inferencelayer 0.2.10

//! SpeechT5 TTS (`microsoft/speecht5_tts`) — text → mel spectrogram → waveform. Pure Rust, CPU.
//!
//! The engine's first synthesis model. Three stages, each gated against a `transformers` oracle
//! (`tests/fixtures/export_speecht5.py`):
//!   1. char-level SentencePiece tokenizer + text encoder — a **post-LN** transformer (attn →
//!      add → LN; ff → add → LN) with an initial LayerNorm before layer 0 and a relative position
//!      bias in self-attention: `score(i,j) += (q_i·scale)·pe_k[clamp(i-j, ±160)]` (encoder only).
//!      Token embeddings get a *scaled* sinusoidal positional encoding (interleaved sin/cos,
//!      learned scalar `alpha`); the pe tables are dumped by the export, not recomputed here.
//!   2. autoregressive mel decoder (prenet + cross-attention + feat_out/prob_out + postnet).
//!   3. HiFi-GAN vocoder (mel → waveform).
//!
//! ⚠️ HF's decoder prenet applies dropout even at inference (deliberate, for prosody variation).
//! The oracle is exported with that dropout disabled and this port implements none — generation
//! here is DETERMINISTIC by design.

use std::collections::HashMap;
use std::path::Path;

use anyhow::{Context, Result};
use rayon::prelude::*;

use crate::weights::LazySt;

/// `y[m,n] = x[m,k]·Wᵀ + b`, W is `[n,k]` row-major (HF Linear). Parallel over rows.
fn linear(x: &[f32], w: &[f32], b: &[f32], m: usize, n: usize, k: usize) -> Vec<f32> {
    let mut out = vec![0f32; m * n];
    out.par_chunks_mut(n).enumerate().for_each(|(i, orow)| {
        let xr = &x[i * k..][..k];
        for o in 0..n {
            let wr = &w[o * k..][..k];
            let mut acc = b[o];
            for c in 0..k {
                acc += xr[c] * wr[c];
            }
            orow[o] = acc;
        }
    });
    out
}

/// Row-wise LayerNorm (biased variance, eps), `[m, d]` → new vec.
fn layer_norm(x: &[f32], d: usize, w: &[f32], b: &[f32], eps: f32) -> Vec<f32> {
    let mut out = vec![0f32; x.len()];
    out.par_chunks_mut(d)
        .zip(x.par_chunks(d))
        .for_each(|(orow, row)| {
            let mean = row.iter().sum::<f32>() / d as f32;
            let var = row.iter().map(|v| (v - mean) * (v - mean)).sum::<f32>() / d as f32;
            let inv = 1.0 / (var + eps).sqrt();
            for i in 0..d {
                orow[i] = (row[i] - mean) * inv * w[i] + b[i];
            }
        });
    out
}

/// Exact (erf) GELU — HF's `"gelu"`.
fn gelu(v: f32) -> f32 {
    0.5 * v * (1.0 + libm::erff(v * std::f32::consts::FRAC_1_SQRT_2))
}

struct Linear {
    w: Vec<f32>,
    b: Vec<f32>,
    n: usize,
    k: usize,
}
impl Linear {
    fn load(st: &LazySt, prefix: &str, n: usize, k: usize) -> Result<Self> {
        let w = st.tensor_f32(&format!("{prefix}.weight"))?;
        let b = st.tensor_f32(&format!("{prefix}.bias"))?;
        anyhow::ensure!(w.len() == n * k, "{prefix}.weight {} != {n}x{k}", w.len());
        Ok(Self { w, b, n, k })
    }
    fn forward(&self, x: &[f32], m: usize) -> Vec<f32> {
        linear(x, &self.w, &self.b, m, self.n, self.k)
    }
}

struct Norm {
    w: Vec<f32>,
    b: Vec<f32>,
    eps: f32,
}
impl Norm {
    fn load(st: &LazySt, prefix: &str, eps: f32) -> Result<Self> {
        Ok(Self {
            w: st.tensor_f32(&format!("{prefix}.weight"))?,
            b: st.tensor_f32(&format!("{prefix}.bias"))?,
            eps,
        })
    }
    fn forward(&self, x: &[f32], d: usize) -> Vec<f32> {
        layer_norm(x, d, &self.w, &self.b, self.eps)
    }
}

struct EncBlock {
    q: Linear,
    k: Linear,
    v: Linear,
    out: Linear,
    ln1: Norm, // layer_norm (after attention residual — post-LN)
    fc1: Linear,
    fc2: Linear,
    ln2: Norm, // final_layer_norm (after ff residual)
}

struct DecBlock {
    sa_q: Linear,
    sa_k: Linear,
    sa_v: Linear,
    sa_out: Linear,
    sa_ln: Norm, // self_attn_layer_norm (post-LN)
    ca_q: Linear,
    ca_k: Linear,
    ca_v: Linear,
    ca_out: Linear,
    ca_ln: Norm, // encoder_attn_layer_norm
    fc1: Linear,
    fc2: Linear,
    ln2: Norm, // final_layer_norm
}

/// Conv1d + BatchNorm1d (+ tanh, except the last layer) — one postnet layer. Conv has no bias.
struct PostnetLayer {
    w: Vec<f32>, // [out, in, k]
    bn_w: Vec<f32>,
    bn_b: Vec<f32>,
    bn_rm: Vec<f32>,
    bn_rv: Vec<f32>,
    out_ch: usize,
    in_ch: usize,
    k: usize,
    tanh: bool,
}

/// Plain Conv1d (stride 1, zero-pad, optional dilation), weights `[out, in, k]`.
struct Conv1d {
    w: Vec<f32>,
    b: Vec<f32>,
    out_ch: usize,
    in_ch: usize,
    k: usize,
    dilation: usize,
    pad: usize,
}
impl Conv1d {
    fn load(
        st: &LazySt,
        prefix: &str,
        out_ch: usize,
        in_ch: usize,
        k: usize,
        dilation: usize,
    ) -> Result<Self> {
        let w = st.tensor_f32(&format!("{prefix}.weight"))?;
        anyhow::ensure!(
            w.len() == out_ch * in_ch * k,
            "{prefix}.weight {} != {out_ch}x{in_ch}x{k}",
            w.len()
        );
        Ok(Self {
            w,
            b: st.tensor_f32(&format!("{prefix}.bias"))?,
            out_ch,
            in_ch,
            k,
            dilation,
            pad: (k * dilation - dilation) / 2,
        })
    }
    /// `x` channel-major `[in_ch, t]` → `[out_ch, t]` (stride 1, "same" length).
    fn forward(&self, x: &[f32], t: usize) -> Vec<f32> {
        let mut y = vec![0f32; self.out_ch * t];
        y.par_chunks_mut(t).enumerate().for_each(|(o, orow)| {
            for (ti, oval) in orow.iter_mut().enumerate() {
                let mut acc = self.b[o];
                for ic in 0..self.in_ch {
                    let xrow = &x[ic * t..][..t];
                    let wrow = &self.w[(o * self.in_ch + ic) * self.k..][..self.k];
                    for (kk, wv) in wrow.iter().enumerate() {
                        let src = ti + kk * self.dilation;
                        if src >= self.pad && src - self.pad < t {
                            acc += xrow[src - self.pad] * wv;
                        }
                    }
                }
                *oval = acc;
            }
        });
        y
    }
}

/// ConvTranspose1d (the HiFi-GAN upsampler), weights `[in, out, k]` (torch layout).
struct ConvT1d {
    w: Vec<f32>,
    b: Vec<f32>,
    out_ch: usize,
    in_ch: usize,
    k: usize,
    stride: usize,
    pad: usize,
}
impl ConvT1d {
    fn load(
        st: &LazySt,
        prefix: &str,
        in_ch: usize,
        out_ch: usize,
        k: usize,
        stride: usize,
    ) -> Result<Self> {
        let w = st.tensor_f32(&format!("{prefix}.weight"))?;
        anyhow::ensure!(
            w.len() == in_ch * out_ch * k,
            "{prefix}.weight {} != {in_ch}x{out_ch}x{k}",
            w.len()
        );
        Ok(Self {
            w,
            b: st.tensor_f32(&format!("{prefix}.bias"))?,
            out_ch,
            in_ch,
            k,
            stride,
            pad: (k - stride) / 2,
        })
    }
    /// `x` `[in_ch, t]` → `[out_ch, (t-1)·stride + k − 2·pad]`.
    fn forward(&self, x: &[f32], t: usize) -> (Vec<f32>, usize) {
        let t_out = (t - 1) * self.stride + self.k - 2 * self.pad;
        let mut y = vec![0f32; self.out_ch * t_out];
        y.par_chunks_mut(t_out).enumerate().for_each(|(o, orow)| {
            for (ti, oval) in orow.iter_mut().enumerate() {
                let mut acc = self.b[o];
                for kk in 0..self.k {
                    let up = ti + self.pad;
                    if up < kk || !(up - kk).is_multiple_of(self.stride) {
                        continue;
                    }
                    let src = (up - kk) / self.stride;
                    if src >= t {
                        continue;
                    }
                    for ic in 0..self.in_ch {
                        acc += x[ic * t + src] * self.w[(ic * self.out_ch + o) * self.k + kk];
                    }
                }
                *oval = acc;
            }
        });
        (y, t_out)
    }
}

/// One MRF residual block: pairs of (dilated conv1, plain conv2), each pair pre-activated with
/// leaky-relu and residually added.
struct ResBlock {
    convs1: Vec<Conv1d>,
    convs2: Vec<Conv1d>,
}
impl ResBlock {
    fn forward(&self, x: &[f32], ch: usize, t: usize, slope: f32) -> Vec<f32> {
        let lrelu = |v: &mut Vec<f32>| {
            for e in v.iter_mut() {
                if *e < 0.0 {
                    *e *= slope;
                }
            }
        };
        let mut x = x.to_vec();
        for (c1, c2) in self.convs1.iter().zip(&self.convs2) {
            let mut h = x.clone();
            lrelu(&mut h);
            let mut h = c1.forward(&h, t);
            lrelu(&mut h);
            let h = c2.forward(&h, t);
            for i in 0..ch * t {
                x[i] += h[i];
            }
        }
        x
    }
}

/// HiFi-GAN vocoder (`microsoft/speecht5_hifigan`): log-mel `[t, 80]` → waveform `[t·256]`.
/// Feed-forward (not autoregressive). The input is normalized with the checkpoint's per-bin
/// mean/scale stats (`normalize_before`).
struct HifiGan {
    mean: Vec<f32>,
    scale: Vec<f32>,
    conv_pre: Conv1d,
    upsampler: Vec<ConvT1d>,
    resblocks: Vec<ResBlock>, // upsampler.len() × num_kernels
    conv_post: Conv1d,
    num_kernels: usize,
    slope: f32,
}

impl HifiGan {
    fn load(st: &LazySt, cfg: &serde_json::Value) -> Result<Self> {
        let g = |k: &str| cfg.get(k).and_then(|x| x.as_u64()).unwrap_or(0) as usize;
        let arr = |k: &str| -> Vec<usize> {
            cfg.get(k)
                .and_then(|x| x.as_array())
                .map(|a| {
                    a.iter()
                        .filter_map(|v| v.as_u64().map(|u| u as usize))
                        .collect()
                })
                .unwrap_or_default()
        };
        let ch0 = g("upsample_initial_channel");
        let rates = arr("upsample_rates");
        let kernels = arr("upsample_kernel_sizes");
        let res_k = arr("resblock_kernel_sizes");
        let dils: Vec<Vec<usize>> = cfg
            .get("resblock_dilation_sizes")
            .and_then(|x| x.as_array())
            .map(|a| {
                a.iter()
                    .map(|row| {
                        row.as_array()
                            .map(|r| {
                                r.iter()
                                    .filter_map(|v| v.as_u64().map(|u| u as usize))
                                    .collect()
                            })
                            .unwrap_or_default()
                    })
                    .collect()
            })
            .unwrap_or_default();
        let upsampler = rates
            .iter()
            .zip(&kernels)
            .enumerate()
            .map(|(i, (&r, &k))| {
                ConvT1d::load(
                    st,
                    &format!("hifigan.upsampler.{i}"),
                    ch0 >> i,
                    ch0 >> (i + 1),
                    k,
                    r,
                )
            })
            .collect::<Result<Vec<_>>>()?;
        let mut resblocks = Vec::new();
        for i in 0..rates.len() {
            let ch = ch0 >> (i + 1);
            for (j, (&k, dil)) in res_k.iter().zip(&dils).enumerate() {
                let b = format!("hifigan.resblocks.{}", i * res_k.len() + j);
                resblocks.push(ResBlock {
                    convs1: dil
                        .iter()
                        .enumerate()
                        .map(|(n, &d)| Conv1d::load(st, &format!("{b}.convs1.{n}"), ch, ch, k, d))
                        .collect::<Result<Vec<_>>>()?,
                    convs2: dil
                        .iter()
                        .enumerate()
                        .map(|(n, _)| Conv1d::load(st, &format!("{b}.convs2.{n}"), ch, ch, k, 1))
                        .collect::<Result<Vec<_>>>()?,
                });
            }
        }
        let last_ch = ch0 >> rates.len();
        Ok(Self {
            mean: st.tensor_f32("hifigan.mean")?,
            scale: st.tensor_f32("hifigan.scale")?,
            conv_pre: Conv1d::load(st, "hifigan.conv_pre", ch0, g("model_in_dim"), 7, 1)?,
            upsampler,
            resblocks,
            conv_post: Conv1d::load(st, "hifigan.conv_post", 1, last_ch, 7, 1)?,
            num_kernels: res_k.len(),
            slope: cfg
                .get("leaky_relu_slope")
                .and_then(|x| x.as_f64())
                .unwrap_or(0.1) as f32,
        })
    }

    /// mel `[t, n_mels]` (time-major, as the decoder emits) → waveform samples.
    fn forward(&self, mel: &[f32], t: usize) -> Vec<f32> {
        let nm = self.mean.len();
        // normalize_before + transpose to channel-major [n_mels, t]
        let mut x = vec![0f32; nm * t];
        for i in 0..t {
            for c in 0..nm {
                x[c * t + i] = (mel[i * nm + c] - self.mean[c]) / self.scale[c];
            }
        }
        let mut x = self.conv_pre.forward(&x, t);
        let mut t = t;
        for (i, up) in self.upsampler.iter().enumerate() {
            for e in x.iter_mut() {
                if *e < 0.0 {
                    *e *= self.slope;
                }
            }
            let (xu, tu) = up.forward(&x, t);
            t = tu;
            let ch = up.out_ch;
            let mut acc = self.resblocks[i * self.num_kernels].forward(&xu, ch, t, self.slope);
            for j in 1..self.num_kernels {
                let r = self.resblocks[i * self.num_kernels + j].forward(&xu, ch, t, self.slope);
                for (a, b) in acc.iter_mut().zip(r) {
                    *a += b;
                }
            }
            let inv = 1.0 / self.num_kernels as f32;
            for a in acc.iter_mut() {
                *a *= inv;
            }
            x = acc;
        }
        // ⚠️ HF calls this leaky_relu WITHOUT a slope — torch's default 0.01, NOT the config 0.1
        for e in x.iter_mut() {
            if *e < 0.0 {
                *e *= 0.01;
            }
        }
        let y = self.conv_post.forward(&x, t);
        y.iter().map(|v| v.tanh()).collect()
    }
}

/// Everything the mel decoder produces, per step — the parity gates check each tensor.
pub struct MelTrace {
    /// what entered the decoder each step (prenet output incl. speaker), `[steps, hidden]`
    pub prenet: Vec<f32>,
    /// decoder output hidden state per step, `[steps, hidden]`
    pub dec_h: Vec<f32>,
    /// stop logits (pre-sigmoid) per step, `[steps, reduction_factor]`
    pub stop_logits: Vec<f32>,
    /// mel before the postnet, `[steps·reduction_factor, n_mels]`
    pub mel_pre: Vec<f32>,
    /// final mel (`mel_pre + postnet(mel_pre)`), same shape
    pub mel_post: Vec<f32>,
    pub steps: usize,
}

pub struct SpeechT5Tts {
    // tokenizer (char-level SentencePiece: dummy '▁' prefix, ' ' → '▁', one piece per char)
    vocab: HashMap<String, u32>,
    unk_id: u32,
    eos_id: u32,
    // text encoder
    embed_tokens: Vec<f32>, // [vocab, hidden]
    enc_alpha: f32,
    enc_pe: Vec<f32>, // [max_text_positions, hidden] sinusoidal (interleaved sin/cos)
    enc_ln0: Norm,    // wrapped_encoder.layer_norm — applied once BEFORE layer 0
    pe_k: Vec<f32>,   // [2·max_rel, head_dim] relative-position embedding
    max_rel: i64,
    enc_blocks: Vec<EncBlock>,
    // mel decoder
    dec_alpha: f32,
    dec_pe: Vec<f32>,      // [max_speech_positions, hidden]
    prenet: Vec<Linear>,   // n_mels → units → units, ReLU (dropout NOT ported — see module doc)
    prenet_final: Linear,  // units → hidden
    speaker_layer: Linear, // hidden + speaker_dim → hidden, ReLU
    dec_blocks: Vec<DecBlock>,
    feat_out: Linear, // hidden → n_mels·reduction_factor
    prob_out: Linear, // hidden → reduction_factor (stop logits)
    postnet: Vec<PostnetLayer>,
    hifigan: HifiGan,
    n_mels: usize,
    reduction: usize,
    hidden: usize,
    heads: usize,
    hd: usize,
    sampling_rate: u32,
}

impl SpeechT5Tts {
    pub fn load(dir: &Path) -> Result<Self> {
        let v: serde_json::Value =
            serde_json::from_slice(&std::fs::read(dir.join("config.json"))?).context("config")?;
        let g = |k: &str| v.get(k).and_then(|x| x.as_u64()).unwrap_or(0) as usize;
        let st = LazySt::open(dir)?;
        let hidden = g("hidden_size");
        let heads = g("encoder_attention_heads");
        let inter = g("encoder_ffn_dim");
        let eps = v
            .get("layer_norm_eps")
            .and_then(|x| x.as_f64())
            .unwrap_or(1e-5) as f32;
        let vocab = v
            .get("vocab")
            .and_then(|m| m.as_object())
            .context("config vocab")?
            .iter()
            .filter_map(|(k, val)| val.as_u64().map(|i| (k.clone(), i as u32)))
            .collect();
        let enc = "speecht5.encoder";
        let enc_blocks = (0..g("encoder_layers"))
            .map(|i| {
                let b = format!("{enc}.wrapped_encoder.layers.{i}");
                Ok(EncBlock {
                    q: Linear::load(&st, &format!("{b}.attention.q_proj"), hidden, hidden)?,
                    k: Linear::load(&st, &format!("{b}.attention.k_proj"), hidden, hidden)?,
                    v: Linear::load(&st, &format!("{b}.attention.v_proj"), hidden, hidden)?,
                    out: Linear::load(&st, &format!("{b}.attention.out_proj"), hidden, hidden)?,
                    ln1: Norm::load(&st, &format!("{b}.layer_norm"), eps)?,
                    fc1: Linear::load(
                        &st,
                        &format!("{b}.feed_forward.intermediate_dense"),
                        inter,
                        hidden,
                    )?,
                    fc2: Linear::load(
                        &st,
                        &format!("{b}.feed_forward.output_dense"),
                        hidden,
                        inter,
                    )?,
                    ln2: Norm::load(&st, &format!("{b}.final_layer_norm"), eps)?,
                })
            })
            .collect::<Result<Vec<_>>>()?;
        let dec = "speecht5.decoder";
        let dec_blocks = (0..g("decoder_layers"))
            .map(|i| {
                let b = format!("{dec}.wrapped_decoder.layers.{i}");
                Ok(DecBlock {
                    sa_q: Linear::load(&st, &format!("{b}.self_attn.q_proj"), hidden, hidden)?,
                    sa_k: Linear::load(&st, &format!("{b}.self_attn.k_proj"), hidden, hidden)?,
                    sa_v: Linear::load(&st, &format!("{b}.self_attn.v_proj"), hidden, hidden)?,
                    sa_out: Linear::load(&st, &format!("{b}.self_attn.out_proj"), hidden, hidden)?,
                    sa_ln: Norm::load(&st, &format!("{b}.self_attn_layer_norm"), eps)?,
                    ca_q: Linear::load(&st, &format!("{b}.encoder_attn.q_proj"), hidden, hidden)?,
                    ca_k: Linear::load(&st, &format!("{b}.encoder_attn.k_proj"), hidden, hidden)?,
                    ca_v: Linear::load(&st, &format!("{b}.encoder_attn.v_proj"), hidden, hidden)?,
                    ca_out: Linear::load(
                        &st,
                        &format!("{b}.encoder_attn.out_proj"),
                        hidden,
                        hidden,
                    )?,
                    ca_ln: Norm::load(&st, &format!("{b}.encoder_attn_layer_norm"), eps)?,
                    fc1: Linear::load(
                        &st,
                        &format!("{b}.feed_forward.intermediate_dense"),
                        g("decoder_ffn_dim"),
                        hidden,
                    )?,
                    fc2: Linear::load(
                        &st,
                        &format!("{b}.feed_forward.output_dense"),
                        hidden,
                        g("decoder_ffn_dim"),
                    )?,
                    ln2: Norm::load(&st, &format!("{b}.final_layer_norm"), eps)?,
                })
            })
            .collect::<Result<Vec<_>>>()?;
        let (n_mels, units, rf) = (
            g("num_mel_bins"),
            g("speech_decoder_prenet_units"),
            g("reduction_factor"),
        );
        let prenet = (0..g("speech_decoder_prenet_layers"))
            .map(|i| {
                Linear::load(
                    &st,
                    &format!("{dec}.prenet.layers.{i}"),
                    units,
                    if i == 0 { n_mels } else { units },
                )
            })
            .collect::<Result<Vec<_>>>()?;
        let post_units = g("speech_decoder_postnet_units");
        let post_layers = g("speech_decoder_postnet_layers");
        let postnet = (0..post_layers)
            .map(|i| {
                let b = format!("speech_decoder_postnet.layers.{i}");
                let (in_ch, out_ch) = (
                    if i == 0 { n_mels } else { post_units },
                    if i == post_layers - 1 {
                        n_mels
                    } else {
                        post_units
                    },
                );
                let w = st.tensor_f32(&format!("{b}.conv.weight"))?;
                let k = g("speech_decoder_postnet_kernel");
                anyhow::ensure!(w.len() == out_ch * in_ch * k, "{b}.conv.weight shape");
                Ok(PostnetLayer {
                    w,
                    bn_w: st.tensor_f32(&format!("{b}.batch_norm.weight"))?,
                    bn_b: st.tensor_f32(&format!("{b}.batch_norm.bias"))?,
                    bn_rm: st.tensor_f32(&format!("{b}.batch_norm.running_mean"))?,
                    bn_rv: st.tensor_f32(&format!("{b}.batch_norm.running_var"))?,
                    out_ch,
                    in_ch,
                    k,
                    tanh: i < post_layers - 1,
                })
            })
            .collect::<Result<Vec<_>>>()?;
        Ok(Self {
            vocab,
            unk_id: g("unk_token_id") as u32,
            eos_id: g("eos_token_id") as u32,
            embed_tokens: st.tensor_f32(&format!("{enc}.prenet.embed_tokens.weight"))?,
            enc_alpha: st.tensor_f32(&format!("{enc}.prenet.encode_positions.alpha"))?[0],
            enc_pe: st.tensor_f32("enc_pe")?,
            enc_ln0: Norm::load(&st, &format!("{enc}.wrapped_encoder.layer_norm"), eps)?,
            pe_k: st.tensor_f32(&format!(
                "{enc}.wrapped_encoder.embed_positions.pe_k.weight"
            ))?,
            max_rel: g("encoder_max_relative_position") as i64,
            enc_blocks,
            dec_alpha: st.tensor_f32(&format!("{dec}.prenet.encode_positions.alpha"))?[0],
            dec_pe: st.tensor_f32("dec_pe")?,
            prenet,
            prenet_final: Linear::load(&st, &format!("{dec}.prenet.final_layer"), hidden, units)?,
            speaker_layer: Linear::load(
                &st,
                &format!("{dec}.prenet.speaker_embeds_layer"),
                hidden,
                hidden + g("speaker_embedding_dim"),
            )?,
            dec_blocks,
            feat_out: Linear::load(&st, "speech_decoder_postnet.feat_out", n_mels * rf, hidden)?,
            prob_out: Linear::load(&st, "speech_decoder_postnet.prob_out", rf, hidden)?,
            postnet,
            hifigan: HifiGan::load(&st, v.get("hifigan").context("config hifigan")?)?,
            n_mels,
            reduction: rf,
            hidden,
            heads,
            hd: hidden / heads,
            sampling_rate: v
                .get("hifigan")
                .and_then(|h| h.get("sampling_rate"))
                .and_then(|x| x.as_u64())
                .unwrap_or(16000) as u32,
        })
    }

    pub fn sampling_rate(&self) -> u32 {
        self.sampling_rate
    }

    /// Text → token ids, matching `SpeechT5Tokenizer` (char-level SentencePiece + `</s>`).
    /// SentencePiece's default normalizer trims + collapses whitespace runs and prepends a dummy
    /// `▁`; every remaining char is its own piece (NFKC is not applied — callers feed plain text;
    /// exotic codepoints map to `<unk>` exactly as they would if absent from the 81-piece vocab).
    pub fn tokenize(&self, text: &str) -> Vec<u32> {
        let sep = self.vocab.get("▁").copied().unwrap_or(self.unk_id);
        let mut ids = Vec::with_capacity(text.len() + 2);
        let mut pending_sep = true; // dummy prefix at start; whitespace runs collapse to one '▁'
        for ch in text.chars() {
            if ch.is_whitespace() {
                pending_sep = true;
                continue;
            }
            if pending_sep {
                ids.push(sep);
                pending_sep = false;
            }
            ids.push(
                *self
                    .vocab
                    .get(ch.to_string().as_str())
                    .unwrap_or(&self.unk_id),
            );
        }
        ids.push(self.eos_id); // trailing whitespace never flushes — stripped like SP does
        ids
    }

    /// Encoder self-attention with the relative position bias (encoder only):
    /// `score(i,j) = (q_i·k_j)·s + (q_i·s)·pe_k[clamp(i−j, −R, R−1) + R]`, s = 1/√hd.
    fn enc_attention(&self, blk: &EncBlock, x: &[f32], t: usize) -> Vec<f32> {
        let (h, heads, hd) = (self.hidden, self.heads, self.hd);
        let q = blk.q.forward(x, t);
        let k = blk.k.forward(x, t);
        let v = blk.v.forward(x, t);
        let scale = 1.0 / (hd as f32).sqrt();
        let ctx: Vec<Vec<f32>> = (0..heads)
            .into_par_iter()
            .map(|head| {
                let mut oh = vec![0f32; t * hd];
                let mut srow = vec![0f32; t];
                for i in 0..t {
                    let qi = &q[i * h + head * hd..][..hd];
                    for j in 0..t {
                        let kj = &k[j * h + head * hd..][..hd];
                        let rel = (i as i64 - j as i64).clamp(-self.max_rel, self.max_rel - 1)
                            + self.max_rel;
                        let pe = &self.pe_k[rel as usize * hd..][..hd];
                        let mut dot = 0f32;
                        let mut bias = 0f32;
                        for c in 0..hd {
                            dot += qi[c] * kj[c];
                            bias += qi[c] * pe[c];
                        }
                        srow[j] = (dot + bias) * scale;
                    }
                    let max = srow.iter().cloned().fold(f32::NEG_INFINITY, f32::max);
                    let mut sum = 0.0;
                    for s in srow.iter_mut() {
                        *s = (*s - max).exp();
                        sum += *s;
                    }
                    let inv = 1.0 / sum;
                    let orow = &mut oh[i * hd..][..hd];
                    for j in 0..t {
                        let w = srow[j] * inv;
                        let vj = &v[j * h + head * hd..][..hd];
                        for c in 0..hd {
                            orow[c] += w * vj[c];
                        }
                    }
                }
                oh
            })
            .collect();
        let mut merged = vec![0f32; t * h];
        for (head, oh) in ctx.iter().enumerate() {
            for i in 0..t {
                merged[i * h + head * hd..][..hd].copy_from_slice(&oh[i * hd..][..hd]);
            }
        }
        blk.out.forward(&merged, t)
    }

    /// Token ids → encoder hidden `[t, hidden]` (embeddings + scaled positions → initial LN →
    /// post-LN blocks with relative attention bias).
    pub fn encode_text(&self, ids: &[u32]) -> Vec<f32> {
        let (h, t) = (self.hidden, ids.len());
        let mut x = vec![0f32; t * h];
        for (i, &id) in ids.iter().enumerate() {
            let emb = &self.embed_tokens[id as usize * h..][..h];
            let pe = &self.enc_pe[i * h..][..h];
            for c in 0..h {
                x[i * h + c] = emb[c] + self.enc_alpha * pe[c];
            }
        }
        let mut x = self.enc_ln0.forward(&x, h);
        for blk in &self.enc_blocks {
            let a = self.enc_attention(blk, &x, t);
            for (xi, ai) in x.iter_mut().zip(a) {
                *xi += ai;
            }
            x = blk.ln1.forward(&x, h);
            let mut mid = blk.fc1.forward(&x, t);
            for m in mid.iter_mut() {
                *m = gelu(*m);
            }
            let f = blk.fc2.forward(&mid, t);
            for (xi, fi) in x.iter_mut().zip(f) {
                *xi += fi;
            }
            x = blk.ln2.forward(&x, h);
        }
        x
    }

    /// One query attending over a cached `[t, hidden]` K/V (decoder self- and cross-attention;
    /// no relative bias in the decoder).
    fn attend_one(&self, q: &[f32], kc: &[f32], vc: &[f32], t: usize) -> Vec<f32> {
        let (h, hd) = (self.hidden, self.hd);
        let scale = 1.0 / (hd as f32).sqrt();
        let mut out = vec![0f32; h];
        out.par_chunks_mut(hd).enumerate().for_each(|(head, orow)| {
            let qh = &q[head * hd..][..hd];
            let mut srow = vec![0f32; t];
            for j in 0..t {
                let kj = &kc[j * h + head * hd..][..hd];
                srow[j] = qh.iter().zip(kj).map(|(a, b)| a * b).sum::<f32>() * scale;
            }
            let max = srow.iter().cloned().fold(f32::NEG_INFINITY, f32::max);
            let mut sum = 0.0;
            for s in srow.iter_mut() {
                *s = (*s - max).exp();
                sum += *s;
            }
            let inv = 1.0 / sum;
            for j in 0..t {
                let w = srow[j] * inv;
                let vj = &vc[j * h + head * hd..][..hd];
                for c in 0..hd {
                    orow[c] += w * vj[c];
                }
            }
        });
        out
    }

    /// Postnet: mel `[t, n_mels]` → residual refinement (5× Conv1d+BatchNorm(+tanh), eval-mode
    /// batchnorm with the checkpoint's running stats). Returns `mel + refinement`.
    fn postnet_forward(&self, mel: &[f32], t: usize) -> Vec<f32> {
        let nm = self.n_mels;
        // channel-major [ch, t] for the convs
        let mut x = vec![0f32; nm * t];
        for i in 0..t {
            for c in 0..nm {
                x[c * t + i] = mel[i * nm + c];
            }
        }
        for layer in &self.postnet {
            let pad = (layer.k - 1) / 2;
            let mut y = vec![0f32; layer.out_ch * t];
            y.par_chunks_mut(t).enumerate().for_each(|(o, orow)| {
                let inv = 1.0 / (layer.bn_rv[o] + 1e-5).sqrt(); // torch BatchNorm1d default eps
                for (ti, oval) in orow.iter_mut().enumerate() {
                    let mut acc = 0f32;
                    for ic in 0..layer.in_ch {
                        let xrow = &x[ic * t..][..t];
                        let wrow = &layer.w[(o * layer.in_ch + ic) * layer.k..][..layer.k];
                        for (kk, wv) in wrow.iter().enumerate() {
                            let src = ti + kk;
                            if src >= pad && src - pad < t {
                                acc += xrow[src - pad] * wv;
                            }
                        }
                    }
                    let bn = (acc - layer.bn_rm[o]) * inv * layer.bn_w[o] + layer.bn_b[o];
                    *oval = if layer.tanh { bn.tanh() } else { bn };
                }
            });
            x = y;
        }
        let mut out = mel.to_vec();
        for i in 0..t {
            for c in 0..nm {
                out[i * nm + c] += x[c * t + i];
            }
        }
        out
    }

    /// Autoregressive mel generation, mirroring HF `_generate_speech` (deterministic — no prenet
    /// dropout): zero frame → prenet(+speaker) → KV-cached decoder step → `reduction_factor`
    /// frames + stop logits per step; stop when Σ sigmoid(stop) ≥ 0.5 (or maxlen = ⌊t_enc·20/rf⌋);
    /// postnet refines the whole mel once at the end.
    pub fn decode_mel(&self, enc: &[f32], speaker: &[f32]) -> MelTrace {
        let (h, nm, rf) = (self.hidden, self.n_mels, self.reduction);
        let t_enc = enc.len() / h;
        let maxlen = t_enc * 20 / rf;
        // F.normalize(speaker): x / max(‖x‖₂, 1e-12)
        let norm = speaker.iter().map(|v| v * v).sum::<f32>().sqrt().max(1e-12);
        let spk: Vec<f32> = speaker.iter().map(|v| v / norm).collect();
        // cross-attention K/V — computed once per layer
        let cross_kv: Vec<(Vec<f32>, Vec<f32>)> = self
            .dec_blocks
            .iter()
            .map(|b| (b.ca_k.forward(enc, t_enc), b.ca_v.forward(enc, t_enc)))
            .collect();
        let mut sa_k: Vec<Vec<f32>> = vec![Vec::new(); self.dec_blocks.len()];
        let mut sa_v: Vec<Vec<f32>> = vec![Vec::new(); self.dec_blocks.len()];
        let mut tr = MelTrace {
            prenet: Vec::new(),
            dec_h: Vec::new(),
            stop_logits: Vec::new(),
            mel_pre: Vec::new(),
            mel_post: Vec::new(),
            steps: 0,
        };
        let mut prev = vec![0f32; nm]; // the all-zeros start frame
        for step in 0..maxlen {
            // prenet on the fed-back frame (position = step), then the speaker projection
            let mut x = prev.clone();
            for l in &self.prenet {
                x = l.forward(&x, 1);
                for v in x.iter_mut() {
                    *v = v.max(0.0);
                }
            }
            let mut x = self.prenet_final.forward(&x, 1);
            for (c, xv) in x.iter_mut().enumerate() {
                *xv += self.dec_alpha * self.dec_pe[step * h + c];
            }
            x.extend_from_slice(&spk);
            let mut x = self.speaker_layer.forward(&x, 1);
            for v in x.iter_mut() {
                *v = v.max(0.0);
            }
            tr.prenet.extend_from_slice(&x);
            // decoder layers (post-LN), single query over the growing cache
            for (li, blk) in self.dec_blocks.iter().enumerate() {
                let q = blk.sa_q.forward(&x, 1);
                sa_k[li].extend(blk.sa_k.forward(&x, 1));
                sa_v[li].extend(blk.sa_v.forward(&x, 1));
                let a = self.attend_one(&q, &sa_k[li], &sa_v[li], step + 1);
                let a = blk.sa_out.forward(&a, 1);
                for (xi, ai) in x.iter_mut().zip(a) {
                    *xi += ai;
                }
                x = blk.sa_ln.forward(&x, h);
                let q = blk.ca_q.forward(&x, 1);
                let a = self.attend_one(&q, &cross_kv[li].0, &cross_kv[li].1, t_enc);
                let a = blk.ca_out.forward(&a, 1);
                for (xi, ai) in x.iter_mut().zip(a) {
                    *xi += ai;
                }
                x = blk.ca_ln.forward(&x, h);
                let mut mid = blk.fc1.forward(&x, 1);
                for m in mid.iter_mut() {
                    *m = gelu(*m);
                }
                let f = blk.fc2.forward(&mid, 1);
                for (xi, fi) in x.iter_mut().zip(f) {
                    *xi += fi;
                }
                x = blk.ln2.forward(&x, h);
            }
            tr.dec_h.extend_from_slice(&x);
            let feat = self.feat_out.forward(&x, 1); // [rf, n_mels] row-major
            let stop = self.prob_out.forward(&x, 1); // [rf] logits
            tr.mel_pre.extend_from_slice(&feat);
            tr.stop_logits.extend_from_slice(&stop);
            prev.copy_from_slice(&feat[(rf - 1) * nm..]); // feed back the LAST frame only
            tr.steps = step + 1;
            let prob_sum: f32 = stop.iter().map(|l| 1.0 / (1.0 + (-l).exp())).sum();
            if prob_sum >= 0.5 {
                break;
            }
        }
        tr.mel_post = self.postnet_forward(&tr.mel_pre, tr.steps * rf);
        tr
    }

    /// HiFi-GAN: final mel `[t, n_mels]` → waveform (`hop` = ∏ upsample_rates samples per frame).
    pub fn vocode(&self, mel: &[f32]) -> Vec<f32> {
        self.hifigan.forward(mel, mel.len() / self.n_mels)
    }

    /// End-to-end synthesis: text + raw 512-d speaker x-vector → 16 kHz waveform. Deterministic.
    pub fn synthesize(&self, text: &str, speaker: &[f32]) -> Vec<f32> {
        let ids = self.tokenize(text);
        let enc = self.encode_text(&ids);
        let tr = self.decode_mel(&enc, speaker);
        self.vocode(&tr.mel_post)
    }
}