fuel-transformers 0.19.1

// SPDX-License-Identifier: MIT OR Apache-2.0
//! GGUF-quantized Qwen3-MoE ported to the lazy-graph API.
//!
//! Qwen3-MoE with Q4_0 block-quantized Linear weights. The forward
//! path is identical to [`crate::models::lazy_qwen3_moe::Qwen3MoeModel`]; only
//! the storage of the per-layer attention projections, the per-layer
//! FFN matrices (dense `gate_proj` / `up_proj` / `down_proj` and the
//! per-expert variants when the layer is MoE), and the final
//! `lm_head` change from F32/BF16 to `WeightStorage::Q4_0`. RmsNorm
//! gains (including the per-head QK-norm gains), attention biases
//! (when `cfg.attention_bias`), the router gate (`mlp.gate.weight`),
//! and the token embedding stay in F32.
//!
//! Construction paths:
//! - [`QuantizedQwen3MoeModel::from_f32_bake`] — take f32 source
//!   weights (same `[in, out]` layout as `Qwen3MoeWeights`) and
//!   quantize on the fly. Used by tests and by callers that have
//!   unquantized weights in memory already.
//! - [`QuantizedQwen3MoeModel::from_gguf`] — load directly from a
//!   Qwen3-MoE GGUF file (`general.architecture` typically
//!   `"qwen3moe"`), keeping Q4_0 tensors quantized and dequantizing
//!   other GGML dtypes to F32. Mirrors the SmolLM3 / Phi-2 loader
//!   convention.
//!
//! GGUF tensor naming convention (llama.cpp):
//!   - `token_embd.weight` — token embedding table
//!   - `output_norm.weight` — final RmsNorm gain
//!   - `output.weight` — `lm_head` (tied to `token_embd.weight` when
//!     absent)
//!   - `blk.{i}.attn_q.weight` / `attn_k.weight` / `attn_v.weight` /
//!     `attn_output.weight` — attention Linear projections
//!   - `blk.{i}.attn_q.bias` / `attn_k.bias` / `attn_v.bias` —
//!     attention biases (Qwen3 base ships them, MoE follows the same
//!     convention)
//!   - `blk.{i}.attn_q_norm.weight` / `attn_k_norm.weight` — per-head
//!     QK-norm gains
//!   - `blk.{i}.attn_norm.weight` / `ffn_norm.weight` — RmsNorm gains
//!   - Dense FFN: `blk.{i}.ffn_gate.weight` / `ffn_up.weight` /
//!     `ffn_down.weight`
//!   - MoE FFN: `blk.{i}.ffn_gate_inp.weight` (router) and
//!     `blk.{i}.ffn_gate.{e}.weight` / `ffn_up.{e}.weight` /
//!     `ffn_down.{e}.weight` per expert.

use crate::models::lazy_qwen3_moe::{
    Qwen3MoeConfig, Qwen3MoeExpertWeights, Qwen3MoeFfn, Qwen3MoeLayerWeights, Qwen3MoeModel,
    Qwen3MoeWeights,
};
#[cfg(test)]
use fuel_ir::Shape;
use fuel_ir::error::Result;
#[cfg(test)]
use fuel_tensor::Device;
use fuel_tensor::{lazy::Tensor, lazy::WeightStorage};
use std::sync::Arc;

/// GGUF-quantized Qwen3-MoE causal language model. Wraps a plain
/// [`Qwen3MoeModel`] whose attention Linear weights, per-layer FFN
/// matrices, per-expert FFN matrices, and `lm_head` are
/// `WeightStorage::Q4_0`. RmsNorm gains, the per-head QK-norm gains,
/// attention biases (when present), the router gate, and the token
/// embedding stay in F32.
#[derive(Debug, Clone)]
pub struct QuantizedQwen3MoeModel {
    inner: Qwen3MoeModel,
}

impl QuantizedQwen3MoeModel {
    /// Forward over a token-ID sequence. Returns logits with shape
    /// `(1, seq, vocab_size)`.
    pub fn forward(&self, tokens: &[u32], start_pos: usize) -> Result<Tensor> {
        self.inner.forward(tokens, start_pos)
    }

    /// Forward over pre-computed embeddings, skipping the
    /// token-embedding lookup. Embeds must have shape
    /// `(1, seq, hidden_size)`.
    pub fn forward_embeds(&self, embeds: &Tensor, start_pos: usize) -> Result<Tensor> {
        self.inner.forward_embeds(embeds, start_pos)
    }

    /// Per-token hidden states up to the final RmsNorm, shape
    /// `(1, seq, hidden_size)`.
    pub fn forward_hidden(&self, tokens: &[u32], start_pos: usize) -> Result<Tensor> {
        self.inner.forward_hidden(tokens, start_pos)
    }

    /// Pre-embedded variant of [`Self::forward_hidden`].
    pub fn forward_hidden_embeds(&self, embeds: &Tensor, start_pos: usize) -> Result<Tensor> {
        self.inner.forward_hidden_embeds(embeds, start_pos)
    }

    /// Build per-token embeddings without running the decoder.
    pub fn embed_tokens_anchored(&self, anchor: &Tensor, tokens: &[u32]) -> Result<Tensor> {
        self.inner.embed_tokens_anchored(anchor, tokens)
    }

    /// Decode/prefill through a pre-allocated KV cache (GAP-029 increment 3).
    ///
    /// **Delegates ONE level, and that is the only level there is** —
    /// `Qwen3MoeModel` is a leaf (`config` + `weights`). Recorded because the sibling
    /// `QuantizedLlama3Model` has precedent for reaching `self.inner.inner` and
    /// silently losing LLaMA-3.1 RoPE scaling; that trap needs two levels AND a
    /// scaled-frequency variant, and this family has neither.
    pub fn forward_with_kv_context(
        &self,
        tokens: &[u32],
        cache: &mut fuel_tensor::inference_context::KvCache,
        ctx: &mut fuel_tensor::inference_context::InferenceContext,
    ) -> Result<Vec<f32>> {
        self.inner.forward_with_kv_context(tokens, cache, ctx)
    }

    /// Plan-once persistent decode.
    pub fn forward_with_kv_context_persistent(
        &self,
        tokens: &[u32],
        cache: &mut fuel_tensor::inference_context::KvCache,
        ctx: &mut fuel_tensor::inference_context::InferenceContext,
        session: &mut Option<fuel_tensor::inference_context::DecodeSession>,
    ) -> Result<Vec<f32>> {
        self.inner
            .forward_with_kv_context_persistent(tokens, cache, ctx, session)
    }

    /// Persistent decode with the session owned by the `InferenceContext`.
    pub fn forward_decode_step(
        &self,
        tokens: &[u32],
        cache: &mut fuel_tensor::inference_context::KvCache,
        ctx: &mut fuel_tensor::inference_context::InferenceContext,
    ) -> Result<Vec<f32>> {
        self.inner.forward_decode_step(tokens, cache, ctx)
    }

    /// Model configuration.
    pub fn config(&self) -> &Qwen3MoeConfig {
        &self.inner.config
    }

    /// Underlying [`Qwen3MoeModel`] for direct access to the lazy
    /// graph API. The wrapper exists solely to label the quantization
    /// origin.
    pub fn inner(&self) -> &Qwen3MoeModel {
        &self.inner
    }

    /// Convenience: load f32 Qwen3-MoE weights from HF safetensors
    /// and quantize each Linear weight to Q4_0. Equivalent to
    /// `Self::from_f32_bake(cfg, Qwen3MoeWeights::load_from_mmapped(st, &cfg)?)`.
    pub fn load_from_mmapped(
        st: &fuel_loaders::safetensors::MmapedSafetensors,
        cfg: Qwen3MoeConfig,
    ) -> Result<Self> {
        let src = Qwen3MoeWeights::load_from_mmapped(st, &cfg)?;
        Self::from_f32_bake(cfg, src)
    }

    /// Construct from in-memory f32 source weights, quantizing each
    /// Linear weight to Q4_0. Norm gains, biases, the router gate,
    /// and the token-embedding table stay in F32 — matching the GGUF
    /// convention used by llama.cpp for Qwen3-MoE releases.
    ///
    /// Q4_0 blocks run along each Linear's *in_features* axis, so
    /// every dimension that serves as an in_features must be a
    /// multiple of the Q4_0 block size (32): `cfg.hidden_size`
    /// (attn_q/k/v, dense + expert ffn_gate/up, output),
    /// `cfg.intermediate_size` (dense ffn_down),
    /// `cfg.moe_intermediate_size` (expert ffn_down), and
    /// `cfg.num_attention_heads * cfg.head_dim` (attn_o).
    /// `num_key_value_heads * head_dim` only ever appears as an
    /// out_features (attn_k / attn_v) and carries no divisibility
    /// requirement. Source weights follow the same
    /// `[in_features, out_features]` row-major layout as
    /// [`Qwen3MoeWeights`].
    pub fn from_f32_bake(cfg: Qwen3MoeConfig, src: Qwen3MoeWeights) -> Result<Self> {
        let h = cfg.hidden_size;
        let inter = cfg.intermediate_size;
        let moe_int = cfg.moe_intermediate_size;
        let kv = cfg.num_key_value_heads * cfg.head_dim;
        let q_dim = cfg.num_attention_heads * cfg.head_dim;
        // Gate exactly the dims that serve as a Linear in_features —
        // Q4_0 blocks run along K only. kv (= num_key_value_heads *
        // head_dim) is only ever an out_features (attn_k / attn_v)
        // and needs no divisibility (mirrors the gemma3 gate).
        check_q4_0_divisible(
            "hidden_size (attn_q/k/v, ffn_gate/up, output in-features)",
            h,
        )?;
        check_q4_0_divisible("intermediate_size (dense ffn_down in-features)", inter)?;
        check_q4_0_divisible(
            "moe_intermediate_size (expert ffn_down in-features)",
            moe_int,
        )?;
        check_q4_0_divisible("num_attention_heads * head_dim (attn_o in-features)", q_dim)?;

        let quantize_linear = |w: &WeightStorage,

                               in_features: usize,
                               out_features: usize|

         -> Result<WeightStorage> {
            let f32_in_out = match w {
                WeightStorage::F32(a) => a.to_vec(),
                _ => return Err(fuel_ir::error::Error::Msg(
                    "QuantizedQwen3MoeModel::from_f32_bake: source weights must be WeightStorage::F32".into(),
                ).bt()),
            };
            if f32_in_out.len() != in_features * out_features {
                return Err(fuel_ir::error::Error::Msg(format!(
                    "QuantizedQwen3MoeModel::from_f32_bake: weight has {} elems, expected {}×{}",
                    f32_in_out.len(),
                    in_features,
                    out_features,
                ))
                .bt());
            }
            quantize_in_out_to_q4_0(&f32_in_out, in_features, out_features)
        };

        let mut layers: Vec<Qwen3MoeLayerWeights> = Vec::with_capacity(cfg.num_hidden_layers);
        for (idx, layer) in src.layers.into_iter().enumerate() {
            let attn_q = quantize_linear(&layer.attn_q, h, q_dim)
                .map_err(|e| layer_err(idx, "attn_q", e))?;
            let attn_k =
                quantize_linear(&layer.attn_k, h, kv).map_err(|e| layer_err(idx, "attn_k", e))?;
            let attn_v =
                quantize_linear(&layer.attn_v, h, kv).map_err(|e| layer_err(idx, "attn_v", e))?;
            let attn_o = quantize_linear(&layer.attn_o, q_dim, h)
                .map_err(|e| layer_err(idx, "attn_o", e))?;

            let ffn = match layer.ffn {
                Qwen3MoeFfn::Dense {
                    gate_w,
                    up_w,
                    down_w,
                } => {
                    let gate_w = quantize_linear(&gate_w, h, inter)
                        .map_err(|e| layer_err(idx, "ffn_gate (dense)", e))?;
                    let up_w = quantize_linear(&up_w, h, inter)
                        .map_err(|e| layer_err(idx, "ffn_up (dense)", e))?;
                    let down_w = quantize_linear(&down_w, inter, h)
                        .map_err(|e| layer_err(idx, "ffn_down (dense)", e))?;
                    Qwen3MoeFfn::Dense {
                        gate_w,
                        up_w,
                        down_w,
                    }
                }
                Qwen3MoeFfn::Moe { router_w, experts } => {
                    // Router gate stays F32 — it's small and stays
                    // in dense float per llama.cpp convention.
                    let mut q_experts: Vec<Qwen3MoeExpertWeights> =
                        Vec::with_capacity(experts.len());
                    for (ei, ew) in experts.into_iter().enumerate() {
                        let gate_w = quantize_linear(&ew.gate_w, h, moe_int)
                            .map_err(|e| layer_err(idx, &format!("ffn_gate expert {ei}"), e))?;
                        let up_w = quantize_linear(&ew.up_w, h, moe_int)
                            .map_err(|e| layer_err(idx, &format!("ffn_up expert {ei}"), e))?;
                        let down_w = quantize_linear(&ew.down_w, moe_int, h)
                            .map_err(|e| layer_err(idx, &format!("ffn_down expert {ei}"), e))?;
                        q_experts.push(Qwen3MoeExpertWeights {
                            gate_w,
                            up_w,
                            down_w,
                        });
                    }
                    Qwen3MoeFfn::Moe {
                        router_w,
                        experts: q_experts,
                    }
                }
            };

            layers.push(Qwen3MoeLayerWeights {
                attn_norm_gain: layer.attn_norm_gain,
                ffn_norm_gain: layer.ffn_norm_gain,
                attn_q,
                attn_q_bias: layer.attn_q_bias,
                attn_k,
                attn_k_bias: layer.attn_k_bias,
                attn_v,
                attn_v_bias: layer.attn_v_bias,
                attn_o,
                q_norm_gain: layer.q_norm_gain,
                k_norm_gain: layer.k_norm_gain,
                ffn,
            });
        }

        let output = quantize_linear(&src.output, h, cfg.vocab_size)
            .map_err(|e| fuel_ir::error::Error::Msg(format!("output: {e}")).bt())?;

        let inner = Qwen3MoeModel {
            config: cfg,
            weights: Qwen3MoeWeights {
                instance: fuel_tensor::decode_shape::ModelInstanceId::next(),
                token_embedding: src.token_embedding,
                layers,
                final_norm_gain: src.final_norm_gain,
                output,
            },
        };
        Ok(Self { inner })
    }

    /// Load a GGUF-quantized Qwen3-MoE checkpoint. Q4_0 weight tensors
    /// stay quantized; norms / biases / embeddings / the router gate
    /// dequantize to F32. Other GGML dtypes for Linear weights are
    /// dequantized to F32 — matching the Phi-2 / SmolLM3 loader
    /// policy. Embedding and `lm_head` share storage when
    /// `output.weight` is absent (tied embeddings).
    pub fn from_gguf<P: AsRef<std::path::Path>>(path: P, cfg: &Qwen3MoeConfig) -> Result<Self> {
        use fuel_loaders::quantized::gguf_mmap::MmapedContent;
        let mc = MmapedContent::from_path(path)?;
        let content = mc.content();
        let mmap_arc = mc.mmap();
        let mmap_bytes: &[u8] = &mmap_arc[..];
        let data_off = content.tensor_data_offset as usize;

        let get_tensor_bytes =
            |name: &str| -> Result<(&[u8], fuel_loaders::quantized::GgmlDType, Vec<usize>)> {
                let info = content.tensor_infos.get(name).ok_or_else(|| {
                    fuel_ir::error::Error::Msg(format!("gguf: missing tensor {name:?}"))
                })?;
                let elems = info.shape.elem_count();
                let block_size = info.ggml_dtype.block_size();
                let bytes_len = elems / block_size * info.ggml_dtype.type_size();
                let start = data_off + info.offset as usize;
                Ok((
                    &mmap_bytes[start..start + bytes_len],
                    info.ggml_dtype,
                    info.shape.dims().to_vec(),
                ))
            };

        let load_f32 = |name: &str| -> Result<Vec<f32>> {
            let (bytes, dt, _) = get_tensor_bytes(name)?;
            dequant_bytes_to_f32(bytes, dt, name)
        };

        let load_weight = |name: &str,

                           out_features: usize,
                           in_features: usize|

         -> Result<WeightStorage> {
            let (bytes, dt, dims) = get_tensor_bytes(name)?;
            let expected = out_features * in_features;
            let actual: usize = dims.iter().product();
            if actual != expected {
                return Err(fuel_ir::error::Error::Msg(format!(
                    "gguf: tensor {name:?} has {actual} elements, expected {expected} for [{out_features}, {in_features}]",
                )).bt());
            }
            match dt {
                fuel_loaders::quantized::GgmlDType::Q4_0 => Ok(WeightStorage::Q4_0 {
                    words: bytes_to_u32_arc(bytes),
                    bytes_len: bytes.len(),
                    in_features,
                    out_features,
                }),
                _ => {
                    // Dequantize to F32 and transpose [out, in] → [in, out].
                    let f32_out_in = dequant_bytes_to_f32(bytes, dt, name)?;
                    let mut f32_in_out = vec![0.0_f32; expected];
                    for o in 0..out_features {
                        for j in 0..in_features {
                            f32_in_out[j * out_features + o] = f32_out_in[o * in_features + j];
                        }
                    }
                    Ok(WeightStorage::F32(Arc::from(f32_in_out)))
                }
            }
        };

        let h = cfg.hidden_size;
        let inter = cfg.intermediate_size;
        let moe_int = cfg.moe_intermediate_size;
        let kv = cfg.num_key_value_heads * cfg.head_dim;
        let q_dim = cfg.num_attention_heads * cfg.head_dim;

        let token_embedding = load_f32("token_embd.weight")?;
        if token_embedding.len() != cfg.vocab_size * h {
            return Err(fuel_ir::error::Error::Msg(format!(
                "gguf token_embd.weight: {} elems, expected {}×{}",
                token_embedding.len(),
                cfg.vocab_size,
                h,
            ))
            .bt());
        }

        let mut layers: Vec<Qwen3MoeLayerWeights> = Vec::with_capacity(cfg.num_hidden_layers);
        for idx in 0..cfg.num_hidden_layers {
            let prefix = format!("blk.{idx}");

            let attn_q = load_weight(&format!("{prefix}.attn_q.weight"), q_dim, h)?;
            let attn_k = load_weight(&format!("{prefix}.attn_k.weight"), kv, h)?;
            let attn_v = load_weight(&format!("{prefix}.attn_v.weight"), kv, h)?;
            let attn_o = load_weight(&format!("{prefix}.attn_output.weight"), h, q_dim)?;

            let attn_norm_gain: Arc<[f32]> =
                Arc::from(load_f32(&format!("{prefix}.attn_norm.weight"))?);
            let ffn_norm_gain: Arc<[f32]> =
                Arc::from(load_f32(&format!("{prefix}.ffn_norm.weight"))?);

            // Per-head QK-norm gains. llama.cpp Qwen3 GGUF uses
            // `attn_q_norm.weight` / `attn_k_norm.weight` for the
            // per-head RMSNorm gains living on a `head_dim` slice.
            let q_norm_gain: Arc<[f32]> =
                Arc::from(load_f32(&format!("{prefix}.attn_q_norm.weight"))?);
            let k_norm_gain: Arc<[f32]> =
                Arc::from(load_f32(&format!("{prefix}.attn_k_norm.weight"))?);

            let bias = |name: &str, len: usize| -> Option<Arc<[f32]>> {
                load_f32(name).ok().and_then(|v| {
                    if v.len() == len {
                        Some(Arc::from(v))
                    } else {
                        None
                    }
                })
            };
            let attn_q_bias = if cfg.attention_bias {
                bias(&format!("{prefix}.attn_q.bias"), q_dim)
            } else {
                None
            };
            let attn_k_bias = if cfg.attention_bias {
                bias(&format!("{prefix}.attn_k.bias"), kv)
            } else {
                None
            };
            let attn_v_bias = if cfg.attention_bias {
                bias(&format!("{prefix}.attn_v.bias"), kv)
            } else {
                None
            };

            let ffn = if cfg.layer_uses_moe(idx) {
                // Router gate: `[num_experts, hidden]` in llama.cpp
                // convention; transpose to `[hidden, num_experts]`
                // for our matmul layout.
                let router_out_in = load_f32(&format!("{prefix}.ffn_gate_inp.weight"))?;
                if router_out_in.len() != cfg.num_experts * h {
                    return Err(fuel_ir::error::Error::Msg(format!(
                        "gguf {prefix}.ffn_gate_inp.weight: {} elems, expected {}×{}",
                        router_out_in.len(),
                        cfg.num_experts,
                        h,
                    ))
                    .bt());
                }
                let mut router_in_out = vec![0.0_f32; h * cfg.num_experts];
                for e in 0..cfg.num_experts {
                    for j in 0..h {
                        router_in_out[j * cfg.num_experts + e] = router_out_in[e * h + j];
                    }
                }
                let router_w: Arc<[f32]> = Arc::from(router_in_out);

                let mut experts: Vec<Qwen3MoeExpertWeights> = Vec::with_capacity(cfg.num_experts);
                for e in 0..cfg.num_experts {
                    let gate_w = load_weight(&format!("{prefix}.ffn_gate.{e}.weight"), moe_int, h)?;
                    let up_w = load_weight(&format!("{prefix}.ffn_up.{e}.weight"), moe_int, h)?;
                    let down_w = load_weight(&format!("{prefix}.ffn_down.{e}.weight"), h, moe_int)?;
                    experts.push(Qwen3MoeExpertWeights {
                        gate_w,
                        up_w,
                        down_w,
                    });
                }
                Qwen3MoeFfn::Moe { router_w, experts }
            } else {
                let gate_w = load_weight(&format!("{prefix}.ffn_gate.weight"), inter, h)?;
                let up_w = load_weight(&format!("{prefix}.ffn_up.weight"), inter, h)?;
                let down_w = load_weight(&format!("{prefix}.ffn_down.weight"), h, inter)?;
                Qwen3MoeFfn::Dense {
                    gate_w,
                    up_w,
                    down_w,
                }
            };

            layers.push(Qwen3MoeLayerWeights {
                attn_norm_gain,
                ffn_norm_gain,
                attn_q,
                attn_q_bias,
                attn_k,
                attn_k_bias,
                attn_v,
                attn_v_bias,
                attn_o,
                q_norm_gain,
                k_norm_gain,
                ffn,
            });
        }

        let final_norm_gain: Arc<[f32]> = Arc::from(load_f32("output_norm.weight")?);

        // Tied embeddings: Qwen3-MoE may reuse the embedding for
        // `lm_head`. If `output.weight` is present we honor it;
        // otherwise we synthesize an F32 lm_head from the
        // (already-dequantized) token embedding by transposing
        // [vocab, h] → [h, vocab].
        let output = if content.tensor_infos.contains_key("output.weight") {
            load_weight("output.weight", cfg.vocab_size, h)?
        } else {
            let mut f32_in_out = vec![0.0_f32; h * cfg.vocab_size];
            for v in 0..cfg.vocab_size {
                for j in 0..h {
                    f32_in_out[j * cfg.vocab_size + v] = token_embedding[v * h + j];
                }
            }
            WeightStorage::F32(Arc::from(f32_in_out))
        };

        let inner = Qwen3MoeModel {
            config: cfg.clone(),
            weights: Qwen3MoeWeights {
                instance: fuel_tensor::decode_shape::ModelInstanceId::next(),
                token_embedding: Arc::from(token_embedding),
                layers,
                final_norm_gain,
                output,
            },
        };
        Ok(Self { inner })
    }
}

fn layer_err(idx: usize, name: &str, e: fuel_ir::error::Error) -> fuel_ir::error::Error {
    fuel_ir::error::Error::Msg(format!("layer {idx} {name}: {e}")).bt()
}

fn check_q4_0_divisible(name: &str, n: usize) -> Result<()> {
    const QK4_0: usize = 32;
    if !n.is_multiple_of(QK4_0) {
        return Err(fuel_ir::error::Error::Msg(format!(
            "QuantizedQwen3MoeModel::from_f32_bake: {name} ({n}) must be divisible by Q4_0 block size ({QK4_0})"
        )).bt());
    }
    Ok(())
}

/// Quantize a `[in_features, out_features]` row-major F32 weight matrix
/// into a `WeightStorage::Q4_0` keeping GGUF's native `[out, in]` block
/// layout. The implementation does the `[in, out] → [out, in]` transpose
/// first, then runs the per-row Q4_0 quantization.
fn quantize_in_out_to_q4_0(
    f32_in_out: &[f32],
    in_features: usize,
    out_features: usize,
) -> Result<WeightStorage> {
    use fuel_quantized::{BlockQ4_0, GgmlType};
    const QK4_0: usize = 32;
    if !in_features.is_multiple_of(QK4_0) {
        return Err(fuel_ir::error::Error::Msg(format!(
            "Q4_0 quantize: in_features ({in_features}) must be divisible by {QK4_0}"
        ))
        .bt());
    }

    // Transpose [in, out] → [out, in] so each row is contiguous in K.
    let mut f32_out_in = vec![0.0_f32; out_features * in_features];
    for o in 0..out_features {
        for j in 0..in_features {
            f32_out_in[o * in_features + j] = f32_in_out[j * out_features + o];
        }
    }

    let n_blocks = out_features * in_features / QK4_0;
    let mut blocks: Vec<BlockQ4_0> = vec![BlockQ4_0::zeros(); n_blocks];
    BlockQ4_0::from_float(&f32_out_in, &mut blocks);

    // BlockQ4_0 is repr(C) and exactly 18 bytes; reinterpret as bytes.
    let bytes_len = n_blocks * std::mem::size_of::<BlockQ4_0>();
    let byte_slice: &[u8] = fuel_quantized::blocks_as_bytes(&blocks);
    // Q4_0 storage holds u32 words; pad bytes to a multiple of 4 by
    // copying into a Vec<u8> first.
    let padded_len = bytes_len.div_ceil(4) * 4;
    let mut padded = vec![0_u8; padded_len];
    padded[..bytes_len].copy_from_slice(byte_slice);
    let words: Vec<u32> = padded
        .as_chunks::<4>()
        .0
        .iter()
        .map(|c| u32::from_le_bytes([c[0], c[1], c[2], c[3]]))
        .collect();
    Ok(WeightStorage::Q4_0 {
        words: Arc::from(words),
        bytes_len,
        in_features,
        out_features,
    })
}

fn bytes_to_u32_arc(bytes: &[u8]) -> Arc<[u32]> {
    let padded_len = bytes.len().div_ceil(4) * 4;
    let mut padded = vec![0_u8; padded_len];
    padded[..bytes.len()].copy_from_slice(bytes);
    let words: Vec<u32> = padded
        .as_chunks::<4>()
        .0
        .iter()
        .map(|c| u32::from_le_bytes([c[0], c[1], c[2], c[3]]))
        .collect();
    Arc::from(words)
}

/// Dequantize a raw GGUF byte slice of the given GGML dtype to F32.
/// Kept private to this module so this file stays self-contained per
/// the existing `lazy_quantized_*` convention.
fn dequant_bytes_to_f32(
    bytes: &[u8],
    dt: fuel_loaders::quantized::GgmlDType,
    name: &str,
) -> Result<Vec<f32>> {
    fuel_quantized::dequant_ggml_bytes(bytes, dt, name)
}

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

    fn test_cfg() -> Qwen3MoeConfig {
        // All Q4_0-quantized dimensions (hidden_size, intermediate_size,
        // moe_intermediate_size, kv_dim) must be multiples of 32.
        // decoder_sparse_step = 2 → layers 1 and 3 (0-indexed) use MoE.
        Qwen3MoeConfig {
            vocab_size: 32,
            hidden_size: 32,
            intermediate_size: 64,
            moe_intermediate_size: 32,
            num_experts: 2,
            num_experts_per_tok: 1,
            decoder_sparse_step: 2,
            num_hidden_layers: 4,
            num_attention_heads: 4,
            num_key_value_heads: 4,
            head_dim: 8,
            max_position_embeddings: 64,
            sliding_window: None,
            max_window_layers: 0,
            use_sliding_window: false,
            rope_theta: 10_000.0,
            rms_norm_eps: 1e-5,
            attention_bias: false,
        }
    }

    fn tiny_weights(cfg: &Qwen3MoeConfig) -> Qwen3MoeWeights {
        let mut s: u32 = 24680;
        let mut next = move || -> f32 {
            s = s.wrapping_mul(1103515245).wrapping_add(12345);
            ((s >> 16) as u16 as f32 / 65535.0 - 0.5) * 0.05
        };
        let mut vec_of =
            |n: usize| -> Arc<[f32]> { Arc::from((0..n).map(|_| next()).collect::<Vec<_>>()) };
        let h = cfg.hidden_size;
        let inter = cfg.intermediate_size;
        let moe_int = cfg.moe_intermediate_size;
        let kv = cfg.num_key_value_heads * cfg.head_dim;
        let q_dim = cfg.num_attention_heads * cfg.head_dim;
        let token_embedding = vec_of(cfg.vocab_size * h);
        let layers: Vec<Qwen3MoeLayerWeights> = (0..cfg.num_hidden_layers)
            .map(|li| {
                let ffn = if cfg.layer_uses_moe(li) {
                    let router_w = vec_of(h * cfg.num_experts);
                    let experts: Vec<Qwen3MoeExpertWeights> = (0..cfg.num_experts)
                        .map(|_| Qwen3MoeExpertWeights {
                            gate_w: WeightStorage::F32(vec_of(h * moe_int)),
                            up_w: WeightStorage::F32(vec_of(h * moe_int)),
                            down_w: WeightStorage::F32(vec_of(moe_int * h)),
                        })
                        .collect();
                    Qwen3MoeFfn::Moe { router_w, experts }
                } else {
                    Qwen3MoeFfn::Dense {
                        gate_w: WeightStorage::F32(vec_of(h * inter)),
                        up_w: WeightStorage::F32(vec_of(h * inter)),
                        down_w: WeightStorage::F32(vec_of(inter * h)),
                    }
                };
                Qwen3MoeLayerWeights {
                    attn_norm_gain: Arc::from(vec![1.0_f32; h]),
                    ffn_norm_gain: Arc::from(vec![1.0_f32; h]),
                    attn_q: WeightStorage::F32(vec_of(h * q_dim)),
                    attn_q_bias: None,
                    attn_k: WeightStorage::F32(vec_of(h * kv)),
                    attn_k_bias: None,
                    attn_v: WeightStorage::F32(vec_of(h * kv)),
                    attn_v_bias: None,
                    attn_o: WeightStorage::F32(vec_of(q_dim * h)),
                    q_norm_gain: Arc::from(vec![1.0_f32; cfg.head_dim]),
                    k_norm_gain: Arc::from(vec![1.0_f32; cfg.head_dim]),
                    ffn,
                }
            })
            .collect();
        let final_norm_gain = Arc::from(vec![1.0_f32; h]);
        let output = WeightStorage::F32(vec_of(h * cfg.vocab_size));
        Qwen3MoeWeights {
            instance: fuel_tensor::decode_shape::ModelInstanceId::next(),
            token_embedding,
            layers,
            final_norm_gain,
            output,
        }
    }

    /// **GAP-029 increment 3 — the quantized wrapper's decode delegation.**
    ///
    /// The module's `test_cfg()` is deliberately dense (`sliding_window: None`),
    /// so this overrides it to a MIXED window over 4 layers with
    /// `decoder_sparse_step: 2` — layers 0 and 1 windowed, 2 and 3 dense, with
    /// MoE on layers 1 and 3. That puts a windowed MoE layer, a windowed dense
    /// layer, and their unwindowed counterparts in ONE decode graph, over Q4_0
    /// storage. Three decode steps, so the assertions reach the REBIND path.
    #[test]
    fn quantized_qwen3_moe_windowed_decode_matches_quantized_forward() {
        use fuel_ir::DType;
        use fuel_tensor::Device;
        use fuel_tensor::{
            inference_context::DecodeSession, inference_context::InferenceContext,
            inference_context::KvCache,
        };

        let cfg = Qwen3MoeConfig {
            sliding_window: Some(4),
            max_window_layers: 2,
            use_sliding_window: true,
            ..test_cfg()
        };
        assert!(
            !cfg.layer_uses_moe(0) && cfg.layer_uses_moe(1),
            "the windowed span must cover both a dense and a MoE layer"
        );
        let model = QuantizedQwen3MoeModel::from_f32_bake(cfg.clone(), tiny_weights(&cfg)).unwrap();
        let tokens: Vec<u32> = vec![1, 2, 3, 4, 5, 6];
        assert!(
            tokens.len() > 4,
            "non-vacuity: the window must actually bite"
        );
        let prefill = 3;

        let dev = Device::cpu();
        let mut cache = KvCache::with_capacity(
            cfg.num_hidden_layers,
            cfg.num_key_value_heads,
            cfg.head_dim,
            tokens.len(),
            DType::F32,
            &dev,
        )
        .expect("with_capacity");
        let mut ctx = InferenceContext::new(dev);
        let mut session: Option<DecodeSession> = None;

        model
            .forward_with_kv_context_persistent(
                &tokens[..prefill],
                &mut cache,
                &mut ctx,
                &mut session,
            )
            .expect("prefill");

        for pos in prefill..tokens.len() {
            let got = model
                .forward_with_kv_context_persistent(
                    &tokens[pos..=pos],
                    &mut cache,
                    &mut ctx,
                    &mut session,
                )
                .expect("decode");
            let full = model.forward(&tokens[..=pos], 0).unwrap().realize_f32();
            let expected = &full[pos * cfg.vocab_size..(pos + 1) * cfg.vocab_size];
            let worst = got
                .iter()
                .zip(expected.iter())
                .map(|(a, b)| (a - b).abs())
                .fold(0.0_f32, f32::max);
            assert!(
                worst < 1e-5,
                "quantized Qwen3Moe windowed decode at position {pos} diverged by {worst}",
            );
        }
        assert_eq!(cache.cached_len, tokens.len());
    }

    #[test]
    fn forward_shape_finite_with_q4_0_weights() {
        let cfg = test_cfg();
        let src = tiny_weights(&cfg);
        let model = QuantizedQwen3MoeModel::from_f32_bake(cfg.clone(), src).unwrap();

        // Layer 0 is dense (decoder_sparse_step = 2, (0+1)%2 != 0).
        let l0 = &model.inner().weights.layers[0];
        assert!(matches!(l0.attn_q, WeightStorage::Q4_0 { .. }));
        assert!(matches!(l0.attn_k, WeightStorage::Q4_0 { .. }));
        assert!(matches!(l0.attn_v, WeightStorage::Q4_0 { .. }));
        assert!(matches!(l0.attn_o, WeightStorage::Q4_0 { .. }));
        match &l0.ffn {
            Qwen3MoeFfn::Dense {
                gate_w,
                up_w,
                down_w,
            } => {
                assert!(matches!(gate_w, WeightStorage::Q4_0 { .. }));
                assert!(matches!(up_w, WeightStorage::Q4_0 { .. }));
                assert!(matches!(down_w, WeightStorage::Q4_0 { .. }));
            }
            _ => panic!("expected layer 0 to be Dense"),
        }

        // Layer 1 is MoE.
        let l1 = &model.inner().weights.layers[1];
        match &l1.ffn {
            Qwen3MoeFfn::Moe { router_w, experts } => {
                assert_eq!(router_w.len(), cfg.hidden_size * cfg.num_experts);
                assert_eq!(experts.len(), cfg.num_experts);
                for ew in experts {
                    assert!(matches!(ew.gate_w, WeightStorage::Q4_0 { .. }));
                    assert!(matches!(ew.up_w, WeightStorage::Q4_0 { .. }));
                    assert!(matches!(ew.down_w, WeightStorage::Q4_0 { .. }));
                }
            }
            _ => panic!("expected layer 1 to be MoE"),
        }

        assert!(matches!(
            model.inner().weights.output,
            WeightStorage::Q4_0 { .. }
        ));

        let logits = model.forward(&[1, 2, 3], 0).unwrap();
        assert_eq!(logits.shape().dims(), &[1, 3, cfg.vocab_size]);
        for &v in &logits.realize_f32() {
            assert!(v.is_finite(), "non-finite logit: {v}");
        }
    }

    #[test]
    fn forward_embeds_matches_forward_after_token_lookup() {
        let cfg = test_cfg();
        let src = tiny_weights(&cfg);
        let model = QuantizedQwen3MoeModel::from_f32_bake(cfg.clone(), src).unwrap();
        let tokens: Vec<u32> = vec![1, 2, 3];
        let logits_ref = model.forward(&tokens, 0).unwrap().realize_f32();
        let anchor =
            Tensor::from_f32(vec![0.0_f32], Shape::from_dims(&[1]), &Device::cpu()).unwrap();
        let embeds = model.embed_tokens_anchored(&anchor, &tokens).unwrap();
        let logits_via_embeds = model.forward_embeds(&embeds, 0).unwrap().realize_f32();
        let max_diff = logits_ref
            .iter()
            .zip(logits_via_embeds.iter())
            .map(|(a, b)| (a - b).abs())
            .fold(0.0_f32, f32::max);
        assert!(
            max_diff < 1e-4,
            "Quantized Qwen3-MoE forward vs forward_embeds must agree (max diff {max_diff})"
        );
    }
}