_norm_add_f32_src);

        // Argsort kernel (Story 2.3) — MoE top-K routing
        let argsort_src: &'static str = include_str!("shaders/argsort.metal");
        sources.insert("argsort_desc_f32".into(), argsort_src);

        // Gather / index_select kernel (Story 2.4)
        let gather_src: &'static str = include_str!("shaders/gather.metal");
        sources.insert("gather_f32".into(), gather_src);

        // F32 KV cache copy kernel (Session merge S1+S2)
        let kv_cache_copy_src: &'static str =
            include_str!("shaders/kv_cache_copy.metal");
        sources.insert("kv_cache_copy".into(), kv_cache_copy_src);
        sources.insert("kv_cache_copy_f32".into(), kv_cache_copy_src);

        // Strided copy kernel (Story 2.5)
        let copy_src: &'static str = include_str!("shaders/copy.metal");
        sources.insert("strided_copy_f32".into(), copy_src);
        sources.insert("offset_copy_f32".into(), copy_src);

        // Fused-QKV split kernel (ADR-005 W-5b.18 — replaces hf2q CPU
        // download → triple-loop split → 3× upload round-trip in
        // gpu_delta_net::layer_qkv_deinterleave).
        let qkv_split_src: &'static str = include_str!("shaders/qkv_split.metal");
        sources.insert("qkv_split_f32".into(), qkv_split_src);

        // Tiled-GQA broadcast kernel (ADR-005 W-5b.19 — replaces hf2q CPU
        // tiled-replicate at gpu_delta_net::apply_gated_delta_net_chunk
        // GQA pre-expansion, ~497 ms / 10.4 ms-per-layer at PP4106).
        let repeat_tiled_src: &'static str =
            include_str!("shaders/repeat_tiled.metal");
        sources.insert("repeat_tiled_f32".into(), repeat_tiled_src);

        // Dense F16 GEMM kernel (Story 2.6) — lm_head projection
        let dense_gemm_src: &'static str = include_str!("shaders/dense_gemm.metal");
        sources.insert("dense_gemm_f16".into(), dense_gemm_src);
        sources.insert("dense_matvec_f16".into(), dense_gemm_src);
        sources.insert("dense_matvec_f16w_f32io".into(), dense_gemm_src);
        // BF16-weight mat-vec: BF16 weights × F32 input → F32 output (decode lm_head)
        sources.insert("dense_matvec_bf16w_f32io".into(), dense_gemm_src);
        // Pure F32 mat-vec: F32 weights × F32 input → F32 output (decode lm_head)
        sources.insert("dense_matvec_f32".into(), dense_gemm_src);

        // Standalone FWHT for TurboQuant pre/post-rotation (SIMD shuffle, zero barriers)
        let fwht_src: &'static str = include_str!("shaders/fwht_standalone.metal");
        sources.insert("fwht_standalone_f32_d256".into(), fwht_src);
        sources.insert("fwht_standalone_f32_d512".into(), fwht_src);
        // ADR-007 iter-14 D1 SRHT variants: sign pre-mult (for Q) + sign undo (for output)
        sources.insert("fwht_sign_premult_f32_d256".into(), fwht_src);
        sources.insert("fwht_sign_premult_f32_d512".into(), fwht_src);
        sources.insert("fwht_sign_undo_f32_d256".into(), fwht_src);
        sources.insert("fwht_sign_undo_f32_d512".into(), fwht_src);

        // Fast Hadamard quantize (SIMD shuffle, zero barriers)
        let hq_fast_src: &'static str = include_str!("shaders/hadamard_quantize_kv_fast.metal");
        sources.insert("hadamard_quantize_kv_fast_d256".into(), hq_fast_src);
        sources.insert("hadamard_quantize_kv_fast_d512".into(), hq_fast_src);
        // ADR-028 iter-485 (Phase 7d / H4): fused K+V single-position 4-bit encoder.
        sources.insert("hadamard_quantize_kv_fast_dual_d256".into(), hq_fast_src);
        sources.insert("hadamard_quantize_kv_fast_dual_d512".into(), hq_fast_src);
        // Track B (iter-21): higher-bit (5/6-bit) quantize kernels (byte-packed)
        sources.insert("hadamard_quantize_kv_hb_d256".into(), hq_fast_src);
        sources.insert("hadamard_quantize_kv_hb_d512".into(), hq_fast_src);
        // ADR-028 iter-148: fused K+V single-position HB encoder
        sources.insert("hadamard_quantize_kv_hb_dual_d256".into(), hq_fast_src);
        sources.insert("hadamard_quantize_kv_hb_dual_d512".into(), hq_fast_src);
        // ADR-028 Phase 10e.5 (iter-351): no-FWHT V quantize for hybrid path.
        // Same byte-packed Lloyd-Max codebook output, but skips the Hadamard
        // rotation so dequant in SDPA recovers raw V (no FWHT-undo needed).
        sources.insert("kv_quantize_v_no_fwht_d256".into(), hq_fast_src);
        sources.insert("kv_quantize_v_no_fwht_d512".into(), hq_fast_src);
        // ADR-028 Phase 10c.5 (iter-354): fused F16-K-copy + V-no-FWHT-encode.
        // Saves 30 KV-write dispatches/decode-token at gemma4 30L by combining
        // the per-layer K-cast and V-encode into a single dispatch (Z-dim).
        sources.insert("kv_copy_kf16_quantize_v_no_fwht_d256".into(), hq_fast_src);
        sources.insert("kv_copy_kf16_quantize_v_no_fwht_d512".into(), hq_fast_src);

        // iter-20 Leg F: TQ KV dequantize kernel (nibbles+norms → F32)
        let tq_dq_src: &'static str = include_str!("shaders/tq_dequantize_kv.metal");
        sources.insert("tq_dequantize_kv".into(), tq_dq_src);
        // Track B (iter-21): higher-bit dequantize kernel (byte-packed indices)
        sources.insert("tq_dequantize_hb_kv".into(), tq_dq_src);
        // ADR-027 Phase B iter-30 (hf2q sub-sub-iter 23c-β.1): sequence-batch
        // dequant variant. Same MSL source; new kernel entry point
        // `tq_dequantize_hb_kv_seq` reads positions [start_pos..start_pos+n_tokens)
        // in one dispatch (one threadgroup per (kv_head, position)). Unblocks
        // hf2q's TQ-aware prefill SDPA path (current per-position kernel
        // requires cur_len separate dispatches).
        sources.insert("tq_dequantize_hb_kv_seq".into(), tq_dq_src);

        // iter-24: native higher-bit (5/6/8-bit) TQ SDPA kernel (byte-packed K/V)
        let tq_hb_src: &'static str = include_str!("shaders/flash_attn_vec_tq_hb.metal");
        sources.insert("flash_attn_vec_tq_hb_dk256".into(), tq_hb_src);
        sources.insert("flash_attn_vec_tq_hb_dk512".into(), tq_hb_src);

        // ADR-028 §iter-485 (Phase 7d H3): fused TQ-HB reduce + FWHT-sign-undo.
        // Combines flash_attn_vec_reduce + fwht_sign_undo_f32 into a single
        // dispatch, saving 1 dispatch + 1 forced barrier per layer per decode
        // token. Gated by env flag `HF2Q_TQ_HB_OUT_FUSED=1` in forward_mlx.rs.
        let reduce_undo_src: &'static str = include_str!("shaders/flash_attn_vec_reduce_tq_hb_undo.metal");
        sources.insert("flash_attn_vec_reduce_tq_hb_undo_dk256".into(), reduce_undo_src);
        sources.insert("flash_attn_vec_reduce_tq_hb_undo_dk512".into(), reduce_undo_src);

        // ADR-028 Phase 10d (iter-349): hybrid F16-K + TQ-HB-V SDPA kernel.
        // Same V-side codebook as flash_attn_vec_tq_hb (5/6/8-bit Lloyd-Max);
        // K-side reads F16 dense — peer-equivalent layout, no codebook lookup.
        let hybrid_src: &'static str = include_str!("shaders/flash_attn_vec_hybrid.metal");
        sources.insert("flash_attn_vec_hybrid_dk256".into(), hybrid_src);
        sources.insert("flash_attn_vec_hybrid_dk512".into(), hybrid_src);
        // ADR-040 M4 — batched multi-seq decode flash (same source file).
        sources.insert("flash_attn_vec_hybrid_batched_dk256".into(), hybrid_src);
        sources.insert("flash_attn_vec_hybrid_batched_dk512".into(), hybrid_src);

        // ADR-029: verbatim llama.cpp peer port.
        // F16-K + F16-V, DK=DV=256, NWG=1, NSG=1, NE=1. No function constants — baked.
        let peer_port_src: &'static str = include_str!("shaders/flash_attn_vec_peer_port_f16.metal");
        sources.insert("flash_attn_vec_peer_port_f16_dk256_dv256".into(), peer_port_src);

        // ADR-029 iter-134: peer reduce kernel (verbatim port of ggml-metal.metal 7235-7275).
        // Pairs with the NWG=32 vec kernel to match peer's actual runtime dispatch.
        let peer_port_reduce_src: &'static str =
            include_str!("shaders/flash_attn_vec_peer_port_f16_reduce.metal");
        sources.insert(
            "flash_attn_vec_peer_port_f16_reduce_dv256_nwg32".into(),
            peer_port_reduce_src,
        );

        // ADR-029 iter-135: NWG=32 variant of the verbatim peer port. Same body as
        // flash_attn_vec_peer_port_f16.metal with NWG=1→32. Pairs with iter-134 reduce kernel.
        let peer_port_nwg32_src: &'static str =
            include_str!("shaders/flash_attn_vec_peer_port_f16_nwg32.metal");
        sources.insert(
            "flash_attn_vec_peer_port_f16_nwg32_dk256_dv256".into(),
            peer_port_nwg32_src,
        );

        // GPU sampling kernels — eliminate logits readback (Phase 6)
        let argmax_src: &'static str = include_str!("shaders/argmax.metal");
        sources.insert("argmax_f32".into(), argmax_src);
        let softmax_sample_src: &'static str =
            include_str!("shaders/softmax_sample.metal");
        sources.insert("softmax_sample_f32".into(), softmax_sample_src);
        // Top-K kernel for Q8 rerank: avoids full-logits readback.
        let top_k_src: &'static str = include_str!("shaders/top_k.metal");
        sources.insert("top_k_f32".into(), top_k_src);

        // MoE GPU routing + weighted reduce (ADR-013 P13.3 perf).
        // Replaces CPU softmax+topk round-trip and CPU weighted accumulate.
        let moe_stk_src: &'static str =
            include_str!("shaders/moe_softmax_topk.metal");
        sources.insert("moe_softmax_topk_f32".into(), moe_stk_src);
        let moe_wr_src: &'static str =
            include_str!("shaders/moe_weighted_reduce.metal");
        sources.insert("moe_weighted_reduce_f32".into(), moe_wr_src);
        let sdpa_decode_src: &'static str =
            include_str!("shaders/sdpa_decode.metal");
        sources.insert("sdpa_decode".into(), sdpa_decode_src);

        Self {
            cache: HashMap::new(),
            sources,
            precompiled_lib: None,
            precompiled_load_attempted: false,
        }
    }

    /// Try to obtain the precompiled `default.metallib` Library, loading it
    /// lazily on first call.  Returns `None` when:
    /// - `MLX_PRECOMPILED_METALLIB` is unset (default)
    /// - The embedded blob is empty (build.rs skipped metallib build)
    /// - `device.new_library_with_data` failed previously
    /// - The previous load attempt already failed (no retry)
    fn try_precompiled_lib(
        &mut self,
        device: &metal::DeviceRef,
    ) -> Option<&metal::LibraryRef> {
        if !precompiled_enabled() {
            return None;
        }
        if !self.precompiled_load_attempted {
            self.precompiled_load_attempted = true;
            if EMBEDDED_METALLIB.is_empty() {
                return None;
            }
            // Apple's `newLibraryWithData:` expects a dispatch_data_t.
            // metal-rs wraps this via `new_library_with_data` which takes
            // a `&[u8]`.
            match device.new_library_with_data(EMBEDDED_METALLIB) {
                Ok(lib) => self.precompiled_lib = Some(lib),
                Err(_) => self.precompiled_lib = None,
            }
        }
        self.precompiled_lib.as_deref()
    }

    /// Register a shader source at runtime (useful for testing and dynamic
    /// kernel generation).
    pub fn register_source(&mut self, name: impl Into<String>, source: &'static str) {
        let name = name.into();
        // Invalidate any cached pipeline for this name since the source changed.
        self.cache.remove(&name);
        self.sources.insert(name, source);
    }

    /// ADR-033 §Pi Task #20 iter 11 (2026-05-23) — eagerly compile a list
    /// of kernel pipelines to move first-call JIT/PSO-creation cost out of
    /// the prefill hot path and into the model-load window.
    ///
    /// The profiler showed that on Qwen3.6 35B-A3B MoE prefill at seq=553,
    /// the FIRST FA layer + FIRST FFN layer take ~40ms each (vs ~14µs
    /// warm) — that's 80ms of the 221ms prefill, dominated by Metal
    /// pipeline state creation. Pre-creating these pipelines at load time
    /// (when 3.3s is already being spent on model parse + upload) is a
    /// strict perf win for measured prefill throughput.
    ///
    /// Best-effort: silently skips kernels that aren't registered (e.g.,
    /// list contains a kernel name for an arch this build doesn't use).
    /// Logs at debug level on failure to keep load-path quiet.
    ///
    /// Returns the count of pipelines successfully prewarmed.
    pub fn prewarm_pipelines(
        &mut self,
        device: &metal::DeviceRef,
        names: &[&str],
    ) -> usize {
        let mut warmed = 0_usize;
        for name in names {
            // Skip if already cached.
            if self.cache.contains_key(*name) {
                warmed += 1;
                continue;
            }
            // Skip if no source registered for this name.
            if !self.sources.contains_key(*name) {
                continue;
            }
            // Best-effort: ignore errors so one broken kernel doesn't
            // poison the whole prewarm pass.
            if self.get_pipeline(name, device).is_ok() {
                warmed += 1;
            }
        }
        warmed
    }

    /// ADR-033 §Pi Task #20 iter 12 (2026-05-23) — prewarm pipelines that
    /// require `[[function_constant]]` specialization. Each entry is
    /// `(name, &[(constant_index, bool_value)])`. Mirrors
    /// `prewarm_pipelines` but routes through
    /// `get_pipeline_with_bool_constants` so kernels declaring
    /// `function_constant` decls without defaults can be safely
    /// prewarmed.
    ///
    /// Use case: hot-path kernels like `flash_attn_prefill_bf16_d256`
    /// (uses bool constants 200/201/300/301/303 for align/mask/causal/blk
    /// flags) cannot be safely prewarmed without specialization — Metal
    /// `validateWithDevice:` asserts and aborts the process. Provide
    /// the constants production uses and prewarming becomes safe.
    ///
    /// Returns count warmed.
    pub fn prewarm_pipelines_with_bool_constants(
        &mut self,
        device: &metal::DeviceRef,
        entries: &[(&str, &[(usize, bool)])],
    ) -> usize {
        let mut warmed = 0_usize;
        for (name, bool_constants) in entries {
            if !self.sources.contains_key(*name) {
                continue;
            }
            if self
                .get_pipeline_with_bool_constants(name, device, bool_constants)
                .is_ok()
            {
                warmed += 1;
            }
        }
        warmed
    }

    /// ADR-033 §Pi Task #20 iter 11 (2026-05-23) — prewarm every registered
    /// kernel source. Useful when the exact set of needed kernels is hard
    /// to enumerate (e.g., serving paths that span multiple arches).
    /// Total cost is bounded by the number of registered kernels times
    /// the per-pipeline PSO creation cost (~5-15ms typical on M-series).
    ///
    /// Returns (warmed, skipped) counts.
    pub fn prewarm_all(&mut self, device: &metal::DeviceRef) -> (usize, usize) {
        let names: Vec<String> = self.sources.keys().cloned().collect();
        let mut warmed = 0_usize;
        let mut skipped = 0_usize;
        for name in &names {
            if self.cache.contains_key(name) {
                warmed += 1;
                continue;
            }
            if self.get_pipeline(name, device).is_ok() {
                warmed += 1;
            } else {
                skipped += 1;
            }
        }
        (warmed, skipped)
    }

    /// Get a compiled compute pipeline for the named kernel function.
    ///
    /// On first call for a given name, this compiles the MSL source into a
    /// Metal library, extracts the named function, and creates a
    /// `ComputePipelineState`.  Subsequent calls return the cached pipeline.
    ///
    /// # Errors
    ///
    /// * `MlxError::KernelNotFound` — no source registered for this name.
    /// * `MlxError::ShaderCompilationError` — MSL compilation or pipeline
    ///   creation failed.
    pub fn get_pipeline(
        &mut self,
        name: &str,
        device: &metal::DeviceRef,
    ) -> Result<&ComputePipelineState> {
        if !self.cache.contains_key(name) {
            // ADR-029 iter-175 Step 1l: precompiled .metallib fast path.
            // When MLX_PRECOMPILED_METALLIB=1 AND the kernel exists in the
            // embedded library, use it.  Otherwise fall through to runtime
            // source compile.  Empirically ~+5.89% faster on q6_K matvec
            // (iter 1k bench).
            let precompiled_function = self
                .try_precompiled_lib(device)
                .and_then(|lib| lib.get_function(name, None).ok());

            let function = match precompiled_function {
                Some(f) => f,
                None => {
                    // Slow path: compile the shader.
                    let source = self.sources.get(name).ok_or_else(|| {
                        MlxError::KernelNotFound(name.to_string())
                    })?;

                    let compile_opts = metal::CompileOptions::new();
                    let library = device
                        .new_library_with_source(source, &compile_opts)
                        .map_err(|msg| MlxError::ShaderCompilationError {
                            name: name.to_string(),
                            message: msg,
                        })?;

                    library
                        .get_function(name, None)
                        .map_err(|msg| MlxError::ShaderCompilationError {
                            name: name.to_string(),
                            message: msg,
                        })?
                }
            };

            // Build the pipeline through a descriptor so we can attach a
            // human-readable label.  The label propagates into Instruments /
            // xctrace Metal System Trace as the per-pipeline identifier
            // (`metal-object-label` schema), giving us per-kernel attribution
            // instead of the generic "Compute Command 0" placeholder.
            //
            // `MTLComputePipelineState.label` is read-only after creation per
            // the Apple Metal spec; the only supported way to set it is via
            // the descriptor before pipeline creation.  ADR-015 iter9b.
            let descriptor = ComputePipelineDescriptor::new();
            descriptor.set_compute_function(Some(&function));
            descriptor.set_label(name);
            // ADR-028 iter-376: threadGroupSizeIsMultipleOfThreadExecutionWidth
            // hint allows the Metal compiler to skip bounds checks and use more
            // aggressive codegen.  Opt-in via HF2Q_PIPELINE_TG_MULT_HINT=1.
            // SAFETY: every dispatched threadgroup MUST be a multiple of 32 at
            // runtime — Apple specifies undefined behavior otherwise.  Our hot
            // kernels use tg_size ∈ {32, 64, 256, 1024} (all multiples of 32).
            if std::env::var("HF2Q_PIPELINE_TG_MULT_HINT").ok().as_deref() == Some("1") {
                descriptor.set_thread_group_size_is_multiple_of_thread_execution_width(true);
            }

            let pipeline = device
                .new_compute_pipeline_state(&descriptor)
                .map_err(|msg| MlxError::ShaderCompilationError {
                    name: name.to_string(),
                    message: msg,
                })?;

            self.cache.insert(name.to_string(), pipeline);
        }

        // At this point the pipeline is guaranteed to be in the cache.
        // We use `ok_or_else` instead of `expect` to satisfy the no-panic policy.
        self.cache.get(name).ok_or_else(|| {
            MlxError::KernelNotFound(name.to_string())
        })
    }

    /// Get a compiled compute pipeline for the named kernel, specialized with
    /// Metal function constants (both bool and i32 in one call).
    ///
    /// `bool_constants` contains `(index, value)` pairs mapping to
    /// `[[function_constant(index)]]` bool declarations in the MSL shader.
    /// `int_constants` contains `(index, value)` pairs mapping to
    /// `[[function_constant(index)]]` int (int32_t) declarations in the MSL
    /// shader.
    ///
    /// Pipelines are cached by a composite key:
    /// `"<name>|<index>:b<0|1>|...|<index>:i<value>|..."`.  The 'b' prefix
    /// marks bool entries and the 'i' prefix marks i32 entries, making the
    /// format unambiguous regardless of constant ordering.  Distinct
    /// `(name, constants)` combinations each compile to a separate pipeline;
    /// the slow compilation path runs at most once per unique combination.
    ///
    /// # Errors
    ///
    /// * `MlxError::KernelNotFound` — no source registered for this name.
    /// * `MlxError::ShaderCompilationError` — MSL compilation, function
    ///   specialisation, or pipeline creation failed.
    pub fn get_pipeline_with_constants(
        &mut self,
        name: &str,
        device: &metal::DeviceRef,
        bool_constants: &[(usize, bool)],
        int_constants: &[(usize, i32)],
    ) -> Result<&ComputePipelineState> {
        // Build a composite cache key so distinct constant combinations each
        // compile to their own pipeline.  Bool entries use the 'b' type marker
        // and i32 entries use 'i'; this prevents a collision between, e.g.,
        // bool index 5 value 1 and int index 5 value 1.
        let mut cache_key = name.to_string();
        for &(index, value) in bool_constants {
            cache_key.push('|');
            cache_key.push_str(&index.to_string());
            cache_key.push_str(if value { ":b1" } else { ":b0" });
        }
        for &(index, value) in int_constants {
            cache_key.push('|');
            cache_key.push_str(&index.to_string());
            cache_key.push(':');
            cache_key.push('i');
            cache_key.push_str(&value.to_string());
        }

        if !self.cache.contains_key(&cache_key) {
            // Build the FunctionConstantValues object with all bool and i32
            // constants.  Metal's set_constant_value_at_index reads the value
            // through a raw pointer; the pointed-to bytes must match the size
            // declared in the MSL shader (1 byte for bool, 4 bytes for int).
            let fcv = FunctionConstantValues::new();

            for &(index, value) in bool_constants {
                // MTLDataType::Bool = 53 (metal-rs argument.rs).
                // The Metal runtime reads it as an Objective-C BOOL (uint8_t).
                let v: u8 = if value { 1 } else { 0 };
                fcv.set_constant_value_at_index(
                    (&v as *const u8).cast::<std::ffi::c_void>(),
                    MTLDataType::Bool,
                    index as u64,
                );
            }

            for &(index, value) in int_constants {
                // MTLDataType::Int = 29 (metal-rs argument.rs).
                // The Metal runtime reads 4 bytes as a signed 32-bit integer,
                // matching the Metal shader type `constant int`.
                fcv.set_constant_value_at_index(
                    (&value as *const i32).cast::<std::ffi::c_void>(),
                    MTLDataType::Int,
                    index as u64,
                );
            }

            // ADR-029 iter-175 Step 1l: try precompiled .metallib first.
            // Step 1m: gated separately on MLX_PRECOMPILED_METALLIB_FCV=1
            // so we can isolate whether the integration regression at
            // Step 1l (tg50 95.4 → 62.1) lives in this FCV path vs the
            // no-FCV path in `get_pipeline`.
            //
            // get_function with FCV takes ownership of the FCV, so we
            // build a separate one for the precompiled probe.
            let precompiled_function = if precompiled_fcv_enabled() {
                let probe_fcv = FunctionConstantValues::new();
                for &(index, value) in bool_constants {
                    let v: u8 = if value { 1 } else { 0 };
                    probe_fcv.set_constant_value_at_index(
                        (&v as *const u8).cast::<std::ffi::c_void>(),
                        MTLDataType::Bool,
                        index as u64,
                    );
                }
                for &(index, value) in int_constants {
                    probe_fcv.set_constant_value_at_index(
                        (&value as *const i32).cast::<std::ffi::c_void>(),
                        MTLDataType::Int,
                        index as u64,
                    );
                }
                self.try_precompiled_lib(device)
                    .and_then(|lib| lib.get_function(name, Some(probe_fcv)).ok())
            } else {
                None
            };

            let function = match precompiled_function {
                Some(f) => f,
                None => {
                    // Slow path: compile the shader with function constant specialisation.
                    let source = self.sources.get(name).ok_or_else(|| {
                        MlxError::KernelNotFound(name.to_string())
                    })?;

                    let compile_opts = metal::CompileOptions::new();
                    let library = device
                        .new_library_with_source(source, &compile_opts)
                        .map_err(|msg| MlxError::ShaderCompilationError {
                            name: name.to_string(),
                            message: msg,
                        })?;

                    library
                        .get_function(name, Some(fcv))
                        .map_err(|msg| MlxError::ShaderCompilationError {
                            name: name.to_string(),
                            message: msg,
                        })?
                }
            };

            // Label this specialisation with the full composite cache key
            // (e.g. `kernel_mul_mv_q4_0_f32|0:b1|3:i32`) so xctrace Metal
            // System Trace shows each function-constant variant as a distinct
            // pipeline.  Without this, all specialisations share a generic
            // "Compute Command 0" identifier and we cannot attribute µs/token
            // to a specific (kernel, constants) combination.  ADR-015 iter9b.
            let descriptor = ComputePipelineDescriptor::new();
            descriptor.set_compute_function(Some(&function));
            descriptor.set_label(&cache_key);
            // ADR-028 iter-376: same hint as primary pipeline path.
            if std::env::var("HF2Q_PIPELINE_TG_MULT_HINT").ok().as_deref() == Some("1") {
                descriptor.set_thread_group_size_is_multiple_of_thread_execution_width(true);
            }

            let pipeline = device
                .new_compute_pipeline_state(&descriptor)
                .map_err(|msg| MlxError::ShaderCompilationError {
                    name: name.to_string(),
                    message: msg,
                })?;

            self.cache.insert(cache_key.clone(), pipeline);
        }

        self.cache.get(&cache_key).ok_or_else(|| {
            MlxError::KernelNotFound(name.to_string())
        })
    }

    /// Get a compiled compute pipeline for the named kernel, specialized with
    /// Metal bool function constants.
    ///
    /// The `bool_constants` slice contains `(index, value)` pairs.  Each pair
    /// maps to a `[[function_constant(index)]]` declaration in the MSL shader.
    ///
    /// This is a thin wrapper around [`get_pipeline_with_constants`] that
    /// passes an empty `int_constants` slice.  Existing callers continue to
    /// work without modification; the cache-key format for pure-bool pipelines
    /// is compatible (bool entries carry the 'b' type marker, which is the
    /// only format ever written by this wrapper).
    ///
    /// # Errors
    ///
    /// * `MlxError::KernelNotFound` — no source registered for this name.
    /// * `MlxError::ShaderCompilationError` — MSL compilation, function
    ///   specialisation, or pipeline creation failed.
    pub fn get_pipeline_with_bool_constants(
        &mut self,
        name: &str,
        device: &metal::DeviceRef,
        bool_constants: &[(usize, bool)],
    ) -> Result<&ComputePipelineState> {
        self.get_pipeline_with_constants(name, device, bool_constants, &[])
    }

    /// Check if a pipeline for the given name is already compiled and cached.
    pub fn is_cached(&self, name: &str) -> bool {
        self.cache.contains_key(name)
    }

    /// Number of compiled pipelines currently in the cache.
    pub fn cached_count(&self) -> usize {
        self.cache.len()
    }

    /// Number of registered shader sources.
    pub fn source_count(&self) -> usize {
        self.sources.len()
    }
}

impl Default for KernelRegistry {
    fn default() -> Self {
        Self::new()
    }
}

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

    /// Minimal Metal shader that uses a single int function constant.
    ///
    /// The kernel writes the constant value N into the first element of the
    /// output buffer, allowing the test to verify that the Metal compiler
    /// actually sees distinct specialisations for N=4 and N=8.
    ///
    /// The shader is intentionally trivial — we only need it to *compile* with
    /// an int function constant; correctness of the kernel logic is not under
    /// test here.
    const INT_FC_TEST_SHADER: &str = r#"
#include <metal_stdlib>
using namespace metal;

constant int test_N [[function_constant(100)]];

kernel void int_fc_test_kernel(
    device int* out [[buffer(0)]],
    uint tid [[thread_position_in_grid]])
{
    if (tid == 0) {
        out[0] = test_N;
    }
}
"#;

    /// Verify that `get_pipeline_with_constants` produces distinct cached
    /// pipelines for different i32 function-constant values, and that
    /// `get_pipeline_with_bool_constants` (the backward-compat wrapper) still
    /// works correctly with the new 'b'-prefixed cache-key format.
    ///
    /// This test requires a real Metal device and is therefore marked
    /// `#[ignore]` on non-Apple platforms, but runs unconditionally on macOS.
    #[test]
    fn test_int_fc_distinct_pipelines_and_bool_compat() {
        let device = metal::Device::system_default()
            .expect("no Metal device — run on Apple Silicon or x86 Mac with Metal support");

        let mut registry = KernelRegistry::new();

        // Register the inline test shader under a name that cannot collide with
        // any production kernel.
        registry.register_source("int_fc_test_kernel", INT_FC_TEST_SHADER);

        // Compile with N=4.
        let p4_ptr = registry
            .get_pipeline_with_constants(
                "int_fc_test_kernel",
                &device,
                &[],                  // no bool constants
                &[(100, 4_i32)],      // int constant index 100 = 4
            )
            .expect("pipeline N=4 should compile") as *const _;

        // Cache must now have exactly 1 entry for this kernel.
        // (Other production kernels may already be in cache from new(); here
        // we check that the N=4 key was inserted.)
        let count_after_n4 = registry.cached_count();

        // Compile with N=8 — must produce a SEPARATE pipeline.
        let p8_ptr = registry
            .get_pipeline_with_constants(
                "int_fc_test_kernel",
                &device,
                &[],
                &[(100, 8_i32)],
            )
            .expect("pipeline N=8 should compile") as *const _;

        // Cache must have grown by exactly 1.
        assert_eq!(
            registry.cached_count(),
            count_after_n4 + 1,
            "N=8 must produce a new cache entry"
        );

        // The two pipelines must be distinct objects in the cache.
        assert_ne!(
            p4_ptr, p8_ptr,
            "N=4 and N=8 specialisations must be separate ComputePipelineState objects"
        );

        // A second call with N=4 must return the SAME pipeline (cache hit, no
        // new compilation).
        let p4_again_ptr = registry
            .get_pipeline_with_constants(
                "int_fc_test_kernel",
                &device,
                &[],
                &[(100, 4_i32)],
            )
            .expect("pipeline N=4 cache hit should succeed") as *const _;

        assert_eq!(
            registry.cached_count(),
            count_after_n4 + 1,
            "repeated N=4 call must be a cache hit, not a new entry"
        );
        assert_eq!(
            p4_ptr, p4_again_ptr,
            "repeated N=4 call must return the same pipeline pointer"
        );

        // Verify backward compatibility: get_pipeline_with_bool_constants must
        // still route through get_pipeline_with_constants and produce a cached
        // pipeline without panicking.
        //
        // We register a separate bool-constant shader that does NOT use a bool
        // function constant (so the Metal compiler ignores missing FCs for
        // this trivial case) — but the call path and cache-key format are what
        // matter here.  We reuse the int_fc_test_kernel source; the bool FC is
        // simply unused by the shader (Metal allows unused FCs when the shader
        // declares them with `function_constant` but the value is never read).
        //
        // To avoid a Metal compiler error for an undeclared function constant,
        // we register a separate bare-kernel shader for the bool wrapper test.
        const BARE_SHADER: &str = r#"
#include <metal_stdlib>
using namespace metal;
kernel void bare_kernel(device int* out [[buffer(0)]], uint tid [[thread_position_in_grid]]) {
    if (tid == 0) { out[0] = 42; }
}
"#;
        registry.register_source("bare_kernel", BARE_SHADER);

        let count_before_bool = registry.cached_count();
        let _bool_pipeline = registry
            .get_pipeline_with_bool_constants("bare_kernel", &device, &[])
            .expect("bool-constants wrapper with empty slice must succeed");

        assert_eq!(
            registry.cached_count(),
            count_before_bool + 1,
            "bool-constants wrapper must insert one new cache entry"
        );
    }

    /// Verify that the `MTLComputePipelineState.label` produced by
    /// `get_pipeline` and `get_pipeline_with_constants` actually propagates
    /// from the descriptor to the resulting pipeline state.
    ///
    /// This is the in-process smoke check for ADR-015 iter9b: we cannot
    /// reach into xctrace from Rust, but we can read back the same `label`
    /// property xctrace consumes via `ComputePipelineStateRef::label()`.
    /// If labels are missing or wrong here, the MST trace will also show
    /// generic identifiers — so this test gates the iter9 retry's
    /// per-Q4_0-kernel attribution.
    #[test]
    fn test_pipeline_labels_propagate_for_mst() {
        let device = metal::Device::system_default()
            .expect("no Metal device — run on Apple Silicon or x86 Mac with Metal support");

        let mut registry = KernelRegistry::new();

        // Reuse the same trivial shaders as the int-FC test.
        registry.register_source("int_fc_test_kernel", INT_FC_TEST_SHADER);

        const BARE_SHADER_LABEL_TEST: &str = r#"
#include <metal_stdlib>
using namespace metal;
kernel void label_smoke_kernel(device int* out [[buffer(0)]], uint tid [[thread_position_in_grid]]) {
    if (tid == 0) { out[0] = 7; }
}
"#;
        registry.register_source("label_smoke_kernel", BARE_SHADER_LABEL_TEST);

        // Plain get_pipeline path — label must equal the kernel name.
        // Capture as owned String so the cache borrow is released before
        // the next get_pipeline_with_constants call below.
        let plain_label = registry
            .get_pipeline("label_smoke_kernel", &device)
            .expect("plain pipeline must compile")
            .label()
            .to_string();
        assert_eq!(
            plain_label, "label_smoke_kernel",
            "get_pipeline must label the pipeline with the kernel name (xctrace MST attribution)"
        );

        // Constants path — label must equal the composite cache key so each
        // function-constant variant is individually attributable in MST.
        // We capture the label as an owned String to release the borrow on
        // the cache before fetching the next specialisation.
        let label_v7 = registry
            .get_pipeline_with_constants(
                "int_fc_test_kernel",
                &device,
                &[],
                &[(100, 7_i32)],
            )
            .expect("specialised pipeline must compile")
            .label()
            .to_string();
        assert_eq!(
            label_v7, "int_fc_test_kernel|100:i7",
            "get_pipeline_with_constants must label with the cache_key so each \
             specialisation is distinct in xctrace MST"
        );

        // A second specialisation must produce a different label.
        let label_v13 = registry
            .get_pipeline_with_constants(
                "int_fc_test_kernel",
                &device,
                &[],
                &[(100, 13_i32)],
            )
            .expect("second specialised pipeline must compile")
            .label()
            .to_string();
        assert_eq!(label_v13, "int_fc_test_kernel|100:i13");
        assert_ne!(
            label_v7, label_v13,
            "distinct constant values must yield distinct pipeline labels"
        );
    }
}