fuel-transformers 0.20.0

// SPDX-License-Identifier: MIT OR Apache-2.0
//! GGUF-quantized Gemma 3 ported to the lazy-graph API.
//!
//! Gemma 3 with Q4_0 block-quantized Linear weights. The forward path is
//! identical to [`crate::models::lazy_gemma3::Gemma3Model`]; only the weight
//! storage of `attn_{q,k,v,o}` / `ffn_{gate,up,down}` changes from
//! F32/BF16 to `WeightStorage::Q4_0`. All norm gains (input, post-attn,
//! pre-FFN, post-FFN, per-head Q/K, final), biases (when present), and
//! the token-embedding table stay in F32. Gemma 3 ties `lm_head` to
//! `token_embedding`, so there is no separate output projection to
//! quantize.
//!
//! Gemma-3-specific structural notes the loader honors:
//! - Independent attention/embedding dims: `q_dim = num_attention_heads
//!   * head_dim` is **not** required to equal `hidden_size`. The Q
//!   projection is `[q_dim, hidden_size]`; o_proj inverts that to
//!   `[hidden_size, q_dim]`. K/V project to `kv_dim = num_key_value_heads
//!   * head_dim`.
//! - Per-head Q/K RmsNorm on `head_dim` (offset `(gain + 1)`) sits
//!   between the QK projection split and RoPE — these gains stay F32.
//! - Tied `lm_head`: the GGUF loader synthesizes the output matrix from
//!   the token embedding when `output.weight` is absent (the common case
//!   for Gemma releases).
//!
//! Construction paths:
//! - [`QuantizedGemma3Model::load_from_mmapped`] — convenience wrapper
//!   over [`Gemma3Weights::load_from_mmapped`] + [`QuantizedGemma3Model::from_f32_bake`].
//! - [`QuantizedGemma3Model::from_f32_bake`] — take f32 source weights
//!   (same `[in_features, out_features]` layout as `Gemma3Weights`) and
//!   quantize each Linear weight to Q4_0 on the fly. Used by tests and
//!   by callers that have unquantized weights in memory already.
//! - [`QuantizedGemma3Model::from_gguf`] — load directly from a Gemma 3
//!   GGUF file, keeping Q4_0 tensors quantized and dequantizing other
//!   GGML dtypes for Linear weights to F32. Mirrors the SmolLM3 loader
//!   policy.

use crate::models::lazy_gemma3::{Gemma3Config, Gemma3LayerWeights, Gemma3Model, Gemma3Weights};
use fuel_ir::error::Result;
use fuel_tensor::{lazy::Tensor, lazy::WeightStorage};
use std::sync::Arc;

/// GGUF-quantized Gemma 3 causal language model. Wraps a plain
/// [`Gemma3Model`] whose Linear weights are `WeightStorage::Q4_0`. The
/// tied `lm_head` continues to share storage with `token_embedding`
/// (kept in F32), matching the eager Gemma 3 convention.
#[derive(Debug, Clone)]
pub struct QuantizedGemma3Model {
    inner: Gemma3Model,
}

impl QuantizedGemma3Model {
    /// 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). The caller is responsible for the
    /// `sqrt(hidden_size)` scaling — Gemma 3 expects pre-scaled
    /// embeddings here.
    pub fn forward_embeds(&self, scaled_embeds: &Tensor, start_pos: usize) -> Result<Tensor> {
        self.inner.forward_embeds(scaled_embeds, start_pos)
    }

    /// Per-token hidden states up to the final offset 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,
        scaled_embeds: &Tensor,
        start_pos: usize,
    ) -> Result<Tensor> {
        self.inner.forward_hidden_embeds(scaled_embeds, start_pos)
    }

    /// Build per-token embeddings without running the decoder. Used by
    /// multimodal compositions to obtain text-side embeddings that will
    /// be spliced with vision features before [`Self::forward_embeds`].
    /// The caller is responsible for the `sqrt(hidden_size)` scaling.
    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).
    ///
    /// **Delegates ONE level** — `Gemma3Model` is a leaf (`config` + `weights`),
    /// so the `inner.inner` trap that cost `QuantizedLlama3Model` its RoPE
    /// scaling has no shape to take here.
    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) -> &Gemma3Config {
        &self.inner.config
    }

    /// Underlying [`Gemma3Model`] for direct access to the lazy graph
    /// API. The wrapper exists solely to label the quantization origin.
    pub fn inner(&self) -> &Gemma3Model {
        &self.inner
    }

    /// Convenience: load f32 Gemma 3 weights from HF safetensors and
    /// quantize each Linear weight to Q4_0. Equivalent to
    /// `Self::from_f32_bake(cfg, Gemma3Weights::load_from_mmapped(st, &cfg)?)`.
    pub fn load_from_mmapped(
        st: &fuel_loaders::safetensors::MmapedSafetensors,
        cfg: Gemma3Config,
    ) -> Result<Self> {
        let src = Gemma3Weights::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, and the
    /// token-embedding table stay in F32 — matching the GGUF convention
    /// used by llama.cpp for Gemma 3 releases.
    ///
    /// Q4_0 requires `in_features % 32 == 0` for every quantized
    /// matrix. The relevant Gemma-3 dims are checked up front:
    /// - `attn_q` / `ffn_gate` / `ffn_up`: in = `hidden_size`
    /// - `attn_k` / `attn_v`: in = `hidden_size`
    /// - `attn_o`: in = `q_dim = num_attention_heads * head_dim`
    /// - `ffn_down`: in = `intermediate_size`
    ///
    /// Source weights follow the same `[in_features, out_features]`
    /// row-major layout as [`Gemma3Weights`].
    pub fn from_f32_bake(cfg: Gemma3Config, src: Gemma3Weights) -> Result<Self> {
        let h = cfg.hidden_size;
        let i_dim = cfg.intermediate_size;
        let q_dim = cfg.num_attention_heads * cfg.head_dim;
        let kv_dim = cfg.num_key_value_heads * cfg.head_dim;
        // attn_q in=h; attn_o in=q_dim; ffn_down in=i_dim. K/V also in=h.
        check_q4_0_divisible("hidden_size (attn_q/k/v/ffn_gate/ffn_up in-features)", h)?;
        check_q4_0_divisible("num_attention_heads * head_dim (attn_o in-features)", q_dim)?;
        check_q4_0_divisible("intermediate_size (ffn_down in-features)", i_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(
                    "QuantizedGemma3Model::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!(
                    "QuantizedGemma3Model::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<Gemma3LayerWeights> = 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_dim)
                .map_err(|e| layer_err(idx, "attn_k", e))?;
            let attn_v = quantize_linear(&layer.attn_v, h, kv_dim)
                .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_gate = quantize_linear(&layer.ffn_gate, h, i_dim)
                .map_err(|e| layer_err(idx, "ffn_gate", e))?;
            let ffn_up = quantize_linear(&layer.ffn_up, h, i_dim)
                .map_err(|e| layer_err(idx, "ffn_up", e))?;
            let ffn_down = quantize_linear(&layer.ffn_down, i_dim, h)
                .map_err(|e| layer_err(idx, "ffn_down", e))?;
            layers.push(Gemma3LayerWeights {
                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,
                attn_o_bias: layer.attn_o_bias,
                q_norm_gain: layer.q_norm_gain,
                k_norm_gain: layer.k_norm_gain,
                input_norm_gain: layer.input_norm_gain,
                post_attn_norm_gain: layer.post_attn_norm_gain,
                pre_ffn_norm_gain: layer.pre_ffn_norm_gain,
                post_ffn_norm_gain: layer.post_ffn_norm_gain,
                ffn_gate,
                ffn_up,
                ffn_down,
            });
        }

        let inner = Gemma3Model {
            config: cfg,
            weights: Gemma3Weights {
                instance: fuel_tensor::decode_shape::ModelInstanceId::next(),
                token_embedding: src.token_embedding,
                layers,
                final_norm_gain: src.final_norm_gain,
            },
        };
        Ok(Self { inner })
    }

    /// Load a GGUF-quantized Gemma 3 checkpoint. Q4_0 weight tensors
    /// stay quantized; norms / biases / embeddings dequantize to F32.
    /// Other GGML dtypes for Linear weights are dequantized to F32 —
    /// matching the SmolLM3 loader policy. Because Gemma 3 ties
    /// `lm_head` to the token embedding, no `output.weight` lookup is
    /// performed — the embedding table is the lm_head.
    ///
    /// GGUF tensor names follow the llama.cpp convention:
    ///   - `token_embd.weight`       → `token_embedding`
    ///   - `output_norm.weight`      → `final_norm_gain`
    ///   - `blk.{i}.attn_q.weight`   → `attn_q` (also `.bias` if present)
    ///   - `blk.{i}.attn_k.weight`   → `attn_k` (also `.bias` if present)
    ///   - `blk.{i}.attn_v.weight`   → `attn_v` (also `.bias` if present)
    ///   - `blk.{i}.attn_output.weight` → `attn_o` (also `.bias` if present)
    ///   - `blk.{i}.attn_q_norm.weight` → `q_norm_gain` (per-head, head_dim)
    ///   - `blk.{i}.attn_k_norm.weight` → `k_norm_gain` (per-head, head_dim)
    ///   - `blk.{i}.attn_norm.weight`            → `input_norm_gain`
    ///   - `blk.{i}.post_attention_norm.weight`  → `post_attn_norm_gain`
    ///   - `blk.{i}.ffn_norm.weight`             → `pre_ffn_norm_gain`
    ///   - `blk.{i}.post_ffw_norm.weight`        → `post_ffn_norm_gain`
    ///   - `blk.{i}.ffn_gate.weight` → `ffn_gate`
    ///   - `blk.{i}.ffn_up.weight`   → `ffn_up`
    ///   - `blk.{i}.ffn_down.weight` → `ffn_down`
    pub fn from_gguf<P: AsRef<std::path::Path>>(path: P, cfg: &Gemma3Config) -> 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 i_dim = cfg.intermediate_size;
        let q_dim = cfg.num_attention_heads * cfg.head_dim;
        let kv_dim = cfg.num_key_value_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<Gemma3LayerWeights> = 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_dim, h)?;
            let attn_v = load_weight(&format!("{prefix}.attn_v.weight"), kv_dim, h)?;
            let attn_o = load_weight(&format!("{prefix}.attn_output.weight"), h, q_dim)?;
            let ffn_gate = load_weight(&format!("{prefix}.ffn_gate.weight"), i_dim, h)?;
            let ffn_up = load_weight(&format!("{prefix}.ffn_up.weight"), i_dim, h)?;
            let ffn_down = load_weight(&format!("{prefix}.ffn_down.weight"), h, i_dim)?;

            // Per-head Q/K RmsNorm on head_dim — present on every Gemma 3 layer.
            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"))?);

            // Four block-level norms (input, post-attn, pre-FFN, post-FFN).
            let input_norm_gain: Arc<[f32]> =
                Arc::from(load_f32(&format!("{prefix}.attn_norm.weight"))?);
            let post_attn_norm_gain: Arc<[f32]> =
                Arc::from(load_f32(&format!("{prefix}.post_attention_norm.weight"))?);
            let pre_ffn_norm_gain: Arc<[f32]> =
                Arc::from(load_f32(&format!("{prefix}.ffn_norm.weight"))?);
            let post_ffn_norm_gain: Arc<[f32]> =
                Arc::from(load_f32(&format!("{prefix}.post_ffw_norm.weight"))?);

            // Biases are optional and only present when attention_bias is set
            // upstream; honor whatever the file ships.
            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
                    }
                })
            };

            layers.push(Gemma3LayerWeights {
                attn_q,
                attn_q_bias: bias(&format!("{prefix}.attn_q.bias"), q_dim),
                attn_k,
                attn_k_bias: bias(&format!("{prefix}.attn_k.bias"), kv_dim),
                attn_v,
                attn_v_bias: bias(&format!("{prefix}.attn_v.bias"), kv_dim),
                attn_o,
                attn_o_bias: bias(&format!("{prefix}.attn_output.bias"), h),
                q_norm_gain,
                k_norm_gain,
                input_norm_gain,
                post_attn_norm_gain,
                pre_ffn_norm_gain,
                post_ffn_norm_gain,
                ffn_gate,
                ffn_up,
                ffn_down,
            });
        }

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

        // Gemma 3 has tied embeddings — there is no `output.weight` to
        // load (and no separate output field on Gemma3Weights). The
        // forward path reuses `token_embedding` as the lm_head.
        let _ = kv_dim; // suppress dead-binding warning if no caller path uses it post-bias

        let inner = Gemma3Model {
            config: cfg.clone(),
            weights: Gemma3Weights {
                instance: fuel_tensor::decode_shape::ModelInstanceId::next(),
                token_embedding: Arc::from(token_embedding),
                layers,
                final_norm_gain,
            },
        };
        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!(
            "QuantizedGemma3Model::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.
/// Mirrors the SmolLM3 helper but lives here so this module stays
/// independent of the SmolLM3 internals.
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::*;
    use crate::models::lazy_gemma3::GemmaActivation;
    use fuel_ir::Shape;
    use fuel_tensor::Device;

    fn test_cfg() -> Gemma3Config {
        // Pick num_heads * head_dim != hidden_size to exercise
        // independent attention/embedding dims like real Gemma3.
        // All Q4_0 in-features must be multiples of 32:
        //   hidden_size = 32, q_dim = 32, intermediate_size = 64.
        Gemma3Config {
            vocab_size: 32,
            hidden_size: 32,
            intermediate_size: 64,
            num_hidden_layers: 4,
            num_attention_heads: 4,
            num_key_value_heads: 2,
            head_dim: 8, // q_dim = 32, kv_dim = 16
            rms_norm_eps: 1e-6,
            rope_theta: 10_000.0,
            rope_local_base_freq: 10_000.0,
            max_position_embeddings: 64,
            sliding_window: 3,
            sliding_window_pattern: 3,
            attention_bias: false,
            hidden_activation: GemmaActivation::GeluPytorchTanh,
            attn_logit_softcapping: Some(50.0),
            final_logit_softcapping: Some(30.0),
        }
    }

    fn tiny_weights(cfg: &Gemma3Config) -> Gemma3Weights {
        let mut s: u32 = 5151;
        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 i_dim = cfg.intermediate_size;
        let q_dim = cfg.num_attention_heads * cfg.head_dim;
        let kv_dim = cfg.num_key_value_heads * cfg.head_dim;
        let token_embedding = vec_of(cfg.vocab_size * h);
        let layers: Vec<Gemma3LayerWeights> = (0..cfg.num_hidden_layers)
            .map(|_| Gemma3LayerWeights {
                attn_q: WeightStorage::F32(vec_of(h * q_dim)),
                attn_q_bias: None,
                attn_k: WeightStorage::F32(vec_of(h * kv_dim)),
                attn_k_bias: None,
                attn_v: WeightStorage::F32(vec_of(h * kv_dim)),
                attn_v_bias: None,
                attn_o: WeightStorage::F32(vec_of(q_dim * h)),
                attn_o_bias: None,
                q_norm_gain: Arc::from(vec![0.05_f32; cfg.head_dim]),
                k_norm_gain: Arc::from(vec![0.05_f32; cfg.head_dim]),
                input_norm_gain: Arc::from(vec![0.05_f32; h]),
                post_attn_norm_gain: Arc::from(vec![0.05_f32; h]),
                pre_ffn_norm_gain: Arc::from(vec![0.05_f32; h]),
                post_ffn_norm_gain: Arc::from(vec![0.05_f32; h]),
                ffn_gate: WeightStorage::F32(vec_of(h * i_dim)),
                ffn_up: WeightStorage::F32(vec_of(h * i_dim)),
                ffn_down: WeightStorage::F32(vec_of(i_dim * h)),
            })
            .collect();
        let final_norm_gain = Arc::from(vec![0.05_f32; h]);
        Gemma3Weights {
            instance: fuel_tensor::decode_shape::ModelInstanceId::next(),
            token_embedding,
            layers,
            final_norm_gain,
        }
    }

    /// **GAP-029 — the quantized wrapper's decode delegation**, over Q4_0
    /// storage with BOTH per-layer axes live.
    ///
    /// ⚠️ **This module's `test_cfg()` is RoPE-degenerate too** — it sets
    /// `rope_local_base_freq: 10_000.0` alongside `rope_theta: 10_000.0`, so a
    /// test on the unmodified fixture would exercise one RoPE table and pass
    /// under a single-base port. The override, and the `n_variants() == 2`
    /// assertions, are what make this a real dual-base test rather than a
    /// fixture-shaped one.
    #[test]
    fn quantized_gemma3_per_layer_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 = Gemma3Config {
            rope_local_base_freq: 1_000.0,
            ..test_cfg()
        };
        let model = QuantizedGemma3Model::from_f32_bake(cfg.clone(), tiny_weights(&cfg)).unwrap();
        assert_ne!(cfg.rope_local_base_freq, cfg.rope_theta, "fixture guard");
        assert_eq!(
            model.inner().decode_rope_plan().n_variants(),
            2,
            "dual base must be live"
        );
        assert_eq!(
            model.inner().decode_mask_plan().n_variants(),
            2,
            "mixed mask must be live"
        );

        let tokens: Vec<u32> = vec![1, 2, 3, 4, 5, 6];
        assert!(
            tokens.len() > cfg.sliding_window,
            "non-vacuity: the window must 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 Gemma3 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 = QuantizedGemma3Model::from_f32_bake(cfg.clone(), src).unwrap();
        // Sanity: all Linear projections in layer 0 are now Q4_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 { .. }));
        assert!(matches!(l0.ffn_gate, WeightStorage::Q4_0 { .. }));
        assert!(matches!(l0.ffn_up, WeightStorage::Q4_0 { .. }));
        assert!(matches!(l0.ffn_down, 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 from_f32_bake_rejects_unaligned_dims() {
        // intermediate_size not divisible by 32 → reject (covers ffn_down in-features).
        let mut cfg = test_cfg();
        cfg.intermediate_size = 48;
        let src = tiny_weights(&cfg);
        assert!(QuantizedGemma3Model::from_f32_bake(cfg, src).is_err());
    }

    #[test]
    fn forward_embeds_matches_forward_after_token_lookup() {
        let cfg = test_cfg();
        let src = tiny_weights(&cfg);
        let model = QuantizedGemma3Model::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 scaled = embeds.mul_scalar((cfg.hidden_size as f64).sqrt());
        let logits_via_embeds = model.forward_embeds(&scaled, 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 Gemma3 forward vs forward_embeds (post-scale) must agree (max diff {max_diff})"
        );
    }
}