aria_inference/

session.rs

use crate::bundle::{load_bundle, Bundle, LoadedWeight};
use crate::chat::{apply_chat_template, strip_assistant_visible, ChatTurn};
use crate::family::{effective_rope_theta, graph_hook, require_runnable, ArchClass, Family};
use crate::multimodal::asr_transcribe_pcm16le;
use crate::tensor_names::{
    action_head_names, attn_k_names, attn_k_norm_names, attn_norm_names, attn_o_names,
    attn_post_norm_names, attn_q_names, attn_q_norm_names, attn_v_names, attn_v_norm_names,
    conv_in_proj_names, conv_kernel_names, conv_out_proj_names, emb_names, embed_per_layer_names,
    ffn_down_names, ffn_gate_names, ffn_norm_names, ffn_post_norm_names, ffn_up_names,
    layer_ple_gate_names, layer_ple_post_norm_names, layer_ple_proj_names, layer_scalar_names,
    linear_a_log_names,
    linear_conv1d_names, linear_dt_bias_names, linear_in_proj_a_names, linear_in_proj_b_names,
    linear_in_proj_ba_names, linear_in_proj_qkv_names, linear_in_proj_qkvz_names,
    linear_in_proj_z_names, linear_out_proj_names, moe_expert_down_names, moe_expert_gate_names, moe_expert_up_names,
    moe_router_names, output_names, output_norm_names, per_layer_model_projection_names,
    per_layer_projection_norm_names, pre_feedforward_norm_names, vision_proj_names,
};
use crate::profile::{
    elapsed_ms, load_profile_begin, load_profile_set_cuda_upload, load_profile_set_materialize,
    load_profile_set_mmap, load_profile_take, EngineProfile, GenerateProfile,
};
use crate::tokenizer::{decode_placeholders, encode_naive, BundleTokenizer};
use aria_kernel::{
    attention_causal_with_scale, attention_with_scale, gated_delta_step, geglu, gelu_pytorch_tanh,
    hdm_linear, kv_sliding_view, linear_cpu, moe_topk_route, resolve_compute, rms_norm,
    rms_norm_gemma, rope_half, rope_half_proportional, short_conv_step, silu_vec, softplus, swiglu,
    ComputeBackend, ComputePref, CudaContext, EngineError, GatedDeltaStep,
};
use std::cell::RefCell;
use std::collections::HashMap;
use std::path::Path;
use std::sync::Arc;
use std::time::Instant;

#[derive(Debug, Clone)]
pub struct GenerateOpts {
    pub max_tokens: usize,
    pub temperature: f32,
}

impl Default for GenerateOpts {
    fn default() -> Self {
        Self {
            max_tokens: 16,
            temperature: 0.0,
        }
    }
}

#[derive(Debug, Clone)]
pub struct Generation {
    pub tokens: Vec<u32>,
    pub text: String,
}

#[derive(Clone)]
struct MatWeight {
    data: Arc<Vec<f32>>,
    hdm_seed: Option<i64>,
}

impl MatWeight {
    fn from_loaded(w: LoadedWeight) -> Self {
        Self {
            data: Arc::new(w.data),
            hdm_seed: w.hdm_seed,
        }
    }

    /// Concatenate two `[out, hidden]` matrices along the output axis.
    fn concat_out(a: &Self, b: &Self) -> Self {
        let mut data = Vec::with_capacity(a.data.len() + b.data.len());
        data.extend_from_slice(&a.data);
        data.extend_from_slice(&b.data);
        Self {
            data: Arc::new(data),
            hdm_seed: None,
        }
    }
}

#[derive(Clone, Copy)]
enum GemmAcct {
    Attn,
    Ffn,
    LmHead,
    Other,
}

#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
enum AttnKind {
    Sliding,
    Full,
}

/// Persistent caches for autoregressive decode (KV / short-conv / DeltaNet).
struct DecodeState {
    k_caches: Vec<Vec<f32>>,
    v_caches: Vec<Vec<f32>>,
    last_kv_src: HashMap<AttnKind, usize>,
    conv_states: Vec<Option<Vec<f32>>>,
    delta_states: Vec<Option<Vec<f32>>>,
    /// Next absolute position (RoPE / seq index).
    pos: usize,
}

#[derive(Clone)]
struct AttnWeights {
    wq: MatWeight,
    /// None on Gemma-4 KV-consumer layers (reuse producer cache).
    wk: Option<MatWeight>,
    wv: Option<MatWeight>,
    wo: MatWeight,
    q_norm: Option<Vec<f32>>,
    k_norm: Option<Vec<f32>>,
    v_norm: Option<Vec<f32>>,
    kind: AttnKind,
}

#[derive(Clone)]
struct ConvWeights {
    in_proj: MatWeight,
    out_proj: MatWeight,
    /// Depthwise kernel `[hidden * kernel]`.
    kernel: Vec<f32>,
    kernel_size: usize,
}

/// Qwen3.5 / Bonsai Gated DeltaNet (linear attention).
#[derive(Clone)]
struct DeltaWeights {
    qkvz: MatWeight,
    ba: MatWeight,
    conv: Vec<f32>,
    conv_k: usize,
    out_proj: MatWeight,
    a_log: Vec<f32>,
    dt_bias: Vec<f32>,
    n_k_heads: usize,
    n_v_heads: usize,
    head_k: usize,
    head_v: usize,
}

#[derive(Clone)]
enum LayerOp {
    Attn(AttnWeights),
    Conv(ConvWeights),
    Linear(DeltaWeights),
}

#[derive(Clone)]
struct ExpertWeights {
    gate: MatWeight,
    up: MatWeight,
    down: MatWeight,
}

#[derive(Clone)]
enum FfnWeights {
    Dense {
        gate: MatWeight,
        up: MatWeight,
        down: MatWeight,
    },
    MoE {
        router: MatWeight,
        experts: Vec<ExpertWeights>,
        top_k: usize,
        use_sigmoid: bool,
    },
}

struct LayerPle {
    gate: MatWeight,
    proj: MatWeight,
    post_norm: Vec<f32>,
}

struct PleModel {
    embed: Arc<Vec<f32>>,
    proj: MatWeight,
    proj_norm: Vec<f32>,
    d: usize,
}

struct LayerWeights {
    attn_norm: Vec<f32>,
    ffn_norm: Vec<f32>,
    post_attn_norm: Option<Vec<f32>>,
    post_ffn_norm: Option<Vec<f32>>,
    ple: Option<LayerPle>,
    /// HF `layer_scalar` / JAX `skip_scale`. 1.0 when the bundle omits it.
    layer_scalar: f32,
    op: LayerOp,
    ffn: FfnWeights,
}

struct ModelWeights {
    emb: MatWeight,
    layers: Vec<LayerWeights>,
    output_norm: Vec<f32>,
    output: MatWeight,
    vision: Option<MatWeight>,
    action: Option<MatWeight>,
    ple: Option<PleModel>,
}

pub struct Session {
    family: Family,
    bundle: Bundle,
    weights: ModelWeights,
    conf: crate::bundle::ModelConfig,
    use_gemma_norm: bool,
    use_gemma4: bool,
    use_geglu: bool,
    embed_scale: f32,
    /// HF `final_logit_softcapping` (Gemma-4 default 30). None = disabled.
    final_logit_softcap: Option<f32>,
    tokenizer: Option<BundleTokenizer>,
    /// Cleared at the start of each `generate`; reused across decode steps.
    decode: Option<DecodeState>,
    compute: ComputeBackend,
    compute_label: String,
    cuda: Option<CudaContext>,
    profile_on: bool,
    last_profile: Option<EngineProfile>,
    gen_acc: RefCell<GenerateProfile>,
}

impl std::fmt::Debug for Session {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("Session")
            .field("family", &self.family)
            .field("model", &self.conf.hidden_size)
            .finish()
    }
}

pub struct SessionBuilder {
    path: Option<std::path::PathBuf>,
    family_path: String,
    compute: ComputePref,
    profile: bool,
}

impl SessionBuilder {
    pub fn new() -> Self {
        Self {
            path: None,
            family_path: "gemma/gemma-4-e2b-it".into(),
            compute: ComputePref::Auto,
            profile: false,
        }
    }

    pub fn model(mut self, path: impl AsRef<Path>) -> Self {
        self.path = Some(path.as_ref().to_path_buf());
        self
    }

    pub fn family(mut self, path: impl Into<String>) -> Self {
        self.family_path = path.into();
        self
    }

    pub fn compute(mut self, pref: ComputePref) -> Self {
        self.compute = pref;
        self
    }

    pub fn profile(mut self, on: bool) -> Self {
        self.profile = on;
        self
    }

    pub fn build(self) -> Result<Session, EngineError> {
        let family = require_runnable(&self.family_path)?;
        let _hook = graph_hook(family.arch);
        let path = self
            .path
            .ok_or_else(|| EngineError::InvalidParam("model path required".into()))?;
        let (compute, compute_label) = resolve_compute(self.compute)?;
        load_profile_begin(self.profile);
        let t_mmap = Instant::now();
        let bundle = load_bundle(&path)?;
        load_profile_set_mmap(elapsed_ms(t_mmap));
        let mut conf = bundle.model.clone();
        conf.rope_theta = effective_rope_theta(family.path(), conf.rope_theta);
        // Hub/q4 artifacts may still omit Gemma-4 geometry; fill from HF architecture
        // then validate. Bundle values always win when present.
        fill_gemma4_architecture_defaults(&mut conf, family.path());
        require_gemma4_config(&conf, family.path())?;
        reject_unsupported_geometry(&conf, family)?;
        let tokenizer = BundleTokenizer::try_load(&path)?;
        let t_mat = Instant::now();
        // Materialize with the same config Session uses for RoPE / head_dim /
        // layer_types so AttnKind and KV sharing match the forward path.
        let weights = materialize_with_config(&bundle, family, &conf)?;
        require_gemma4_ple(&weights, &conf, family.path())?;
        load_profile_set_materialize(elapsed_ms(t_mat));
        let mut cuda = None;
        if compute == ComputeBackend::Cuda {
            let t_up = Instant::now();
            let ctx = CudaContext::new()?;
            upload_weights(&ctx, &weights)?;
            load_profile_set_cuda_upload(elapsed_ms(t_up));
            cuda = Some(ctx);
        }
        let act = conf
            .hidden_act
            .as_deref()
            .unwrap_or("")
            .to_ascii_lowercase();
        let use_gemma4 = family.path().contains("gemma-4");
        let use_gemma_norm = family.path().contains("gemma") && !use_gemma4;
        let use_geglu = act.contains("gelu") || use_gemma4;
        // Real Gemma-4 E2B/E4B checkpoints always ship PLE. Tiny/unit fixtures may
        // omit it; materialize already errors if model-level PLE is partial.
        let embed_scale = if family.path().contains("gemma") {
            (conf.hidden_size as f32).sqrt()
        } else {
            1.0
        };
        let final_logit_softcap = if use_gemma4 { Some(30.0) } else { None };
        let load = load_profile_take();
        let last_profile = self.profile.then(|| EngineProfile {
            compute: compute_label.clone(),
            load,
            generate: None,
            ci_fail: false,
        });
        Ok(Session {
            family,
            bundle,
            weights,
            conf,
            use_gemma_norm,
            use_gemma4,
            use_geglu,
            embed_scale,
            final_logit_softcap,
            tokenizer,
            decode: None,
            compute,
            compute_label,
            cuda,
            profile_on: self.profile,
            last_profile,
            gen_acc: RefCell::new(GenerateProfile::default()),
        })
    }
}

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

fn reject_unsupported_geometry(
    conf: &crate::bundle::ModelConfig,
    family: Family,
) -> Result<(), EngineError> {
    let path = family.path();
    // Hybrid linear-attn families must declare linear_attention / delta layers.
    if path.contains("qwen3.5") || path.contains("bonsai") {
        let has_linear = conf
            .layer_types
            .as_ref()
            .map(|t| {
                t.iter().any(|s| {
                    let s = s.to_ascii_lowercase();
                    s.contains("linear_attention") || s.contains("delta")
                })
            })
            .unwrap_or(false);
        if !has_linear {
            return Err(EngineError::Unsupported(format!(
                "{path}: requires model.layer_types with Gated DeltaNet / linear_attention \
                 (dense-only bundles are unsupported until DeltaNet lands)"
            )));
        }
    }
    // MoE families need explicit expert count (router/experts materialize from that).
    if family.is_moe() && conf.num_experts.unwrap_or(0) == 0 {
        return Err(EngineError::Unsupported(format!(
            "{path}: MoE family requires model.num_experts > 0 in bundle config"
        )));
    }
    Ok(())
}

fn layer_type_str(conf: &crate::bundle::ModelConfig, layer: usize) -> String {
    conf.layer_types
        .as_ref()
        .and_then(|t| t.get(layer))
        .map(|s| s.to_ascii_lowercase())
        .unwrap_or_else(|| "full_attention".into())
}

fn layer_is_conv(conf: &crate::bundle::ModelConfig, layer: usize) -> bool {
    layer_type_str(conf, layer).contains("conv")
}

fn layer_is_linear(conf: &crate::bundle::ModelConfig, layer: usize) -> bool {
    let t = layer_type_str(conf, layer);
    t.contains("linear_attention") || t.contains("delta")
}

fn attn_kind(conf: &crate::bundle::ModelConfig, layer: usize) -> AttnKind {
    if layer_type_str(conf, layer).contains("sliding") {
        AttnKind::Sliding
    } else {
        AttnKind::Full
    }
}

fn is_kv_consumer(conf: &crate::bundle::ModelConfig, layer: usize) -> bool {
    let n = conf.num_kv_shared_layers.unwrap_or(0);
    n > 0 && layer >= conf.num_layers.saturating_sub(n)
}

/// HF Gemma-4 E2B/E4B: repeating 4×sliding + 1×full (full at indices 4,9,...,n-1).
fn default_gemma4_layer_types(n: usize) -> Vec<String> {
    (0..n)
        .map(|i| {
            if (i + 1) % 5 == 0 {
                "full_attention".into()
            } else {
                "sliding_attention".into()
            }
        })
        .collect()
}

/// Fill Gemma-4 geometry omitted by published hub q4 (base fields only) or
/// half-updated bundles. Prefer explicit `config_from_hf` values when present.
fn fill_gemma4_architecture_defaults(conf: &mut crate::bundle::ModelConfig, family_path: &str) {
    if !family_path.contains("gemma-4") {
        return;
    }
    if conf.layer_types.as_ref().map(|t| t.len()) != Some(conf.num_layers) {
        conf.layer_types = Some(default_gemma4_layer_types(conf.num_layers));
    }
    if conf.sliding_window.unwrap_or(0) == 0 {
        conf.sliding_window = Some(512);
    }
    if conf.partial_rotary_factor.is_none() {
        conf.partial_rotary_factor = Some(0.25);
    }
    // Full E2B/E4B use head_dim=256 / global_head_dim=512; tiny fixtures derive.
    if conf.hidden_size >= 1024 {
        if conf.head_dim.unwrap_or(0) == 0 {
            conf.head_dim = Some(256);
        }
        if conf.global_head_dim.unwrap_or(0) == 0 {
            conf.global_head_dim = Some(512);
        }
        if conf.num_kv_shared_layers.is_none() {
            conf.num_kv_shared_layers = Some(20);
        }
    } else {
        if conf.head_dim.unwrap_or(0) == 0 && conf.num_attention_heads > 0 {
            conf.head_dim = Some(conf.hidden_size / conf.num_attention_heads);
        }
        if conf.global_head_dim.unwrap_or(0) == 0 {
            conf.global_head_dim = conf.head_dim;
        }
    }
}

/// After defaults, Gemma-4 must have a complete geometry (bundle or filled).
fn require_gemma4_config(
    conf: &crate::bundle::ModelConfig,
    family_path: &str,
) -> Result<(), EngineError> {
    if !family_path.contains("gemma-4") {
        return Ok(());
    }
    let missing = |field: &str| {
        EngineError::Unsupported(format!(
            "{family_path}: model.{field} required after Gemma-4 architecture fill \
             (re-quantize with current model config_from_hf)"
        ))
    };
    match &conf.layer_types {
        None => return Err(missing("layer_types")),
        Some(t) if t.len() != conf.num_layers => {
            return Err(EngineError::Unsupported(format!(
                "{family_path}: model.layer_types length {} != num_layers {}",
                t.len(),
                conf.num_layers
            )));
        }
        Some(_) => {}
    }
    if conf.sliding_window.unwrap_or(0) == 0 {
        return Err(missing("sliding_window"));
    }
    match conf.partial_rotary_factor {
        Some(f) if f > 0.0 && f <= 1.0 => {}
        _ => return Err(missing("partial_rotary_factor")),
    }
    if conf.head_dim.unwrap_or(0) == 0 {
        return Err(missing("head_dim"));
    }
    if conf.global_head_dim.unwrap_or(0) == 0 {
        return Err(missing("global_head_dim"));
    }
    Ok(())
}

fn gemma4_requires_ple(family_path: &str, hidden: usize) -> bool {
    family_path.contains("gemma-4") && hidden >= 1024
}

fn require_gemma4_ple(
    weights: &ModelWeights,
    conf: &crate::bundle::ModelConfig,
    family_path: &str,
) -> Result<(), EngineError> {
    if !gemma4_requires_ple(family_path, conf.hidden_size) {
        return Ok(());
    }
    if weights.ple.is_none() {
        return Err(EngineError::Format(format!(
            "{family_path}: codebook PLE required for Gemma-4 E2B/E4B \
             (embed_tokens_per_layer + per_layer_model_projection + \
             per_layer_projection_norm); refusing silent no-op"
        )));
    }
    Ok(())
}

/// When bundle declares distinct global vs local head dims, prefer q_proj geometry.
fn resolve_attn_kind(
    conf: &crate::bundle::ModelConfig,
    layer: usize,
    q_dim: usize,
    n_heads: usize,
) -> AttnKind {
    if n_heads > 0 && q_dim.is_multiple_of(n_heads) {
        let head_from_q = q_dim / n_heads;
        if let (Some(g), Some(h)) = (
            conf.global_head_dim.filter(|d| *d > 0),
            conf.head_dim.filter(|d| *d > 0),
        ) {
            if g != h {
                if head_from_q == g {
                    return AttnKind::Full;
                }
                if head_from_q == h {
                    return AttnKind::Sliding;
                }
            }
        }
    }
    attn_kind(conf, layer)
}

fn materialize_with_config(
    b: &Bundle,
    family: Family,
    conf: &crate::bundle::ModelConfig,
) -> Result<ModelWeights, EngineError> {
    fn any_mat(b: &Bundle, names: &[String]) -> Result<MatWeight, EngineError> {
        let refs: Vec<&str> = names.iter().map(String::as_str).collect();
        Ok(MatWeight::from_loaded(b.weight_loaded_any(&refs)?))
    }
    fn try_mat(b: &Bundle, names: &[String]) -> Result<Option<MatWeight>, EngineError> {
        match any_mat(b, names) {
            Ok(w) => Ok(Some(w)),
            Err(EngineError::Format(_)) => Ok(None),
            Err(e) => Err(e),
        }
    }
    fn any_vec(b: &Bundle, names: &[String]) -> Result<Vec<f32>, EngineError> {
        Ok((*any_mat(b, names)?.data).clone())
    }
    fn optional_vec(b: &Bundle, names: &[String]) -> Option<Vec<f32>> {
        any_vec(b, names).ok()
    }

    let m = conf;
    let hidden = m.hidden_size;
    let n_heads = m.num_attention_heads;
    let n_experts = m.num_experts.unwrap_or(0);
    let top_k = m.num_experts_per_tok.unwrap_or(1).max(1);
    // LFM MoE uses sigmoid routing; Mixtral/Inkling-style uses softmax.
    let use_sigmoid_router = n_experts > 0 && m.layer_types.is_some();

    let mut layers = Vec::with_capacity(m.num_layers);
    let mut prev_wk: Option<MatWeight> = None;
    let mut prev_wv: Option<MatWeight> = None;
    for layer in 0..m.num_layers {
        let attn_norm = any_vec(b, &attn_norm_names(layer))?;
        let pre_ff = optional_vec(b, &pre_feedforward_norm_names(layer));
        let post_attn_norm = if pre_ff.is_some() {
            optional_vec(b, &attn_post_norm_names(layer))
        } else {
            None
        };
        let post_ffn_norm = optional_vec(b, &ffn_post_norm_names(layer));
        let ffn_norm = if let Some(v) = pre_ff {
            v
        } else {
            any_vec(b, &ffn_norm_names(layer))?
        };

        let op = if layer_is_conv(m, layer) {
            let in_proj = any_mat(b, &conv_in_proj_names(layer))?;
            let out_proj = any_mat(b, &conv_out_proj_names(layer))?;
            let kw = any_mat(b, &conv_kernel_names(layer))?;
            let kernel_size = m.conv_l_cache.unwrap_or(3).max(1);
            if kw.data.len() % hidden != 0 {
                return Err(EngineError::ShapeMismatch(format!(
                    "layer {layer} conv kernel len {} not divisible by hidden {hidden}",
                    kw.data.len()
                )));
            }
            let inferred_k = kw.data.len() / hidden;
            let kernel_size = if inferred_k > 0 {
                inferred_k
            } else {
                kernel_size
            };
            // Accept [H,K] or squeezed [H,1,K] (same flat length).
            if kw.data.len() != hidden * kernel_size {
                return Err(EngineError::ShapeMismatch(format!(
                    "layer {layer} conv kernel len {} != hidden*kernel {hidden}*{kernel_size}",
                    kw.data.len()
                )));
            }
            let kernel = (*kw.data).clone();
            if in_proj.data.len() != 3 * hidden * hidden {
                return Err(EngineError::ShapeMismatch(format!(
                    "layer {layer} conv in_proj len {} != 3*hidden*hidden",
                    in_proj.data.len()
                )));
            }
            if out_proj.data.len() != hidden * hidden {
                return Err(EngineError::ShapeMismatch(format!(
                    "layer {layer} conv out_proj len {} != hidden*hidden",
                    out_proj.data.len()
                )));
            }
            LayerOp::Conv(ConvWeights {
                in_proj,
                out_proj,
                kernel,
                kernel_size,
            })
        } else if layer_is_linear(m, layer) {
            let qkvz = if let Some(w) = try_mat(b, &linear_in_proj_qkvz_names(layer))? {
                w
            } else {
                let qkv = any_mat(b, &linear_in_proj_qkv_names(layer))?;
                let z = any_mat(b, &linear_in_proj_z_names(layer))?;
                MatWeight::concat_out(&qkv, &z)
            };
            let ba = if let Some(w) = try_mat(b, &linear_in_proj_ba_names(layer))? {
                w
            } else {
                let proj_b = any_mat(b, &linear_in_proj_b_names(layer))?;
                let proj_a = any_mat(b, &linear_in_proj_a_names(layer))?;
                MatWeight::concat_out(&proj_b, &proj_a)
            };
            let conv_w = any_mat(b, &linear_conv1d_names(layer))?;
            let out_proj = any_mat(b, &linear_out_proj_names(layer))?;
            let a_log = any_vec(b, &linear_a_log_names(layer))?;
            let dt_bias = any_vec(b, &linear_dt_bias_names(layer))?;
            let n_v_heads = a_log.len();
            if n_v_heads == 0 || dt_bias.len() != n_v_heads {
                return Err(EngineError::ShapeMismatch(format!(
                    "layer {layer} A_log/dt_bias head mismatch"
                )));
            }
            if ba.data.len() % hidden != 0 {
                return Err(EngineError::ShapeMismatch(
                    "linear in_proj_ba not divisible by hidden".into(),
                ));
            }
            if ba.data.len() / hidden != 2 * n_v_heads {
                return Err(EngineError::ShapeMismatch(format!(
                    "layer {layer} in_proj_ba out {} != 2*n_v_heads {}",
                    ba.data.len() / hidden,
                    2 * n_v_heads
                )));
            }
            if qkvz.data.len() % hidden != 0 {
                return Err(EngineError::ShapeMismatch(
                    "linear in_proj_qkvz not divisible by hidden".into(),
                ));
            }
            let qkvz_out = qkvz.data.len() / hidden;
            // Equal k/v dims: qkvz = 2*key + 2*value = 4*key.
            if !qkvz_out.is_multiple_of(4) {
                return Err(EngineError::ShapeMismatch(format!(
                    "layer {layer} qkvz out {qkvz_out} not divisible by 4"
                )));
            }
            let key_dim = qkvz_out / 4;
            let value_dim = key_dim;
            let n_k_heads = n_v_heads;
            if n_k_heads == 0
                || !key_dim.is_multiple_of(n_k_heads)
                || !value_dim.is_multiple_of(n_v_heads)
            {
                return Err(EngineError::ShapeMismatch(format!(
                    "layer {layer} cannot infer DeltaNet head dims"
                )));
            }
            let head_k = key_dim / n_k_heads;
            let head_v = value_dim / n_v_heads;
            let conv_dim = key_dim * 2 + value_dim;
            if conv_w.data.len() % conv_dim != 0 {
                return Err(EngineError::ShapeMismatch(format!(
                    "layer {layer} conv1d len {} not divisible by conv_dim {conv_dim}",
                    conv_w.data.len()
                )));
            }
            let conv_k = conv_w.data.len() / conv_dim;
            if out_proj.data.len() != hidden * value_dim {
                return Err(EngineError::ShapeMismatch(format!(
                    "layer {layer} linear out_proj len {} != hidden*value_dim",
                    out_proj.data.len()
                )));
            }
            LayerOp::Linear(DeltaWeights {
                qkvz,
                ba,
                conv: (*conv_w.data).clone(),
                conv_k,
                out_proj,
                a_log,
                dt_bias,
                n_k_heads,
                n_v_heads,
                head_k,
                head_v,
            })
        } else {
            let consumer = is_kv_consumer(m, layer);
            let (wk, wv) = if consumer {
                (None, None)
            } else {
                let wk = match any_mat(b, &attn_k_names(layer)) {
                    Ok(w) => {
                        prev_wk = Some(w.clone());
                        Some(w)
                    }
                    Err(e) => Some(prev_wk.clone().ok_or_else(|| {
                        EngineError::Format(format!(
                            "missing k_proj for layer {layer} and no prior KV to share ({e})"
                        ))
                    })?),
                };
                let wv = match any_mat(b, &attn_v_names(layer)) {
                    Ok(w) => {
                        prev_wv = Some(w.clone());
                        Some(w)
                    }
                    Err(e) => Some(prev_wv.clone().ok_or_else(|| {
                        EngineError::Format(format!(
                            "missing v_proj for layer {layer} and no prior KV to share ({e})"
                        ))
                    })?),
                };
                (wk, wv)
            };
            let wq = any_mat(b, &attn_q_names(layer))?;
            if wq.data.len() % hidden != 0 {
                return Err(EngineError::ShapeMismatch(format!(
                    "layer {layer} q_proj len {} not divisible by hidden {hidden}",
                    wq.data.len()
                )));
            }
            let q_dim = wq.data.len() / hidden;
            LayerOp::Attn(AttnWeights {
                wq,
                wk,
                wv,
                wo: any_mat(b, &attn_o_names(layer))?,
                q_norm: optional_vec(b, &attn_q_norm_names(layer)),
                k_norm: optional_vec(b, &attn_k_norm_names(layer)),
                v_norm: optional_vec(b, &attn_v_norm_names(layer)),
                kind: resolve_attn_kind(m, layer, q_dim, n_heads),
            })
        };

        let ffn = if n_experts > 0 {
            match any_mat(b, &moe_router_names(layer)) {
                Ok(router) => {
                    if router.data.len() != n_experts * hidden {
                        return Err(EngineError::ShapeMismatch(format!(
                            "layer {layer} MoE router len {} != num_experts*hidden {n_experts}*{hidden}",
                            router.data.len()
                        )));
                    }
                    let mut experts = Vec::with_capacity(n_experts);
                    for e in 0..n_experts {
                        experts.push(ExpertWeights {
                            gate: any_mat(b, &moe_expert_gate_names(layer, e))?,
                            up: any_mat(b, &moe_expert_up_names(layer, e))?,
                            down: any_mat(b, &moe_expert_down_names(layer, e))?,
                        });
                    }
                    FfnWeights::MoE {
                        router,
                        experts,
                        top_k,
                        use_sigmoid: use_sigmoid_router,
                    }
                }
                Err(_) => {
                    // Dense FFN on this layer (e.g. LFM2-A1B first layers).
                    FfnWeights::Dense {
                        gate: any_mat(b, &ffn_gate_names(layer))?,
                        up: any_mat(b, &ffn_up_names(layer))?,
                        down: any_mat(b, &ffn_down_names(layer))?,
                    }
                }
            }
        } else {
            FfnWeights::Dense {
                gate: any_mat(b, &ffn_gate_names(layer))?,
                up: any_mat(b, &ffn_up_names(layer))?,
                down: any_mat(b, &ffn_down_names(layer))?,
            }
        };

        let ple = match (
            any_mat(b, &layer_ple_gate_names(layer)),
            any_mat(b, &layer_ple_proj_names(layer)),
            optional_vec(b, &layer_ple_post_norm_names(layer)),
        ) {
            (Ok(gate), Ok(proj), Some(post_norm)) => Some(LayerPle {
                gate,
                proj,
                post_norm,
            }),
            _ => None,
        };

        let layer_scalar = optional_vec(b, &layer_scalar_names(layer))
            .and_then(|v| v.into_iter().find(|x| x.is_finite()))
            .unwrap_or(1.0);

        layers.push(LayerWeights {
            attn_norm,
            ffn_norm,
            post_attn_norm,
            post_ffn_norm,
            ple,
            layer_scalar,
            op,
            ffn,
        });
    }
    let emb_n = emb_names();
    let out_norm_n = output_norm_names();
    let out_n = output_names();
    let vis_n: Vec<String> = vision_proj_names()
        .iter()
        .map(|s| (*s).to_string())
        .collect();
    let act_n: Vec<String> = action_head_names()
        .iter()
        .map(|s| (*s).to_string())
        .collect();
    let emb = MatWeight::from_loaded(b.weight_loaded_any(&emb_n)?);
    let output = if m.tie_word_embeddings.unwrap_or(false)
        || family.path().contains("gemma-4")
        || (family.path().contains("qwen3") && !family.path().contains("qwen3.5"))
    {
        // Qwen3-0.6B/1.7B and Gemma-4 tie lm_head to embed; prefer embed even
        // if a separate lm_head tensor exists (often a worse-quantized copy).
        emb.clone()
    } else {
        MatWeight::from_loaded(b.weight_loaded_any(&out_n)?)
    };
    let require_ple = gemma4_requires_ple(family.path(), hidden);
    let ple = {
        let embed_n = embed_per_layer_names();
        let proj_n: Vec<String> = per_layer_model_projection_names()
            .iter()
            .map(|s| (*s).to_string())
            .collect();
        let norm_n: Vec<String> = per_layer_projection_norm_names()
            .iter()
            .map(|s| (*s).to_string())
            .collect();
        let embed_res = b.weight_loaded_any(&embed_n);
        let proj_res = any_mat(b, &proj_n);
        let proj_norm = optional_vec(b, &norm_n);
        match (embed_res, proj_res, proj_norm) {
            (Ok(embed), Ok(proj), Some(proj_norm)) => {
                let d = proj_norm.len();
                if d == 0 {
                    return Err(EngineError::ShapeMismatch(
                        "PLE projection norm dim is 0".into(),
                    ));
                }
                Some(PleModel {
                    embed: Arc::new(embed.data),
                    proj,
                    proj_norm,
                    d,
                })
            }
            (embed_res, proj_res, proj_norm) if require_ple => {
                let embed_s = match &embed_res {
                    Ok(_) => "ok".to_string(),
                    Err(e) => e.to_string(),
                };
                let proj_s = match &proj_res {
                    Ok(_) => "ok".to_string(),
                    Err(e) => e.to_string(),
                };
                let norm_s = if proj_norm.is_some() { "ok" } else { "missing" };
                return Err(EngineError::Format(format!(
                    "{}: codebook PLE required (embed_tokens_per_layer={embed_s}, \
                     per_layer_model_projection={proj_s}, per_layer_projection_norm={norm_s})",
                    family.path()
                )));
            }
            _ => None,
        }
    };
    if let Some(ple) = &ple {
        let packed = m.num_layers.saturating_mul(ple.d);
        if packed == 0
            || !ple.embed.len().is_multiple_of(packed)
            || ple.proj.data.len() != packed * hidden
        {
            return Err(EngineError::ShapeMismatch(format!(
                "PLE shapes: embed {} proj {} expected packed={} hidden={hidden}",
                ple.embed.len(),
                ple.proj.data.len(),
                packed
            )));
        }
        for (i, layer) in layers.iter().enumerate() {
            let Some(lp) = &layer.ple else {
                return Err(EngineError::Format(format!(
                    "PLE model tensors present but layer {i} missing gate/proj/norm"
                )));
            };
            if lp.gate.data.len() != ple.d * hidden || lp.proj.data.len() != hidden * ple.d {
                return Err(EngineError::ShapeMismatch(format!(
                    "layer {i} PLE gate/proj shape mismatch (d={}, hidden={hidden})",
                    ple.d
                )));
            }
            // HF `post_per_layer_input_norm` is RMSNorm(hidden_size), not ple_d.
            // A wrong-sized weight would silently chunk-norm and corrupt residuals.
            if lp.post_norm.len() != hidden {
                return Err(EngineError::ShapeMismatch(format!(
                    "layer {i} PLE post_norm len {} != hidden {hidden}",
                    lp.post_norm.len()
                )));
            }
        }
    }
    Ok(ModelWeights {
        emb,
        layers,
        output_norm: b.weight_loaded_any(&out_norm_n)?.data,
        output,
        vision: any_mat(b, &vis_n).ok(),
        action: any_mat(b, &act_n).ok(),
        ple,
    })
}

fn upload_weights(ctx: &CudaContext, w: &ModelWeights) -> Result<(), EngineError> {
    ctx.upload(&w.emb.data)?;
    ctx.upload(&w.output.data)?;
    if let Some(v) = &w.vision {
        ctx.upload(&v.data)?;
    }
    if let Some(a) = &w.action {
        ctx.upload(&a.data)?;
    }
    if let Some(ple) = &w.ple {
        ctx.upload(&ple.embed)?;
        ctx.upload(&ple.proj.data)?;
    }
    for layer in &w.layers {
        match &layer.op {
            LayerOp::Attn(attn) => {
                ctx.upload(&attn.wq.data)?;
                if let Some(wk) = &attn.wk {
                    ctx.upload(&wk.data)?;
                }
                if let Some(wv) = &attn.wv {
                    ctx.upload(&wv.data)?;
                }
                ctx.upload(&attn.wo.data)?;
            }
            LayerOp::Conv(c) => {
                ctx.upload(&c.in_proj.data)?;
                ctx.upload(&c.out_proj.data)?;
            }
            LayerOp::Linear(d) => {
                ctx.upload(&d.qkvz.data)?;
                ctx.upload(&d.ba.data)?;
                ctx.upload(&d.out_proj.data)?;
            }
        }
        if let Some(ple) = &layer.ple {
            ctx.upload(&ple.gate.data)?;
            ctx.upload(&ple.proj.data)?;
        }
        match &layer.ffn {
            FfnWeights::Dense { gate, up, down } => {
                ctx.upload(&gate.data)?;
                ctx.upload(&up.data)?;
                ctx.upload(&down.data)?;
            }
            FfnWeights::MoE { router, experts, .. } => {
                ctx.upload(&router.data)?;
                for e in experts {
                    ctx.upload(&e.gate.data)?;
                    ctx.upload(&e.up.data)?;
                    ctx.upload(&e.down.data)?;
                }
            }
        }
    }
    Ok(())
}

impl Session {
    pub fn family(&self) -> Family {
        self.family
    }

    pub fn model_id(&self) -> &str {
        self.family.path()
    }

    pub fn config(&self) -> &crate::bundle::ModelConfig {
        &self.conf
    }

    pub fn bundle(&self) -> &Bundle {
        &self.bundle
    }

    pub fn compute_label(&self) -> &str {
        &self.compute_label
    }

    pub fn last_profile(&self) -> Option<&EngineProfile> {
        self.last_profile.as_ref()
    }

    fn wmm(
        &self,
        w: &MatWeight,
        x: &[f32],
        out_f: usize,
        in_f: usize,
        acct: GemmAcct,
    ) -> Result<Vec<f32>, EngineError> {
        let t0 = Instant::now();
        let y = if let Some(seed) = w.hdm_seed {
            hdm_linear(x, &w.data, out_f, in_f, Some(seed))?
        } else if self.compute == ComputeBackend::Cuda {
            let ctx = self.cuda.as_ref().ok_or_else(|| {
                EngineError::Unsupported("compute=cuda but CudaContext missing".into())
            })?;
            ctx.linear(x, &w.data, out_f, in_f)?
        } else {
            linear_cpu(x, &w.data, out_f, in_f)?
        };
        if self.profile_on {
            let ms = elapsed_ms(t0);
            let mut g = self.gen_acc.borrow_mut();
            match acct {
                GemmAcct::Attn => g.gemm_attn_ms += ms,
                GemmAcct::Ffn => g.gemm_ffn_ms += ms,
                GemmAcct::LmHead => g.gemm_lm_head_ms += ms,
                GemmAcct::Other => {}
            }
        }
        Ok(y)
    }

    fn can_batch_prefill(&self) -> bool {
        self.weights.layers.iter().all(|layer| {
            matches!(layer.op, LayerOp::Attn(_)) && matches!(layer.ffn, FfnWeights::Dense { .. })
        })
    }

    /// Greedy (temperature<=0) generation from prompt token ids.
    /// Prefills the prompt once, then runs one incremental decode step per new token.
    pub fn generate(
        &mut self,
        prompt: &[u32],
        opts: &GenerateOpts,
    ) -> Result<Generation, EngineError> {
        if opts.max_tokens == 0 {
            return Err(EngineError::InvalidParam("max_tokens must be > 0".into()));
        }
        let mut tokens: Vec<u32> = prompt.to_vec();
        if tokens.is_empty() {
            tokens.push(1);
        }
        self.decode = Some(self.fresh_decode_state());
        *self.gen_acc.borrow_mut() = GenerateProfile::default();
        let result = (|| {
            let t_pre = Instant::now();
            let mut logits = if self.can_batch_prefill() && tokens.len() > 1 {
                self.forward_prompt(&tokens)?
            } else {
                let mut last = Vec::new();
                for &tok in &tokens {
                    last = self.forward_step(tok)?;
                }
                last
            };
            if self.profile_on {
                self.gen_acc.borrow_mut().prefill_ms = elapsed_ms(t_pre);
            }
            let mut generated = Vec::new();
            let t_dec = Instant::now();
            for _ in 0..opts.max_tokens {
                // Stage A: greedy for determinism; temperature reserved.
                let next = argmax(&logits);
                generated.push(next);
                tokens.push(next);
                if self.is_stop_id(next) {
                    generated.pop();
                    break;
                }
                logits = self.forward_step(next)?;
            }
            if self.profile_on {
                self.gen_acc.borrow_mut().decode_ms = elapsed_ms(t_dec);
            }
            let text = self.decode_tokens(&generated);
            Ok(Generation {
                tokens: generated,
                text,
            })
        })();
        if self.profile_on {
            let mut p = self.last_profile.take().unwrap_or(EngineProfile {
                compute: self.compute_label.clone(),
                load: load_profile_take(),
                generate: None,
                ci_fail: false,
            });
            p.generate = Some(self.gen_acc.borrow().clone());
            self.last_profile = Some(p);
        }
        self.decode = None;
        result
    }

    /// Map token ids → UTF-8 via bundle `tokenizer.json` (byte-level when applicable).
    /// Falls back to `<id>` placeholders when no sidecar is present.
    pub fn decode_tokens(&self, ids: &[u32]) -> String {
        match &self.tokenizer {
            Some(tok) => {
                let raw = tok.decode_opts(ids, false);
                strip_assistant_visible(&raw)
            }
            None => decode_placeholders(ids),
        }
    }

    /// Encode with bundle `tokenizer.json` when present; else naive byte fallback.
    pub fn encode_text(&self, text: &str) -> Vec<u32> {
        match &self.tokenizer {
            Some(tok) => match tok.encode(text) {
                Ok(ids) if !ids.is_empty() => ids,
                Ok(_) => encode_naive(text, self.conf.vocab_size as u32),
                Err(_) => encode_naive(text, self.conf.vocab_size as u32),
            },
            None => encode_naive(text, self.conf.vocab_size as u32),
        }
    }

    /// Encode OpenAI-style messages with the family / tokenizer chat template.
    pub fn encode_chat(&self, messages: &[ChatTurn]) -> Vec<u32> {
        // Prefer the session family when it is gemma-4 so a stale tokenizer hint
        // (e.g. gemma-3 `<start_of_turn>`) cannot override `<|turn>` markers.
        let family = if self.family.path().contains("gemma-4") {
            self.family.path()
        } else {
            self.tokenizer
                .as_ref()
                .and_then(|t| t.chat_family_hint())
                .unwrap_or(self.family.path())
        };
        let prompt = apply_chat_template(family, messages);
        self.encode_text(&prompt)
    }

    fn is_stop_id(&self, id: u32) -> bool {
        match &self.tokenizer {
            Some(t) => t.is_stop(id),
            None => id == 0,
        }
    }

    pub fn arch(&self) -> ArchClass {
        self.family.arch
    }

    pub fn graph_hook_name(&self) -> &'static str {
        graph_hook(self.family.arch)
    }

    /// Mean-pool token embeddings (stage C `/v1/embeddings`).
    pub fn embed_text(&self, text: &str) -> Result<Vec<f32>, EngineError> {
        let toks = self.encode_text(text);
        let hidden = self.conf.hidden_size;
        let vocab = self.conf.vocab_size;
        let mut acc = vec![0.0f32; hidden];
        if toks.is_empty() {
            return Ok(acc);
        }
        for &tok in &toks {
            let tid = (tok as usize) % vocab;
            let row = &self.weights.emb.data[tid * hidden..(tid + 1) * hidden];
            for i in 0..hidden {
                acc[i] += row[i];
            }
        }
        let inv = 1.0 / toks.len() as f32;
        for v in &mut acc {
            *v *= inv;
        }
        Ok(acc)
    }

    /// Stage C VL: project RGB via bundle vision weights (no mean-pool stub).
    pub fn vision_prefix(
        &self,
        rgb: &[u8],
        height: usize,
        width: usize,
    ) -> Result<Vec<f32>, EngineError> {
        if !matches!(self.family.arch, ArchClass::VL | ArchClass::VLA) {
            return Err(EngineError::Unsupported(format!(
                "vision_prefix not available for arch {:?}",
                self.family.arch
            )));
        }
        let Some(proj) = &self.weights.vision else {
            return Err(EngineError::Unsupported(format!(
                "{}: no vision projector tensor in bundle",
                self.family.path()
            )));
        };
        let hidden = self.conf.hidden_size;
        if hidden == 0 || proj.data.len() % hidden != 0 {
            return Err(EngineError::ShapeMismatch(
                "vision projector not divisible by hidden_size".into(),
            ));
        }
        let in_f = proj.data.len() / hidden;
        let need = height
            .checked_mul(width)
            .and_then(|n| n.checked_mul(3))
            .ok_or_else(|| EngineError::InvalidParam("vision size overflow".into()))?;
        if rgb.len() < need {
            return Err(EngineError::ShapeMismatch(format!(
                "rgb len {} < {}x{}x3",
                rgb.len(),
                height,
                width
            )));
        }
        let mut feat = vec![0.0f32; in_f];
        let pixels = height * width;
        if in_f == 3 {
            let mut acc = [0.0f32; 3];
            for p in 0..pixels {
                acc[0] += rgb[p * 3] as f32 / 255.0;
                acc[1] += rgb[p * 3 + 1] as f32 / 255.0;
                acc[2] += rgb[p * 3 + 2] as f32 / 255.0;
            }
            let s = 1.0 / pixels.max(1) as f32;
            feat[0] = acc[0] * s;
            feat[1] = acc[1] * s;
            feat[2] = acc[2] * s;
        } else {
            for i in 0..in_f {
                feat[i] = rgb[i % need] as f32 / 255.0;
            }
        }
        self.wmm(proj, &feat, hidden, in_f, GemmAcct::Other)
    }

    /// Stage C VLA: project last-token embedding with bundle action weights.
    pub fn predict_action(&self, prompt: &str, action_dim: usize) -> Result<Vec<f32>, EngineError> {
        if self.family.arch != ArchClass::VLA {
            return Err(EngineError::Unsupported(format!(
                "predict_action requires VLA, got {:?}",
                self.family.arch
            )));
        }
        if action_dim == 0 {
            return Err(EngineError::InvalidParam("action_dim must be > 0".into()));
        }
        let Some(head) = &self.weights.action else {
            return Err(EngineError::Unsupported(format!(
                "{}: no action head tensor in bundle",
                self.family.path()
            )));
        };
        let h = self.embed_text(prompt)?;
        let hidden = self.conf.hidden_size;
        if head.data.len() % hidden != 0 {
            return Err(EngineError::ShapeMismatch(
                "action head not divisible by hidden_size".into(),
            ));
        }
        let out_f = head.data.len() / hidden;
        if out_f != action_dim {
            return Err(EngineError::ShapeMismatch(format!(
                "action head out {out_f} != requested {action_dim}"
            )));
        }
        self.wmm(head, &h, out_f, hidden, GemmAcct::Other)
    }

    /// Stage C ASR stub bound to session vocab.
    pub fn transcribe_pcm16le(&self, pcm: &[u8]) -> Result<String, EngineError> {
        asr_transcribe_pcm16le(pcm, self.conf.vocab_size as u32)
    }

    fn norm(&self, x: &[f32], weight: &[f32]) -> Result<Vec<f32>, EngineError> {
        if self.use_gemma_norm {
            rms_norm_gemma(x, weight, 1e-6)
        } else {
            rms_norm(x, weight, 1e-6)
        }
    }

    fn add_normed_residual(
        &self,
        x: &mut [f32],
        y: &[f32],
        post_norm: Option<&[f32]>,
    ) -> Result<(), EngineError> {
        if y.len() != x.len() {
            return Err(EngineError::ShapeMismatch(
                "residual length mismatch".into(),
            ));
        }
        if let Some(w) = post_norm {
            let yn = self.norm(y, w)?;
            for (a, b) in x.iter_mut().zip(yn.iter()) {
                *a += *b;
            }
        } else {
            for (a, b) in x.iter_mut().zip(y.iter()) {
                *a += *b;
            }
        }
        Ok(())
    }

    fn attn_scale(&self, head_dim: usize) -> f32 {
        if self.use_gemma4 {
            1.0
        } else {
            1.0 / (head_dim as f32).sqrt()
        }
    }

    /// Sliding layers attend to the last `sliding_window` keys (Gemma-4: 512).
    /// Full-attention layers keep the causal prefix. KV cache itself is not cropped
    /// so shared-KV producers still keep full history.
    fn attn_window(&self, kind: AttnKind) -> Option<usize> {
        if kind != AttnKind::Sliding {
            return None;
        }
        self.conf.sliding_window.filter(|w| *w > 0)
    }

    fn layer_rope_params(&self, kind: AttnKind) -> (f32, Option<f32>) {
        if self.use_gemma4 {
            match kind {
                AttnKind::Sliding => (10_000.0, None),
                AttnKind::Full => {
                    // Validated by require_gemma4_config; 1.0 means full RoPE.
                    let factor = self.conf.partial_rotary_factor.unwrap_or(1.0);
                    let partial = if factor > 0.0 && factor < 1.0 {
                        Some(factor)
                    } else {
                        None
                    };
                    (1_000_000.0, partial)
                }
            }
        } else {
            (self.conf.rope_theta, None)
        }
    }

    fn layer_head_dim(
        &self,
        kind: AttnKind,
        q_dim: usize,
        n_heads: usize,
    ) -> Result<usize, EngineError> {
        if n_heads == 0 || !q_dim.is_multiple_of(n_heads) {
            return Err(EngineError::ShapeMismatch(
                "q_dim not divisible by num_attention_heads".into(),
            ));
        }
        let configured = match kind {
            AttnKind::Full => self.conf.global_head_dim.or(self.conf.head_dim),
            AttnKind::Sliding => self.conf.head_dim,
        };
        Ok(configured
            .filter(|d| *d > 0 && q_dim == n_heads * *d)
            .unwrap_or(q_dim / n_heads))
    }

    fn apply_rope(
        x: &mut [f32],
        head_dim: usize,
        pos: usize,
        theta: f32,
        proportional: Option<f32>,
    ) -> Result<(), EngineError> {
        if let Some(factor) = proportional {
            // HF `_compute_proportional_rope_parameters`: rotate leading
            // factor*head_dim/2 pairs; inv_freq denominator is full head_dim.
            rope_half_proportional(x, head_dim, factor, pos, theta)
        } else {
            rope_half(x, head_dim, pos, theta)
        }
    }

    fn apply_v_norm(
        &self,
        v: Vec<f32>,
        v_norm: Option<&[f32]>,
        head_dim: usize,
    ) -> Result<Vec<f32>, EngineError> {
        if let Some(vn) = v_norm {
            if vn.len() != head_dim {
                return Err(EngineError::ShapeMismatch(format!(
                    "v_norm len {} != head_dim {head_dim}",
                    vn.len()
                )));
            }
            rms_norm(&v, vn, 1e-6)
        } else if self.use_gemma4 {
            let ones = vec![1.0f32; head_dim];
            rms_norm(&v, &ones, 1e-6)
        } else {
            Ok(v)
        }
    }

    fn compute_ple_inputs(
        &self,
        toks: &[u32],
        embeds: &[f32],
    ) -> Result<Option<Vec<f32>>, EngineError> {
        let Some(ple) = &self.weights.ple else {
            return Ok(None);
        };
        let hidden = self.conf.hidden_size;
        let n_layers = self.weights.layers.len();
        let d = ple.d;
        let packed = n_layers * d;
        let seq = toks.len();
        if seq == 0 || embeds.len() != seq * hidden {
            return Err(EngineError::ShapeMismatch(
                "PLE embed sequence length mismatch".into(),
            ));
        }
        let scale_lookup = (d as f32).sqrt();
        let ple_vocab = ple.embed.len() / packed;
        if ple_vocab == 0 {
            return Err(EngineError::ShapeMismatch("PLE embed vocab is 0".into()));
        }
        let mut lookup = vec![0.0f32; seq * packed];
        for (t, &tok) in toks.iter().enumerate() {
            let tid = (tok as usize) % ple_vocab;
            let row = &ple.embed[tid * packed..(tid + 1) * packed];
            for i in 0..packed {
                lookup[t * packed + i] = row[i] * scale_lookup;
            }
        }
        let proj_scale = (hidden as f32).sqrt().recip();
        let mut proj = self.wmm(&ple.proj, embeds, packed, hidden, GemmAcct::Other)?;
        for v in &mut proj {
            *v *= proj_scale;
        }
        proj = rms_norm(&proj, &ple.proj_norm, 1e-6)?;
        let inv_sqrt2 = std::f32::consts::FRAC_1_SQRT_2;
        for i in 0..proj.len() {
            proj[i] = (proj[i] + lookup[i]) * inv_sqrt2;
        }
        Ok(Some(proj))
    }

    fn apply_ple(
        &self,
        x: &mut [f32],
        layer: &LayerWeights,
        li: usize,
        ple_tok: Option<&[f32]>,
        hidden: usize,
    ) -> Result<(), EngineError> {
        let (Some(ple), Some(ple_tok)) = (&layer.ple, ple_tok) else {
            return Ok(());
        };
        let d = self
            .weights
            .ple
            .as_ref()
            .map(|p| p.d)
            .ok_or_else(|| EngineError::Format("layer PLE without model PLE".into()))?;
        let n_layers = self.weights.layers.len();
        let seq = x.len() / hidden;
        let gate_out = self.wmm(&ple.gate, x, d, hidden, GemmAcct::Ffn)?;
        let mut gated = vec![0.0f32; seq * d];
        for t in 0..seq {
            for i in 0..d {
                let g = gelu_pytorch_tanh(gate_out[t * d + i]);
                let p = ple_tok[t * n_layers * d + li * d + i];
                gated[t * d + i] = g * p;
            }
        }
        let proj = self.wmm(&ple.proj, &gated, hidden, d, GemmAcct::Ffn)?;
        let nrm = self.norm(&proj, &ple.post_norm)?;
        for (a, b) in x.iter_mut().zip(nrm.iter()) {
            *a += *b;
        }
        Ok(())
    }

    fn apply_layer_scalar(x: &mut [f32], scale: f32) {
        if (scale - 1.0).abs() < 1e-8 {
            return;
        }
        for v in x {
            *v *= scale;
        }
    }

    fn apply_ffn(
        &self,
        layer: &LayerWeights,
        xn2: &[f32],
        hidden: usize,
    ) -> Result<Vec<f32>, EngineError> {
        match &layer.ffn {
            FfnWeights::Dense { gate, up, down } => {
                if gate.data.len() % hidden != 0 {
                    return Err(EngineError::ShapeMismatch(
                        "dense gate len not divisible by hidden".into(),
                    ));
                }
                let inter = gate.data.len() / hidden;
                if inter == 0
                    || up.data.len() != inter * hidden
                    || down.data.len() != hidden * inter
                {
                    return Err(EngineError::ShapeMismatch(
                        "dense FFN weight shape mismatch".into(),
                    ));
                }
                let g = self.wmm(gate, xn2, inter, hidden, GemmAcct::Ffn)?;
                let u = self.wmm(up, xn2, inter, hidden, GemmAcct::Ffn)?;
                let h = if self.use_geglu {
                    geglu(&g, &u)?
                } else {
                    swiglu(&g, &u)?
                };
                self.wmm(down, &h, hidden, inter, GemmAcct::Ffn)
            }
            FfnWeights::MoE {
                router,
                experts,
                top_k,
                use_sigmoid,
            } => {
                let n_exp = experts.len();
                let logits = self.wmm(router, xn2, n_exp, hidden, GemmAcct::Ffn)?;
                let (ids, weights) = moe_topk_route(&logits, *top_k, *use_sigmoid)?;
                let mut acc = vec![0.0f32; hidden];
                for (ei, &w) in ids.iter().zip(weights.iter()) {
                    let ex = &experts[*ei];
                    if ex.gate.data.len() % hidden != 0 {
                        return Err(EngineError::ShapeMismatch(
                            "expert gate len not divisible by hidden".into(),
                        ));
                    }
                    let inter = ex.gate.data.len() / hidden;
                    let g = self.wmm(&ex.gate, xn2, inter, hidden, GemmAcct::Ffn)?;
                    let u = self.wmm(&ex.up, xn2, inter, hidden, GemmAcct::Ffn)?;
                    let h = swiglu(&g, &u)?;
                    let down = self.wmm(&ex.down, &h, hidden, inter, GemmAcct::Ffn)?;
                    for i in 0..hidden {
                        acc[i] += w * down[i];
                    }
                }
                Ok(acc)
            }
        }
    }

    fn fresh_decode_state(&self) -> DecodeState {
        let hidden = self.conf.hidden_size;
        DecodeState {
            k_caches: (0..self.conf.num_layers).map(|_| Vec::new()).collect(),
            v_caches: (0..self.conf.num_layers).map(|_| Vec::new()).collect(),
            last_kv_src: HashMap::new(),
            conv_states: self
                .weights
                .layers
                .iter()
                .map(|layer| match &layer.op {
                    LayerOp::Conv(c) => {
                        let hist = c.kernel_size.saturating_sub(1);
                        Some(vec![0.0f32; hidden * hist])
                    }
                    LayerOp::Linear(d) => {
                        let conv_dim = d.n_k_heads * d.head_k * 2 + d.n_v_heads * d.head_v;
                        let hist = d.conv_k.saturating_sub(1);
                        Some(vec![0.0f32; conv_dim * hist])
                    }
                    LayerOp::Attn(_) => None,
                })
                .collect(),
            delta_states: self
                .weights
                .layers
                .iter()
                .map(|layer| match &layer.op {
                    LayerOp::Linear(d) => Some(vec![0.0f32; d.n_v_heads * d.head_k * d.head_v]),
                    _ => None,
                })
                .collect(),
            pos: 0,
        }
    }

    /// Full-sequence forward (allocates fresh caches). Used by tests / parity checks.
    #[cfg(test)]
    fn forward(&self, tokens: &[u32]) -> Result<Vec<f32>, EngineError> {
        let mut state = self.fresh_decode_state();
        let mut logits = Vec::new();
        for &tok in tokens {
            logits = self.forward_step_with(&mut state, tok)?;
        }
        Ok(logits)
    }

    fn forward_prompt(&mut self, toks: &[u32]) -> Result<Vec<f32>, EngineError> {
        let mut owned = self.decode.take().ok_or_else(|| {
            EngineError::InvalidParam("decode state missing; call generate/prefill first".into())
        })?;
        let logits = self.forward_prompt_with(&mut owned, toks);
        self.decode = Some(owned);
        logits
    }

    fn apply_rope_seq(
        x: &mut [f32],
        seq: usize,
        tok_dim: usize,
        head_dim: usize,
        pos0: usize,
        theta: f32,
        proportional: Option<f32>,
    ) -> Result<(), EngineError> {
        if x.len() != seq * tok_dim {
            return Err(EngineError::ShapeMismatch(
                "rope seq buffer length mismatch".into(),
            ));
        }
        for t in 0..seq {
            Self::apply_rope(
                &mut x[t * tok_dim..(t + 1) * tok_dim],
                head_dim,
                pos0 + t,
                theta,
                proportional,
            )?;
        }
        Ok(())
    }

    fn forward_prompt_with(
        &self,
        state: &mut DecodeState,
        toks: &[u32],
    ) -> Result<Vec<f32>, EngineError> {
        if toks.is_empty() {
            return Err(EngineError::InvalidParam("empty prompt".into()));
        }
        let hidden = self.conf.hidden_size;
        let n_heads = self.conf.num_attention_heads;
        let n_kv = self.conf.num_kv_heads;
        let vocab = self.conf.vocab_size;
        let seq = toks.len();
        if hidden == 0 {
            return Err(EngineError::ShapeMismatch("hidden_size is 0".into()));
        }
        if self.weights.emb.data.len() < vocab.saturating_mul(hidden)
            || !self.weights.emb.data.len().is_multiple_of(hidden)
        {
            return Err(EngineError::ShapeMismatch(format!(
                "embedding length {} not compatible with vocab={vocab} hidden={hidden}",
                self.weights.emb.data.len()
            )));
        }
        let pos0 = state.pos;
        let mut x = vec![0.0f32; seq * hidden];
        for (t, &tok) in toks.iter().enumerate() {
            let tid = (tok as usize) % vocab;
            x[t * hidden..(t + 1) * hidden]
                .copy_from_slice(&self.weights.emb.data[tid * hidden..(tid + 1) * hidden]);
        }
        if self.embed_scale != 1.0 {
            for v in &mut x {
                *v *= self.embed_scale;
            }
        }
        let ple_tok = self.compute_ple_inputs(toks, &x)?;

        for (li, layer) in self.weights.layers.iter().enumerate() {
            let xn = self.norm(&x, &layer.attn_norm)?;
            match &layer.op {
                LayerOp::Attn(attn) => {
                    if attn.wq.data.len() % hidden != 0 {
                        return Err(EngineError::ShapeMismatch(
                            "attn q proj weight not divisible by hidden_size".into(),
                        ));
                    }
                    let q_dim = attn.wq.data.len() / hidden;
                    let head_dim = self.layer_head_dim(attn.kind, q_dim, n_heads)?;
                    if attn.wo.data.len() != hidden * q_dim {
                        return Err(EngineError::ShapeMismatch(
                            "attn output proj weight shape mismatch".into(),
                        ));
                    }
                    let mut q = self.wmm(&attn.wq, &xn, q_dim, hidden, GemmAcct::Attn)?;
                    if let Some(qn) = &attn.q_norm {
                        if qn.len() != head_dim {
                            return Err(EngineError::ShapeMismatch(format!(
                                "q_norm len {} != head_dim {head_dim}",
                                qn.len()
                            )));
                        }
                        q = rms_norm(&q, qn, 1e-6)?;
                    }
                    let (theta, proportional) = self.layer_rope_params(attn.kind);
                    Self::apply_rope_seq(
                        &mut q,
                        seq,
                        q_dim,
                        head_dim,
                        pos0,
                        theta,
                        proportional,
                    )?;

                    let (k_src, v_src) = if let (Some(wk), Some(wv)) = (&attn.wk, &attn.wv) {
                        if wk.data.len() % hidden != 0 || wv.data.len() % hidden != 0 {
                            return Err(EngineError::ShapeMismatch(
                                "attn kv proj weight not divisible by hidden_size".into(),
                            ));
                        }
                        let k_dim = wk.data.len() / hidden;
                        let v_dim = wv.data.len() / hidden;
                        if k_dim != n_kv * head_dim || v_dim != n_kv * head_dim {
                            return Err(EngineError::ShapeMismatch(format!(
                                "kv dims {k_dim}/{v_dim} != n_kv*head_dim {}",
                                n_kv * head_dim
                            )));
                        }
                        let mut k = self.wmm(wk, &xn, k_dim, hidden, GemmAcct::Attn)?;
                        let mut v = self.wmm(wv, &xn, v_dim, hidden, GemmAcct::Attn)?;
                        if let Some(kn) = &attn.k_norm {
                            if kn.len() != head_dim {
                                return Err(EngineError::ShapeMismatch(format!(
                                    "k_norm len {} != head_dim {head_dim}",
                                    kn.len()
                                )));
                            }
                            k = rms_norm(&k, kn, 1e-6)?;
                        }
                        Self::apply_rope_seq(
                            &mut k,
                            seq,
                            k_dim,
                            head_dim,
                            pos0,
                            theta,
                            proportional,
                        )?;
                        v = self.apply_v_norm(v, attn.v_norm.as_deref(), head_dim)?;
                        state.k_caches[li] = k;
                        state.v_caches[li] = v;
                        state.last_kv_src.insert(attn.kind, li);
                        (li, li)
                    } else {
                        let src = state.last_kv_src.get(&attn.kind).copied().ok_or_else(|| {
                            EngineError::Format(format!(
                                "KV-consumer layer {li} has no producer of kind {:?}",
                                attn.kind
                            ))
                        })?;
                        (src, src)
                    };
                    let attn_out = attention_causal_with_scale(
                        &q,
                        &state.k_caches[k_src],
                        &state.v_caches[v_src],
                        n_heads,
                        n_kv,
                        head_dim,
                        self.attn_scale(head_dim),
                        self.attn_window(attn.kind),
                    )?;
                    let ao = self.wmm(&attn.wo, &attn_out, hidden, q_dim, GemmAcct::Attn)?;
                    self.add_normed_residual(&mut x, &ao, layer.post_attn_norm.as_deref())?;
                }
                LayerOp::Conv(_) | LayerOp::Linear(_) => {
                    return Err(EngineError::Unsupported(
                        "batched prefill is only implemented for attention+dense FFN layers"
                            .into(),
                    ));
                }
            }
            let xn2 = self.norm(&x, &layer.ffn_norm)?;
            let down = self.apply_ffn(layer, &xn2, hidden)?;
            self.add_normed_residual(&mut x, &down, layer.post_ffn_norm.as_deref())?;
            self.apply_ple(&mut x, layer, li, ple_tok.as_deref(), hidden)?;
            Self::apply_layer_scalar(&mut x, layer.layer_scalar);
        }
        state.pos = pos0 + seq;
        let last = &x[(seq - 1) * hidden..seq * hidden];
        let xn = self.norm(last, &self.weights.output_norm)?;
        if !self.weights.output.data.len().is_multiple_of(hidden) {
            return Err(EngineError::ShapeMismatch(format!(
                "lm_head len {} not divisible by hidden {hidden}",
                self.weights.output.data.len()
            )));
        }
        let out_rows = self.weights.output.data.len() / hidden;
        let logits = self.wmm(
            &self.weights.output,
            &xn,
            out_rows,
            hidden,
            GemmAcct::LmHead,
        )?;
        Ok(self.softcap_logits(logits))
    }

    fn forward_step(&mut self, tok: u32) -> Result<Vec<f32>, EngineError> {
        let mut owned = self.decode.take().ok_or_else(|| {
            EngineError::InvalidParam("decode state missing; call generate/prefill first".into())
        })?;
        let logits = self.forward_step_with(&mut owned, tok);
        self.decode = Some(owned);
        logits
    }

    fn forward_step_with(
        &self,
        state: &mut DecodeState,
        tok: u32,
    ) -> Result<Vec<f32>, EngineError> {
        let hidden = self.conf.hidden_size;
        let n_heads = self.conf.num_attention_heads;
        let n_kv = self.conf.num_kv_heads;
        let vocab = self.conf.vocab_size;
        if hidden == 0 {
            return Err(EngineError::ShapeMismatch("hidden_size is 0".into()));
        }
        if self.weights.emb.data.len() < vocab.saturating_mul(hidden)
            || !self.weights.emb.data.len().is_multiple_of(hidden)
        {
            return Err(EngineError::ShapeMismatch(format!(
                "embedding length {} not compatible with vocab={vocab} hidden={hidden}",
                self.weights.emb.data.len()
            )));
        }
        let pos = state.pos;
        let tid = (tok as usize) % vocab;
        let mut x = vec![0.0f32; hidden];
        x.copy_from_slice(&self.weights.emb.data[tid * hidden..(tid + 1) * hidden]);
        if self.embed_scale != 1.0 {
            for v in &mut x {
                *v *= self.embed_scale;
            }
        }
        let ple_tok = self.compute_ple_inputs(&[tok], &x)?;

        for (li, layer) in self.weights.layers.iter().enumerate() {
            let xn = self.norm(&x, &layer.attn_norm)?;
            match &layer.op {
                LayerOp::Attn(attn) => {
                    if attn.wq.data.len() % hidden != 0 {
                        return Err(EngineError::ShapeMismatch(
                            "attn q proj weight not divisible by hidden_size".into(),
                        ));
                    }
                    let q_dim = attn.wq.data.len() / hidden;
                    let head_dim = self.layer_head_dim(attn.kind, q_dim, n_heads)?;
                    if attn.wo.data.len() != hidden * q_dim {
                        return Err(EngineError::ShapeMismatch(
                            "attn output proj weight shape mismatch".into(),
                        ));
                    }
                    let mut q = self.wmm(&attn.wq, &xn, q_dim, hidden, GemmAcct::Attn)?;
                    if let Some(qn) = &attn.q_norm {
                        if qn.len() != head_dim {
                            return Err(EngineError::ShapeMismatch(format!(
                                "q_norm len {} != head_dim {head_dim}",
                                qn.len()
                            )));
                        }
                        q = rms_norm(&q, qn, 1e-6)?;
                    }
                    let (theta, proportional) = self.layer_rope_params(attn.kind);
                    Self::apply_rope(&mut q, head_dim, pos, theta, proportional)?;

                    let (k_src, v_src) = if let (Some(wk), Some(wv)) = (&attn.wk, &attn.wv) {
                        if wk.data.len() % hidden != 0 || wv.data.len() % hidden != 0 {
                            return Err(EngineError::ShapeMismatch(
                                "attn kv proj weight not divisible by hidden_size".into(),
                            ));
                        }
                        let k_dim = wk.data.len() / hidden;
                        let v_dim = wv.data.len() / hidden;
                        if k_dim != n_kv * head_dim || v_dim != n_kv * head_dim {
                            return Err(EngineError::ShapeMismatch(format!(
                                "kv dims {k_dim}/{v_dim} != n_kv*head_dim {}",
                                n_kv * head_dim
                            )));
                        }
                        let mut k = self.wmm(wk, &xn, k_dim, hidden, GemmAcct::Attn)?;
                        let mut v = self.wmm(wv, &xn, v_dim, hidden, GemmAcct::Attn)?;
                        if let Some(kn) = &attn.k_norm {
                            if kn.len() != head_dim {
                                return Err(EngineError::ShapeMismatch(format!(
                                    "k_norm len {} != head_dim {head_dim}",
                                    kn.len()
                                )));
                            }
                            k = rms_norm(&k, kn, 1e-6)?;
                        }
                        Self::apply_rope(&mut k, head_dim, pos, theta, proportional)?;
                        v = self.apply_v_norm(v, attn.v_norm.as_deref(), head_dim)?;
                        state.k_caches[li].extend_from_slice(&k);
                        state.v_caches[li].extend_from_slice(&v);
                        state.last_kv_src.insert(attn.kind, li);
                        (li, li)
                    } else {
                        let src = state.last_kv_src.get(&attn.kind).copied().ok_or_else(|| {
                            EngineError::Format(format!(
                                "KV-consumer layer {li} has no producer of kind {:?}",
                                attn.kind
                            ))
                        })?;
                        (src, src)
                    };
                    let kv_dim = n_kv * head_dim;
                    let (k_view, v_view) = kv_sliding_view(
                        &state.k_caches[k_src],
                        &state.v_caches[v_src],
                        kv_dim,
                        self.attn_window(attn.kind),
                    )?;
                    let attn_out = attention_with_scale(
                        &q,
                        k_view,
                        v_view,
                        n_heads,
                        n_kv,
                        head_dim,
                        self.attn_scale(head_dim),
                    )?;
                    let ao = self.wmm(&attn.wo, &attn_out, hidden, q_dim, GemmAcct::Attn)?;
                    self.add_normed_residual(&mut x, &ao, layer.post_attn_norm.as_deref())?;
                }
                LayerOp::Conv(conv) => {
                    let bcx = self.wmm(&conv.in_proj, &xn, 3 * hidden, hidden, GemmAcct::Attn)?;
                    let mut bx = vec![0.0f32; hidden];
                    let mut c_gate = vec![0.0f32; hidden];
                    for i in 0..hidden {
                        let b = bcx[i];
                        let c = bcx[hidden + i];
                        let xx = bcx[2 * hidden + i];
                        bx[i] = b * xx;
                        c_gate[i] = c;
                    }
                    let cstate = state.conv_states[li]
                        .as_mut()
                        .ok_or_else(|| EngineError::ShapeMismatch("missing conv state".into()))?;
                    let conv_y =
                        short_conv_step(&bx, &conv.kernel, cstate, hidden, conv.kernel_size)?;
                    let mut y = vec![0.0f32; hidden];
                    for i in 0..hidden {
                        y[i] = c_gate[i] * conv_y[i];
                    }
                    let ao = self.wmm(&conv.out_proj, &y, hidden, hidden, GemmAcct::Attn)?;
                    self.add_normed_residual(&mut x, &ao, layer.post_attn_norm.as_deref())?;
                }
                LayerOp::Linear(dn) => {
                    let key_dim = dn.n_k_heads * dn.head_k;
                    let value_dim = dn.n_v_heads * dn.head_v;
                    let qkvz_out = 2 * key_dim + 2 * value_dim;
                    let mixed = self.wmm(&dn.qkvz, &xn, qkvz_out, hidden, GemmAcct::Attn)?;
                    let mut q = mixed[0..key_dim].to_vec();
                    let mut k = mixed[key_dim..2 * key_dim].to_vec();
                    let mut v = mixed[2 * key_dim..2 * key_dim + value_dim].to_vec();
                    let z = mixed[2 * key_dim + value_dim..].to_vec();
                    let mut qkv = Vec::with_capacity(key_dim * 2 + value_dim);
                    qkv.extend_from_slice(&q);
                    qkv.extend_from_slice(&k);
                    qkv.extend_from_slice(&v);
                    let conv_dim = qkv.len();
                    let cstate = state.conv_states[li].as_mut().ok_or_else(|| {
                        EngineError::ShapeMismatch("missing delta conv state".into())
                    })?;
                    let mut mixed_c = short_conv_step(&qkv, &dn.conv, cstate, conv_dim, dn.conv_k)?;
                    silu_vec(&mut mixed_c);
                    q.copy_from_slice(&mixed_c[0..key_dim]);
                    k.copy_from_slice(&mixed_c[key_dim..2 * key_dim]);
                    v.copy_from_slice(&mixed_c[2 * key_dim..]);
                    let ba = self.wmm(&dn.ba, &xn, 2 * dn.n_v_heads, hidden, GemmAcct::Attn)?;
                    let mut beta = vec![0.0f32; dn.n_v_heads];
                    let mut g = vec![0.0f32; dn.n_v_heads];
                    for h in 0..dn.n_v_heads {
                        beta[h] = 1.0 / (1.0 + (-ba[h]).exp());
                        let alpha =
                            -dn.a_log[h].exp() * softplus(ba[dn.n_v_heads + h] + dn.dt_bias[h]);
                        g[h] = alpha.exp();
                    }
                    if dn.n_v_heads != dn.n_k_heads {
                        return Err(EngineError::Unsupported(
                            "DeltaNet GQA (n_v != n_k) not implemented".into(),
                        ));
                    }
                    let s = state.delta_states[li].as_mut().ok_or_else(|| {
                        EngineError::ShapeMismatch("missing delta recurrent state".into())
                    })?;
                    let mut core = gated_delta_step(GatedDeltaStep {
                        q: &q,
                        k: &k,
                        v: &v,
                        g: &g,
                        beta: &beta,
                        state: s,
                        n_heads: dn.n_v_heads,
                        dk: dn.head_k,
                        dv: dn.head_v,
                    })?;
                    // RMSNormGated approx: rms(core) * silu(z)
                    let ones = vec![1.0f32; dn.head_v];
                    core = rms_norm(&core, &ones, 1e-6)?;
                    let mut z_act = z;
                    silu_vec(&mut z_act);
                    for i in 0..core.len() {
                        core[i] *= z_act[i];
                    }
                    let ao = self.wmm(&dn.out_proj, &core, hidden, value_dim, GemmAcct::Attn)?;
                    self.add_normed_residual(&mut x, &ao, layer.post_attn_norm.as_deref())?;
                }
            }
            let xn2 = self.norm(&x, &layer.ffn_norm)?;
            let down = self.apply_ffn(layer, &xn2, hidden)?;
            self.add_normed_residual(&mut x, &down, layer.post_ffn_norm.as_deref())?;
            self.apply_ple(&mut x, layer, li, ple_tok.as_deref(), hidden)?;
            Self::apply_layer_scalar(&mut x, layer.layer_scalar);
        }
        state.pos += 1;
        let xn = self.norm(&x, &self.weights.output_norm)?;
        if !self.weights.output.data.len().is_multiple_of(hidden) {
            return Err(EngineError::ShapeMismatch(format!(
                "lm_head len {} not divisible by hidden {hidden}",
                self.weights.output.data.len()
            )));
        }
        let out_rows = self.weights.output.data.len() / hidden;
        let logits = self.wmm(&self.weights.output, &xn, out_rows, hidden, GemmAcct::LmHead)?;
        Ok(self.softcap_logits(logits))
    }

    fn softcap_logits(&self, mut logits: Vec<f32>) -> Vec<f32> {
        if let Some(cap) = self.final_logit_softcap.filter(|c| *c > 0.0) {
            for x in &mut logits {
                *x = (*x / cap).tanh() * cap;
            }
        }
        logits
    }
}

fn argmax(v: &[f32]) -> u32 {
    let mut best = 0usize;
    let mut best_v = f32::NEG_INFINITY;
    for (i, &x) in v.iter().enumerate() {
        if x > best_v {
            best_v = x;
            best = i;
        }
    }
    best as u32
}

/// Optional confidence heuristic for hybrid: mean max-softmax over last logits proxy.
pub fn confidence_from_logits(logits: &[f32]) -> f32 {
    if logits.is_empty() {
        return 0.0;
    }
    let m = logits.iter().cloned().fold(f32::NEG_INFINITY, f32::max);
    let mut sum = 0.0f32;
    let mut maxp = 0.0f32;
    for &x in logits {
        let e = (x - m).exp();
        sum += e;
        if e > maxp {
            maxp = e;
        }
    }
    if sum > 0.0 {
        maxp / sum
    } else {
        0.0
    }
}

#[allow(dead_code)]
pub fn cache_shapes_ok(cache: &HashMap<usize, Vec<f32>>, kv_dim: usize) -> bool {
    cache.values().all(|v| v.len().is_multiple_of(kv_dim))
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::family::{arch_class_representatives, graph_hook, lookup_family, require_stage_b};
    use crate::fixture::write_tiny_q4_bundle;
    use aria_kernel::{resolve_compute, ComputePref};
    use serde_json::{json, Value};

    #[test]
    fn gemma4_fills_hub_bundle_missing_geometry_fields() {
        let dir = tempfile::tempdir().unwrap();
        write_tiny_q4_bundle(dir.path()).unwrap();
        let cfg_path = dir.path().join("config.json");
        let raw = std::fs::read_to_string(&cfg_path).unwrap();
        let mut cfg: Value = serde_json::from_str(&raw).unwrap();
        // Hub q4 still ships base geometry only (no Gemma-4 RoPE / window fields).
        let model = cfg["model"].as_object_mut().unwrap();
        for key in [
            "layer_types",
            "sliding_window",
            "partial_rotary_factor",
            "global_head_dim",
            "head_dim",
        ] {
            model.remove(key);
        }
        std::fs::write(&cfg_path, cfg.to_string()).unwrap();
        let s = SessionBuilder::new()
            .model(dir.path())
            .family("gemma/gemma-4-e2b-it")
            .build()
            .unwrap();
        assert_eq!(s.config().sliding_window, Some(512));
        assert_eq!(s.config().partial_rotary_factor, Some(0.25));
        assert!(s.config().head_dim.unwrap_or(0) > 0);
        assert!(s.config().global_head_dim.unwrap_or(0) > 0);
        assert_eq!(
            s.config().layer_types.as_ref().map(|t| t.len()),
            Some(s.config().num_layers)
        );
    }

    #[test]
    fn generate_tokens() {
        let dir = tempfile::tempdir().unwrap();
        write_tiny_q4_bundle(dir.path()).unwrap();
        let mut s = SessionBuilder::new()
            .model(dir.path())
            .family("gemma/gemma-4-e2b-it")
            .build()
            .unwrap();
        assert_eq!(s.config().sliding_window, Some(512));
        assert_eq!(s.config().partial_rotary_factor, Some(0.25));
        assert_eq!(s.config().head_dim, Some(16));
        assert_eq!(s.config().global_head_dim, Some(16));
        assert_eq!(
            s.config().layer_types,
            Some(vec!["full_attention".into(), "full_attention".into()])
        );
        assert_eq!(
            s.layer_rope_params(AttnKind::Full),
            (1_000_000.0, Some(0.25))
        );
        assert_eq!(s.layer_rope_params(AttnKind::Sliding), (10_000.0, None));
        assert_eq!(s.attn_window(AttnKind::Sliding), Some(512));
        assert_eq!(s.attn_window(AttnKind::Full), None);
        let prompt = s.encode_text("hi");
        let gen = s
            .generate(
                &prompt,
                &GenerateOpts {
                    max_tokens: 4,
                    temperature: 0.0,
                },
            )
            .unwrap();
        assert!(!gen.tokens.is_empty());
        assert!(!gen.text.is_empty());
    }

    #[test]
    fn materialize_accepts_hf_tensor_names() {
        // Minimal HF-named raw bundle matching Qwen-style paths.
        let dir = tempfile::tempdir().unwrap();
        let hidden = 8usize;
        let layers = 1usize;
        let inter = 16usize;
        let vocab = 16usize;
        let n_heads = 2usize;
        let n_kv = 1usize;
        let head_dim = 4usize; // q_dim = 8, k_dim = 4
        let q_dim = n_heads * head_dim;
        let k_dim = n_kv * head_dim;

        let mut tensors = serde_json::Map::new();
        let mut bin = Vec::new();
        let mut add_raw = |name: &str, shape: Vec<usize>, data: &[f32]| {
            let offset = bin.len();
            for &v in data {
                bin.extend_from_slice(&v.to_le_bytes());
            }
            let nbytes = data.len() * 4;
            let mut meta = serde_json::Map::new();
            meta.insert("kind".into(), json!("raw"));
            meta.insert("dtype".into(), json!("f32"));
            meta.insert("shape".into(), json!(shape));
            meta.insert("offsets".into(), json!({ "data": [offset, nbytes] }));
            tensors.insert(name.to_string(), Value::Object(meta));
        };
        let emb: Vec<f32> = (0..vocab * hidden).map(|i| i as f32 * 0.01).collect();
        add_raw("model.embed_tokens.weight", vec![vocab, hidden], &emb);
        let n1 = vec![1.0f32; hidden];
        add_raw("model.layers.0.input_layernorm.weight", vec![hidden], &n1);
        add_raw(
            "model.layers.0.post_attention_layernorm.weight",
            vec![hidden],
            &n1,
        );
        let wq = vec![0.01f32; q_dim * hidden];
        let wk = vec![0.01f32; k_dim * hidden];
        let wv = vec![0.01f32; k_dim * hidden];
        let wo = vec![0.01f32; hidden * q_dim];
        add_raw(
            "model.layers.0.self_attn.q_proj.weight",
            vec![q_dim, hidden],
            &wq,
        );
        add_raw(
            "model.layers.0.self_attn.k_proj.weight",
            vec![k_dim, hidden],
            &wk,
        );
        add_raw(
            "model.layers.0.self_attn.v_proj.weight",
            vec![k_dim, hidden],
            &wv,
        );
        add_raw(
            "model.layers.0.self_attn.o_proj.weight",
            vec![hidden, q_dim],
            &wo,
        );
        let g = vec![0.01f32; inter * hidden];
        let d = vec![0.01f32; hidden * inter];
        add_raw(
            "model.layers.0.mlp.gate_proj.weight",
            vec![inter, hidden],
            &g,
        );
        add_raw("model.layers.0.mlp.up_proj.weight", vec![inter, hidden], &g);
        add_raw(
            "model.layers.0.mlp.down_proj.weight",
            vec![hidden, inter],
            &d,
        );
        add_raw("model.norm.weight", vec![hidden], &n1);
        add_raw("lm_head.weight", vec![vocab, hidden], &emb);

        let cfg = json!({
            "format": "aria-quant-bundle",
            "format_version": 2,
            "quantization": "test",
            "group_size_default": 32,
            "hadamard_seed": 0,
            "model": {
                "hidden_size": hidden,
                "num_layers": layers,
                "num_attention_heads": n_heads,
                "num_kv_heads": n_kv,
                "intermediate_size": inter,
                "vocab_size": vocab,
                "context_length": 32,
                "rope_theta": 10000.0
            },
            "tensors": tensors
        });
        std::fs::write(dir.path().join("config.json"), cfg.to_string()).unwrap();
        std::fs::write(dir.path().join("weight.bin"), &bin).unwrap();

        let mut s = SessionBuilder::new()
            .model(dir.path())
            .family("qwen/qwen3-0.6b")
            .build()
            .unwrap();
        let gen = s
            .generate(
                &[1, 2],
                &GenerateOpts {
                    max_tokens: 2,
                    temperature: 0.0,
                },
            )
            .unwrap();
        assert_eq!(gen.tokens.len(), 2);
    }

    #[test]
    fn materialize_accepts_language_model_prefix_and_pre_ffn_norm() {
        // Gemma-4 / Gemma-3n VL-style HF paths.
        let dir = tempfile::tempdir().unwrap();
        let hidden = 8usize;
        let layers = 1usize;
        let inter = 16usize;
        let vocab = 16usize;
        let n_heads = 2usize;
        let n_kv = 1usize;
        let head_dim = 4usize;
        let q_dim = n_heads * head_dim;
        let k_dim = n_kv * head_dim;

        let mut tensors = serde_json::Map::new();
        let mut bin = Vec::new();
        let mut add_raw = |name: &str, shape: Vec<usize>, data: &[f32]| {
            let offset = bin.len();
            for &v in data {
                bin.extend_from_slice(&v.to_le_bytes());
            }
            let nbytes = data.len() * 4;
            let mut meta = serde_json::Map::new();
            meta.insert("kind".into(), json!("raw"));
            meta.insert("dtype".into(), json!("f32"));
            meta.insert("shape".into(), json!(shape));
            meta.insert("offsets".into(), json!({ "data": [offset, nbytes] }));
            tensors.insert(name.to_string(), Value::Object(meta));
        };
        let emb: Vec<f32> = (0..vocab * hidden).map(|i| i as f32 * 0.01).collect();
        let p = "model.language_model";
        add_raw(
            &format!("{p}.embed_tokens.weight"),
            vec![vocab, hidden],
            &emb,
        );
        let n1 = vec![1.0f32; hidden];
        add_raw(
            &format!("{p}.layers.0.input_layernorm.weight"),
            vec![hidden],
            &n1,
        );
        add_raw(
            &format!("{p}.layers.0.pre_feedforward_layernorm.weight"),
            vec![hidden],
            &n1,
        );
        let wq = vec![0.01f32; q_dim * hidden];
        let wk = vec![0.01f32; k_dim * hidden];
        let wv = vec![0.01f32; k_dim * hidden];
        let wo = vec![0.01f32; hidden * q_dim];
        add_raw(
            &format!("{p}.layers.0.self_attn.q_proj.weight"),
            vec![q_dim, hidden],
            &wq,
        );
        add_raw(
            &format!("{p}.layers.0.self_attn.k_proj.weight"),
            vec![k_dim, hidden],
            &wk,
        );
        add_raw(
            &format!("{p}.layers.0.self_attn.v_proj.weight"),
            vec![k_dim, hidden],
            &wv,
        );
        add_raw(
            &format!("{p}.layers.0.self_attn.o_proj.weight"),
            vec![hidden, q_dim],
            &wo,
        );
        let g = vec![0.01f32; inter * hidden];
        let d = vec![0.01f32; hidden * inter];
        add_raw(
            &format!("{p}.layers.0.mlp.gate_proj.weight"),
            vec![inter, hidden],
            &g,
        );
        add_raw(
            &format!("{p}.layers.0.mlp.up_proj.weight"),
            vec![inter, hidden],
            &g,
        );
        add_raw(
            &format!("{p}.layers.0.mlp.down_proj.weight"),
            vec![hidden, inter],
            &d,
        );
        add_raw(&format!("{p}.norm.weight"), vec![hidden], &n1);
        add_raw("lm_head.weight", vec![vocab, hidden], &emb);

        let cfg = json!({
            "format": "aria-quant-bundle",
            "format_version": 2,
            "quantization": "test",
            "group_size_default": 32,
            "hadamard_seed": 0,
            "model": {
                "hidden_size": hidden,
                "num_layers": layers,
                "num_attention_heads": n_heads,
                "num_kv_heads": n_kv,
                "intermediate_size": inter,
                "vocab_size": vocab,
                "context_length": 32,
                "rope_theta": 10000.0,
                "head_dim": head_dim,
                "global_head_dim": head_dim,
                "sliding_window": 512,
                "partial_rotary_factor": 0.25,
                "layer_types": ["full_attention"]
            },
            "tensors": tensors
        });
        std::fs::write(dir.path().join("config.json"), cfg.to_string()).unwrap();
        std::fs::write(dir.path().join("weight.bin"), &bin).unwrap();

        let mut s = SessionBuilder::new()
            .model(dir.path())
            .family("gemma/gemma-4-e2b-it")
            .build()
            .unwrap();
        let gen = s
            .generate(
                &[1, 2],
                &GenerateOpts {
                    max_tokens: 2,
                    temperature: 0.0,
                },
            )
            .unwrap();
        assert_eq!(gen.tokens.len(), 2);
    }

    #[test]
    fn gemma4_style_double_wide_mlp_and_shared_kv() {
        // Config intermediate_size stays at the narrow width; layer 1 is 2× (KV-shared)
        // and omits k/v projections — reuses producer KV cache (num_kv_shared_layers=1).
        let dir = tempfile::tempdir().unwrap();
        let hidden = 8usize;
        let layers = 2usize;
        let inter = 16usize;
        let inter_wide = 32usize;
        let vocab = 16usize;
        let n_heads = 2usize;
        let n_kv = 1usize;
        let head_dim = 4usize;
        let q_dim = n_heads * head_dim;
        let k_dim = n_kv * head_dim;
        let p = "model.language_model";

        let mut tensors = serde_json::Map::new();
        let mut bin = Vec::new();
        let mut add_raw = |name: &str, shape: Vec<usize>, data: &[f32]| {
            let offset = bin.len();
            for &v in data {
                bin.extend_from_slice(&v.to_le_bytes());
            }
            let nbytes = data.len() * 4;
            let mut meta = serde_json::Map::new();
            meta.insert("kind".into(), json!("raw"));
            meta.insert("dtype".into(), json!("f32"));
            meta.insert("shape".into(), json!(shape));
            meta.insert("offsets".into(), json!({ "data": [offset, nbytes] }));
            tensors.insert(name.to_string(), Value::Object(meta));
        };
        let emb: Vec<f32> = (0..vocab * hidden).map(|i| i as f32 * 0.01).collect();
        add_raw(
            &format!("{p}.embed_tokens.weight"),
            vec![vocab, hidden],
            &emb,
        );
        let n1 = vec![1.0f32; hidden];
        let wq = vec![0.01f32; q_dim * hidden];
        let wk = vec![0.01f32; k_dim * hidden];
        let wv = vec![0.01f32; k_dim * hidden];
        let wo = vec![0.01f32; hidden * q_dim];
        for li in 0..layers {
            let layer_inter = if li == 0 { inter } else { inter_wide };
            add_raw(
                &format!("{p}.layers.{li}.input_layernorm.weight"),
                vec![hidden],
                &n1,
            );
            add_raw(
                &format!("{p}.layers.{li}.pre_feedforward_layernorm.weight"),
                vec![hidden],
                &n1,
            );
            add_raw(
                &format!("{p}.layers.{li}.self_attn.q_proj.weight"),
                vec![q_dim, hidden],
                &wq,
            );
            if li == 0 {
                add_raw(
                    &format!("{p}.layers.{li}.self_attn.k_proj.weight"),
                    vec![k_dim, hidden],
                    &wk,
                );
                add_raw(
                    &format!("{p}.layers.{li}.self_attn.v_proj.weight"),
                    vec![k_dim, hidden],
                    &wv,
                );
            }
            add_raw(
                &format!("{p}.layers.{li}.self_attn.o_proj.weight"),
                vec![hidden, q_dim],
                &wo,
            );
            let g = vec![0.01f32; layer_inter * hidden];
            let d = vec![0.01f32; hidden * layer_inter];
            add_raw(
                &format!("{p}.layers.{li}.mlp.gate_proj.weight"),
                vec![layer_inter, hidden],
                &g,
            );
            add_raw(
                &format!("{p}.layers.{li}.mlp.up_proj.weight"),
                vec![layer_inter, hidden],
                &g,
            );
            add_raw(
                &format!("{p}.layers.{li}.mlp.down_proj.weight"),
                vec![hidden, layer_inter],
                &d,
            );
        }
        add_raw(&format!("{p}.norm.weight"), vec![hidden], &n1);
        add_raw("lm_head.weight", vec![vocab, hidden], &emb);

        let cfg = json!({
            "format": "aria-quant-bundle",
            "format_version": 2,
            "quantization": "test",
            "group_size_default": 32,
            "hadamard_seed": 0,
            "model": {
                "hidden_size": hidden,
                "num_layers": layers,
                "num_attention_heads": n_heads,
                "num_kv_heads": n_kv,
                "intermediate_size": inter,
                "vocab_size": vocab,
                "context_length": 32,
                "rope_theta": 10000.0,
                "num_kv_shared_layers": 1,
                "head_dim": head_dim,
                "global_head_dim": head_dim,
                "sliding_window": 512,
                "partial_rotary_factor": 0.25,
                "layer_types": ["full_attention", "full_attention"]
            },
            "tensors": tensors
        });
        std::fs::write(dir.path().join("config.json"), cfg.to_string()).unwrap();
        std::fs::write(dir.path().join("weight.bin"), &bin).unwrap();

        let mut s = SessionBuilder::new()
            .model(dir.path())
            .family("gemma/gemma-4-e2b-it")
            .build()
            .unwrap();
        let gen = s
            .generate(
                &[1, 2],
                &GenerateOpts {
                    max_tokens: 2,
                    temperature: 0.0,
                },
            )
            .unwrap();
        assert_eq!(gen.tokens.len(), 2);
    }

    #[test]
    fn stage_b_arch_classes_generate() {
        for (path, arch) in arch_class_representatives() {
            if matches!(arch, ArchClass::VL | ArchClass::VLA | ArchClass::TextMoE) {
                continue; // stage C / MoE gated separately
            }
            // Hybrid linear_attention families refuse dense Session until DeltaNet lands.
            if path.contains("qwen3.5") || path.contains("bonsai") {
                let dir = tempfile::tempdir().unwrap();
                write_tiny_q4_bundle(dir.path()).unwrap();
                let err = SessionBuilder::new()
                    .model(dir.path())
                    .family(*path)
                    .build()
                    .unwrap_err();
                assert!(
                    matches!(err, EngineError::Unsupported(_)),
                    "{path}: {err:?}"
                );
                continue;
            }
            assert!(require_stage_b(path).is_ok(), "{path}");
            let dir = tempfile::tempdir().unwrap();
            write_tiny_q4_bundle(dir.path()).unwrap();
            let mut s = SessionBuilder::new()
                .model(dir.path())
                .family(*path)
                .build()
                .unwrap();
            assert_eq!(s.arch(), *arch);
            assert!(!s.graph_hook_name().is_empty());
            let gen = s
                .generate(
                    &s.encode_text("ok"),
                    &GenerateOpts {
                        max_tokens: 2,
                        temperature: 0.0,
                    },
                )
                .unwrap();
            assert!(!gen.tokens.is_empty(), "{path}");
        }
    }

    #[test]
    fn stage_c_vl_vla_hooks() {
        let dir = tempfile::tempdir().unwrap();
        write_tiny_q4_bundle(dir.path()).unwrap();
        let s = SessionBuilder::new()
            .model(dir.path())
            .family("lfm/lfm2-vl-450m")
            .build()
            .unwrap();
        let rgb = vec![10u8; 3 * 4 * 4];
        let err = s.vision_prefix(&rgb, 4, 4).unwrap_err();
        assert!(matches!(err, EngineError::Unsupported(_)));

        let vla = SessionBuilder::new()
            .model(dir.path())
            .family("openvla/openvla-7b")
            .build()
            .unwrap();
        let err = vla.predict_action("move", 7).unwrap_err();
        assert!(matches!(err, EngineError::Unsupported(_)));
        let emb = vla.embed_text("hello").unwrap();
        assert_eq!(emb.len(), vla.config().hidden_size);
    }

    #[test]
    fn unknown_family() {
        let err = SessionBuilder::new()
            .model("/tmp")
            .family("no/such-model")
            .build()
            .unwrap_err();
        assert!(matches!(err, EngineError::UnsupportedFamily(_)));
    }

    #[test]
    fn greedy_deterministic() {
        let dir = tempfile::tempdir().unwrap();
        write_tiny_q4_bundle(dir.path()).unwrap();
        let mut s = SessionBuilder::new()
            .model(dir.path())
            .family("gemma/gemma-4-e2b-it")
            .build()
            .unwrap();
        let prompt = s.encode_text("hi");
        let opts = GenerateOpts {
            max_tokens: 3,
            temperature: 0.0,
        };
        let a = s.generate(&prompt, &opts).unwrap();
        let b = s.generate(&prompt, &opts).unwrap();
        assert_eq!(a.tokens, b.tokens);
        assert_eq!(a.tokens.len(), 3);
    }

    #[test]
    fn encode_chat_is_longer_than_raw_user_text() {
        let dir = tempfile::tempdir().unwrap();
        write_tiny_q4_bundle(dir.path()).unwrap();
        let s = SessionBuilder::new()
            .model(dir.path())
            .family("qwen/qwen3-0.6b")
            .build()
            .unwrap();
        let raw = s.encode_text("Hello");
        let chat = s.encode_chat(&[ChatTurn::new("user", "Hello")]);
        assert!(
            chat.len() > raw.len(),
            "chat template should wrap the user turn (raw={}, chat={})",
            raw.len(),
            chat.len()
        );
        assert!(
            (s.config().rope_theta - 1_000_000.0).abs() < 1.0,
            "Qwen3 must not keep Llama-default rope_theta=10000, got {}",
            s.config().rope_theta
        );
    }

    #[test]
    fn incremental_decode_matches_full_recompute() {
        let dir = tempfile::tempdir().unwrap();
        write_tiny_q4_bundle(dir.path()).unwrap();
        let mut s = SessionBuilder::new()
            .model(dir.path())
            .family("gemma/gemma-4-e2b-it")
            .build()
            .unwrap();
        let prompt = s.encode_text("hi");
        let max_tokens = 5usize;

        // Legacy path: re-run full forward over the growing prefix each step.
        let mut prefix = prompt.clone();
        if prefix.is_empty() {
            prefix.push(1);
        }
        let mut full_tokens = Vec::new();
        for _ in 0..max_tokens {
            let logits = s.forward(&prefix).unwrap();
            let next = argmax(&logits);
            full_tokens.push(next);
            prefix.push(next);
            if s.is_stop_id(next) {
                full_tokens.pop();
                break;
            }
        }

        let incr = s
            .generate(
                &prompt,
                &GenerateOpts {
                    max_tokens,
                    temperature: 0.0,
                },
            )
            .unwrap();
        assert_eq!(
            incr.tokens, full_tokens,
            "incremental decode must match full-recompute greedy tokens"
        );
    }

    #[test]
    fn profile_records_load_and_generate() {
        let dir = tempfile::tempdir().unwrap();
        write_tiny_q4_bundle(dir.path()).unwrap();
        let mut s = SessionBuilder::new()
            .model(dir.path())
            .family("gemma/gemma-4-e2b-it")
            .compute(ComputePref::Cpu)
            .profile(true)
            .build()
            .unwrap();
        assert!(s.compute_label().contains("cpu"));
        let load = s.last_profile().expect("load profile");
        assert!(!load.ci_fail);
        assert!(load.load.materialize_ms >= 0.0);
        s.generate(
            &s.encode_text("hi"),
            &GenerateOpts {
                max_tokens: 2,
                temperature: 0.0,
            },
        )
        .unwrap();
        let p = s.last_profile().expect("generate profile");
        let g = p.generate.as_ref().expect("generate timings");
        assert!(g.prefill_ms >= 0.0);
        assert!(g.decode_ms >= 0.0);
    }

    #[test]
    fn cuda_greedy_matches_cpu_if_available() {
        if resolve_compute(ComputePref::Cuda).is_err() {
            return;
        }
        let dir = tempfile::tempdir().unwrap();
        write_tiny_q4_bundle(dir.path()).unwrap();
        let prompt_text = "hi";
        let opts = GenerateOpts {
            max_tokens: 4,
            temperature: 0.0,
        };
        let mut cpu = SessionBuilder::new()
            .model(dir.path())
            .family("gemma/gemma-4-e2b-it")
            .compute(ComputePref::Cpu)
            .build()
            .unwrap();
        let mut gpu = SessionBuilder::new()
            .model(dir.path())
            .family("gemma/gemma-4-e2b-it")
            .compute(ComputePref::Cuda)
            .build()
            .unwrap();
        assert!(gpu.compute_label().contains("cuda"));
        let prompt = cpu.encode_text(prompt_text);
        let a = cpu.generate(&prompt, &opts).unwrap();
        let b = gpu.generate(&prompt, &opts).unwrap();
        assert_eq!(
            a.tokens, b.tokens,
            "CUDA greedy tokens must match CPU (tiny bundle)"
        );
    }

    #[test]
    fn max_tokens_zero_rejected() {
        let dir = tempfile::tempdir().unwrap();
        write_tiny_q4_bundle(dir.path()).unwrap();
        let mut s = SessionBuilder::new()
            .model(dir.path())
            .family("gemma/gemma-4-e2b-it")
            .build()
            .unwrap();
        let err = s
            .generate(
                &s.encode_text("x"),
                &GenerateOpts {
                    max_tokens: 0,
                    temperature: 0.0,
                },
            )
            .unwrap_err();
        assert!(matches!(err, EngineError::InvalidParam(_)));
    }

    #[test]
    fn moe_family_refuses_dense_stub() {
        let dir = tempfile::tempdir().unwrap();
        write_tiny_q4_bundle(dir.path()).unwrap();
        let err = SessionBuilder::new()
            .model(dir.path())
            .family("lfm/lfm2-8b-a1b")
            .build()
            .unwrap_err();
        assert!(matches!(err, EngineError::Unsupported(_)));
        assert_eq!(
            lookup_family("lfm/lfm2-8b-a1b").unwrap().arch,
            ArchClass::TextMoE
        );
        assert_eq!(graph_hook(ArchClass::TextMoE), "text_moe_decoder");
    }

    #[test]
    fn geometry_gates_conv_and_experts() {
        // layer_types=conv without conv.* weights → Format (not silent dense attn).
        let dir = tempfile::tempdir().unwrap();
        write_tiny_q4_bundle(dir.path()).unwrap();
        let cfg_path = dir.path().join("config.json");
        let mut cfg: Value =
            serde_json::from_str(&std::fs::read_to_string(&cfg_path).unwrap()).unwrap();
        cfg["model"]["layer_types"] = json!(["conv", "full_attention"]);
        std::fs::write(&cfg_path, cfg.to_string()).unwrap();
        let err = SessionBuilder::new()
            .model(dir.path())
            .family("lfm/lfm2-350m")
            .build()
            .unwrap_err();
        assert!(
            matches!(err, EngineError::Format(_)),
            "expected missing conv tensors, got {err:?}"
        );

        // linear_attention still hard-gated.
        let dir2 = tempfile::tempdir().unwrap();
        write_tiny_q4_bundle(dir2.path()).unwrap();
        let cfg_path = dir2.path().join("config.json");
        let mut cfg: Value =
            serde_json::from_str(&std::fs::read_to_string(&cfg_path).unwrap()).unwrap();
        cfg["model"]["layer_types"] = json!(["linear_attention", "full_attention"]);
        std::fs::write(&cfg_path, cfg.to_string()).unwrap();
        let err = SessionBuilder::new()
            .model(dir2.path())
            .family("gemma/gemma-3-270m-it")
            .build()
            .unwrap_err();
        assert!(
            matches!(err, EngineError::Format(_)),
            "expected missing DeltaNet tensors, got {err:?}"
        );
    }

    #[test]
    fn lfm_short_conv_and_attn_generate() {
        let dir = tempfile::tempdir().unwrap();
        let hidden = 8usize;
        let layers = 2usize;
        let inter = 16usize;
        let vocab = 16usize;
        let n_heads = 2usize;
        let n_kv = 1usize;
        let head_dim = 4usize;
        let q_dim = n_heads * head_dim;
        let k_dim = n_kv * head_dim;
        let kernel = 3usize;

        let mut tensors = serde_json::Map::new();
        let mut bin = Vec::new();
        let mut add_raw = |name: &str, shape: Vec<usize>, data: &[f32]| {
            let offset = bin.len();
            for &v in data {
                bin.extend_from_slice(&v.to_le_bytes());
            }
            let nbytes = data.len() * 4;
            let mut meta = serde_json::Map::new();
            meta.insert("kind".into(), json!("raw"));
            meta.insert("dtype".into(), json!("f32"));
            meta.insert("shape".into(), json!(shape));
            meta.insert("offsets".into(), json!({ "data": [offset, nbytes] }));
            tensors.insert(name.to_string(), Value::Object(meta));
        };
        let emb: Vec<f32> = (0..vocab * hidden).map(|i| i as f32 * 0.01).collect();
        add_raw("model.embed_tokens.weight", vec![vocab, hidden], &emb);
        let n1 = vec![1.0f32; hidden];
        // Layer 0: short-conv
        add_raw("model.layers.0.operator_norm.weight", vec![hidden], &n1);
        add_raw("model.layers.0.ffn_norm.weight", vec![hidden], &n1);
        let in_proj = vec![0.02f32; 3 * hidden * hidden];
        let out_proj = vec![0.02f32; hidden * hidden];
        let conv_w = vec![0.1f32; hidden * kernel];
        add_raw(
            "model.layers.0.conv.in_proj.weight",
            vec![3 * hidden, hidden],
            &in_proj,
        );
        add_raw(
            "model.layers.0.conv.out_proj.weight",
            vec![hidden, hidden],
            &out_proj,
        );
        add_raw(
            "model.layers.0.conv.conv.weight",
            vec![hidden, kernel],
            &conv_w,
        );
        let g = vec![0.02f32; inter * hidden];
        let d = vec![0.02f32; hidden * inter];
        add_raw(
            "model.layers.0.mlp.gate_proj.weight",
            vec![inter, hidden],
            &g,
        );
        add_raw("model.layers.0.mlp.up_proj.weight", vec![inter, hidden], &g);
        add_raw(
            "model.layers.0.mlp.down_proj.weight",
            vec![hidden, inter],
            &d,
        );
        // Layer 1: full attention
        add_raw("model.layers.1.operator_norm.weight", vec![hidden], &n1);
        add_raw(
            "model.layers.1.post_attention_layernorm.weight",
            vec![hidden],
            &n1,
        );
        let wq = vec![0.02f32; q_dim * hidden];
        let wk = vec![0.02f32; k_dim * hidden];
        let wv = vec![0.02f32; k_dim * hidden];
        let wo = vec![0.02f32; hidden * q_dim];
        add_raw(
            "model.layers.1.self_attn.q_proj.weight",
            vec![q_dim, hidden],
            &wq,
        );
        add_raw(
            "model.layers.1.self_attn.k_proj.weight",
            vec![k_dim, hidden],
            &wk,
        );
        add_raw(
            "model.layers.1.self_attn.v_proj.weight",
            vec![k_dim, hidden],
            &wv,
        );
        add_raw(
            "model.layers.1.self_attn.o_proj.weight",
            vec![hidden, q_dim],
            &wo,
        );
        add_raw(
            "model.layers.1.mlp.gate_proj.weight",
            vec![inter, hidden],
            &g,
        );
        add_raw("model.layers.1.mlp.up_proj.weight", vec![inter, hidden], &g);
        add_raw(
            "model.layers.1.mlp.down_proj.weight",
            vec![hidden, inter],
            &d,
        );
        add_raw("model.norm.weight", vec![hidden], &n1);
        add_raw("lm_head.weight", vec![vocab, hidden], &emb);

        let cfg = json!({
            "format": "aria-quant-bundle",
            "format_version": 2,
            "quantization": "test",
            "group_size_default": 32,
            "hadamard_seed": 0,
            "model": {
                "hidden_size": hidden,
                "num_layers": layers,
                "num_attention_heads": n_heads,
                "num_kv_heads": n_kv,
                "head_dim": head_dim,
                "intermediate_size": inter,
                "vocab_size": vocab,
                "context_length": 32,
                "rope_theta": 10000.0,
                "conv_l_cache": kernel,
                "layer_types": ["conv", "full_attention"]
            },
            "tensors": tensors
        });
        std::fs::write(dir.path().join("config.json"), cfg.to_string()).unwrap();
        std::fs::write(dir.path().join("weight.bin"), &bin).unwrap();

        let mut s = SessionBuilder::new()
            .model(dir.path())
            .family("lfm/lfm2-350m")
            .build()
            .unwrap();
        let gen = s
            .generate(
                &[1, 2, 3],
                &GenerateOpts {
                    max_tokens: 2,
                    temperature: 0.0,
                },
            )
            .unwrap();
        assert_eq!(gen.tokens.len(), 2);
    }

    #[test]
    fn moe_topk_experts_generate() {
        let dir = tempfile::tempdir().unwrap();
        let hidden = 8usize;
        let layers = 1usize;
        let inter = 16usize;
        let vocab = 16usize;
        let n_heads = 2usize;
        let n_kv = 1usize;
        let head_dim = 4usize;
        let q_dim = n_heads * head_dim;
        let k_dim = n_kv * head_dim;
        let n_experts = 4usize;
        let top_k = 2usize;

        let mut tensors = serde_json::Map::new();
        let mut bin = Vec::new();
        let mut add_raw = |name: &str, shape: Vec<usize>, data: &[f32]| {
            let offset = bin.len();
            for &v in data {
                bin.extend_from_slice(&v.to_le_bytes());
            }
            let nbytes = data.len() * 4;
            let mut meta = serde_json::Map::new();
            meta.insert("kind".into(), json!("raw"));
            meta.insert("dtype".into(), json!("f32"));
            meta.insert("shape".into(), json!(shape));
            meta.insert("offsets".into(), json!({ "data": [offset, nbytes] }));
            tensors.insert(name.to_string(), Value::Object(meta));
        };
        let emb: Vec<f32> = (0..vocab * hidden).map(|i| i as f32 * 0.01).collect();
        add_raw("model.embed_tokens.weight", vec![vocab, hidden], &emb);
        let n1 = vec![1.0f32; hidden];
        add_raw("model.layers.0.input_layernorm.weight", vec![hidden], &n1);
        add_raw(
            "model.layers.0.post_attention_layernorm.weight",
            vec![hidden],
            &n1,
        );
        let wq = vec![0.02f32; q_dim * hidden];
        let wk = vec![0.02f32; k_dim * hidden];
        let wv = vec![0.02f32; k_dim * hidden];
        let wo = vec![0.02f32; hidden * q_dim];
        add_raw(
            "model.layers.0.self_attn.q_proj.weight",
            vec![q_dim, hidden],
            &wq,
        );
        add_raw(
            "model.layers.0.self_attn.k_proj.weight",
            vec![k_dim, hidden],
            &wk,
        );
        add_raw(
            "model.layers.0.self_attn.v_proj.weight",
            vec![k_dim, hidden],
            &wv,
        );
        add_raw(
            "model.layers.0.self_attn.o_proj.weight",
            vec![hidden, q_dim],
            &wo,
        );
        let router: Vec<f32> = (0..n_experts * hidden)
            .map(|i| ((i % n_experts) as f32) * 0.1)
            .collect();
        add_raw(
            "model.layers.0.block_sparse_moe.gate.weight",
            vec![n_experts, hidden],
            &router,
        );
        let g = vec![0.02f32; inter * hidden];
        let d = vec![0.02f32; hidden * inter];
        for e in 0..n_experts {
            add_raw(
                &format!("model.layers.0.block_sparse_moe.experts.{e}.w1.weight"),
                vec![inter, hidden],
                &g,
            );
            add_raw(
                &format!("model.layers.0.block_sparse_moe.experts.{e}.w3.weight"),
                vec![inter, hidden],
                &g,
            );
            add_raw(
                &format!("model.layers.0.block_sparse_moe.experts.{e}.w2.weight"),
                vec![hidden, inter],
                &d,
            );
        }
        add_raw("model.norm.weight", vec![hidden], &n1);
        add_raw("lm_head.weight", vec![vocab, hidden], &emb);

        let cfg = json!({
            "format": "aria-quant-bundle",
            "format_version": 2,
            "quantization": "test",
            "group_size_default": 32,
            "hadamard_seed": 0,
            "model": {
                "hidden_size": hidden,
                "num_layers": layers,
                "num_attention_heads": n_heads,
                "num_kv_heads": n_kv,
                "head_dim": head_dim,
                "intermediate_size": inter,
                "vocab_size": vocab,
                "context_length": 32,
                "rope_theta": 10000.0,
                "num_experts": n_experts,
                "num_experts_per_tok": top_k
            },
            "tensors": tensors
        });
        std::fs::write(dir.path().join("config.json"), cfg.to_string()).unwrap();
        std::fs::write(dir.path().join("weight.bin"), &bin).unwrap();

        let mut s = SessionBuilder::new()
            .model(dir.path())
            .family("inkling/inkling-small")
            .build()
            .unwrap();
        assert_eq!(s.arch(), ArchClass::TextMoE);
        assert_eq!(s.graph_hook_name(), "text_moe_decoder");
        let gen = s
            .generate(
                &[1, 2],
                &GenerateOpts {
                    max_tokens: 2,
                    temperature: 0.0,
                },
            )
            .unwrap();
        assert_eq!(gen.tokens.len(), 2);
    }

    #[test]
    fn tiny_q4_codebook_weights_unrotate_on_load() {
        let dir = tempfile::tempdir().unwrap();
        write_tiny_q4_bundle(dir.path()).unwrap();
        let b = load_bundle(dir.path()).unwrap();
        let w = b.weight_loaded("blk.0.attn_q.weight").unwrap();
        assert!(
            w.hdm_seed.is_none(),
            "reconstruct_weight path stores original-space W for linear()"
        );
        let mut s = SessionBuilder::new()
            .model(dir.path())
            .family("gemma/gemma-4-e2b-it")
            .build()
            .unwrap();
        let gen = s
            .generate(
                &[1, 2],
                &GenerateOpts {
                    max_tokens: 2,
                    temperature: 0.0,
                },
            )
            .unwrap();
        assert_eq!(gen.tokens.len(), 2);
    }

    #[test]
    fn gemma_hidden_act_geglu_and_qk_norm() {
        let dir = tempfile::tempdir().unwrap();
        let hidden = 8usize;
        let layers = 1usize;
        let inter = 16usize;
        let vocab = 16usize;
        let n_heads = 2usize;
        let n_kv = 1usize;
        let head_dim = 4usize;
        let q_dim = n_heads * head_dim;
        let k_dim = n_kv * head_dim;
        let p = "model.language_model";

        let mut tensors = serde_json::Map::new();
        let mut bin = Vec::new();
        let mut add_raw = |name: &str, shape: Vec<usize>, data: &[f32]| {
            let offset = bin.len();
            for &v in data {
                bin.extend_from_slice(&v.to_le_bytes());
            }
            let nbytes = data.len() * 4;
            let mut meta = serde_json::Map::new();
            meta.insert("kind".into(), json!("raw"));
            meta.insert("dtype".into(), json!("f32"));
            meta.insert("shape".into(), json!(shape));
            meta.insert("offsets".into(), json!({ "data": [offset, nbytes] }));
            tensors.insert(name.to_string(), Value::Object(meta));
        };
        let emb: Vec<f32> = (0..vocab * hidden).map(|i| i as f32 * 0.01).collect();
        add_raw(
            &format!("{p}.embed_tokens.weight"),
            vec![vocab, hidden],
            &emb,
        );
        let n1 = vec![1.0f32; hidden];
        let qn = vec![1.0f32; head_dim];
        let kn = vec![1.0f32; head_dim];
        let wq = vec![0.01f32; q_dim * hidden];
        let wk = vec![0.01f32; k_dim * hidden];
        let wv = vec![0.01f32; k_dim * hidden];
        let wo = vec![0.01f32; hidden * q_dim];
        add_raw(
            &format!("{p}.layers.0.input_layernorm.weight"),
            vec![hidden],
            &n1,
        );
        add_raw(
            &format!("{p}.layers.0.pre_feedforward_layernorm.weight"),
            vec![hidden],
            &n1,
        );
        add_raw(
            &format!("{p}.layers.0.self_attn.q_proj.weight"),
            vec![q_dim, hidden],
            &wq,
        );
        add_raw(
            &format!("{p}.layers.0.self_attn.k_proj.weight"),
            vec![k_dim, hidden],
            &wk,
        );
        add_raw(
            &format!("{p}.layers.0.self_attn.v_proj.weight"),
            vec![k_dim, hidden],
            &wv,
        );
        add_raw(
            &format!("{p}.layers.0.self_attn.o_proj.weight"),
            vec![hidden, q_dim],
            &wo,
        );
        add_raw(
            &format!("{p}.layers.0.self_attn.q_norm.weight"),
            vec![head_dim],
            &qn,
        );
        add_raw(
            &format!("{p}.layers.0.self_attn.k_norm.weight"),
            vec![head_dim],
            &kn,
        );
        let g = vec![0.01f32; inter * hidden];
        let d = vec![0.01f32; hidden * inter];
        add_raw(
            &format!("{p}.layers.0.mlp.gate_proj.weight"),
            vec![inter, hidden],
            &g,
        );
        add_raw(
            &format!("{p}.layers.0.mlp.up_proj.weight"),
            vec![inter, hidden],
            &g,
        );
        add_raw(
            &format!("{p}.layers.0.mlp.down_proj.weight"),
            vec![hidden, inter],
            &d,
        );
        add_raw(&format!("{p}.norm.weight"), vec![hidden], &n1);
        add_raw("lm_head.weight", vec![vocab, hidden], &emb);

        let cfg = json!({
            "format": "aria-quant-bundle",
            "format_version": 2,
            "quantization": "test",
            "group_size_default": 32,
            "hadamard_seed": 0,
            "model": {
                "hidden_size": hidden,
                "num_layers": layers,
                "num_attention_heads": n_heads,
                "num_kv_heads": n_kv,
                "head_dim": head_dim,
                "global_head_dim": head_dim,
                "sliding_window": 512,
                "partial_rotary_factor": 0.25,
                "intermediate_size": inter,
                "vocab_size": vocab,
                "context_length": 32,
                "rope_theta": 10000.0,
                "hidden_act": "gelu_pytorch_tanh",
                "layer_types": ["full_attention"]
            },
            "tensors": tensors
        });
        std::fs::write(dir.path().join("config.json"), cfg.to_string()).unwrap();
        std::fs::write(dir.path().join("weight.bin"), &bin).unwrap();

        let mut s = SessionBuilder::new()
            .model(dir.path())
            .family("gemma/gemma-4-e2b-it")
            .build()
            .unwrap();
        assert_eq!(s.config().hidden_act.as_deref(), Some("gelu_pytorch_tanh"));
        let gen = s
            .generate(
                &[1, 2],
                &GenerateOpts {
                    max_tokens: 2,
                    temperature: 0.0,
                },
            )
            .unwrap();
        assert_eq!(gen.tokens.len(), 2);
    }

    #[test]
    fn gated_deltanet_and_full_attn_generate() {
        let dir = tempfile::tempdir().unwrap();
        let hidden = 8usize;
        let inter = 16usize;
        let vocab = 16usize;
        let n_heads = 2usize;
        let n_kv = 1usize;
        let head_dim = 4usize;
        let q_dim = n_heads * head_dim;
        let k_dim = n_kv * head_dim;
        let n_lin = 2usize;
        let hk = 4usize;
        let hv = 4usize;
        let key_dim = n_lin * hk;
        let value_dim = n_lin * hv;
        let conv_k = 4usize;
        let conv_dim = key_dim * 2 + value_dim;

        let mut tensors = serde_json::Map::new();
        let mut bin = Vec::new();
        let mut add_raw = |name: &str, shape: Vec<usize>, data: &[f32]| {
            let offset = bin.len();
            for &v in data {
                bin.extend_from_slice(&v.to_le_bytes());
            }
            let nbytes = data.len() * 4;
            let mut meta = serde_json::Map::new();
            meta.insert("kind".into(), json!("raw"));
            meta.insert("dtype".into(), json!("f32"));
            meta.insert("shape".into(), json!(shape));
            meta.insert("offsets".into(), json!({ "data": [offset, nbytes] }));
            tensors.insert(name.to_string(), Value::Object(meta));
        };
        let emb: Vec<f32> = (0..vocab * hidden).map(|i| i as f32 * 0.01).collect();
        add_raw("model.embed_tokens.weight", vec![vocab, hidden], &emb);
        let n1 = vec![1.0f32; hidden];
        add_raw("model.layers.0.input_layernorm.weight", vec![hidden], &n1);
        add_raw(
            "model.layers.0.post_attention_layernorm.weight",
            vec![hidden],
            &n1,
        );
        let qkvz = vec![0.02f32; (2 * key_dim + 2 * value_dim) * hidden];
        let ba = vec![0.1f32; 2 * n_lin * hidden];
        let conv = vec![0.05f32; conv_dim * conv_k];
        let a_log = vec![0.5f32; n_lin];
        let dt = vec![1.0f32; n_lin];
        let outp = vec![0.02f32; hidden * value_dim];
        add_raw(
            "model.layers.0.linear_attn.in_proj_qkvz.weight",
            vec![2 * key_dim + 2 * value_dim, hidden],
            &qkvz,
        );
        add_raw(
            "model.layers.0.linear_attn.in_proj_ba.weight",
            vec![2 * n_lin, hidden],
            &ba,
        );
        add_raw(
            "model.layers.0.linear_attn.conv1d.weight",
            vec![conv_dim, conv_k],
            &conv,
        );
        add_raw("model.layers.0.linear_attn.A_log", vec![n_lin], &a_log);
        add_raw("model.layers.0.linear_attn.dt_bias", vec![n_lin], &dt);
        add_raw(
            "model.layers.0.linear_attn.out_proj.weight",
            vec![hidden, value_dim],
            &outp,
        );
        let g = vec![0.02f32; inter * hidden];
        let d = vec![0.02f32; hidden * inter];
        add_raw(
            "model.layers.0.mlp.gate_proj.weight",
            vec![inter, hidden],
            &g,
        );
        add_raw("model.layers.0.mlp.up_proj.weight", vec![inter, hidden], &g);
        add_raw(
            "model.layers.0.mlp.down_proj.weight",
            vec![hidden, inter],
            &d,
        );

        add_raw("model.layers.1.input_layernorm.weight", vec![hidden], &n1);
        add_raw(
            "model.layers.1.post_attention_layernorm.weight",
            vec![hidden],
            &n1,
        );
        let wq = vec![0.02f32; q_dim * hidden];
        let wk = vec![0.02f32; k_dim * hidden];
        let wv = vec![0.02f32; k_dim * hidden];
        let wo = vec![0.02f32; hidden * q_dim];
        add_raw(
            "model.layers.1.self_attn.q_proj.weight",
            vec![q_dim, hidden],
            &wq,
        );
        add_raw(
            "model.layers.1.self_attn.k_proj.weight",
            vec![k_dim, hidden],
            &wk,
        );
        add_raw(
            "model.layers.1.self_attn.v_proj.weight",
            vec![k_dim, hidden],
            &wv,
        );
        add_raw(
            "model.layers.1.self_attn.o_proj.weight",
            vec![hidden, q_dim],
            &wo,
        );
        add_raw(
            "model.layers.1.mlp.gate_proj.weight",
            vec![inter, hidden],
            &g,
        );
        add_raw("model.layers.1.mlp.up_proj.weight", vec![inter, hidden], &g);
        add_raw(
            "model.layers.1.mlp.down_proj.weight",
            vec![hidden, inter],
            &d,
        );
        add_raw("model.norm.weight", vec![hidden], &n1);
        add_raw("lm_head.weight", vec![vocab, hidden], &emb);

        let cfg = json!({
            "format": "aria-quant-bundle",
            "format_version": 2,
            "quantization": "test",
            "hadamard_seed": 0,
            "model": {
                "hidden_size": hidden,
                "num_layers": 2,
                "num_attention_heads": n_heads,
                "num_kv_heads": n_kv,
                "head_dim": head_dim,
                "intermediate_size": inter,
                "vocab_size": vocab,
                "context_length": 32,
                "rope_theta": 10000.0,
                "layer_types": ["linear_attention", "full_attention"]
            },
            "tensors": tensors
        });
        std::fs::write(dir.path().join("config.json"), cfg.to_string()).unwrap();
        std::fs::write(dir.path().join("weight.bin"), &bin).unwrap();
        let mut s = SessionBuilder::new()
            .model(dir.path())
            .family("qwen/qwen3.5-2b")
            .build()
            .unwrap();
        let gen = s
            .generate(
                &[1, 2, 3],
                &GenerateOpts {
                    max_tokens: 2,
                    temperature: 0.0,
                },
            )
            .unwrap();
        assert_eq!(gen.tokens.len(), 2);
    }

    #[test]
    fn gated_deltanet_split_qwen35_projections_generate() {
        let dir = tempfile::tempdir().unwrap();
        let hidden = 8usize;
        let inter = 16usize;
        let vocab = 16usize;
        let n_heads = 2usize;
        let n_kv = 1usize;
        let head_dim = 4usize;
        let q_dim = n_heads * head_dim;
        let k_dim = n_kv * head_dim;
        let n_lin = 2usize;
        let hk = 4usize;
        let hv = 4usize;
        let key_dim = n_lin * hk;
        let value_dim = n_lin * hv;
        let conv_k = 4usize;
        let conv_dim = key_dim * 2 + value_dim;

        let mut tensors = serde_json::Map::new();
        let mut bin = Vec::new();
        let mut add_raw = |name: &str, shape: Vec<usize>, data: &[f32]| {
            let offset = bin.len();
            for &v in data {
                bin.extend_from_slice(&v.to_le_bytes());
            }
            let nbytes = data.len() * 4;
            let mut meta = serde_json::Map::new();
            meta.insert("kind".into(), json!("raw"));
            meta.insert("dtype".into(), json!("f32"));
            meta.insert("shape".into(), json!(shape));
            meta.insert("offsets".into(), json!({ "data": [offset, nbytes] }));
            tensors.insert(name.to_string(), Value::Object(meta));
        };
        let emb: Vec<f32> = (0..vocab * hidden).map(|i| i as f32 * 0.01).collect();
        add_raw("model.embed_tokens.weight", vec![vocab, hidden], &emb);
        let n1 = vec![1.0f32; hidden];
        add_raw("model.layers.0.input_layernorm.weight", vec![hidden], &n1);
        add_raw(
            "model.layers.0.post_attention_layernorm.weight",
            vec![hidden],
            &n1,
        );
        let qkv = vec![0.02f32; (2 * key_dim + value_dim) * hidden];
        let z = vec![0.02f32; value_dim * hidden];
        let proj_b = vec![0.1f32; n_lin * hidden];
        let proj_a = vec![0.1f32; n_lin * hidden];
        let conv = vec![0.05f32; conv_dim * conv_k];
        let a_log = vec![0.5f32; n_lin];
        let dt = vec![1.0f32; n_lin];
        let outp = vec![0.02f32; hidden * value_dim];
        add_raw(
            "model.layers.0.linear_attn.in_proj_qkv.weight",
            vec![2 * key_dim + value_dim, hidden],
            &qkv,
        );
        add_raw(
            "model.layers.0.linear_attn.in_proj_z.weight",
            vec![value_dim, hidden],
            &z,
        );
        add_raw(
            "model.layers.0.linear_attn.in_proj_b.weight",
            vec![n_lin, hidden],
            &proj_b,
        );
        add_raw(
            "model.layers.0.linear_attn.in_proj_a.weight",
            vec![n_lin, hidden],
            &proj_a,
        );
        add_raw(
            "model.layers.0.linear_attn.conv1d.weight",
            vec![conv_dim, conv_k],
            &conv,
        );
        add_raw("model.layers.0.linear_attn.A_log", vec![n_lin], &a_log);
        add_raw("model.layers.0.linear_attn.dt_bias", vec![n_lin], &dt);
        add_raw(
            "model.layers.0.linear_attn.out_proj.weight",
            vec![hidden, value_dim],
            &outp,
        );
        let g = vec![0.02f32; inter * hidden];
        let d = vec![0.02f32; hidden * inter];
        add_raw(
            "model.layers.0.mlp.gate_proj.weight",
            vec![inter, hidden],
            &g,
        );
        add_raw("model.layers.0.mlp.up_proj.weight", vec![inter, hidden], &g);
        add_raw(
            "model.layers.0.mlp.down_proj.weight",
            vec![hidden, inter],
            &d,
        );

        add_raw("model.layers.1.input_layernorm.weight", vec![hidden], &n1);
        add_raw(
            "model.layers.1.post_attention_layernorm.weight",
            vec![hidden],
            &n1,
        );
        let wq = vec![0.02f32; q_dim * hidden];
        let wk = vec![0.02f32; k_dim * hidden];
        let wv = vec![0.02f32; k_dim * hidden];
        let wo = vec![0.02f32; hidden * q_dim];
        add_raw(
            "model.layers.1.self_attn.q_proj.weight",
            vec![q_dim, hidden],
            &wq,
        );
        add_raw(
            "model.layers.1.self_attn.k_proj.weight",
            vec![k_dim, hidden],
            &wk,
        );
        add_raw(
            "model.layers.1.self_attn.v_proj.weight",
            vec![k_dim, hidden],
            &wv,
        );
        add_raw(
            "model.layers.1.self_attn.o_proj.weight",
            vec![hidden, q_dim],
            &wo,
        );
        add_raw("model.layers.1.mlp.gate_proj.weight", vec![inter, hidden], &g);
        add_raw("model.layers.1.mlp.up_proj.weight", vec![inter, hidden], &g);
        add_raw(
            "model.layers.1.mlp.down_proj.weight",
            vec![hidden, inter],
            &d,
        );
        add_raw("model.norm.weight", vec![hidden], &n1);
        add_raw("lm_head.weight", vec![vocab, hidden], &emb);

        let cfg = json!({
            "format": "aria-quant-bundle",
            "format_version": 2,
            "quantization": "test",
            "hadamard_seed": 0,
            "model": {
                "hidden_size": hidden,
                "num_layers": 2,
                "num_attention_heads": n_heads,
                "num_kv_heads": n_kv,
                "head_dim": head_dim,
                "intermediate_size": inter,
                "vocab_size": vocab,
                "context_length": 32,
                "rope_theta": 10000.0,
                "layer_types": ["linear_attention", "full_attention"]
            },
            "tensors": tensors
        });
        std::fs::write(dir.path().join("config.json"), cfg.to_string()).unwrap();
        std::fs::write(dir.path().join("weight.bin"), &bin).unwrap();
        let mut s = SessionBuilder::new()
            .model(dir.path())
            .family("qwen/qwen3.5-0.8b")
            .build()
            .unwrap();
        let gen = s
            .generate(
                &[1, 2, 3],
                &GenerateOpts {
                    max_tokens: 2,
                    temperature: 0.0,
                },
            )
            .unwrap();
        assert_eq!(gen.tokens.len(), 2);
    }

    #[test]
    fn vision_and_action_consume_bundle_weights() {
        let dir = tempfile::tempdir().unwrap();
        write_tiny_q4_bundle(dir.path()).unwrap();
        let cfg_path = dir.path().join("config.json");
        let mut cfg: Value =
            serde_json::from_str(&std::fs::read_to_string(&cfg_path).unwrap()).unwrap();
        let hidden = cfg["model"]["hidden_size"].as_u64().unwrap() as usize;
        let mut tensors = cfg["tensors"].as_object().cloned().unwrap();
        let mut bin = std::fs::read(dir.path().join("weight.bin")).unwrap();
        let mut add_raw = |name: &str, shape: Vec<usize>, data: &[f32]| {
            let offset = bin.len();
            for &v in data {
                bin.extend_from_slice(&v.to_le_bytes());
            }
            let nbytes = data.len() * 4;
            tensors.insert(
                name.to_string(),
                json!({
                    "kind": "raw",
                    "dtype": "f32",
                    "shape": shape,
                    "offsets": { "data": [offset, nbytes] }
                }),
            );
        };
        let vis = vec![0.1f32; hidden * 3];
        add_raw("mm_projector.weight", vec![hidden, 3], &vis);
        let act_dim = 7usize;
        let act = vec![0.05f32; act_dim * hidden];
        add_raw("action_head.weight", vec![act_dim, hidden], &act);
        cfg["tensors"] = Value::Object(tensors);
        std::fs::write(&cfg_path, cfg.to_string()).unwrap();
        std::fs::write(dir.path().join("weight.bin"), &bin).unwrap();

        let s = SessionBuilder::new()
            .model(dir.path())
            .family("lfm/lfm2-vl-450m")
            .build()
            .unwrap();
        let rgb = vec![10u8; 3 * 4 * 4];
        let pref = s.vision_prefix(&rgb, 4, 4).unwrap();
        assert_eq!(pref.len(), hidden);

        let vla = SessionBuilder::new()
            .model(dir.path())
            .family("openvla/openvla-7b")
            .build()
            .unwrap();
        let a = vla.predict_action("move", act_dim).unwrap();
        assert_eq!(a.len(), act_dim);
    }

    #[test]
    fn load_real_hf_named_bundle_if_present() {
        // Optional local smoke: ARIA_SMOKE_BUNDLE=/path/to/qwen3-0.6b_q4 or gemma-4-e2b-it_q4
        let Ok(path) = std::env::var("ARIA_SMOKE_BUNDLE") else {
            return;
        };
        let path = std::path::Path::new(&path);
        if !path.join("config.json").is_file() {
            return;
        }
        let family = if path.to_string_lossy().contains("gemma-4") {
            "gemma/gemma-4-e2b-it"
        } else {
            "qwen/qwen3-0.6b"
        };
        let s = SessionBuilder::new()
            .model(path)
            .family(family)
            .build()
            .unwrap_or_else(|e| panic!("{family} bundle should materialize: {e}"));
        assert!(s.config().num_layers > 0);
        assert!(s.config().hidden_size > 0);
        if family.contains("gemma-4") && s.config().hidden_size >= 1024 {
            assert!(
                s.weights.ple.is_some(),
                "real Gemma-4 q4 must load codebook PLE"
            );
            let hidden = s.config().hidden_size;
            let vocab = s.config().vocab_size;
            assert!(
                s.weights.emb.data.len() >= vocab.saturating_mul(hidden),
                "embed table too small for vocab={vocab} hidden={hidden}"
            );
        }
    }

    #[test]
    fn gemma4_four_norm_ple_and_tied_embed_generate() {
        let dir = tempfile::tempdir().unwrap();
        let hidden = 8usize;
        let layers = 1usize;
        let inter = 16usize;
        let vocab = 16usize;
        let n_heads = 2usize;
        let n_kv = 1usize;
        let head_dim = 4usize;
        let q_dim = n_heads * head_dim;
        let k_dim = n_kv * head_dim;
        let ple_d = 4usize;
        let p = "model.language_model";

        let mut tensors = serde_json::Map::new();
        let mut bin = Vec::new();
        let mut add_raw = |name: &str, shape: Vec<usize>, data: &[f32]| {
            let offset = bin.len();
            for &v in data {
                bin.extend_from_slice(&v.to_le_bytes());
            }
            let nbytes = data.len() * 4;
            let mut meta = serde_json::Map::new();
            meta.insert("kind".into(), json!("raw"));
            meta.insert("dtype".into(), json!("f32"));
            meta.insert("shape".into(), json!(shape));
            meta.insert("offsets".into(), json!({ "data": [offset, nbytes] }));
            tensors.insert(name.to_string(), Value::Object(meta));
        };
        let emb: Vec<f32> = (0..vocab * hidden).map(|i| i as f32 * 0.01).collect();
        add_raw(
            &format!("{p}.embed_tokens.weight"),
            vec![vocab, hidden],
            &emb,
        );
        let n1 = vec![1.0f32; hidden];
        add_raw(
            &format!("{p}.layers.0.input_layernorm.weight"),
            vec![hidden],
            &n1,
        );
        add_raw(
            &format!("{p}.layers.0.post_attention_layernorm.weight"),
            vec![hidden],
            &n1,
        );
        add_raw(
            &format!("{p}.layers.0.pre_feedforward_layernorm.weight"),
            vec![hidden],
            &n1,
        );
        add_raw(
            &format!("{p}.layers.0.post_feedforward_layernorm.weight"),
            vec![hidden],
            &n1,
        );
        // HF skip_scale → layer_scalar (not ones at the published checkpoint).
        add_raw(&format!("{p}.layers.0.layer_scalar"), vec![1], &[0.5f32]);
        let wq = vec![0.01f32; q_dim * hidden];
        let wk = vec![0.01f32; k_dim * hidden];
        let wv = vec![0.01f32; k_dim * hidden];
        let wo = vec![0.01f32; hidden * q_dim];
        add_raw(
            &format!("{p}.layers.0.self_attn.q_proj.weight"),
            vec![q_dim, hidden],
            &wq,
        );
        add_raw(
            &format!("{p}.layers.0.self_attn.k_proj.weight"),
            vec![k_dim, hidden],
            &wk,
        );
        add_raw(
            &format!("{p}.layers.0.self_attn.v_proj.weight"),
            vec![k_dim, hidden],
            &wv,
        );
        add_raw(
            &format!("{p}.layers.0.self_attn.o_proj.weight"),
            vec![hidden, q_dim],
            &wo,
        );
        let g = vec![0.01f32; inter * hidden];
        let d = vec![0.01f32; hidden * inter];
        add_raw(
            &format!("{p}.layers.0.mlp.gate_proj.weight"),
            vec![inter, hidden],
            &g,
        );
        add_raw(
            &format!("{p}.layers.0.mlp.up_proj.weight"),
            vec![inter, hidden],
            &g,
        );
        add_raw(
            &format!("{p}.layers.0.mlp.down_proj.weight"),
            vec![hidden, inter],
            &d,
        );
        let packed = layers * ple_d;
        let ple_emb = vec![0.02f32; vocab * packed];
        add_raw(
            &format!("{p}.embed_tokens_per_layer.weight"),
            vec![vocab, packed],
            &ple_emb,
        );
        let ple_proj = vec![0.01f32; packed * hidden];
        add_raw(
            &format!("{p}.per_layer_model_projection.weight"),
            vec![packed, hidden],
            &ple_proj,
        );
        let ple_pn = vec![1.0f32; ple_d];
        add_raw(
            &format!("{p}.per_layer_projection_norm.weight"),
            vec![ple_d],
            &ple_pn,
        );
        let ple_gate = vec![0.01f32; ple_d * hidden];
        let ple_out = vec![0.01f32; hidden * ple_d];
        add_raw(
            &format!("{p}.layers.0.per_layer_input_gate.weight"),
            vec![ple_d, hidden],
            &ple_gate,
        );
        add_raw(
            &format!("{p}.layers.0.per_layer_projection.weight"),
            vec![hidden, ple_d],
            &ple_out,
        );
        add_raw(
            &format!("{p}.layers.0.post_per_layer_input_norm.weight"),
            vec![hidden],
            &n1,
        );
        add_raw(&format!("{p}.norm.weight"), vec![hidden], &n1);
        add_raw("lm_head.weight", vec![vocab, hidden], &emb);

        let cfg = json!({
            "format": "aria-quant-bundle",
            "format_version": 2,
            "quantization": "test",
            "group_size_default": 32,
            "hadamard_seed": 0,
            "model": {
                "hidden_size": hidden,
                "num_layers": layers,
                "num_attention_heads": n_heads,
                "num_kv_heads": n_kv,
                "intermediate_size": inter,
                "vocab_size": vocab,
                "context_length": 32,
                "rope_theta": 10000.0,
                "hidden_act": "gelu_pytorch_tanh",
                "tie_word_embeddings": true,
                "head_dim": head_dim,
                "global_head_dim": head_dim,
                "sliding_window": 512,
                "partial_rotary_factor": 0.25,
                "layer_types": ["full_attention"]
            },
            "tensors": tensors
        });
        std::fs::write(dir.path().join("config.json"), cfg.to_string()).unwrap();
        std::fs::write(dir.path().join("weight.bin"), &bin).unwrap();

        let mut s = SessionBuilder::new()
            .model(dir.path())
            .family("gemma/gemma-4-e2b-it")
            .build()
            .unwrap();
        assert!((s.embed_scale - (hidden as f32).sqrt()).abs() < 1e-5);
        assert!(s.weights.ple.is_some());
        assert!((s.weights.layers[0].layer_scalar - 0.5).abs() < 1e-6);
        assert!(s.weights.layers[0].post_attn_norm.is_some());
        assert!(s.weights.layers[0].post_ffn_norm.is_some());
        let prompt = vec![1u32, 2];
        let batched = s
            .generate(
                &prompt,
                &GenerateOpts {
                    max_tokens: 3,
                    temperature: 0.0,
                },
            )
            .unwrap();
        let step = s
            .generate(
                &prompt,
                &GenerateOpts {
                    max_tokens: 3,
                    temperature: 0.0,
                },
            )
            .unwrap();
        assert_eq!(batched.tokens, step.tokens);
        assert_eq!(batched.tokens.len(), 3);
        assert_eq!(s.config().sliding_window, Some(512));
    }

    #[test]
    fn gemma4_ple_required_gate() {
        assert!(!gemma4_requires_ple("gemma/gemma-4-e2b-it", 64));
        assert!(gemma4_requires_ple("gemma/gemma-4-e2b-it", 1024));
        assert!(gemma4_requires_ple("gemma/gemma-4-e2b-it", 1536));
        assert!(!gemma4_requires_ple("qwen/qwen3-0.6b", 1536));
    }

    #[test]
    fn gemma4_e2b_scale_missing_ple_is_hard_error() {
        let dir = tempfile::tempdir().unwrap();
        let hidden = 1024usize;
        let layers = 1usize;
        let vocab = 8usize;
        let n_heads = 8usize;
        let n_kv = 1usize;
        let head_dim = 128usize;
        let q_dim = n_heads * head_dim;
        let k_dim = n_kv * head_dim;
        let inter = 32usize;
        let p = "model.language_model";

        let mut tensors = serde_json::Map::new();
        let mut bin = Vec::new();
        let mut add_raw = |name: &str, shape: Vec<usize>, data: &[f32]| {
            let offset = bin.len();
            for &v in data {
                bin.extend_from_slice(&v.to_le_bytes());
            }
            let nbytes = data.len() * 4;
            tensors.insert(
                name.to_string(),
                json!({
                    "kind": "raw",
                    "dtype": "f32",
                    "shape": shape,
                    "offsets": { "data": [offset, nbytes] }
                }),
            );
        };
        let emb = vec![0.01f32; vocab * hidden];
        add_raw(&format!("{p}.embed_tokens.weight"), vec![vocab, hidden], &emb);
        let ones = vec![1.0f32; hidden];
        add_raw(
            &format!("{p}.layers.0.input_layernorm.weight"),
            vec![hidden],
            &ones,
        );
        add_raw(
            &format!("{p}.layers.0.pre_feedforward_layernorm.weight"),
            vec![hidden],
            &ones,
        );
        add_raw(
            &format!("{p}.layers.0.post_attention_layernorm.weight"),
            vec![hidden],
            &ones,
        );
        add_raw(
            &format!("{p}.layers.0.post_feedforward_layernorm.weight"),
            vec![hidden],
            &ones,
        );
        add_raw(&format!("{p}.norm.weight"), vec![hidden], &ones);
        let q = vec![0.01f32; q_dim * hidden];
        let k = vec![0.01f32; k_dim * hidden];
        add_raw(
            &format!("{p}.layers.0.self_attn.q_proj.weight"),
            vec![q_dim, hidden],
            &q,
        );
        add_raw(
            &format!("{p}.layers.0.self_attn.k_proj.weight"),
            vec![k_dim, hidden],
            &k,
        );
        add_raw(
            &format!("{p}.layers.0.self_attn.v_proj.weight"),
            vec![k_dim, hidden],
            &k,
        );
        add_raw(
            &format!("{p}.layers.0.self_attn.o_proj.weight"),
            vec![hidden, q_dim],
            &q,
        );
        let qn = vec![1.0f32; head_dim];
        add_raw(
            &format!("{p}.layers.0.self_attn.q_norm.weight"),
            vec![head_dim],
            &qn,
        );
        add_raw(
            &format!("{p}.layers.0.self_attn.k_norm.weight"),
            vec![head_dim],
            &qn,
        );
        let g = vec![0.01f32; inter * hidden];
        add_raw(
            &format!("{p}.layers.0.mlp.gate_proj.weight"),
            vec![inter, hidden],
            &g,
        );
        add_raw(
            &format!("{p}.layers.0.mlp.up_proj.weight"),
            vec![inter, hidden],
            &g,
        );
        add_raw(
            &format!("{p}.layers.0.mlp.down_proj.weight"),
            vec![hidden, inter],
            &g,
        );
        let cfg = json!({
            "format": "aria-quant-bundle",
            "format_version": 2,
            "quantization": "test",
            "group_size_default": 32,
            "hadamard_seed": 0,
            "model": {
                "hidden_size": hidden,
                "num_layers": layers,
                "num_attention_heads": n_heads,
                "num_kv_heads": n_kv,
                "intermediate_size": inter,
                "vocab_size": vocab,
                "context_length": 32,
                "rope_theta": 10000.0,
                "hidden_act": "gelu_pytorch_tanh",
                "tie_word_embeddings": true,
                "head_dim": head_dim,
                "global_head_dim": head_dim,
                "sliding_window": 512,
                "partial_rotary_factor": 0.25,
                "num_kv_shared_layers": 0,
                "layer_types": ["full_attention"]
            },
            "tensors": tensors
        });
        std::fs::write(dir.path().join("config.json"), cfg.to_string()).unwrap();
        std::fs::write(dir.path().join("weight.bin"), &bin).unwrap();

        let err = SessionBuilder::new()
            .model(dir.path())
            .family("gemma/gemma-4-e2b-it")
            .build()
            .unwrap_err();
        let msg = err.to_string();
        assert!(
            msg.contains("PLE") && msg.contains("embed_tokens_per_layer"),
            "{msg}"
        );
    }

    #[test]
    fn gemma4_sliding_window_config_and_generate() {
        let dir_wide = tempfile::tempdir().unwrap();
        write_tiny_q4_bundle(dir_wide.path()).unwrap();
        let dir_narrow = tempfile::tempdir().unwrap();
        write_tiny_q4_bundle(dir_narrow.path()).unwrap();
        let patch = |path: &std::path::Path, window: usize| {
            let cfg_path = path.join("config.json");
            let raw = std::fs::read_to_string(&cfg_path).unwrap();
            let mut cfg: Value = serde_json::from_str(&raw).unwrap();
            cfg["model"]["sliding_window"] = json!(window);
            cfg["model"]["layer_types"] = json!(["sliding_attention", "sliding_attention"]);
            std::fs::write(&cfg_path, cfg.to_string()).unwrap();
        };
        patch(dir_wide.path(), 512);
        patch(dir_narrow.path(), 1);
        let wide = SessionBuilder::new()
            .model(dir_wide.path())
            .family("gemma/gemma-4-e2b-it")
            .build()
            .unwrap();
        let mut narrow = SessionBuilder::new()
            .model(dir_narrow.path())
            .family("gemma/gemma-4-e2b-it")
            .build()
            .unwrap();
        assert_eq!(wide.config().sliding_window, Some(512));
        assert_eq!(narrow.config().sliding_window, Some(1));
        assert_eq!(wide.attn_window(AttnKind::Sliding), Some(512));
        assert_eq!(narrow.attn_window(AttnKind::Sliding), Some(1));
        for layer in &narrow.weights.layers {
            if let LayerOp::Attn(attn) = &layer.op {
                assert_eq!(attn.kind, AttnKind::Sliding);
            }
        }
        let prompt = vec![1u32, 2, 3, 4];
        let gen = narrow
            .generate(
                &prompt,
                &GenerateOpts {
                    max_tokens: 3,
                    temperature: 0.0,
                },
            )
            .unwrap();
        assert_eq!(gen.tokens.len(), 3);

        // Prefill window slice must match stepwise decode (same mask).
        let mut incr = SessionBuilder::new()
            .model(dir_narrow.path())
            .family("gemma/gemma-4-e2b-it")
            .build()
            .unwrap();
        let again = incr
            .generate(
                &prompt,
                &GenerateOpts {
                    max_tokens: 3,
                    temperature: 0.0,
                },
            )
            .unwrap();
        assert_eq!(gen.tokens, again.tokens);
    }
}