ferrum-models 0.7.7

//! Qwen3-TTS Talker model — generates speech codec tokens from text.
//!
//! Architecture: Qwen3 backbone (20 layers, 1024 hidden, 16 heads, 2 KV heads)
//! with text projection (2048→1024) and SubTalker code predictor (31 codebooks).
//!
//! candle loads weights from safetensors; forward pass is ours for Metal/CPU.
//!
//! WIP: the Backend<B> port (Phase F) is in flight, so the candle-based path
//! still carries placeholder fields and debug-only variables. Crate-level
//! allows keep CI green without forcing cosmetic churn on code that's about
//! to be rewritten.

#![allow(dead_code, unused_imports, unused_variables, unused_mut, unused_parens)]

use candle_core::{DType, Device as CandleDevice, IndexOp, Module, Tensor, D};
use candle_nn::{Embedding, Linear, RmsNorm, VarBuilder};
use ferrum_types::{FerrumError, Result};
use parking_lot::Mutex;
use std::collections::HashMap;
use std::sync::OnceLock;
use tracing::info;

use super::repeat_kv;

// ── Config ──────────────────────────────────────────────────────────────

/// Talker LM config (from config.json talker_config section).
#[derive(Debug, Clone)]
pub struct TalkerConfig {
    pub vocab_size: usize,                  // 3072 (codec token vocabulary)
    pub hidden_size: usize,                 // 1024
    pub intermediate_size: usize,           // 2816
    pub num_hidden_layers: usize,           // 20
    pub num_attention_heads: usize,         // 16
    pub num_key_value_heads: usize,         // 2
    pub head_dim: usize,                    // 64
    pub max_position_embeddings: usize,     // 32768
    pub rope_theta: f64,                    // 1000000.0
    pub rms_norm_eps: f64,                  // 1e-6
    pub text_vocab_size: usize,             // 151936
    pub text_hidden_size: usize,            // 2048
    pub num_code_groups: usize,             // 32
    pub codec_eos_token_id: u32,            // 4198
    pub codec_pad_id: u32,                  // 4196
    pub codec_bos_id: u32,                  // 4197
    pub codec_think_id: u32,                // 4202
    pub codec_nothink_id: u32,              // 4203
    pub codec_think_bos_id: u32,            // 4204
    pub codec_think_eos_id: u32,            // 4205
    pub tts_bos_token_id: u32,              // 151672
    pub tts_eos_token_id: u32,              // 151673
    pub tts_pad_token_id: u32,              // 151671
    pub code_predictor_vocab_size: usize,   // 2048
    pub code_predictor_hidden_size: usize,  // 1024
    pub code_predictor_num_layers: usize,   // typically 4
    pub code_predictor_num_heads: usize,    // 16
    pub code_predictor_num_kv_heads: usize, // 8
    pub code_predictor_head_dim: usize,     // 128 (explicit — not hidden/num_heads)
    /// Speaker ID mapping (speaker_name → token_id)
    pub spk_id: HashMap<String, Vec<u32>>,
    /// Language ID mapping (language_name → token_id)
    pub codec_language_id: HashMap<String, u32>,
}

impl TalkerConfig {
    /// Parse from the config.json's talker_config section.
    pub fn from_json(v: &serde_json::Value) -> Result<Self> {
        let tc = v
            .get("talker_config")
            .ok_or_else(|| FerrumError::model("missing talker_config"))?;

        let get_usize = |key: &str, default: usize| -> usize {
            tc.get(key)
                .and_then(|v| v.as_u64())
                .map(|v| v as usize)
                .unwrap_or(default)
        };
        let get_f64 = |key: &str, default: f64| -> f64 {
            tc.get(key).and_then(|v| v.as_f64()).unwrap_or(default)
        };
        let get_u32 = |key: &str, default: u32| -> u32 {
            tc.get(key)
                .and_then(|v| v.as_u64())
                .map(|v| v as u32)
                .unwrap_or(default)
        };

        let mut spk_id = HashMap::new();
        if let Some(obj) = tc.get("spk_id").and_then(|v| v.as_object()) {
            for (k, v) in obj {
                if let Some(arr) = v.as_array() {
                    let ids: Vec<u32> = arr
                        .iter()
                        .filter_map(|x| x.as_u64().map(|n| n as u32))
                        .collect();
                    spk_id.insert(k.clone(), ids);
                }
            }
        }

        let mut codec_language_id = HashMap::new();
        if let Some(obj) = tc.get("codec_language_id").and_then(|v| v.as_object()) {
            for (k, v) in obj {
                if let Some(id) = v.as_u64() {
                    codec_language_id.insert(k.clone(), id as u32);
                }
            }
        }

        Ok(Self {
            vocab_size: get_usize("vocab_size", 3072),
            hidden_size: get_usize("hidden_size", 1024),
            intermediate_size: get_usize("intermediate_size", 3072),
            num_hidden_layers: get_usize("num_hidden_layers", 28),
            num_attention_heads: get_usize("num_attention_heads", 16),
            num_key_value_heads: get_usize("num_key_value_heads", 8),
            head_dim: get_usize("head_dim", 128),
            max_position_embeddings: get_usize("max_position_embeddings", 32768),
            rope_theta: get_f64("rope_theta", 1000000.0),
            rms_norm_eps: get_f64("rms_norm_eps", 1e-6),
            text_vocab_size: get_usize("text_vocab_size", 151936),
            text_hidden_size: get_usize("text_hidden_size", 2048),
            num_code_groups: get_usize("num_code_groups", 16),
            codec_eos_token_id: get_u32("codec_eos_token_id", 2150),
            codec_pad_id: get_u32("codec_pad_id", 2148),
            codec_bos_id: get_u32("codec_bos_id", 2149),
            codec_think_id: get_u32("codec_think_id", 2154),
            codec_nothink_id: get_u32("codec_nothink_id", 2155),
            codec_think_bos_id: get_u32("codec_think_bos_id", 2156),
            codec_think_eos_id: get_u32("codec_think_eos_id", 2157),
            tts_bos_token_id: v
                .get("tts_bos_token_id")
                .and_then(|v| v.as_u64())
                .map(|v| v as u32)
                .unwrap_or(151672),
            tts_eos_token_id: v
                .get("tts_eos_token_id")
                .and_then(|v| v.as_u64())
                .map(|v| v as u32)
                .unwrap_or(151673),
            tts_pad_token_id: v
                .get("tts_pad_token_id")
                .and_then(|v| v.as_u64())
                .map(|v| v as u32)
                .unwrap_or(151671),
            code_predictor_vocab_size: {
                let cp = tc.get("code_predictor_config");
                cp.and_then(|c| c.get("vocab_size"))
                    .and_then(|v| v.as_u64())
                    .map(|v| v as usize)
                    .unwrap_or(2048)
            },
            code_predictor_hidden_size: {
                let cp = tc.get("code_predictor_config");
                cp.and_then(|c| c.get("hidden_size"))
                    .and_then(|v| v.as_u64())
                    .map(|v| v as usize)
                    .unwrap_or(1024)
            },
            code_predictor_num_layers: {
                let cp = tc.get("code_predictor_config");
                cp.and_then(|c| c.get("num_hidden_layers"))
                    .and_then(|v| v.as_u64())
                    .map(|v| v as usize)
                    .unwrap_or(5)
            },
            code_predictor_num_heads: {
                let cp = tc.get("code_predictor_config");
                cp.and_then(|c| c.get("num_attention_heads"))
                    .and_then(|v| v.as_u64())
                    .map(|v| v as usize)
                    .unwrap_or(16)
            },
            code_predictor_num_kv_heads: {
                let cp = tc.get("code_predictor_config");
                cp.and_then(|c| c.get("num_key_value_heads"))
                    .and_then(|v| v.as_u64())
                    .map(|v| v as usize)
                    .unwrap_or(8)
            },
            code_predictor_head_dim: {
                let cp = tc.get("code_predictor_config");
                cp.and_then(|c| c.get("head_dim"))
                    .and_then(|v| v.as_u64())
                    .map(|v| v as usize)
                    .unwrap_or(128)
            },
            spk_id,
            codec_language_id,
        })
    }
}

// ── Rotary Embedding ────────────────────────────────────────────────────

#[derive(Debug, Clone)]
struct RotaryEmbedding {
    sin: Tensor,
    cos: Tensor,
}

impl RotaryEmbedding {
    fn new(dtype: DType, cfg: &TalkerConfig, dev: &CandleDevice) -> candle_core::Result<Self> {
        let dim = cfg.head_dim;
        let max_seq_len = cfg.max_position_embeddings;
        let inv_freq: Vec<_> = (0..dim)
            .step_by(2)
            .map(|i| 1f32 / cfg.rope_theta.powf(i as f64 / dim as f64) as f32)
            .collect();
        let inv_freq_len = inv_freq.len();
        let inv_freq = Tensor::from_vec(inv_freq, (1, inv_freq_len), dev)?.to_dtype(dtype)?;
        let t = Tensor::arange(0u32, max_seq_len as u32, dev)?
            .to_dtype(dtype)?
            .reshape((max_seq_len, 1))?;
        let freqs = t.matmul(&inv_freq)?;
        Ok(Self {
            sin: freqs.sin()?,
            cos: freqs.cos()?,
        })
    }

    fn apply(
        &self,
        q: &Tensor,
        k: &Tensor,
        offset: usize,
    ) -> candle_core::Result<(Tensor, Tensor)> {
        // Match reference project: narrow + manual rotation (not rope_slow)
        let (_, _, seq_len, d) = q.dims4()?;
        let cos = self
            .cos
            .narrow(0, offset, seq_len)?
            .unsqueeze(0)?
            .unsqueeze(0)?
            .to_dtype(q.dtype())?;
        let sin = self
            .sin
            .narrow(0, offset, seq_len)?
            .unsqueeze(0)?
            .unsqueeze(0)?
            .to_dtype(q.dtype())?;

        fn rope_rotate(x: &Tensor, cos: &Tensor, sin: &Tensor) -> candle_core::Result<Tensor> {
            let d = x.dim(candle_core::D::Minus1)?;
            let x1 = x.narrow(candle_core::D::Minus1, 0, d / 2)?;
            let x2 = x.narrow(candle_core::D::Minus1, d / 2, d / 2)?;
            let cos = cos.broadcast_as(x1.shape())?;
            let sin = sin.broadcast_as(x1.shape())?;
            Tensor::cat(
                &[
                    &(x1.mul(&cos)? - x2.mul(&sin)?)?,
                    &(x2.mul(&cos)? + x1.mul(&sin)?)?,
                ],
                candle_core::D::Minus1,
            )
        }

        let q_embed = rope_rotate(q, &cos, &sin)?;
        let k_embed = rope_rotate(k, &cos, &sin)?;
        Ok((q_embed, k_embed))
    }
}

// ── RMSNorm (manual ops for Metal compatibility) ────────────────────────

#[derive(Debug, Clone)]
struct ManualRmsNorm {
    weight: Tensor,
    eps: f64,
}

impl ManualRmsNorm {
    fn new(weight: Tensor, eps: f64) -> Self {
        Self { weight, eps }
    }
}

impl Module for ManualRmsNorm {
    fn forward(&self, x: &Tensor) -> candle_core::Result<Tensor> {
        // Match candle_nn::RmsNorm formula: x / (sqrt(mean(x²)) + eps)
        // NOT x / sqrt(mean(x²) + eps) — eps placement matters!
        let x_f32 = x.to_dtype(DType::F32)?;
        let hidden_size = x_f32.dim(candle_core::D::Minus1)?;
        let norm_x =
            (x_f32.sqr()?.sum_keepdim(candle_core::D::Minus1)? / hidden_size as f64)?.sqrt()?;
        let normed = x_f32.broadcast_div(&(norm_x + self.eps)?)?;
        let normed = normed.to_dtype(x.dtype())?;
        normed.broadcast_mul(&self.weight)
    }
}

fn rms_norm(size: usize, eps: f64, vb: VarBuilder) -> candle_core::Result<ManualRmsNorm> {
    let w = vb.get(size, "weight")?;
    Ok(ManualRmsNorm::new(w, eps))
}

// ── MLP ─────────────────────────────────────────────────────────────────

#[derive(Debug, Clone)]
struct MLP {
    gate_proj: Linear,
    up_proj: Linear,
    down_proj: Linear,
    intermediate_size: usize,
}

impl MLP {
    fn new(cfg: &TalkerConfig, vb: VarBuilder) -> candle_core::Result<Self> {
        let h = cfg.hidden_size;
        let i = cfg.intermediate_size;
        let gate_proj = candle_nn::linear_no_bias(h, i, vb.pp("gate_proj"))?;
        let up_proj = candle_nn::linear_no_bias(h, i, vb.pp("up_proj"))?;
        let down_proj = candle_nn::linear_no_bias(i, h, vb.pp("down_proj"))?;
        Ok(Self {
            gate_proj,
            up_proj,
            down_proj,
            intermediate_size: i,
        })
    }
}

impl Module for MLP {
    fn forward(&self, x: &Tensor) -> candle_core::Result<Tensor> {
        let gate = x.apply(&self.gate_proj)?.silu()?;
        let up = x.apply(&self.up_proj)?;
        (gate * up)?.apply(&self.down_proj)
    }
}

// ── Self-Attention with KV Cache ────────────────────────────────────────

#[derive(Debug, Clone)]
struct Attention {
    q_proj: Linear,
    k_proj: Linear,
    v_proj: Linear,
    o_proj: Linear,
    q_norm: ManualRmsNorm,
    k_norm: ManualRmsNorm,
    num_heads: usize,
    num_kv_heads: usize,
    head_dim: usize,
    rotary: RotaryEmbedding,
    kv_cache: Option<(Tensor, Tensor)>,
}

impl Attention {
    fn new(
        cfg: &TalkerConfig,
        rotary: RotaryEmbedding,
        vb: VarBuilder,
    ) -> candle_core::Result<Self> {
        let h = cfg.hidden_size;
        let hd = cfg.head_dim;
        let nh = cfg.num_attention_heads;
        let nkv = cfg.num_key_value_heads;
        let q_proj = candle_nn::linear_no_bias(h, nh * hd, vb.pp("q_proj"))?;
        let k_proj = candle_nn::linear_no_bias(h, nkv * hd, vb.pp("k_proj"))?;
        let v_proj = candle_nn::linear_no_bias(h, nkv * hd, vb.pp("v_proj"))?;
        let o_proj = candle_nn::linear_no_bias(nh * hd, h, vb.pp("o_proj"))?;
        let q_norm = rms_norm(hd, cfg.rms_norm_eps, vb.pp("q_norm"))?;
        let k_norm = rms_norm(hd, cfg.rms_norm_eps, vb.pp("k_norm"))?;
        Ok(Self {
            q_proj,
            k_proj,
            v_proj,
            o_proj,
            q_norm,
            k_norm,
            num_heads: nh,
            num_kv_heads: nkv,
            head_dim: hd,
            rotary,
            kv_cache: None,
        })
    }

    fn forward(&mut self, x: &Tensor, pos_offset: usize) -> candle_core::Result<Tensor> {
        let (b, seq_len, _) = x.dims3()?;
        let hd = self.head_dim;

        let q = x.apply(&self.q_proj)?;
        let k = x.apply(&self.k_proj)?;
        let v = x.apply(&self.v_proj)?;

        // Reshape to [b, seq, heads, hd] for QK norm (matching reference: norm BEFORE transpose)
        let q = q.reshape((b, seq_len, self.num_heads, hd))?;
        let k = k.reshape((b, seq_len, self.num_kv_heads, hd))?;
        let v = v.reshape((b, seq_len, self.num_kv_heads, hd))?;

        // QK norm on [b, seq, heads, hd] — BEFORE transpose
        let q = q.apply(&self.q_norm)?;
        let k = k.apply(&self.k_norm)?;

        // Transpose to [b, heads, seq, hd]
        let q = q.transpose(1, 2)?;
        let k = k.transpose(1, 2)?;
        let v = v.transpose(1, 2)?;

        // RoPE
        let (q, k) = self.rotary.apply(&q, &k, pos_offset)?;

        // KV cache
        let (k, v) = if let Some((prev_k, prev_v)) = &self.kv_cache {
            (
                Tensor::cat(&[prev_k, &k], 2)?,
                Tensor::cat(&[prev_v, &v], 2)?,
            )
        } else {
            (k, v)
        };
        self.kv_cache = Some((k.clone(), v.clone()));

        // GQA: repeat KV heads to match Q heads
        let kv_len = k.dim(2)?;
        let n_rep = self.num_heads / self.num_kv_heads;
        let k = if n_rep > 1 {
            k.unsqueeze(2)?
                .expand((b, self.num_kv_heads, n_rep, kv_len, hd))?
                .reshape((b, self.num_heads, kv_len, hd))?
        } else {
            k
        };
        let v = if n_rep > 1 {
            v.unsqueeze(2)?
                .expand((b, self.num_kv_heads, n_rep, kv_len, hd))?
                .reshape((b, self.num_heads, kv_len, hd))?
        } else {
            v
        };

        // Standard attention using candle ops (matching reference project)
        let scale = (hd as f64).powf(-0.5);
        let attn_weights = (q.matmul(&k.transpose(2, 3)?)? * scale)?;

        // Causal mask (only for prefill, not decode)
        let attn_weights = if seq_len > 1 {
            let mut mask_data = vec![0.0f32; seq_len * kv_len];
            for i in 0..seq_len {
                let attend_up_to = pos_offset + i + 1;
                for j in attend_up_to..kv_len {
                    mask_data[i * kv_len + j] = f32::NEG_INFINITY;
                }
            }
            let mask = Tensor::from_vec(mask_data, (1, 1, seq_len, kv_len), x.device())?;
            attn_weights.broadcast_add(&mask)?
        } else {
            attn_weights
        };

        // Use softmax_last_dim (fused single-pass, matches reference project)
        // Falls back to decomposed softmax on Metal (no Metal impl)
        let attn_weights = if x.device().is_cpu() {
            candle_nn::ops::softmax_last_dim(&attn_weights)?
        } else {
            candle_nn::ops::softmax(&attn_weights, candle_core::D::Minus1)?
        };
        let out = attn_weights.matmul(&v)?;

        let out = out
            .transpose(1, 2)?
            .reshape((b, seq_len, self.num_heads * hd))?;
        out.apply(&self.o_proj)
    }

    fn reset_cache(&mut self) {
        self.kv_cache = None;
    }
}

// ── Transformer Layer ───────────────────────────────────────────────────

#[derive(Debug, Clone)]
struct TransformerLayer {
    self_attn: Attention,
    mlp: MLP,
    input_layernorm: ManualRmsNorm,
    post_attention_layernorm: ManualRmsNorm,
}

impl TransformerLayer {
    fn new(
        cfg: &TalkerConfig,
        rotary: RotaryEmbedding,
        vb: VarBuilder,
    ) -> candle_core::Result<Self> {
        let self_attn = Attention::new(cfg, rotary, vb.pp("self_attn"))?;
        let mlp = MLP::new(cfg, vb.pp("mlp"))?;
        let input_layernorm =
            rms_norm(cfg.hidden_size, cfg.rms_norm_eps, vb.pp("input_layernorm"))?;
        let post_attention_layernorm = rms_norm(
            cfg.hidden_size,
            cfg.rms_norm_eps,
            vb.pp("post_attention_layernorm"),
        )?;
        Ok(Self {
            self_attn,
            mlp,
            input_layernorm,
            post_attention_layernorm,
        })
    }

    fn forward(&mut self, x: &Tensor, pos_offset: usize) -> candle_core::Result<Tensor> {
        let residual = x.clone();
        let x = x.apply(&self.input_layernorm)?;

        // Debug: dump layernorm output for layer 0 comparison
        if pos_offset == 0 && x.dim(1).unwrap_or(0) > 1 {
            if let Ok(vals) = x
                .narrow(0, 0, 1)
                .and_then(|t| {
                    let sl = t.dim(1).unwrap_or(1);
                    t.narrow(1, sl - 1, 1)
                })
                .and_then(|t| t.narrow(2, 0, 5))
                .and_then(|t| t.flatten_all())
                .and_then(|t| t.to_vec1::<f32>())
            {
                tracing::info!("  layernorm pos -1 first 5: {:?}", vals);
            }
        }

        let x = self.self_attn.forward(&x, pos_offset)?;
        let x = (residual + x)?;
        let residual = x.clone();
        let x = x.apply(&self.post_attention_layernorm)?;
        let x = x.apply(&self.mlp)?;
        residual + x
    }

    fn reset_cache(&mut self) {
        self.self_attn.reset_cache();
    }
}

// ── Text Projection (ResizeMLP: text_hidden → hidden) ───────────────────

#[derive(Debug, Clone)]
struct TextProjection {
    linear1: Linear,
    linear2: Linear,
}

impl TextProjection {
    fn new(text_hidden: usize, hidden: usize, vb: VarBuilder) -> candle_core::Result<Self> {
        // Keys: talker.text_projection.linear_fc1.weight/bias, linear_fc2.weight/bias
        let linear1 = candle_nn::linear(text_hidden, text_hidden, vb.pp("linear_fc1"))?;
        let linear2 = candle_nn::linear(text_hidden, hidden, vb.pp("linear_fc2"))?;
        Ok(Self { linear1, linear2 })
    }

    fn forward(&self, x: &Tensor) -> candle_core::Result<Tensor> {
        x.apply(&self.linear1)?.silu()?.apply(&self.linear2)
    }
}

// ── Talker Model (main LM) ─────────────────────────────────────────────

/// Qwen3-TTS Talker: text → speech codec tokens.
pub struct Qwen3TTSTalker {
    text_embedding: Embedding,
    text_projection: TextProjection,
    codec_embedding: Embedding,
    layers: Vec<TransformerLayer>,
    norm: ManualRmsNorm,
    codec_head: Linear,
    config: TalkerConfig,
    device: CandleDevice,
    tokens_generated: usize,
    fused: ferrum_kernels::attention::FusedTransformer,
    /// Optional Backend<B> transformer stack. When set, `forward_step`
    /// routes through this instead of `fused`. Used on Linux + CUDA where
    /// `fused` would silently fall back to a broken naive-fp64 CPU matmul.
    /// See `multimodal::qwen3_tts_backbone`.
    backend_override: Option<Box<dyn crate::multimodal::qwen3_tts_backbone::TalkerBackboneForward>>,
}

impl Qwen3TTSTalker {
    pub fn load(cfg: &TalkerConfig, vb: VarBuilder, device: CandleDevice) -> Result<Self> {
        let dtype = vb.dtype();
        let model_vb = vb.pp("talker").pp("model");

        let text_embedding = candle_nn::embedding(
            cfg.text_vocab_size,
            cfg.text_hidden_size,
            model_vb.pp("text_embedding"),
        )
        .map_err(|e| FerrumError::model(format!("text_embedding: {e}")))?;

        let text_projection = TextProjection::new(
            cfg.text_hidden_size,
            cfg.hidden_size,
            vb.pp("talker").pp("text_projection"),
        )
        .map_err(|e| FerrumError::model(format!("text_projection: {e}")))?;

        let codec_embedding = candle_nn::embedding(
            cfg.vocab_size,
            cfg.hidden_size,
            model_vb.pp("codec_embedding"),
        )
        .map_err(|e| FerrumError::model(format!("codec_embedding: {e}")))?;

        let rotary = RotaryEmbedding::new(dtype, cfg, &device)
            .map_err(|e| FerrumError::model(format!("rotary: {e}")))?;

        let mut layers = Vec::with_capacity(cfg.num_hidden_layers);
        for i in 0..cfg.num_hidden_layers {
            let layer =
                TransformerLayer::new(cfg, rotary.clone(), model_vb.pp(format!("layers.{i}")))
                    .map_err(|e| FerrumError::model(format!("layer {i}: {e}")))?;
            layers.push(layer);
        }

        let norm = rms_norm(cfg.hidden_size, cfg.rms_norm_eps, model_vb.pp("norm"))
            .map_err(|e| FerrumError::model(format!("norm: {e}")))?;

        let codec_head = candle_nn::linear_no_bias(
            cfg.hidden_size,
            cfg.vocab_size,
            vb.pp("talker").pp("codec_head"),
        )
        .map_err(|e| FerrumError::model(format!("codec_head: {e}")))?;

        // Build fused transformer (Metal or CPU, bypasses candle for precision)
        let to_cpu_vec = |t: &Tensor| -> candle_core::Result<Vec<f32>> {
            t.to_device(&candle_core::Device::Cpu)?
                .to_dtype(DType::F32)?
                .flatten_all()?
                .to_vec1()
        };
        let get_w = |vb: &VarBuilder, shape: candle_core::Shape, name: &str| -> Result<Vec<f32>> {
            let t = vb
                .get(shape, name)
                .map_err(|e| FerrumError::model(format!("w {name}: {e}")))?;
            to_cpu_vec(&t).map_err(|e| FerrumError::model(format!("vec {name}: {e}")))
        };

        let mut fused_layers = Vec::with_capacity(cfg.num_hidden_layers);
        for i in 0..cfg.num_hidden_layers {
            let lv = model_vb.pp(format!("layers.{i}"));
            let av = lv.pp("self_attn");
            let mv = lv.pp("mlp");
            fused_layers.push(ferrum_kernels::attention::LayerWeights {
                input_ln_w: get_w(&lv.pp("input_layernorm"), cfg.hidden_size.into(), "weight")?,
                q_proj_w: get_w(
                    &av.pp("q_proj"),
                    (cfg.num_attention_heads * cfg.head_dim, cfg.hidden_size).into(),
                    "weight",
                )?,
                k_proj_w: get_w(
                    &av.pp("k_proj"),
                    (cfg.num_key_value_heads * cfg.head_dim, cfg.hidden_size).into(),
                    "weight",
                )?,
                v_proj_w: get_w(
                    &av.pp("v_proj"),
                    (cfg.num_key_value_heads * cfg.head_dim, cfg.hidden_size).into(),
                    "weight",
                )?,
                o_proj_w: get_w(
                    &av.pp("o_proj"),
                    (cfg.hidden_size, cfg.num_attention_heads * cfg.head_dim).into(),
                    "weight",
                )?,
                q_norm_w: get_w(&av.pp("q_norm"), cfg.head_dim.into(), "weight")?,
                k_norm_w: get_w(&av.pp("k_norm"), cfg.head_dim.into(), "weight")?,
                post_ln_w: get_w(
                    &lv.pp("post_attention_layernorm"),
                    cfg.hidden_size.into(),
                    "weight",
                )?,
                gate_proj_w: get_w(
                    &mv.pp("gate_proj"),
                    (cfg.intermediate_size, cfg.hidden_size).into(),
                    "weight",
                )?,
                up_proj_w: get_w(
                    &mv.pp("up_proj"),
                    (cfg.intermediate_size, cfg.hidden_size).into(),
                    "weight",
                )?,
                down_proj_w: get_w(
                    &mv.pp("down_proj"),
                    (cfg.hidden_size, cfg.intermediate_size).into(),
                    "weight",
                )?,
                attn_layer_scale: None,
                mlp_layer_scale: None,
            });
        }
        let norm_w =
            to_cpu_vec(&norm.weight).map_err(|e| FerrumError::model(format!("norm_w: {e}")))?;

        let fused = ferrum_kernels::attention::FusedTransformer::new(
            ferrum_kernels::attention::TransformerConfig {
                hidden_size: cfg.hidden_size,
                intermediate_size: cfg.intermediate_size,
                num_heads: cfg.num_attention_heads,
                num_kv_heads: cfg.num_key_value_heads,
                head_dim: cfg.head_dim,
                num_layers: cfg.num_hidden_layers,
                rms_norm_eps: cfg.rms_norm_eps,
                rope_theta: cfg.rope_theta,
                max_position_embeddings: cfg.max_position_embeddings,
            },
            fused_layers,
            norm_w,
        );

        info!(
            "Qwen3TTSTalker loaded: hidden={}, layers={}, heads={}/{}, vocab={} (fused transformer ready)",
            cfg.hidden_size, cfg.num_hidden_layers, cfg.num_attention_heads, cfg.num_key_value_heads, cfg.vocab_size,
        );

        Ok(Self {
            text_embedding,
            text_projection,
            codec_embedding,
            layers,
            norm,
            codec_head,
            config: cfg.clone(),
            device,
            tokens_generated: 0,
            fused,
            backend_override: None,
        })
    }

    /// Install a Backend<B>-backed transformer to bypass `fused`. Used by
    /// the executor on CUDA where `fused`'s fallback is broken.
    pub fn set_backend_override(
        &mut self,
        backend: Box<dyn crate::multimodal::qwen3_tts_backbone::TalkerBackboneForward>,
    ) {
        self.backend_override = Some(backend);
        self.tokens_generated = 0;
    }

    pub fn has_backend_override(&self) -> bool {
        self.backend_override.is_some()
    }

    /// Embed text token IDs through text_embedding + text_projection.
    pub fn embed_text(&self, text_ids: &Tensor) -> Result<Tensor> {
        let embeds = text_ids
            .apply(&self.text_embedding)
            .map_err(|e| FerrumError::model(format!("text_embed: {e}")))?;
        self.text_projection
            .forward(&embeds)
            .map_err(|e| FerrumError::model(format!("text_proj: {e}")))
    }

    /// Embed codec token IDs through codec_embedding.
    pub fn embed_codec(&self, codec_ids: &Tensor) -> Result<Tensor> {
        codec_ids
            .apply(&self.codec_embedding)
            .map_err(|e| FerrumError::model(format!("codec_embed: {e}")))
    }

    /// Forward through transformer layers. Uses fused path by default,
    /// set FERRUM_USE_CANDLE=1 to use candle's native ops (for precision testing).
    pub fn forward_step(&mut self, input_embeds: &Tensor) -> Result<Tensor> {
        let use_candle = qwen3_tts_use_candle();

        // If a Backend<B> override is installed, use it. Returns post-norm
        // hidden so semantics match `fused.forward`.
        if let Some(ref mut backend) = self.backend_override {
            let seq_len = input_embeds
                .dim(1)
                .map_err(|e| FerrumError::model(format!("dim: {e}")))?;
            let h = self.config.hidden_size;
            let input_data: Vec<f32> = input_embeds
                .to_device(&candle_core::Device::Cpu)
                .and_then(|t| t.to_dtype(DType::F32))
                .and_then(|t| t.flatten_all())
                .and_then(|t| t.to_vec1())
                .map_err(|e| FerrumError::model(format!("input extract: {e}")))?;

            let output = backend.forward(&input_data, seq_len);
            self.tokens_generated += seq_len;

            return Tensor::from_vec(output, (1, seq_len, h), &candle_core::Device::Cpu)
                .and_then(|t| t.to_device(&self.device))
                .map_err(|e| FerrumError::model(format!("output tensor: {e}")));
        }

        if use_candle {
            let pos_offset = self.tokens_generated;
            let seq_len = input_embeds
                .dim(1)
                .map_err(|e| FerrumError::model(format!("dim: {e}")))?;
            let mut hidden = input_embeds.clone();
            for (li, layer) in self.layers.iter_mut().enumerate() {
                hidden = layer
                    .forward(&hidden, pos_offset)
                    .map_err(|e| FerrumError::model(format!("layer {li}: {e}")))?;
            }
            hidden = hidden
                .apply(&self.norm)
                .map_err(|e| FerrumError::model(format!("norm: {e}")))?;
            self.tokens_generated += seq_len;
            return Ok(hidden);
        }

        {
            // Fused path: Metal GPU or CPU custom ops.
            let seq_len = input_embeds
                .dim(1)
                .map_err(|e| FerrumError::model(format!("dim: {e}")))?;
            let h = self.config.hidden_size;

            // Extract input to CPU (needed for both GPU and CPU paths)
            let input_data: Vec<f32> = input_embeds
                .to_device(&candle_core::Device::Cpu)
                .and_then(|t| t.to_dtype(DType::F32))
                .and_then(|t| t.flatten_all())
                .and_then(|t| t.to_vec1())
                .map_err(|e| FerrumError::model(format!("input extract: {e}")))?;

            // Try GPU path with GPU-side norm (avoids CPU norm overhead)
            #[cfg(feature = "metal")]
            if let Some(data) = self.fused.forward_gpu_to_vec(&input_data, seq_len) {
                self.tokens_generated += seq_len;
                return Tensor::from_vec(data, (1, seq_len, h), &candle_core::Device::Cpu)
                    .and_then(|t| t.to_device(&self.device))
                    .map_err(|e| FerrumError::model(format!("output tensor: {e}")));
            }

            // CPU fallback
            let output = self.fused.forward(&input_data, seq_len);
            self.tokens_generated += seq_len;

            Tensor::from_vec(output, (1, seq_len, h), &candle_core::Device::Cpu)
                .and_then(|t| t.to_device(&self.device))
                .map_err(|e| FerrumError::model(format!("output tensor: {e}")))
        }
    }

    /// Get logits from hidden states.
    pub fn logits(&self, hidden: &Tensor) -> Result<Tensor> {
        hidden
            .apply(&self.codec_head)
            .map_err(|e| FerrumError::model(format!("codec_head: {e}")))
    }

    pub fn reset(&mut self) {
        self.tokens_generated = 0;
        self.fused.reset();
        if let Some(ref mut backend) = self.backend_override {
            backend.reset();
        }
        for layer in &mut self.layers {
            layer.reset_cache();
        }
    }

    pub fn config(&self) -> &TalkerConfig {
        &self.config
    }

    pub fn device(&self) -> &CandleDevice {
        &self.device
    }
}

fn qwen3_tts_use_candle() -> bool {
    static FLAG: OnceLock<bool> = OnceLock::new();
    *FLAG.get_or_init(|| {
        std::env::vars().any(|(name, value)| name == "FERRUM_USE_CANDLE" && value == "1")
    })
}

// ── SubTalker (Code Predictor) ──────────────────────────────────────────
//
// Predicts codec tokens 1..num_code_groups-1 given the talker hidden state
// and the first codec token embedding. 5-layer transformer with per-codebook
// lm_head and embedding.

pub struct SubTalker {
    layers: Vec<TransformerLayer>,
    norm: ManualRmsNorm,
    /// Per-codebook embeddings: codec_embedding[i] maps token → hidden for codebook i+1
    pub codec_embeddings: Vec<Embedding>,
    /// Per-codebook prediction heads: lm_head[i] maps hidden → logits for codebook i+1
    lm_heads: Vec<Linear>,
    /// Cached raw weights for zero-overhead predict loop
    lm_raw: Vec<Vec<f32>>, // [n_extra][vocab * hidden]
    pub(crate) emb_raw: Vec<Vec<f32>>, // [n_extra][vocab * emb_dim]
    vocab_size: usize,
    pub(crate) emb_dim: usize,
    /// Projection from talker hidden to subtalker hidden (if sizes differ)
    projection: Option<Linear>,
    /// Cached projection weights for fast CPU matmul (avoids GPU→CPU transfer per step)
    proj_w_raw: Option<Vec<f32>>, // [out_dim, in_dim] row-major
    proj_b_raw: Option<Vec<f32>>, // [out_dim]
    proj_out_dim: usize,
    num_code_groups: usize,
    tokens_generated: usize,
    /// Fused transformer (bypasses candle for precision)
    fused: ferrum_kernels::attention::FusedTransformer,
    fused_hidden_size: usize,
    /// Optional Backend<B> transformer stack that supersedes `fused` on
    /// CUDA. Same motivation as `Qwen3TTSTalker::backend_override`.
    backend_override: Option<Box<dyn crate::multimodal::qwen3_tts_backbone::TalkerBackboneForward>>,
}

impl SubTalker {
    pub fn load(cfg: &TalkerConfig, vb: VarBuilder, device: CandleDevice) -> Result<Self> {
        let dtype = vb.dtype();
        let cp_vb = vb.pp("talker").pp("code_predictor");
        let model_vb = cp_vb.pp("model");

        let cp_cfg = TalkerConfig {
            hidden_size: cfg.code_predictor_hidden_size,
            intermediate_size: cfg.code_predictor_hidden_size * 3, // ~3072
            num_hidden_layers: cfg.code_predictor_num_layers,
            num_attention_heads: cfg.code_predictor_num_heads,
            num_key_value_heads: cfg.code_predictor_num_kv_heads,
            ..cfg.clone()
        };

        let rotary = RotaryEmbedding::new(dtype, &cp_cfg, &device)
            .map_err(|e| FerrumError::model(format!("subtalker rotary: {e}")))?;

        let mut layers = Vec::new();
        for i in 0..cfg.code_predictor_num_layers {
            layers.push(
                TransformerLayer::new(&cp_cfg, rotary.clone(), model_vb.pp(format!("layers.{i}")))
                    .map_err(|e| FerrumError::model(format!("subtalker layer {i}: {e}")))?,
            );
        }

        let norm = rms_norm(
            cfg.code_predictor_hidden_size,
            cfg.rms_norm_eps,
            model_vb.pp("norm"),
        )
        .map_err(|e| FerrumError::model(format!("subtalker norm: {e}")))?;

        // Per-codebook embeddings (num_code_groups - 1)
        let n_extra = cfg.num_code_groups - 1;
        let mut codec_embeddings = Vec::new();
        for i in 0..n_extra {
            codec_embeddings.push(
                candle_nn::embedding(
                    cfg.code_predictor_vocab_size,
                    cfg.hidden_size, // embedding dim = talker hidden size
                    model_vb.pp(format!("codec_embedding.{i}")),
                )
                .map_err(|e| FerrumError::model(format!("subtalker codec_embedding.{i}: {e}")))?,
            );
        }

        // Per-codebook lm_heads
        let mut lm_heads = Vec::new();
        for i in 0..n_extra {
            lm_heads.push(
                candle_nn::linear_no_bias(
                    cfg.code_predictor_hidden_size,
                    cfg.code_predictor_vocab_size,
                    cp_vb.pp(format!("lm_head.{i}")),
                )
                .map_err(|e| FerrumError::model(format!("subtalker lm_head.{i}: {e}")))?,
            );
        }

        // Projection if hidden sizes differ
        let projection = if cfg.hidden_size != cfg.code_predictor_hidden_size {
            Some(
                candle_nn::linear(
                    cfg.hidden_size,
                    cfg.code_predictor_hidden_size,
                    cp_vb.pp("small_to_mtp_projection"),
                )
                .map_err(|e| FerrumError::model(format!("subtalker projection: {e}")))?,
            )
        } else {
            None
        };

        // Build fused transformer for SubTalker (bypasses candle)
        let to_cpu_vec = |t: &Tensor| -> candle_core::Result<Vec<f32>> {
            t.to_device(&candle_core::Device::Cpu)?
                .to_dtype(DType::F32)?
                .flatten_all()?
                .to_vec1()
        };
        let get_w = |vb: &VarBuilder, shape: candle_core::Shape, name: &str| -> Result<Vec<f32>> {
            let t = vb
                .get(shape, name)
                .map_err(|e| FerrumError::model(format!("st w {name}: {e}")))?;
            to_cpu_vec(&t).map_err(|e| FerrumError::model(format!("st vec {name}: {e}")))
        };

        let st_h = cp_cfg.hidden_size;
        let st_im = cp_cfg.intermediate_size;
        let st_nh = cp_cfg.num_attention_heads;
        let st_nkv = cp_cfg.num_key_value_heads;
        let st_hd = cp_cfg.head_dim;

        let mut fused_layers = Vec::with_capacity(cfg.code_predictor_num_layers);
        for i in 0..cfg.code_predictor_num_layers {
            let lv = model_vb.pp(format!("layers.{i}"));
            let av = lv.pp("self_attn");
            let mv = lv.pp("mlp");
            fused_layers.push(ferrum_kernels::attention::LayerWeights {
                input_ln_w: get_w(&lv.pp("input_layernorm"), st_h.into(), "weight")?,
                q_proj_w: get_w(&av.pp("q_proj"), (st_nh * st_hd, st_h).into(), "weight")?,
                k_proj_w: get_w(&av.pp("k_proj"), (st_nkv * st_hd, st_h).into(), "weight")?,
                v_proj_w: get_w(&av.pp("v_proj"), (st_nkv * st_hd, st_h).into(), "weight")?,
                o_proj_w: get_w(&av.pp("o_proj"), (st_h, st_nh * st_hd).into(), "weight")?,
                q_norm_w: get_w(&av.pp("q_norm"), st_hd.into(), "weight")?,
                k_norm_w: get_w(&av.pp("k_norm"), st_hd.into(), "weight")?,
                post_ln_w: get_w(&lv.pp("post_attention_layernorm"), st_h.into(), "weight")?,
                gate_proj_w: get_w(&mv.pp("gate_proj"), (st_im, st_h).into(), "weight")?,
                up_proj_w: get_w(&mv.pp("up_proj"), (st_im, st_h).into(), "weight")?,
                down_proj_w: get_w(&mv.pp("down_proj"), (st_h, st_im).into(), "weight")?,
                attn_layer_scale: None,
                mlp_layer_scale: None,
            });
        }
        let st_norm_w =
            to_cpu_vec(&norm.weight).map_err(|e| FerrumError::model(format!("st norm_w: {e}")))?;

        let fused = ferrum_kernels::attention::FusedTransformer::new(
            ferrum_kernels::attention::TransformerConfig {
                hidden_size: st_h,
                intermediate_size: st_im,
                num_heads: st_nh,
                num_kv_heads: st_nkv,
                head_dim: st_hd,
                num_layers: cfg.code_predictor_num_layers,
                rms_norm_eps: cfg.rms_norm_eps,
                rope_theta: cfg.rope_theta,
                max_position_embeddings: cfg.max_position_embeddings,
            },
            fused_layers,
            st_norm_w,
        );

        info!(
            "SubTalker loaded: layers={}, heads={}/{}, codebooks={} (fused transformer ready)",
            cfg.code_predictor_num_layers,
            cfg.code_predictor_num_heads,
            cfg.code_predictor_num_kv_heads,
            n_extra,
        );

        // Pre-extract raw weights for zero-overhead predict loop
        let lm_raw: Vec<Vec<f32>> = lm_heads
            .iter()
            .map(|lm| lm.weight().flatten_all().unwrap().to_vec1::<f32>().unwrap())
            .collect();
        let emb_raw: Vec<Vec<f32>> = codec_embeddings
            .iter()
            .map(|e| {
                e.embeddings()
                    .flatten_all()
                    .unwrap()
                    .to_vec1::<f32>()
                    .unwrap()
            })
            .collect();
        let vocab_size = cfg.code_predictor_vocab_size;
        let emb_dim = if !emb_raw.is_empty() {
            emb_raw[0].len() / vocab_size
        } else {
            st_h
        };

        // Cache projection weights for fast CPU matmul
        let (proj_w_raw, proj_b_raw, proj_out_dim) = if let Some(ref proj) = projection {
            let w: Vec<f32> = proj
                .weight()
                .to_device(&candle_core::Device::Cpu)
                .and_then(|w| w.flatten_all())
                .and_then(|w| w.to_vec1())
                .unwrap_or_default();
            let out_dim = proj.weight().dim(0).unwrap_or(st_h);
            let b: Option<Vec<f32>> = proj
                .bias()
                .map(|b| {
                    b.to_device(&candle_core::Device::Cpu)
                        .and_then(|b| b.to_vec1())
                        .ok()
                })
                .flatten();
            (Some(w), b, out_dim)
        } else {
            (None, None, 0)
        };

        Ok(Self {
            layers,
            norm,
            codec_embeddings,
            lm_heads,
            lm_raw,
            emb_raw,
            vocab_size,
            emb_dim,
            projection,
            proj_w_raw,
            proj_b_raw,
            proj_out_dim,
            num_code_groups: cfg.num_code_groups,
            tokens_generated: 0,
            fused,
            fused_hidden_size: st_h,
            backend_override: None,
        })
    }

    /// Install a Backend<B>-backed transformer for this SubTalker. Used
    /// on CUDA where the legacy fused path would route to broken Linux
    /// fp64 naive matmul. Mirrors `Qwen3TTSTalker::set_backend_override`.
    pub fn set_backend_override(
        &mut self,
        backend: Box<dyn crate::multimodal::qwen3_tts_backbone::TalkerBackboneForward>,
    ) {
        self.backend_override = Some(backend);
        self.tokens_generated = 0;
    }

    pub fn has_backend_override(&self) -> bool {
        self.backend_override.is_some()
    }

    /// Predict codec tokens 1..num_code_groups-1 given talker hidden state and first codec token.
    /// Optimized: entire loop runs on raw f32 — no Tensor allocation in the hot path.
    pub fn predict(
        &mut self,
        talker_hidden: &Tensor,
        first_codec_embed: &Tensor,
        temperature: f32,
        top_k: usize,
    ) -> Result<Vec<u32>> {
        self.reset();
        let h = self.fused_hidden_size;
        let n_extra = self.num_code_groups - 1;

        // Concat [talker_hidden, first_codec_embed] then project (matching reference)
        let combined = Tensor::cat(&[talker_hidden, first_codec_embed], 1)
            .map_err(|e| FerrumError::model(format!("predict cat: {e}")))?;
        let combined = if let Some(ref proj) = self.projection {
            combined
                .apply(proj)
                .map_err(|e| FerrumError::model(format!("predict proj: {e}")))?
        } else {
            combined
        };

        // Extract to raw f32: [2*h] floats
        let input_data: Vec<f32> = combined
            .flatten_all()
            .and_then(|t| t.to_vec1())
            .map_err(|e| FerrumError::model(format!("input extract: {e}")))?;

        // Prefill (2 tokens through fused transformer or Backend<B> override)
        let output = if let Some(ref mut backend) = self.backend_override {
            backend.forward(&input_data, 2)
        } else {
            self.fused.forward(&input_data, 2)
        };
        // Last position = output[h..2h]
        let mut last_hidden = output[h..2 * h].to_vec();

        // Use pre-cached raw weights (extracted at init time)
        let vocab = self.vocab_size;
        let emb_dim = self.emb_dim;

        #[cfg(target_os = "macos")]
        extern "C" {
            fn cblas_sgemm(
                order: i32,
                ta: i32,
                tb: i32,
                m: i32,
                n: i32,
                k: i32,
                alpha: f32,
                a: *const f32,
                lda: i32,
                b: *const f32,
                ldb: i32,
                beta: f32,
                c: *mut f32,
                ldc: i32,
            );
        }

        let mut predicted_tokens = Vec::with_capacity(n_extra);
        let mut logits_buf = vec![0.0f32; vocab];

        for i in 0..n_extra {
            // lm_head: logits = last_hidden @ lm_weights[i]^T
            #[cfg(target_os = "macos")]
            unsafe {
                cblas_sgemm(
                    101,
                    111,
                    112,
                    1,
                    vocab as i32,
                    h as i32,
                    1.0,
                    last_hidden.as_ptr(),
                    h as i32,
                    self.lm_raw[i].as_ptr(),
                    h as i32,
                    0.0,
                    logits_buf.as_mut_ptr(),
                    vocab as i32,
                );
            }
            #[cfg(not(target_os = "macos"))]
            for j in 0..vocab {
                let mut s = 0.0f32;
                for k in 0..h {
                    s += last_hidden[k] * self.lm_raw[i][j * h + k];
                }
                logits_buf[j] = s;
            }

            let token = crate::executor::tts_executor::sample_token(&logits_buf, 0.0, top_k, 1.0);
            predicted_tokens.push(token);

            if i < n_extra - 1 {
                // Raw embedding lookup
                let t = token as usize;
                let embed = &self.emb_raw[i][t * emb_dim..(t + 1) * emb_dim];

                // Project if needed (1.7B: 2048→1024) using cached weights + cblas
                let embed = if let Some(ref w) = self.proj_w_raw {
                    let od = self.proj_out_dim;
                    let mut projected = vec![0.0f32; od];
                    #[cfg(target_os = "macos")]
                    unsafe {
                        // cblas_sgemv: y = alpha * A * x + beta * y
                        // A=[od, emb_dim] row-major, x=[emb_dim], y=[od]
                        extern "C" {
                            fn cblas_sgemv(
                                order: i32,
                                trans: i32,
                                m: i32,
                                n: i32,
                                alpha: f32,
                                a: *const f32,
                                lda: i32,
                                x: *const f32,
                                incx: i32,
                                beta: f32,
                                y: *mut f32,
                                incy: i32,
                            );
                        }
                        cblas_sgemv(
                            101,
                            111,
                            od as i32,
                            emb_dim as i32,
                            1.0,
                            w.as_ptr(),
                            emb_dim as i32,
                            embed.as_ptr(),
                            1,
                            0.0,
                            projected.as_mut_ptr(),
                            1,
                        );
                    }
                    #[cfg(not(target_os = "macos"))]
                    for j in 0..od {
                        let mut s = 0.0f32;
                        for k in 0..emb_dim {
                            s += embed[k] * w[j * emb_dim + k];
                        }
                        projected[j] = s;
                    }
                    if let Some(ref b) = self.proj_b_raw {
                        for j in 0..od {
                            projected[j] += b[j];
                        }
                    }
                    projected
                } else {
                    embed.to_vec()
                };

                // Forward through Backend<B> override if installed, else
                // fused transformer (1 token). Metal GPU path shortcut still
                // applies on the fused fallback.
                if let Some(ref mut backend) = self.backend_override {
                    last_hidden = backend.forward(&embed, 1);
                } else {
                    #[cfg(feature = "metal")]
                    {
                        if let Some(output) = self.fused.forward_gpu_to_vec(&embed, 1) {
                            last_hidden = output;
                        } else {
                            last_hidden = self.fused.forward(&embed, 1);
                        }
                    }
                    #[cfg(not(feature = "metal"))]
                    {
                        last_hidden = self.fused.forward(&embed, 1);
                    }
                }
            }
        }

        Ok(predicted_tokens)
    }

    fn forward_layers(&mut self, input: &Tensor) -> Result<Tensor> {
        let seq_len = input
            .dim(1)
            .map_err(|e| FerrumError::model(format!("dim: {e}")))?;
        let h = self.fused_hidden_size;

        // Fast path: if already on CPU f32, avoid extra copies
        let input_data = if input.device().is_cpu() && input.dtype() == DType::F32 {
            input
                .flatten_all()
                .and_then(|t| t.to_vec1())
                .map_err(|e| FerrumError::model(format!("st extract: {e}")))?
        } else {
            input
                .to_device(&candle_core::Device::Cpu)
                .and_then(|t| t.to_dtype(DType::F32))
                .and_then(|t| t.flatten_all())
                .and_then(|t| t.to_vec1())
                .map_err(|e| FerrumError::model(format!("st extract: {e}")))?
        };

        let output = if let Some(ref mut backend) = self.backend_override {
            backend.forward(&input_data, seq_len)
        } else {
            self.fused.forward(&input_data, seq_len)
        };

        Tensor::from_vec(output, (1, seq_len, h), &candle_core::Device::Cpu)
            .map_err(|e| FerrumError::model(format!("st output: {e}")))
    }

    pub fn reset(&mut self) {
        self.tokens_generated = 0;
        self.fused.reset();
        if let Some(ref mut backend) = self.backend_override {
            backend.reset();
        }
        for layer in &mut self.layers {
            layer.reset_cache();
        }
    }
}