cortiq_engine/

dsv4.rs

//! DeepSeek-V4 blocks that no other supported architecture has.
//!
//! Transcribed from the reference `inference/model.py` + `inference/kernel.py`
//! shipped with the checkpoint, not inferred from tensor names — the pieces
//! below have enough hidden structure (a second normalization on the heads, a
//! bias that steers selection but not weights, a mixing matrix normalized by
//! Sinkhorn) that guessing produces a model which *almost* answers.
//!
//! Each function is the smallest unit the reference defines, so it can be
//! checked on its own. The forward that stitches them together comes after
//! the attention and compressor land.

/// Hyper-connections. The hidden state of this model is not a vector: it is
/// `hc` copies of one (`hc_mult`, 4 in the release). A block folds them to
/// one, runs attention or the FFN, then expands back — there is no ordinary
/// residual anywhere in the stack.
///
/// `mixes` is the per-token projection `F.linear(x.flatten(), hc_fn) * rsqrt`
/// of length `(2 + hc) * hc`; it splits into three parts:
///   * `pre[j]`  — how much of copy `j` goes into the folded vector,
///   * `post[j]` — how much of the block's output returns to copy `j`,
///   * `comb`    — an `hc x hc` matrix mixing the old copies into the new.
///
/// `comb` is made doubly stochastic by Sinkhorn: a row softmax, then
/// alternating row/column normalization. The reference runs the column step
/// once before the loop and `iters - 1` times inside it, which is why the
/// loop below starts from the column-normalized matrix.
pub fn hc_split_sinkhorn(
    mixes: &[f32],
    hc_scale: &[f32; 3],
    hc_base: &[f32],
    hc: usize,
    iters: usize,
    eps: f32,
    pre: &mut [f32],
    post: &mut [f32],
    comb: &mut [f32],
) {
    debug_assert_eq!(mixes.len(), (2 + hc) * hc);
    debug_assert_eq!(comb.len(), hc * hc);
    for j in 0..hc {
        pre[j] = sigmoid(mixes[j] * hc_scale[0] + hc_base[j]) + eps;
        // The post weights carry a factor 2 in the reference — with a
        // sigmoid alone the block's output could never exceed the residual.
        post[j] = 2.0 * sigmoid(mixes[j + hc] * hc_scale[1] + hc_base[j + hc]);
    }
    for j in 0..hc {
        for k in 0..hc {
            let idx = j * hc + k + hc * 2;
            comb[j * hc + k] = mixes[idx] * hc_scale[2] + hc_base[idx];
        }
    }
    // row softmax + eps
    for j in 0..hc {
        let row = &mut comb[j * hc..(j + 1) * hc];
        let m = row.iter().fold(f32::NEG_INFINITY, |a, &b| a.max(b));
        let mut sum = 0.0;
        for v in row.iter_mut() {
            *v = (*v - m).exp();
            sum += *v;
        }
        for v in row.iter_mut() {
            *v = *v / sum + eps;
        }
    }
    // one column normalization, then (iters - 1) row/column rounds
    normalize_cols(comb, hc, eps);
    for _ in 0..iters.saturating_sub(1) {
        normalize_rows(comb, hc, eps);
        normalize_cols(comb, hc, eps);
    }
}

fn normalize_rows(m: &mut [f32], n: usize, eps: f32) {
    for j in 0..n {
        let s: f32 = m[j * n..(j + 1) * n].iter().sum::<f32>() + eps;
        for v in m[j * n..(j + 1) * n].iter_mut() {
            *v /= s;
        }
    }
}

fn normalize_cols(m: &mut [f32], n: usize, eps: f32) {
    for k in 0..n {
        let mut s = eps;
        for j in 0..n {
            s += m[j * n + k];
        }
        for j in 0..n {
            m[j * n + k] /= s;
        }
    }
}

#[inline]
fn sigmoid(x: f32) -> f32 {
    1.0 / (1.0 + (-x).exp())
}

/// The projection feeding `hc_split_sinkhorn`: the `hc` copies are flattened
/// to one `hc*dim` vector, RMS-scaled (no learned weight — the reference uses
/// a bare rsqrt of the mean square), and projected by `hc_fn` `[mix_hc, hc*dim]`.
pub fn hc_mixes(
    x_flat: &[f32],
    hc_fn: &[f32],
    mix_hc: usize,
    eps: f32,
    pool: Option<&crate::pool::Pool>,
    out: &mut [f32],
) {
    let n = x_flat.len();
    debug_assert_eq!(hc_fn.len(), mix_hc * n);
    debug_assert_eq!(out.len(), mix_hc);
    let ms = x_flat.iter().map(|v| v * v).sum::<f32>() / n as f32;
    let rsqrt = 1.0 / (ms + eps).sqrt();
    // A dense f32 matvec of mix_hc rows over hc*dim — 1.6 MB read per call on
    // the release, and TWO calls per layer, so 135 MB a token. It ran on one
    // thread and cost more than the whole attention block.
    match pool {
        Some(p) if n >= 4096 => {
            let addr = crate::pool::SendMut::new(out.as_mut_ptr());
            p.run_rows(mix_hc, &|start, end| {
                for i in start..end {
                    let row = &hc_fn[i * n..(i + 1) * n];
                    let v = row.iter().zip(x_flat).map(|(a, b)| a * b).sum::<f32>() * rsqrt;
                    unsafe { *addr.at(i) = v };
                }
            });
        }
        _ => {
            for (i, o) in out.iter_mut().enumerate() {
                let row = &hc_fn[i * n..(i + 1) * n];
                *o = row.iter().zip(x_flat).map(|(a, b)| a * b).sum::<f32>() * rsqrt;
            }
        }
    }
}

/// Fold `hc` copies into one vector: `y = Σ_j pre[j] · x[j]`.
pub fn hc_fold(x: &[f32], pre: &[f32], hc: usize, dim: usize, out: &mut [f32]) {
    debug_assert_eq!(x.len(), hc * dim);
    out.fill(0.0);
    for j in 0..hc {
        let w = pre[j];
        let src = &x[j * dim..(j + 1) * dim];
        for (o, v) in out.iter_mut().zip(src) {
            *o += w * v;
        }
    }
}

/// Expand the block's output back into `hc` copies:
/// `y[j] = post[j] · out + Σ_k comb[k][j] · residual[k]`.
///
/// Note the transpose: the reference sums over the SECOND-to-last axis of
/// `comb.unsqueeze(-1) * residual.unsqueeze(-2)`, i.e. copy `k` of the
/// residual contributes to new copy `j` with weight `comb[k][j]`.
pub fn hc_expand(
    block_out: &[f32],
    residual: &[f32],
    post: &[f32],
    comb: &[f32],
    hc: usize,
    dim: usize,
    out: &mut [f32],
) {
    debug_assert_eq!(residual.len(), hc * dim);
    debug_assert_eq!(out.len(), hc * dim);
    for j in 0..hc {
        let dst = &mut out[j * dim..(j + 1) * dim];
        let p = post[j];
        for (d, o) in dst.iter_mut().enumerate() {
            *o = p * block_out[d];
        }
        for k in 0..hc {
            let w = comb[k * hc + j];
            let src = &residual[k * dim..(k + 1) * dim];
            for (o, v) in dst.iter_mut().zip(src) {
                *o += w * v;
            }
        }
    }
}

/// The head fold, run once after the last layer: same shape as `hc_fold`'s
/// weights but WITHOUT Sinkhorn — a plain sigmoid gate per copy.
pub fn hc_head_pre(mixes: &[f32], scale: f32, base: &[f32], hc: usize, eps: f32, pre: &mut [f32]) {
    for j in 0..hc {
        pre[j] = sigmoid(mixes[j] * scale + base[j]) + eps;
    }
}

/// MoE routing. Three details decide whether this model answers or merely
/// produces fluent text:
///   * the score is `sqrt(softplus(x))`, not a softmax or a sigmoid;
///   * the selection bias shifts WHICH experts win but never the weights —
///     those come from the pre-bias scores;
///   * the weights are renormalized over the chosen experts, then scaled.
///
/// `bias` is `None` on the hash layers, where `indices` come from a
/// token-id table instead (see `hash_route`).
/// `forced` fixes the chosen experts (the hash layers' token-id table). They
/// have to be known here rather than swapped in afterwards: the weights are
/// the scores gathered at whichever indices win, so substituting the indices
/// later leaves every weight attached to a different expert.
pub fn route(
    scores_in: &[f32],
    bias: Option<&[f32]>,
    top_k: usize,
    route_scale: f32,
    forced: Option<&[usize]>,
    mask: Option<&[bool]>,
    indices: &mut Vec<usize>,
    weights: &mut Vec<f32>,
) {
    let n = scores_in.len();
    let mut scores = Vec::with_capacity(n);
    for &s in scores_in {
        // softplus, guarded like the reference's F.softplus (linear past 20)
        let sp = if s > 20.0 { s } else { (1.0 + s.exp()).ln() };
        scores.push(sp.sqrt());
    }
    indices.clear();
    weights.clear();
    match forced {
        Some(f) => indices.extend(f.iter().copied()),
        None => {
            let mut shifted: Vec<f32> = match bias {
                Some(b) => scores.iter().zip(b).map(|(s, b)| s + b).collect(),
                None => scores.clone(),
            };
            if let Some(m) = mask {
                for (i, s) in shifted.iter_mut().enumerate() {
                    if !m.get(i).copied().unwrap_or(true) {
                        *s = f32::NEG_INFINITY;
                    }
                }
            }
            for _ in 0..top_k.min(n) {
                let mut best = 0usize;
                let mut bv = f32::NEG_INFINITY;
                for (i, &v) in shifted.iter().enumerate() {
                    if v > bv {
                        bv = v;
                        best = i;
                    }
                }
                if !bv.is_finite() {
                    break;
                }
                indices.push(best);
                shifted[best] = f32::NEG_INFINITY;
            }
        }
    }
    // The weight is always the PRE-bias score of the chosen expert.
    for &i in indices.iter() {
        weights.push(scores.get(i).copied().unwrap_or(0.0));
    }
    let sum: f32 = weights.iter().sum();
    if sum > 0.0 {
        for w in weights.iter_mut() {
            *w = *w / sum * route_scale;
        }
    }
}

/// Hash layers: the experts of token `tid` are a row of the `tid2eid` table,
/// and the router does not run at all. Their weights still come from the
/// scored path (the reference gathers `original_scores` at those indices).
pub fn hash_route(tid2eid: &[f32], vocab: usize, top_k: usize, tid: u32) -> Vec<usize> {
    let row = (tid as usize).min(vocab.saturating_sub(1)) * top_k;
    (0..top_k)
        .map(|k| tid2eid.get(row + k).copied().unwrap_or(0.0) as usize)
        .collect()
}

/// Rotary on the LAST `rd` dims only — the rest of the head carries no
/// position. `inverse` runs the rotation backwards, which the reference
/// applies to the attention OUTPUT before the o-projection (the value
/// stream carries the same rope-tagged tail as the keys, and it has to be
/// untagged again). Missing that step leaves a model that reads fluently
/// and attends to the wrong offsets.
pub fn rope_tail(v: &mut [f32], inv_freq: &[f32], pos: usize, rd: usize, inverse: bool) {
    let n = v.len();
    debug_assert!(
        rd <= n && rd % 2 == 0,
        "rope tail {rd} wider than the vector {n}"
    );
    // A tail wider than the vector is a configuration mistake, and `n - rd`
    // would wrap into an index in the billions rather than say so.
    let rd = rd.min(n) & !1;
    let base = n - rd;
    // ADJACENT pairs, not halves. The reference forms its complex numbers
    // with `unflatten(-1, (-1, 2))` + `view_as_complex`, i.e. (x0,x1),
    // (x2,x3), … — the interleaved convention. Half-split pairing agrees
    // with it exactly at position 0, where the rotation is the identity,
    // and disagrees everywhere else. That is why short answers came out
    // right and everything longer drifted, repeated itself and could not
    // count: every position past the first was rotated into the wrong
    // basis.
    for i in 0..rd / 2 {
        let theta = pos as f32 * inv_freq[i];
        let (s, c) = (theta.sin(), theta.cos());
        let s = if inverse { -s } else { s };
        let a = v[base + 2 * i];
        let b = v[base + 2 * i + 1];
        v[base + 2 * i] = a * c - b * s;
        v[base + 2 * i + 1] = a * s + b * c;
    }
}

/// RMS normalize in place with no learned weight — the reference applies
/// this to each attention head AFTER `wq_b`, on top of the `q_norm` that
/// already normalized the LoRA rank. Two normalizations, not one.
pub fn rms_inplace(v: &mut [f32], eps: f32) {
    let ms = v.iter().map(|x| x * x).sum::<f32>() / v.len() as f32;
    let inv = 1.0 / (ms + eps).sqrt();
    for x in v.iter_mut() {
        *x *= inv;
    }
}

/// Attention over an explicit position LIST (window ⊕ compressed), with a
/// learned per-head sink. The sink is an extra logit with no value vector:
/// it lets a head attend to "nothing", so its softmax denominator carries
/// `exp(sink - max)` while contributing no output. Index `usize::MAX`
/// marks a masked slot (the reference writes -1 into topk_idxs).
pub fn sparse_attend(
    q: &[f32],
    kv: &[f32],
    idxs: &[usize],
    sink: f32,
    scale: f32,
    head_dim: usize,
    out: &mut [f32],
) {
    let mut m = sink;
    let mut scores = Vec::with_capacity(idxs.len());
    for &p in idxs {
        if p == usize::MAX {
            scores.push(f32::NEG_INFINITY);
            continue;
        }
        let k = &kv[p * head_dim..(p + 1) * head_dim];
        let dot: f32 = q.iter().zip(k).map(|(a, b)| a * b).sum::<f32>() * scale;
        m = m.max(dot);
        scores.push(dot);
    }
    let mut denom = (sink - m).exp();
    out.fill(0.0);
    for (&p, &s) in idxs.iter().zip(&scores) {
        if p == usize::MAX {
            continue;
        }
        let w = (s - m).exp();
        denom += w;
        let v = &kv[p * head_dim..(p + 1) * head_dim];
        for (o, x) in out.iter_mut().zip(v) {
            *o += w * x;
        }
    }
    if std::env::var("CMF_ATTN_DEBUG").is_ok() {
        eprintln!(
            "    [порт] позиций={} score={:?} sink={sink:.4} denom={denom:.4} |q|={:.3}",
            idxs.iter().filter(|&&p| p != usize::MAX).count(),
            scores
                .iter()
                .map(|x| (x * 10000.0).round() / 10000.0)
                .collect::<Vec<_>>(),
            q.iter().map(|x| x * x).sum::<f32>().sqrt()
        );
    }
    let inv = 1.0 / denom;
    for o in out.iter_mut() {
        *o *= inv;
    }
}

/// The grouped low-rank output projection: heads are split into `groups`,
/// each group's slice is projected to `lora` by its own block of `wo_a`,
/// and the concatenation goes through `wo_b`. `wo_a` is stored
/// `[groups, lora, per_group]`.
/// `wo_a_row` is `(row, x) -> dot`, reading one row of `wo_a` against the
/// slice of `attn` its group owns; `wo_b` is the plain projection of the
/// concatenated groups. Both arrive as closures so the caller can serve them
/// straight from quantized tensors.
pub fn o_project(
    attn: &[f32],
    wo_a_row: &(dyn Fn(usize, &[f32], &mut [f32]) -> f32 + Sync),
    scratch_len: usize,
    wo_b: &dyn Fn(&[f32], &mut [f32]),
    groups: usize,
    lora: usize,
    pool: Option<&crate::pool::Pool>,
    out: &mut [f32],
) {
    let per_group = attn.len() / groups;
    let mut mid = vec![0.0f32; groups * lora];
    let slice_of = |i: usize| {
        let g = i / lora;
        &attn[g * per_group..(g + 1) * per_group]
    };
    match pool {
        // Each row of `mid` is one dot product against its group's slice —
        // independent, so the rows split cleanly. This is the largest
        // single-threaded cost in the decode otherwise: on the release
        // checkpoint wo_a is 33M weights, read once per layer per token.
        Some(p) if mid.len() >= 256 => {
            let addr = crate::pool::SendMut::new(mid.as_mut_ptr());
            p.run_rows(mid.len(), &|start, end| {
                let mut sc = vec![0.0f32; scratch_len];
                for i in start..end {
                    let v = wo_a_row(i, slice_of(i), &mut sc);
                    unsafe { *addr.at(i) = v };
                }
            });
        }
        _ => {
            let mut sc = vec![0.0f32; scratch_len];
            for (i, m) in mid.iter_mut().enumerate() {
                *m = wo_a_row(i, slice_of(i), &mut sc);
            }
        }
    }
    wo_b(&mid, out);
}

pub fn compress_window(
    kv: &[f32],
    score: &[f32],
    ape: &[f32],
    ratio: usize,
    width: usize,
    out: &mut [f32],
) {
    debug_assert_eq!(kv.len(), ratio * width);
    debug_assert_eq!(ape.len(), ratio * width);
    let biased: Vec<f32> = score.iter().zip(ape).map(|(s, a)| s + a).collect();
    pool_by_score(kv, &biased, ratio, width, out);
}

/// Softmax over the `slots` axis, per dimension, then the weighted sum —
/// the pooling both the plain and the overlapping compressor end in.
/// `-inf` scores are how an absent slot votes for nothing, so the
/// max-subtraction has to survive a whole column of them.
pub fn pool_by_score(kv: &[f32], score: &[f32], slots: usize, width: usize, out: &mut [f32]) {
    debug_assert_eq!(kv.len(), slots * width);
    debug_assert_eq!(score.len(), slots * width);
    out.fill(0.0);
    for d in 0..width {
        let mut m = f32::NEG_INFINITY;
        for t in 0..slots {
            m = m.max(score[t * width + d]);
        }
        if !m.is_finite() {
            continue;
        }
        let mut denom = 0.0;
        for t in 0..slots {
            denom += (score[t * width + d] - m).exp();
        }
        if denom <= 0.0 {
            continue;
        }
        for t in 0..slots {
            out[d] += ((score[t * width + d] - m).exp() / denom) * kv[t * width + d];
        }
    }
}

/// The overlapping compressor (the release uses it wherever the ratio is 4).
///
/// Each token contributes `2*d` values: the first half belongs to the window
/// that started half a stride earlier, the second half to the current one.
/// At fold time the reference pools `2*ratio` entries of width `d` — the
/// PREVIOUS window's slots taking their first half, the current window's
/// slots taking their second half — then the current window becomes the
/// previous one. An absent previous window votes with `-inf`.
#[allow(clippy::too_many_arguments)]
pub fn compress_window_overlap(
    prev_kv: &[f32],
    prev_score: &[f32],
    cur_kv: &[f32],
    cur_score: &[f32],
    ratio: usize,
    d: usize,
    out: &mut [f32],
) {
    let slots = 2 * ratio;
    let mut kv = vec![0.0f32; slots * d];
    let mut sc = vec![f32::NEG_INFINITY; slots * d];
    let have_prev = prev_kv.len() == ratio * 2 * d;
    for t in 0..ratio {
        if have_prev {
            // the previous window's slots, first half of the dimensions
            kv[t * d..(t + 1) * d].copy_from_slice(&prev_kv[t * 2 * d..t * 2 * d + d]);
            sc[t * d..(t + 1) * d].copy_from_slice(&prev_score[t * 2 * d..t * 2 * d + d]);
        }
        // the current window's slots, second half
        let src = t * 2 * d + d;
        let dst = (ratio + t) * d;
        kv[dst..dst + d].copy_from_slice(&cur_kv[src..src + d]);
        sc[dst..dst + d].copy_from_slice(&cur_score[src..src + d]);
    }
    pool_by_score(&kv, &sc, slots, d, out);
}

/// The sparse indexer's scoring pass. For each query it ranks the
/// compressed positions and keeps the best `topk`.
///
/// Three details from the reference that a shape-only reading misses:
///   * the query comes from the SHARED LoRA output `qr` (the output of
///     `q_norm(wq_a(x))`, before attention's own `wq_b`), through the
///     indexer's own `wq_b` — not from attention's queries;
///   * scores are **relu'd** before the per-head weighting, so a head can
///     only ever vote for a position, never against it;
///   * the per-head weights are a projection of the hidden state scaled by
///     `head_dim^-0.5 * n_heads^-0.5`.
///
/// `causal_limit` is the number of compressed positions this query may see
/// (`(pos + 1) / ratio`); anything at or past it is masked.
#[allow(clippy::too_many_arguments)]
pub fn index_scores(
    q_heads: &[f32],
    kv: &[f32],
    head_weights: &[f32],
    n_heads: usize,
    head_dim: usize,
    n_pos: usize,
    causal_limit: usize,
    pool: Option<&crate::pool::Pool>,
    out: &mut Vec<f32>,
) {
    out.clear();
    out.resize(n_pos, 0.0);
    let score_at = |t: usize| -> f32 {
        if t >= causal_limit {
            return f32::NEG_INFINITY;
        }
        let k = &kv[t * head_dim..(t + 1) * head_dim];
        let mut acc = 0.0;
        for h in 0..n_heads {
            let q = &q_heads[h * head_dim..(h + 1) * head_dim];
            let dot: f32 = q.iter().zip(k).map(|(a, b)| a * b).sum();
            // relu BEFORE weighting: a head votes for a position or abstains
            acc += dot.max(0.0) * head_weights[h];
        }
        acc
    };
    // Positions are independent, and their number grows with the context —
    // this was the one loop in the attention step still walking the whole
    // compressed axis on one thread.
    match pool {
        Some(p) if n_pos >= 64 => {
            let addr = crate::pool::SendMut::new(out.as_mut_ptr());
            p.run_rows(n_pos, &|start, end| {
                for t in start..end {
                    unsafe { *addr.at(t) = score_at(t) };
                }
            });
        }
        _ => {
            for (t, o) in out.iter_mut().enumerate() {
                *o = score_at(t);
            }
        }
    }
}

/// Top-`k` positions by score, ties broken by the lower index so the choice
/// is deterministic across backends. Masked slots (-inf) never win, and a
/// short history simply returns fewer than `k`.
pub fn top_k_positions(scores: &[f32], k: usize, out: &mut Vec<usize>) {
    out.clear();
    // When k reaches the whole list there is nothing to choose: every finite
    // position wins, and they come out in index order anyway. The general
    // path is k rounds of argmax — O(k·n) — and at index_topk = 512 against a
    // compressed axis that is still shorter than that, it was doing 160k
    // comparisons a layer to arrive at "all of them". This grows with the
    // context, which is exactly when it hurts.
    if k >= scores.len() {
        out.extend(
            scores
                .iter()
                .enumerate()
                .filter(|(_, v)| v.is_finite())
                .map(|(i, _)| i),
        );
        return;
    }
    let mut taken = vec![false; scores.len()];
    for _ in 0..k.min(scores.len()) {
        let mut best = usize::MAX;
        let mut bv = f32::NEG_INFINITY;
        for (i, &v) in scores.iter().enumerate() {
            if !taken[i] && v > bv && v.is_finite() {
                bv = v;
                best = i;
            }
        }
        if best == usize::MAX {
            break;
        }
        taken[best] = true;
        out.push(best);
    }
    out.sort_unstable();
}

/// SwiGLU expert: `w2(silu(w1(x)) * w3(x))`, with the routing weight folded
/// in before the down projection exactly as the reference does.
///
/// `limit` is the reference's `swiglu_limit` (10.0 in the release), and its
/// asymmetry is not a typo: `up` is clamped on BOTH sides, `gate` only from
/// above — the reference leaves silu's negative tail alone. A limit of 0
/// disables the clamp, which is also what the reference does.
#[allow(clippy::too_many_arguments)]
pub fn expert_swiglu(
    x: &[f32],
    w1: &dyn Fn(&[f32], &mut [f32]),
    w3: &dyn Fn(&[f32], &mut [f32]),
    w2: &dyn Fn(&[f32], &mut [f32]),
    inter: usize,
    weight: f32,
    limit: f32,
    out: &mut [f32],
) {
    let mut gate = vec![0.0f32; inter];
    let mut up = vec![0.0f32; inter];
    w1(x, &mut gate);
    w3(x, &mut up);
    if limit > 0.0 {
        for u in up.iter_mut() {
            *u = u.clamp(-limit, limit);
        }
        for g in gate.iter_mut() {
            *g = g.min(limit);
        }
    }
    for (g, u) in gate.iter_mut().zip(&up) {
        let silu = *g / (1.0 + (-*g).exp());
        *g = silu * u * weight;
    }
    w2(&gate, out);
}

/// Everything one layer needs that is not a plain matrix: the shapes and
/// scalars the reference reads out of `ModelArgs`.
#[derive(Debug, Clone, Copy)]
pub struct Dsv4Cfg {
    pub dim: usize,
    pub n_heads: usize,
    pub head_dim: usize,
    pub rope_head_dim: usize,
    pub q_lora_rank: usize,
    pub o_lora_rank: usize,
    pub o_groups: usize,
    pub hc_mult: usize,
    pub hc_sinkhorn_iters: usize,
    pub hc_eps: f32,
    pub norm_eps: f32,
    pub n_routed_experts: usize,
    pub top_k: usize,
    pub moe_inter: usize,
    pub route_scale: f32,
    /// The reference's `swiglu_limit`; 0 disables the clamp.
    pub swiglu_limit: f32,
    /// Sliding-window size (`window_size`, 128 in the release).
    pub window: usize,
    pub index_topk: usize,
    pub vocab: usize,
}

/// The per-block hyper-connection cycle, which is the same shape around
/// attention and around the FFN: fold the copies, normalize, run the
/// block, expand back. `block` sees a plain `dim`-vector and knows nothing
/// about the copies — that separation is what keeps attention and the MoE
/// free of hyper-connection bookkeeping.
///
/// `hc_fn` is `[mix_hc, hc*dim]`, `hc_base` is `[mix_hc]`, `hc_scale` is 3.
#[allow(clippy::too_many_arguments)]
#[allow(clippy::too_many_arguments)]
pub fn hc_block<F: FnMut(&[f32], &mut [f32])>(
    state: &mut [f32],
    hc_fn: &[f32],
    hc_scale: &[f32; 3],
    hc_base: &[f32],
    norm_w: &[f32],
    cfg: &Dsv4Cfg,
    scratch: &mut HcScratch,
    pool: Option<&crate::pool::Pool>,
    mut block: F,
) {
    let (hc, dim) = (cfg.hc_mult, cfg.dim);
    let mix_hc = (2 + hc) * hc;
    hc_mixes(state, hc_fn, mix_hc, cfg.norm_eps, pool, &mut scratch.mixes);
    hc_split_sinkhorn(
        &scratch.mixes,
        hc_scale,
        hc_base,
        hc,
        cfg.hc_sinkhorn_iters,
        cfg.hc_eps,
        &mut scratch.pre,
        &mut scratch.post,
        &mut scratch.comb,
    );
    hc_fold(state, &scratch.pre, hc, dim, &mut scratch.folded);
    // RMSNorm with the layer's learned weight, on the folded vector.
    let ms = scratch.folded.iter().map(|v| v * v).sum::<f32>() / dim as f32;
    let inv = 1.0 / (ms + cfg.norm_eps).sqrt();
    for (v, w) in scratch.folded.iter_mut().zip(norm_w) {
        *v = *v * inv * w;
    }
    block(&scratch.folded, &mut scratch.block_out);
    scratch.residual.copy_from_slice(state);
    hc_expand(
        &scratch.block_out,
        &scratch.residual,
        &scratch.post,
        &scratch.comb,
        hc,
        dim,
        state,
    );
}

/// Reusable buffers for `hc_block` — one allocation per pipeline, not per
/// layer per token.
pub struct HcScratch {
    pub mixes: Vec<f32>,
    pub pre: Vec<f32>,
    pub post: Vec<f32>,
    pub comb: Vec<f32>,
    pub folded: Vec<f32>,
    pub block_out: Vec<f32>,
    pub residual: Vec<f32>,
}

impl HcScratch {
    pub fn new(cfg: &Dsv4Cfg) -> Self {
        let (hc, dim) = (cfg.hc_mult, cfg.dim);
        Self {
            mixes: vec![0.0; (2 + hc) * hc],
            pre: vec![0.0; hc],
            post: vec![0.0; hc],
            comb: vec![0.0; hc * hc],
            folded: vec![0.0; dim],
            block_out: vec![0.0; dim],
            residual: vec![0.0; hc * dim],
        }
    }
}

/// The final fold, after the last layer: `hc` copies to one vector, with a
/// plain sigmoid gate (no Sinkhorn), then the model's output norm.
pub fn hc_head_fold(
    state: &[f32],
    hc_fn: &[f32],
    hc_scale: f32,
    hc_base: &[f32],
    cfg: &Dsv4Cfg,
    pool: Option<&crate::pool::Pool>,
    out: &mut [f32],
) {
    let (hc, dim) = (cfg.hc_mult, cfg.dim);
    let mut mixes = vec![0.0f32; hc];
    hc_mixes(state, hc_fn, hc, cfg.norm_eps, pool, &mut mixes);
    let mut pre = vec![0.0f32; hc];
    hc_head_pre(&mixes, hc_scale, hc_base, hc, cfg.hc_eps, &mut pre);
    hc_fold(state, &pre, hc, dim, out);
}

/// One layer's weights. Everything quantized rides as `QTensor` so the
/// existing kernels (and the mmap) serve them; the small fp32 pieces —
/// norms, the hyper-connection projections, the sink, the compressor's
/// position bias — are plain vectors, exactly as the reference keeps them
/// in fp32 regardless of the checkpoint's storage dtype.
pub struct Dsv4Layer {
    pub attn_norm: Vec<f32>,
    pub ffn_norm: Vec<f32>,
    // attention: the double LoRA, the compressed KV, the grouped output
    pub wq_a: crate::qtensor::QTensor,
    pub q_norm: Vec<f32>,
    pub wq_b: crate::qtensor::QTensor,
    pub wkv: crate::qtensor::QTensor,
    pub kv_norm: Vec<f32>,
    pub wo_a: crate::qtensor::QTensor,
    pub wo_b: crate::qtensor::QTensor,
    pub attn_sink: Vec<f32>,
    /// `None` on the pure sliding-window layers (`compress_ratio == 0`).
    pub compressor: Option<Dsv4Compressor>,
    /// Only on the layers whose ratio is 4.
    pub indexer: Option<Dsv4Indexer>,
    // hyper-connections, one set for the attention half and one for the FFN
    pub hc_attn_fn: Vec<f32>,
    pub hc_attn_base: Vec<f32>,
    pub hc_attn_scale: [f32; 3],
    pub hc_ffn_fn: Vec<f32>,
    pub hc_ffn_base: Vec<f32>,
    pub hc_ffn_scale: [f32; 3],
    // MoE
    pub gate: crate::qtensor::QTensor,
    /// noaux_tc selection bias — `None` on the hash layers.
    pub gate_bias: Option<Vec<f32>>,
    /// Token-id → expert table on the hash layers, `None` elsewhere.
    pub tid2eid: Option<Vec<f32>>,
    pub experts: Vec<Dsv4Expert>,
    pub shared: Dsv4Expert,
    /// Task-conditional restriction over the routed experts
    /// (`CMF_MOE_MASK` + `CMF_MOE_MASK_COVER`): `false` experts are not
    /// selectable and the weights renormalize over what remains. `None` on
    /// the hash layers — their table names specific experts, so masking
    /// there would silently reroute rather than restrict.
    pub mask: Option<Vec<bool>>,
}

pub struct Dsv4Expert {
    pub w1: crate::qtensor::QTensor,
    pub w2: crate::qtensor::QTensor,
    pub w3: crate::qtensor::QTensor,
}

pub struct Dsv4Compressor {
    pub wkv: crate::qtensor::QTensor,
    pub wgate: crate::qtensor::QTensor,
    pub norm: Vec<f32>,
    /// `[ratio, coff*head_dim]` — the in-window position bias.
    pub ape: Vec<f32>,
    pub ratio: usize,
    /// Overlapping windows (the reference sets this when ratio == 4), which
    /// doubles the projection width.
    pub overlap: bool,
}

pub struct Dsv4Indexer {
    pub wq_b: crate::qtensor::QTensor,
    pub weights_proj: crate::qtensor::QTensor,
    pub compressor: Dsv4Compressor,
}

/// Model-global pieces: the embedding, the output head and the final
/// hyper-connection fold.
pub struct Dsv4Globals {
    /// RoPE frequencies for the layers that carry a KV compressor: base
    /// `compress_rope_theta` (160 000 in the release) WITH YaRN.
    pub inv_freq_compress: Vec<f32>,
    /// …and for the pure sliding-window layers: base `rope_theta` (10 000)
    /// with YaRN OFF. The reference picks per layer:
    ///   if compress_ratio { original_seq_len, compress_rope_theta }
    ///   else              { 0, rope_theta }   // "disable YaRN"
    /// One shared table gets both groups wrong — the model still retrieves
    /// facts, because attention still attends, but every position is rotated
    /// by the wrong angle, so it repeats itself and cannot count.
    pub inv_freq_window: Vec<f32>,
    pub embed: crate::qtensor::QTensor,
    pub norm: Vec<f32>,
    pub head: crate::qtensor::QTensor,
    pub hc_head_fn: Vec<f32>,
    pub hc_head_base: Vec<f32>,
    pub hc_head_scale: f32,
}

/// Per-sequence state. The compressor and the indexer each keep their own
/// compressed cache and a partial window, so decode picks up mid-window
/// exactly where prefill left off.
pub struct Dsv4State {
    /// Sliding-window KV per layer, `[window, head_dim]` ring.
    pub window: Vec<Vec<f32>>,
    /// Compressed KV per layer, appended once per `ratio` tokens.
    pub compressed: Vec<Vec<f32>>,
    /// The indexer's own compressed cache per layer.
    pub index_kv: Vec<Vec<f32>>,
    /// Partial window being accumulated, per layer: kv and score streams.
    pub pending_kv: Vec<Vec<f32>>,
    pub pending_score: Vec<Vec<f32>>,
    /// The window before it, kept only by the overlapping compressor —
    /// its fold reads half its dimensions from the previous stride.
    pub prev_kv: Vec<Vec<f32>>,
    pub prev_score: Vec<Vec<f32>>,
    /// The indexer's compressor runs alongside the attention one and keeps
    /// its own window — same shape, different width and different weights.
    pub pending_ix_kv: Vec<Vec<f32>>,
    pub pending_ix_score: Vec<Vec<f32>>,
    pub prev_ix_kv: Vec<Vec<f32>>,
    pub prev_ix_score: Vec<Vec<f32>>,
    pub pos: usize,
    /// Identifies this sequence's caches on the device. A fresh state gets a
    /// fresh id, so a device buffer left over from the previous conversation
    /// can never be read as if it belonged to this one.
    pub kv_id: u64,
    /// When the token graph owns a layer's caches, the CONTENTS live on the
    /// card and only these counts stay here — how much of the window is
    /// filled, and how many compressed entries each cache holds. All three
    /// follow from the position, so keeping them costs nothing and reading
    /// them back would cost a round trip.
    pub dev_filled: Vec<usize>,
    pub dev_n_comp: Vec<usize>,
    pub dev_n_ix: Vec<usize>,
    /// True once this sequence has run a layer on the card with the device
    /// owning its state. The host copies above are stale from then on, so
    /// the CPU path must not be used for that layer again.
    pub dev_owned: bool,
    /// The device-layer set of the FIRST chained token. If it ever differs,
    /// some layer's caches are on the wrong side and the answer would be
    /// quietly wrong — the loop refuses instead.
    pub dev_set: Vec<bool>,
    /// Which layers run their MoE on the card from a PARTIAL expert pack.
    /// Their walk attention must stay on the host: the device attention
    /// frame and the device MoE frame of one layer share pooled slots and
    /// poison each other across tokens (see `attention_step`).
    pub partial_set: Vec<bool>,
    /// More than one layer walks past the device prefix. The stale-slot
    /// poison needs a CHAIN of walk frames handing state through the pooled
    /// slots; a single tail layer (the canonical shape) never chains and
    /// its device attention is measured exact.
    pub split_deep: bool,
}

impl Dsv4State {
    pub fn new(layers: usize) -> Self {
        use std::sync::atomic::{AtomicU64, Ordering};
        static NEXT: AtomicU64 = AtomicU64::new(1);
        Self {
            kv_id: NEXT.fetch_add(1, Ordering::Relaxed),
            dev_filled: vec![0; layers],
            dev_n_comp: vec![0; layers],
            dev_n_ix: vec![0; layers],
            dev_owned: false,
            dev_set: Vec::new(),
            partial_set: Vec::new(),
            split_deep: false,
            window: vec![Vec::new(); layers],
            compressed: vec![Vec::new(); layers],
            index_kv: vec![Vec::new(); layers],
            pending_kv: vec![Vec::new(); layers],
            pending_score: vec![Vec::new(); layers],
            prev_kv: vec![Vec::new(); layers],
            prev_score: vec![Vec::new(); layers],
            pending_ix_kv: vec![Vec::new(); layers],
            pending_ix_score: vec![Vec::new(); layers],
            prev_ix_kv: vec![Vec::new(); layers],
            prev_ix_score: vec![Vec::new(); layers],
            pos: 0,
        }
    }
}

/// One attention block for a single position. `hidden` is the folded,
/// normalized vector `hc_block` hands over; the result goes back to it.
///
/// The order matters and is the reference's: q through the LoRA pair with
/// a normalization at each end, kv compressed to one head's width, rope on
/// the tails, the window and the compressed positions concatenated into
/// one index list, sparse attention with the sink, the INVERSE rope on the
/// output, then the grouped low-rank projection.
#[allow(clippy::too_many_arguments)]
/// Advance one compressor by a token and return its folded entry when the
/// window closes. Both the attention compressor and the indexer's own run
/// through here — the indexer's was simply never called, so its cache stayed
/// empty and every layer that has an indexer selected ZERO compressed
/// positions, discarding a correctly-built long-range memory.
#[allow(clippy::too_many_arguments)]
fn compressor_step(
    cp: &Dsv4Compressor,
    hidden: &[f32],
    pos: usize,
    rd: usize,
    norm_eps: f32,
    inv_freq: &[f32],
    pool: Option<&crate::pool::Pool>,
    pending_kv: &mut Vec<f32>,
    pending_score: &mut Vec<f32>,
    prev_kv: &mut Vec<f32>,
    prev_score: &mut Vec<f32>,
) -> Option<Vec<f32>> {
    let width = cp.wkv.rows();
    let ew = if cp.overlap { width / 2 } else { width };
    let mut ckv = vec![0.0f32; width];
    let mut cscore = vec![0.0f32; width];
    // Same input, so one dispatch instead of two — and this runs twice a
    // layer (the compressor and the indexer's own), 43 layers a token.
    crate::qtensor::QTensor::matvec_many(
        [&cp.wkv, &cp.wgate],
        hidden,
        [&mut ckv, &mut cscore],
        pool,
    );
    if cp.overlap {
        // The reference biases the score as the token arrives and keeps it
        // biased across the shift, so ape is added ONCE, here.
        let slot = pos % cp.ratio;
        for (c, a) in cscore
            .iter_mut()
            .zip(&cp.ape[slot * width..(slot + 1) * width])
        {
            *c += a;
        }
    }
    pending_kv.extend_from_slice(&ckv);
    pending_score.extend_from_slice(&cscore);
    if pending_kv.len() / width < cp.ratio {
        return None;
    }
    let mut folded = vec![0.0f32; ew];
    if cp.overlap {
        compress_window_overlap(
            prev_kv,
            prev_score,
            pending_kv,
            pending_score,
            cp.ratio,
            ew,
            &mut folded,
        );
        *prev_kv = std::mem::take(pending_kv);
        *prev_score = std::mem::take(pending_score);
    } else {
        compress_window(
            pending_kv,
            pending_score,
            &cp.ape,
            cp.ratio,
            width,
            &mut folded,
        );
    }
    rms_weighted(&mut folded, &cp.norm, norm_eps);
    // The entry carries the same rope-tagged tail as a window key, at the
    // position of the window's first token.
    rope_tail(&mut folded, inv_freq, pos + 1 - cp.ratio, rd, false);
    pending_kv.clear();
    pending_score.clear();
    Some(folded)
}

/// `CMF_DSV4_PROFILE=1` accumulates wall time per stage and prints the split
/// when the process ends. Guessing which half of a layer costs what is how
/// one ends up optimising the cheap one: the fused attention block came out a
/// wash on the release checkpoint, and no amount of reasoning about MAC
/// counts settles whether that is because attention was already cheap or
/// because the device arm was slow.
pub(crate) mod prof {
    use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};

    pub static ATTN_NS: AtomicU64 = AtomicU64::new(0);
    pub static MOE_NS: AtomicU64 = AtomicU64::new(0);
    pub static CALLS: AtomicU64 = AtomicU64::new(0);
    /// Everything in a layer that is neither attention nor the experts: the
    /// hyper-connection fold and expand, the two norms, the residual.
    pub static HC_NS: AtomicU64 = AtomicU64::new(0);
    /// The head: final norm plus lm_head over 129280 rows.
    pub static HEAD_NS: AtomicU64 = AtomicU64::new(0);
    /// Host prep of the device layer: attention_step's CPU half (indexer,
    /// compressor, qr) — suspected owner of the unaccounted milliseconds.
    pub static PREP_NS: AtomicU64 = AtomicU64::new(0);
    /// The per-layer KV/window cache uploads before the frame.
    pub static CACHEW_NS: AtomicU64 = AtomicU64::new(0);
    /// The whole forward, so the buckets can be checked against a total
    /// instead of against a guess. 78 ms of measured work in a 108 ms token
    /// left 30 ms that no counter had ever looked at.
    pub static ALL_NS: AtomicU64 = AtomicU64::new(0);
    pub static TOKENS: AtomicU64 = AtomicU64::new(0);

    /// One token = one visit to layer zero. Counting `moe_step` calls instead
    /// counts layers.
    pub fn note_layer(li: usize) {
        CALLS.fetch_add(1, Ordering::Relaxed);
        if li == 0 {
            // The first token pays for the whole expert set reaching the card
            // — tens of seconds of it. Left in, that one-time cost is divided
            // by every later call and reads as a per-call price: it is what
            // made "the host encodes for 4.45 ms a layer" out of an upload
            // that happens once. Everything measured before the SECOND token
            // starts is therefore thrown away, and the report describes
            // steady state, which is the only thing worth optimising.
            // `swap` and not a TOKENS comparison: resetting TOKENS to 1 made
            // the test true again on every later token, so the report
            // described one token instead of the run.
            if TOKENS.fetch_add(1, Ordering::Relaxed) == 1 && !ZEROED.swap(true, Ordering::Relaxed)
            {
                for a in [
                    &ATTN_NS, &MOE_NS, &HC_NS, &HEAD_NS, &ALL_NS, &CALLS, &PREP_NS, &CACHEW_NS,
                ] {
                    a.store(0, Ordering::Relaxed);
                }
                TOKENS.store(1, Ordering::Relaxed);
                #[cfg(feature = "gpu")]
                for a in [
                    &crate::gpu_wgpu::MOE_ENC_NS,
                    &crate::gpu_wgpu::MOE_WAIT_NS,
                    &crate::gpu_wgpu::MOE_BUFS_NS,
                    &crate::gpu_wgpu::MOE_UP_NS,
                    &crate::gpu_wgpu::MOE_PASS_NS,
                    &crate::gpu_wgpu::ATT_ENC_NS,
                    &crate::gpu_wgpu::ATT_WAIT_NS,
                    &crate::gpu_wgpu::CHAIN_ENC_NS,
                    &crate::gpu_wgpu::CHAIN_WAIT_NS,
                    &crate::gpu_wgpu::CHAIN_LAYERS,
                    &crate::gpu_wgpu::CHAIN_RUNS,
                    &crate::gpu_wgpu::SUBMITS,
                    &crate::gpu_wgpu::PASSES,
                ] {
                    a.store(0, Ordering::Relaxed);
                }
            }
        }
    }
    static REPORT: AtomicBool = AtomicBool::new(false);
    /// The one-time "drop the first token's numbers" latch.
    static ZEROED: AtomicBool = AtomicBool::new(false);

    pub fn on() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| std::env::var("CMF_DSV4_PROFILE").is_ok_and(|v| v != "0"))
    }

    /// Print once, from wherever the last caller happens to be — a process
    /// that exits through several paths would otherwise report zero or twice.
    pub fn report() {
        if !on() || REPORT.swap(true, Ordering::Relaxed) {
            return;
        }
        // CALLS counts layer visits, not tokens — dividing by it and calling
        // the result "per token" is off by the layer count, which is 43 on
        // the release and reads as a plausible number either way.
        let calls = CALLS.load(Ordering::Relaxed).max(1);
        let toks = TOKENS.load(Ordering::Relaxed).max(1);
        let (a, m) = (
            ATTN_NS.load(Ordering::Relaxed) as f64 / 1e6,
            MOE_NS.load(Ordering::Relaxed) as f64 / 1e6,
        );
        let all = ALL_NS.load(Ordering::Relaxed) as f64 / 1e6;
        // HC_NS wraps the FFN half's hc_block WHOLE, and moe_step runs
        // inside that block — so the raw counter double-counts every MoE
        // millisecond as hyper-connection time. Reported as the difference:
        // the glue alone. (This inflation is what made moving the
        // hyper-connections to the card look like a 19 ms win when the glue
        // is ~4.)
        let hc = (HC_NS.load(Ordering::Relaxed) as f64 / 1e6
            - MOE_NS.load(Ordering::Relaxed) as f64 / 1e6)
            .max(0.0);
        let hd = HEAD_NS.load(Ordering::Relaxed) as f64 / 1e6;
        let prep = PREP_NS.load(Ordering::Relaxed) as f64 / 1e6;
        let cw = CACHEW_NS.load(Ordering::Relaxed) as f64 / 1e6;
        eprintln!(
            "[dsv4-профиль] ХОСТ-ПРЕП слоя: {:.0} мс/токен, KV-заливки: {:.0} мс/токен",
            prep / toks as f64,
            cw / toks as f64
        );
        #[cfg(feature = "gpu")]
        {
            let f = crate::gpu_wgpu::DSV4_FILLS.load(Ordering::Relaxed);
            let fb = crate::gpu_wgpu::DSV4_FILL_BYTES.load(Ordering::Relaxed);
            eprintln!(
                "[dsv4-профиль] ЗАЛИВКИ СЛОТОВ: {:.1} эксп/токен, {:.0} МБ/токен",
                f as f64 / toks as f64,
                fb as f64 / 1e6 / toks as f64
            );
        }
        eprintln!(
            "[dsv4-профиль] {calls} вызовов слоя за {toks} токенов | \
             на токен: внимание {:.0} мс, MoE {:.0} мс, гипер-связи+нормы {:.0} мс, \
             голова {:.0} мс | на вызов: внимание {:.2}, MoE {:.2}, связи {:.2}",
            a / toks as f64,
            m / toks as f64,
            hc / toks as f64,
            hd / toks as f64,
            a / calls as f64,
            m / calls as f64,
            hc / calls as f64,
        );
        eprintln!(
            "[dsv4-профиль] весь проход {:.0} мс на токен; вне счётчиков {:.0} мс",
            all / toks as f64,
            (all - a - m - hd) / toks as f64,
        );
        #[cfg(feature = "gpu")]
        {
            let ae = crate::gpu_wgpu::ATT_ENC_NS.load(Ordering::Relaxed) as f64 / 1e6;
            let aw = crate::gpu_wgpu::ATT_WAIT_NS.load(Ordering::Relaxed) as f64 / 1e6;
            if ae + aw > 0.0 {
                eprintln!(
                    "[dsv4-профиль] кадр внимания на вызов: кодирование {:.2} мс, \
                     отправка и ожидание {:.2} мс",
                    ae / calls as f64,
                    aw / calls as f64,
                );
            }
            // At the OUTER level on purpose: this used to sit inside the MoE
            // frame's own report, and the chain does not use the MoE frame —
            // so the one number that says where a chained token goes was
            // printed only when the chain was not running.
            let ub = crate::gpu_wgpu::UPLOAD_BYTES.load(Ordering::Relaxed);
            let un = crate::gpu_wgpu::UPLOAD_NS.load(Ordering::Relaxed);
            if ub > 0 && un > 0 {
                eprintln!(
                    "[dsv4-профиль] ЗАЛИВКА весов: {:.1} ГБ за {:.1} с ({:.0} МБ/с)",
                    ub as f64 / 1e9,
                    un as f64 / 1e9,
                    ub as f64 / (un as f64 / 1e9) / 1e6,
                );
            }
            let sub = crate::gpu_wgpu::SUBMITS.load(Ordering::Relaxed);
            if sub > 0 {
                eprintln!(
                    "[dsv4-профиль] ОТПРАВОК на карту: {:.1} на токен, ПРОХОДОВ {:.0} \
                     ({:.1} на слой)",
                    sub as f64 / toks as f64,
                    crate::gpu_wgpu::PASSES.load(Ordering::Relaxed) as f64 / toks as f64,
                    crate::gpu_wgpu::PASSES.load(Ordering::Relaxed) as f64 / calls as f64,
                );
            }
            let cl = crate::gpu_wgpu::CHAIN_LAYERS.load(Ordering::Relaxed);
            if cl > 0 {
                let toks2 = toks.max(1) as f64;
                eprintln!(
                    "[dsv4-профиль] ЦЕПОЧКА на токен: кодирование {:.2} мс, \
                     ожидание {:.2} мс ({} слоёв, {} отправок)",
                    crate::gpu_wgpu::CHAIN_ENC_NS.load(Ordering::Relaxed) as f64 / 1e6 / toks2,
                    crate::gpu_wgpu::CHAIN_WAIT_NS.load(Ordering::Relaxed) as f64 / 1e6 / toks2,
                    cl / toks.max(1),
                    crate::gpu_wgpu::CHAIN_RUNS.load(Ordering::Relaxed) / toks.max(1),
                );
            }
            let e = crate::gpu_wgpu::MOE_ENC_NS.load(Ordering::Relaxed) as f64 / 1e6;
            let wt = crate::gpu_wgpu::MOE_WAIT_NS.load(Ordering::Relaxed) as f64 / 1e6;
            if e + wt > 0.0 {
                let ns = |a: &std::sync::atomic::AtomicU64| {
                    a.load(Ordering::Relaxed) as f64 / 1e6 / calls as f64
                };
                eprintln!(
                    "[dsv4-профиль] кадр MoE на вызов: кодирование {:.2} мс, \
                     отправка и ожидание {:.2} мс",
                    e / calls as f64,
                    wt / calls as f64,
                );
                let an = crate::gpu_wgpu::ATT_GPU_N.load(Ordering::Relaxed);
                if an > 0 {
                    let g = |i: usize| {
                        crate::gpu_wgpu::ATT_GPU_NS[i].load(Ordering::Relaxed) as f64
                            / 1e6
                            / an as f64
                    };
                    eprintln!(
                        "[dsv4-профиль]   ВНИМАНИЕ НА КАРТЕ на вызов: одиночное {:.3} мс, \
                         оценки {:.3} мс, применение {:.3} мс",
                        g(0),
                        g(1),
                        g(2),
                    );
                }
                let gn = crate::gpu_wgpu::MOE_GPU_N.load(Ordering::Relaxed);
                let gns = crate::gpu_wgpu::MOE_GPU_NS[0].load(Ordering::Relaxed);
                if gn > 0 && gns > 0 {
                    eprintln!(
                        "[dsv4-профиль]   MoE НА КАРТЕ: {:.3} мс на вызов ({gn} замеров)",
                        gns as f64 / 1e6 / gn as f64,
                    );
                } else if gn > 0 {
                    // Zero across thousands of samples is a broken query, not
                    // an instant kernel, and printing it as a time is how a
                    // profile starts lying.
                    eprintln!(
                        "[dsv4-профиль]   MoE НА КАРТЕ: метки вернули НОЛЬ на {gn} замерах — \
                         запрос времени не сработал, число не использовать"
                    );
                }
                eprintln!(
                    "[dsv4-профиль]   из кодирования: буферы экспертов {:.2} мс, \
                     загрузки {:.2} мс, проходы {:.2} мс",
                    ns(&crate::gpu_wgpu::MOE_BUFS_NS),
                    ns(&crate::gpu_wgpu::MOE_UP_NS),
                    ns(&crate::gpu_wgpu::MOE_PASS_NS),
                );
            }
        }
    }
}

/// Print the per-token split, if `CMF_DSV4_PROFILE` asked for one.
pub fn profile_report() {
    prof::report();
}

/// `CMF_DSV4_GPU_ATTN=1` moves the attention block onto the device as one
/// submission. Off by default: it needs every attention weight in q4tp and a
/// working wgpu context, and a frame that declines mid-layer after the state
/// has been advanced would be worse than one that never ran.
fn gpu_attn_enabled() -> bool {
    #[cfg(feature = "gpu")]
    {
        use std::sync::OnceLock;
        static ON: OnceLock<bool> = OnceLock::new();
        *ON.get_or_init(|| {
            let want = std::env::var("CMF_DSV4_GPU_ATTN")
                .map(|v| v != "0")
                .unwrap_or(true);
            let have = want && crate::gpu::backend_available();
            if want && !have && std::env::var("CMF_DSV4_GPU_ATTN").is_ok() {
                tracing::warn!(
                    "CMF_DSV4_GPU_ATTN задан, но устройства нет — блок внимания                      остаётся на CPU. Проверьте CMF_GPU=wgpu и Vulkan-ICD."
                );
            }
            if std::env::var("CMF_DSV4_FRAME_DEBUG").is_ok() {
                eprintln!("кадр dsv4: запрошен={want} доступен={have}");
            }
            have
        })
    }
    #[cfg(not(feature = "gpu"))]
    {
        false
    }
}

/// The device half of `attention_step`. Returns false — having changed
/// nothing — whenever it cannot do the whole block, so the caller's CPU path
/// is still correct to run.
#[cfg(feature = "gpu")]
#[allow(clippy::too_many_arguments)]
fn attn_frame(
    l: &Dsv4Layer,
    cfg: &Dsv4Cfg,
    st: &Dsv4State,
    li: usize,
    hidden: &[f32],
    qn: &[f32],
    idxs: &[usize],
    inv_freq: &[f32],
    pos: usize,
    win_len: usize,
    scale: f32,
    // Present: the frame also does this layer's hyper-connection handover
    // and leaves the MoE half's input on the card. `out` may then be empty.
    hc: Option<&crate::gpu_wgpu::Dsv4HcTail>,
    out: &mut [f32],
) -> bool {
    let hd = cfg.head_dim;
    let (Some(wq_a), Some(wq_b), Some(wo_a), Some(wo_b)) = (
        l.wq_a.model_idx(),
        l.wq_b.model_idx(),
        l.wo_a.model_idx(),
        l.wo_b.model_idx(),
    ) else {
        return false;
    };
    let Some(model) = l.wq_b.model_arc() else {
        return false;
    };
    // Fixed window region, then the compressed tail — so a token writes one
    // window slot's worth of movement and whatever the compressor just added,
    // not the whole cache. `cap` has to cover the longest run this sequence
    // will reach; the compressed axis grows by one entry per `ratio` tokens.
    let n_comp = st.compressed[li].len() / hd;
    let cap = (cfg.window + n_comp.next_power_of_two().max(64)) * hd;
    let kv_id = st.kv_id;
    // The window is rewritten whole. A ring would write one slot instead of
    // 128 — 2 KB against 256 — and was tried: it bought NOTHING (the cost is
    // per-dispatch driver bookkeeping, not the copy) and moved perplexity by
    // 6e-5 because the attended positions arrive in a different order and the
    // softmax accumulates differently. Not a trade worth making.
    if !crate::gpu_wgpu::dsv4_cache_write(kv_id, li, 0, &st.window[li], cap) {
        return false;
    }
    // The compressed axis only ever grows, so write the TAIL. Rewriting it
    // whole was 22 MB a token at 1024 positions — the cache write, not the
    // arithmetic, was what the attention block had left to pay.
    // The compressed tail is written WHOLE every token. Writing only the new
    // part was tried and gave nothing measurable, and the bookkeeping it
    // needs — a per-layer tail count invalidated by every buffer growth — is
    // exactly the kind of state that drifts silently and shows up as a model
    // that stops early. Not worth carrying for zero.
    if n_comp > 0
        && !crate::gpu_wgpu::dsv4_cache_write(kv_id, li, cfg.window * hd, &st.compressed[li], cap)
    {
        return false;
    }
    let idx32: Vec<u32> = idxs
        .iter()
        .map(|&p| {
            if p < win_len {
                p as u32
            } else {
                (cfg.window + (p - win_len)) as u32
            }
        })
        .collect();
    let w = crate::gpu_wgpu::Dsv4AttnW {
        wq_a,
        wq_b,
        wo_a,
        wo_b,
        q_norm: &l.q_norm,
        sink: &l.attn_sink,
    };
    let g = crate::gpu_wgpu::Dsv4AttnGeom {
        dim: cfg.dim,
        nh: cfg.n_heads,
        hd,
        rd: cfg.rope_head_dim,
        q_lora: cfg.q_lora_rank,
        o_lora: cfg.o_lora_rank,
        o_groups: cfg.o_groups,
        eps: cfg.norm_eps,
        scale,
    };
    // The host fold, explicitly. The frame used to read this half's input
    // from the pooled x2 slot — which a device MoE frame of the SAME layer
    // overwrites each token with the NEXT layer's input, so the second
    // token of any chain+partial configuration attended over garbage
    // (perplexity 5.3 against the 4.578 gold on every budget small enough
    // to split a layer). The host has the exact vector either way; one
    // hidden-width upload per call is what correctness costs.
    crate::gpu_wgpu::dsv4_attn_frame(
        &model,
        &w,
        g,
        hidden,
        Some(qn),
        kv_id,
        li,
        &idx32,
        inv_freq,
        pos,
        hc,
        out,
    )
}

/// What the host still owes the device before a layer frame can run: the
/// shared LoRA vector the indexer reads, and the attended position list.
#[derive(Default)]
pub struct AttnPrep {
    pub qr: Vec<f32>,
    pub idxs: Vec<usize>,
    pub win_len: usize,
}

#[allow(clippy::too_many_arguments)]
pub fn attention_step(
    hidden: &[f32],
    l: &Dsv4Layer,
    cfg: &Dsv4Cfg,
    st: &mut Dsv4State,
    li: usize,
    // Chosen by the caller from the layer's kind — see Dsv4Globals.
    inv_freq: &[f32],
    pool: Option<&crate::pool::Pool>,
    // When set, stop once the caches are advanced and the index list is
    // built, and hand those back instead of running attention: the layer
    // frame does the rest on the device.
    prep_out: Option<&mut AttnPrep>,
    out: &mut [f32],
) {
    let _t0 = prof::on().then(std::time::Instant::now);
    let _guard = scopeguard_attn(_t0);
    let (hd, rd) = (cfg.head_dim, cfg.rope_head_dim);
    let pos = st.pos;
    if std::env::var("CMF_FREQ_DEBUG").is_ok() && li == 0 && pos == 0 {
        eprintln!(
            "    [порт] rd={rd} частот={} inv_freq[0..4]={:?}",
            inv_freq.len(),
            &inv_freq[..4.min(inv_freq.len())]
        );
    }

    // ── q and kv: both read the same hidden state, so they go out as ONE
    // dispatch. The norms after them differ, and they stay separate.
    // (q: wq_a → q_norm → wq_b → per-head norm → rope tail;
    //  kv: one head's width, shared by every query head.)
    let mut qr = vec![0.0f32; cfg.q_lora_rank];
    let mut kv = vec![0.0f32; hd];
    crate::qtensor::QTensor::matvec_many([&l.wq_a, &l.wkv], hidden, [&mut qr, &mut kv], pool);
    rms_weighted(&mut qr, &l.q_norm, cfg.norm_eps);
    // The queries are built further down, after the frame has had its chance
    // at the whole block. `qr` is needed either way: the indexer reads it.
    // A PARTIAL layer walks its attention on the host. Its device MoE
    // frame refills the pooled walk slots (x2, the hyper-connection state)
    // each token with the NEXT layer's values, so the same layer's device
    // attention frame attends over the previous token's leftovers on the
    // second token — measured as perplexity 5.3 against the 4.578 gold on
    // every budget small enough to split a layer, and exact the moment
    // that one layer's attention walks on the host. Layers whose MoE runs
    // on the HOST keep their device attention: nothing refills their
    // slots mid-walk, and the MAX_LI ladder measures them bit-exact.
    // …and it spreads: the partial layer's MoE frame cycles slots that the
    // FOLLOWING host-MoE layers' device attention also reads, so in any
    // configuration that holds a partial layer, every layer past the chain
    // prefix walks its attention on the host. A configuration with no
    // partial layer keeps device attention everywhere — the canonical
    // stand and the MAX_LI ladder both measure that bit-exact.
    let split_config = st.partial_set.iter().any(|&p| p) && st.split_deep;
    let past_chain =
        st.dev_owned && (li >= st.dev_set.len() || !st.dev_set.get(li).copied().unwrap_or(false));
    if std::env::var("CMF_DSV4_GATE_DBG").is_ok() {
        eprintln!(
            "[gate] li={li} pos={} split={split_config} past={past_chain} dev_owned={} set_len={} part_len={}",
            st.pos,
            st.dev_owned,
            st.dev_set.len(),
            st.partial_set.len()
        );
    }
    let on_gpu = gpu_attn_enabled() && !(split_config && past_chain);

    rms_weighted(&mut kv, &l.kv_norm, cfg.norm_eps);
    rope_tail(&mut kv, inv_freq, pos, rd, false);

    // ── the compressor: accumulate `ratio` tokens, then fold them into
    // one compressed entry. The reference fires when (pos+1) % ratio == 0,
    // so a partial window simply waits — which is why the state carries
    // the pending streams across tokens.
    if let Some(cp) = &l.compressor {
        let mut pk = std::mem::take(&mut st.pending_kv[li]);
        let mut ps = std::mem::take(&mut st.pending_score[li]);
        let mut qk = std::mem::take(&mut st.prev_kv[li]);
        let mut qs = std::mem::take(&mut st.prev_score[li]);
        let entry = compressor_step(
            cp,
            hidden,
            pos,
            rd,
            cfg.norm_eps,
            inv_freq,
            pool,
            &mut pk,
            &mut ps,
            &mut qk,
            &mut qs,
        );
        st.pending_kv[li] = pk;
        st.pending_score[li] = ps;
        st.prev_kv[li] = qk;
        st.prev_score[li] = qs;
        if let Some(e) = entry {
            st.compressed[li].extend_from_slice(&e);
        }
    }
    // The indexer scores against ITS OWN compressed cache, built by its own
    // compressor. Without this the cache is empty, `n_ix` is zero, and every
    // indexer layer picks no compressed positions at all — the long-range
    // memory is built and then never read.
    if let Some(ix) = &l.indexer {
        let mut pk = std::mem::take(&mut st.pending_ix_kv[li]);
        let mut ps = std::mem::take(&mut st.pending_ix_score[li]);
        let mut qk = std::mem::take(&mut st.prev_ix_kv[li]);
        let mut qs = std::mem::take(&mut st.prev_ix_score[li]);
        let entry = compressor_step(
            &ix.compressor,
            hidden,
            pos,
            rd,
            cfg.norm_eps,
            inv_freq,
            pool,
            &mut pk,
            &mut ps,
            &mut qk,
            &mut qs,
        );
        st.pending_ix_kv[li] = pk;
        st.pending_ix_score[li] = ps;
        st.prev_ix_kv[li] = qk;
        st.prev_ix_score[li] = qs;
        if let Some(e) = entry {
            st.index_kv[li].extend_from_slice(&e);
        }
    }

    st.window[li].extend_from_slice(&kv);
    // The reference keeps the window in a ring of `window_size`; holding the
    // last N in order is the same set, and without this the "window" grows
    // for the whole generation — wrong attention AND unbounded memory.
    let cap = cfg.window * hd;
    if st.window[li].len() > cap {
        let drop = st.window[li].len() - cap;
        st.window[li].drain(..drop);
    }
    let win_len = st.window[li].len() / hd;
    let n_pos = win_len + st.compressed[li].len() / hd;

    // Index list: every window position, plus whatever the indexer picked
    // (or, without an indexer, every compressed position).
    //
    // CMF_DSV4_NO_COMPRESSED=1 attends to the sliding window ALONE. That is
    // not a mode anyone should serve — it drops the model's long-range
    // memory — but it separates two failure modes that look identical from
    // the outside: output that degrades because the compressed path is
    // wrong, and output that degrades because the weights are too coarse.
    let mut idxs: Vec<usize> = (0..win_len).collect();
    if !st.compressed[li].is_empty() && !no_compressed() {
        let n_comp = st.compressed[li].len() / hd;
        match &l.indexer {
            Some(ix) => {
                // The indexer scores from the SHARED LoRA output through
                // its own wq_b — not from attention's queries — and its
                // per-head weights are a projection of the hidden state,
                // scaled by head_dim^-0.5 * n_heads^-0.5 as the reference
                // folds into `weights_proj`'s output.
                //
                // The reference also applies a randomized Hadamard rotation
                // to the queries here and to the keys in the indexer's
                // compressor, then simulates FP4 on both. That transform is
                // orthogonal (`hadamard_transform` scaled by d^-0.5) and it
                // hits BOTH sides of the same dot product, so it cancels:
                // its purpose is to condition the FP4 quantization, which we
                // do not do either. Omitting the pair is exact, and keeping
                // f32 is strictly more accurate than the reference — not an
                // approximation to be fixed later.
                let ih = ix.weights_proj.rows();
                let idim = ix.wq_b.rows() / ih.max(1);
                let mut qi = vec![0.0f32; ix.wq_b.rows()];
                ix.wq_b.matvec(&qr, &mut qi, pool);
                for h in 0..ih {
                    rope_tail(&mut qi[h * idim..(h + 1) * idim], inv_freq, pos, rd, false);
                }
                let mut hw = vec![0.0f32; ih];
                ix.weights_proj.matvec(hidden, &mut hw, pool);
                let sc_factor = (idim as f32).powf(-0.5) * (ih as f32).powf(-0.5);
                for w in hw.iter_mut() {
                    *w *= sc_factor;
                }
                let n_ix = st.index_kv[li].len() / idim.max(1);
                let mut sc = Vec::new();
                index_scores(
                    &qi,
                    &st.index_kv[li],
                    &hw,
                    ih,
                    idim,
                    n_ix.min(n_comp),
                    n_ix.min(n_comp),
                    pool,
                    &mut sc,
                );
                let mut picked = Vec::new();
                top_k_positions(&sc, cfg.index_topk, &mut picked);
                idxs.extend(picked.into_iter().map(|p| win_len + p));
            }
            None => idxs.extend((0..n_comp).map(|p| win_len + p)),
        }
    }
    debug_assert!(idxs.iter().all(|&p| p < n_pos));
    if let Some(p) = prep_out {
        p.qr = qr;
        p.idxs = idxs;
        p.win_len = win_len;
        return;
    }

    // ── the whole block on the device, or nothing ──
    let scale = (hd as f32).powf(-0.5);
    #[cfg(feature = "gpu")]
    if on_gpu
        && {
            if std::env::var("CMF_DSV4_XCHK").is_ok() {
                // The frame reads this half's input from the card's x2
                // slot; the host walked its own. Disagreement = the
                // chain→walk handoff, and the number says by how much.
                if let Some(card) = crate::gpu_wgpu::dsv4_dbg_read_tag(45, 0, hidden.len()) {
                    let md = hidden
                        .iter()
                        .zip(card.iter())
                        .map(|(a, b)| (a - b).abs())
                        .fold(0.0f32, f32::max);
                    eprintln!("[xchk] li={li} pos={pos} x2 maxdiff={md:.3e}");
                }
            }
            true
        }
        && attn_frame(
            l, cfg, st, li, hidden, &qr, &idxs, inv_freq, pos, win_len, scale, None, out,
        )
    {
        return;
    }

    // ── queries: wq_b, then a norm and the rope tail per head ──
    let mut q = vec![0.0f32; cfg.n_heads * hd];
    l.wq_b.matvec(&qr, &mut q, pool);
    for h in 0..cfg.n_heads {
        let head = &mut q[h * hd..(h + 1) * hd];
        rms_inplace(head, cfg.norm_eps);
        rope_tail(head, inv_freq, pos, rd, false);
    }
    let mut cache: Vec<f32> = st.window[li].clone();
    cache.extend_from_slice(&st.compressed[li]);

    // ── sparse attention per head, then the inverse rope ──
    let mut attn = vec![0.0f32; cfg.n_heads * hd];
    for h in 0..cfg.n_heads {
        let qh = &q[h * hd..(h + 1) * hd];
        // Straight into this head's slice of the output: the scratch vector
        // that used to sit here was an allocation and a copy per head, so 64
        // of each per layer per token, for a value that was never read
        // anywhere else.
        let oh = &mut attn[h * hd..(h + 1) * hd];
        sparse_attend(qh, &cache, &idxs, l.attn_sink[h], scale, hd, oh);
        rope_tail(oh, inv_freq, pos, rd, true);
    }

    // ── grouped low-rank output ──
    // Read the two blocks through the quantized readers. Materializing them
    // here instead costs ~270 MB of dequantization per layer per token on
    // the release checkpoint (wo_a and wo_b are 33M weights each), which is
    // the difference between decoding and not.
    o_project(
        &attn,
        &|r, x, sc| l.wo_a.row_dot(r, x, sc),
        l.wo_a.cols(),
        &|mid, dst| l.wo_b.matvec(mid, dst, pool),
        cfg.o_groups,
        cfg.o_lora_rank,
        pool,
        out,
    );
}

/// RMSNorm with a learned weight, in place.
pub fn rms_weighted(v: &mut [f32], w: &[f32], eps: f32) {
    let ms = v.iter().map(|x| x * x).sum::<f32>() / v.len() as f32;
    let inv = 1.0 / (ms + eps).sqrt();
    for (x, g) in v.iter_mut().zip(w) {
        *x = *x * inv * g;
    }
}

// The MoE half of a block: route, run the chosen experts plus the shared one,
// and sum. `token_id` is only read on the hash layers. Per-layer expert
// routing mass is also recorded here for task-conditional expert sets
// (`CMF_MOE_STATS`). An older implementation counted every top-k winner as
// one. That is the wrong quantity for DeepSeek-V4: a weak eighth route and
// the dominant route then consume the same `cover` budget, so a compact mask
// can retain frequent noise while dropping a rarer expert that carries much
// more of the block output. Accumulate the normalized route weights as
// fixed-point integers instead. The JSON stays the same `{layer: [u64]}`
// shape and old count files remain valid inputs because the mask builder only
// compares relative mass within a layer.
//
// Decode drives this from one thread; the pool parallelizes inside the
// matvecs, below this point.
thread_local! {
    static ROUTE_COUNTS: std::cell::RefCell<Vec<Vec<u64>>> =
        const { std::cell::RefCell::new(Vec::new()) };
}

fn route_stats_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| std::env::var("CMF_MOE_STATS").is_ok())
}

fn record_route(li: usize, n_layers_hint: usize, n_experts: usize, routed: &[(usize, f32)]) {
    ROUTE_COUNTS.with(|c| {
        let mut c = c.borrow_mut();
        if c.len() <= li.max(n_layers_hint) {
            c.resize(li.max(n_layers_hint) + 1, Vec::new());
        }
        let row = &mut c[li];
        if row.len() < n_experts {
            row.resize(n_experts, 0);
        }
        for &(e, weight) in routed {
            if e < row.len() {
                // One unit keeps a finite selected route visible even if a
                // future quantized router rounds an extremely small weight
                // below the fixed-point scale.
                let mass = (weight.abs() as f64 * 1_000_000.0).round() as u64;
                row[e] = row[e].saturating_add(mass.max(1));
            }
        }
    });
}

/// Take the recorded routing field, leaving the counters empty.
pub fn take_route_counts() -> Vec<Vec<u64>> {
    ROUTE_COUNTS.with(|c| std::mem::take(&mut *c.borrow_mut()))
}

/// Charge elapsed time to a counter when it goes out of scope — the two
/// steps have several early returns each, and a timer that only stops on the
/// long path measures the short one as free.
struct Charge(
    Option<std::time::Instant>,
    &'static std::sync::atomic::AtomicU64,
);
impl Drop for Charge {
    fn drop(&mut self) {
        if let Some(t) = self.0 {
            self.1.fetch_add(
                t.elapsed().as_nanos() as u64,
                std::sync::atomic::Ordering::Relaxed,
            );
        }
    }
}
fn scopeguard_attn(t: Option<std::time::Instant>) -> Charge {
    Charge(t, &prof::ATTN_NS)
}
fn scopeguard_moe(t: Option<std::time::Instant>, li: usize) -> Charge {
    if t.is_some() {
        prof::note_layer(li);
    }
    Charge(t, &prof::MOE_NS)
}

/// The whole token, one submission per layer. Returns false having changed
/// nothing if the device declines any layer — the caller's loop is then still
/// correct to run.
#[cfg(feature = "gpu")]
#[allow(clippy::too_many_arguments)]
fn dsv4_layer_loop(
    state: &mut [f32],
    layers: &[Dsv4Layer],
    g: &Dsv4Globals,
    cfg: &Dsv4Cfg,
    st: &mut Dsv4State,
    token_id: u32,
    inv_freq: &[f32],
    pool: Option<&crate::pool::Pool>,
    scratch: &mut HcScratch,
) -> bool {
    let dim = cfg.dim;
    let freqs_of = |l: &Dsv4Layer| -> &[f32] {
        let f = if l.compressor.is_some() {
            &g.inv_freq_compress
        } else {
            &g.inv_freq_window
        };
        if f.is_empty() { inv_freq } else { f.as_slice() }
    };
    // PRE-FLIGHT. The prep inside the loop advances the window and the
    // compressor caches, so a refusal halfway leaves state that the CPU
    // fallback would advance a SECOND time — which is not a slow answer but a
    // wrong one. Everything that can decline is therefore asked before the
    // first byte of state moves. The expert upload happens here too, which is
    // where it belonged anyway.
    // The head goes to the card BEFORE the experts ask for room. It is the
    // single most-used tensor in the file — every token reads all of it —
    // and it is a rounding error next to the expert stack: 265 MB against
    // ninety-odd gigabytes on the release. Uploaded in first-touch order it
    // arrived last, after the budget was gone, and stayed on the host for
    // the life of the process.
    {
        static SAID: std::sync::atomic::AtomicBool = std::sync::atomic::AtomicBool::new(false);
        if !SAID.swap(true, std::sync::atomic::Ordering::Relaxed) {
            if let (Some(idx), Some(model)) = (g.head.model_idx(), g.head.model_arc()) {
                let ok = crate::gpu_wgpu::dsv4_weight_ready(&model, idx);
                tracing::info!("dsv4: голова на карте: {}", if ok { "да" } else { "нет" });
            }
        }
    }
    let mut on_dev = vec![false; layers.len()];
    let mut partial_dev = vec![false; layers.len()];
    let mut attn_ready = vec![false; layers.len()];
    // Two phases are essential for the unified pool: first pin every small
    // attention/compressor skeleton, then give all remaining weight budget
    // to the one expert arena. Allocating the arena after layer zero alone
    // would honestly fit it, but starve layer one's attention weights.
    for (li, l) in layers.iter().enumerate() {
        if l.wq_a.model_idx().is_none()
            || l.wq_b.model_idx().is_none()
            || l.wo_a.model_idx().is_none()
            || l.wo_b.model_idx().is_none()
        {
            return false;
        }
        let Some(model) = l.experts.first().and_then(|e| e.w1.model_arc()) else {
            return false;
        };
        // A layer whose experts do not fit is not a reason to abandon the
        // token: 100 GB of experts against a 98 GB card means SOME layer will
        // always miss. Those run on the host, with the state fetched and put
        // back around them — two transfers for the few that need it.
        // The attention weights have to be asked for too. Experts fill the
        // card first, and a wo_b that misses at layer 11 used to surface as a
        // mid-loop refusal — after the caches had advanced, which the CPU
        // fallback then advanced again.
        // …and, when the layer is to prepare itself, everything that
        // preparation reads: the KV projection, both compressors and the
        // indexer. Leaving them out is how the chain came to refuse ninety
        // times a token on the release — the experts had taken the card by
        // the time `dsv4_encode_prep` asked, and it declined silently into a
        // fallback that looked like "the chain simply does not help".
        let mut want = vec![
            l.wq_a.model_idx(),
            l.wq_b.model_idx(),
            l.wo_a.model_idx(),
            l.wo_b.model_idx(),
        ];
        if chain_enabled() {
            want.push(l.wkv.model_idx());
            if let Some(cp) = &l.compressor {
                want.push(cp.wkv.model_idx());
                want.push(cp.wgate.model_idx());
            }
            if let Some(ix) = &l.indexer {
                want.push(ix.wq_b.model_idx());
                want.push(ix.weights_proj.model_idx());
                want.push(ix.compressor.wkv.model_idx());
                want.push(ix.compressor.wgate.model_idx());
            }
        }
        attn_ready[li] = want
            .into_iter()
            .flatten()
            .all(|i| crate::gpu_wgpu::dsv4_weight_ready(&model, i));
    }
    for (li, l) in layers.iter().enumerate() {
        let Some(model) = l.experts.first().and_then(|e| e.w1.model_arc()) else {
            return false;
        };
        let gu_q2 = l
            .experts
            .first()
            .is_some_and(|e| e.w1.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP));
        // Size expert storage only after EVERY layer's skeleton is resident.
        // The old per-layer packs could interleave these allocations; one
        // global allocation cannot, so the ordering is now explicit.
        let pk = pack_for(l, cfg, li);
        if let Some(pk) = pk {
            let dn_q2 = l
                .experts
                .first()
                .is_some_and(|e| e.w2.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP));
            let experts_ok = if pk.global.is_some() {
                crate::gpu_wgpu::dsv4_global_moe_ready(&model)
            } else {
                crate::gpu_wgpu::dsv4_experts_ready(
                    &model,
                    &pk.tensors,
                    cfg.moe_inter,
                    dim,
                    gu_q2,
                    dn_q2,
                )
            };
            on_dev[li] = attn_ready[li] && experts_ok && pk.route_complete();
            partial_dev[li] = attn_ready[li] && experts_ok && !pk.route_complete();
        }
    }
    let active_dev: Vec<bool> = on_dev
        .iter()
        .zip(&partial_dev)
        .map(|(&full, &partial)| full || partial)
        .collect();
    if !active_dev.iter().any(|&x| x) {
        return false;
    }
    // The attention gate below needs to know about partial layers BEFORE
    // the decode path commits the device set — a perplexity run only ever
    // prefills, and with this left empty every split budget scored the
    // model wrong (measured; see `attention_step`).
    if st.partial_set.len() != partial_dev.len() || st.partial_set != partial_dev {
        st.partial_set = partial_dev.clone();
        st.split_deep = active_dev
            .iter()
            .zip(&partial_dev)
            .filter(|(a, p)| !**a || **p)
            .count()
            > 1;
    }

    // Which layers the card actually took, said once. A layer that falls to
    // the host costs an order of magnitude more than one that does not, and
    // "the GPU path is on" hid the difference between all of them and most.
    {
        static SAID: std::sync::atomic::AtomicBool = std::sync::atomic::AtomicBool::new(false);
        if !SAID.swap(true, std::sync::atomic::Ordering::Relaxed) {
            let host: Vec<usize> = active_dev
                .iter()
                .enumerate()
                .filter(|&(_, d)| !*d)
                .map(|(i, _)| i)
                .collect();
            let partial: Vec<(usize, usize)> = partial_dev
                .iter()
                .enumerate()
                .filter(|&(_, d)| *d)
                .filter_map(|(li, _)| pack_for(&layers[li], cfg, li).map(|p| (li, p.globals.len())))
                .collect();
            if host.is_empty() && partial.is_empty() {
                tracing::info!("dsv4: все {} слоёв на карте", on_dev.len());
            } else {
                tracing::info!(
                    "dsv4: {} из {} слоёв используют карту; частичные {:?}; на хосте {:?}",
                    active_dev.len() - host.len(),
                    on_dev.len(),
                    partial,
                    host,
                );
            }
        }
    }

    // Layer zero's opening fold has no frame before it to have prepared it.
    let (mut folded, post0, comb0) = hc_fold_norm(
        state,
        &layers[0].hc_attn_fn,
        &layers[0].hc_attn_scale,
        &layers[0].hc_attn_base,
        &layers[0].attn_norm,
        cfg,
        pool,
    );
    if !crate::gpu_wgpu::dsv4_state_write(state) || !crate::gpu_wgpu::dsv4_hc_write(&post0, &comb0)
    {
        return false;
    }
    // The device-owned set must not move once a token has run on it — but
    // the two directions are not the same risk. At a tight budget the set
    // GROWS between tokens as more weights finish uploading, and a layer that
    // merely joined can be left on the host: its caches are there and nothing
    // is inconsistent. Refusing on that was costing the whole fast path once
    // per token — 125 times in a 48-token run on an emulated 24 GB card, on
    // which the engine is slow enough already.
    //
    // A layer LEAVING the set is the dangerous direction: its caches are on
    // the card and the host would advance its own. That still refuses.
    if st.dev_owned && st.dev_set != active_dev {
        let left: Vec<usize> = (0..active_dev.len().min(st.dev_set.len()))
            .filter(|&i| st.dev_set[i] && !active_dev[i])
            .collect();
        if !left.is_empty() {
            tracing::warn!("слои {left:?} ушли с карты — кеши на разных сторонах");
            return false;
        }
        // A layer that was active remains device-owned. Its full/partial mode
        // is still derived from the current pack; only cache ownership is
        // sticky across tokens.
    }
    let chain = chain_enabled();
    // CMF_DSV4_LAYERS_PROBE=N — TIMING ONLY, the answer is garbage. Runs the
    // first N layers and leaves the rest alone. Decode time against N is a
    // line whose SLOPE is the per-layer cost and whose intercept is
    // everything that happens once a token. Unlike the skip probe it does
    // not change what a layer does — which on a MoE model is the difference
    // between a measurement and an artefact, because dropping any stage
    // changes the routing and the routing changes what the experts cost.
    let layer_cap = {
        static N: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
        *N.get_or_init(|| {
            std::env::var("CMF_DSV4_LAYERS_PROBE")
                .ok()
                .and_then(|v| v.parse::<usize>().ok())
                .unwrap_or(usize::MAX)
        })
    };
    let mut run: Vec<usize> = Vec::new();
    let mut sink_out = vec![0.0f32; dim];
    // `state` starts current on both sides. A device run makes the host copy
    // stale unless that same run carries it home. Tracking this explicitly
    // avoids a separate state fence before a host layer and, for a final host
    // layer, the old upload-immediately-followed-by-readback pair.
    let mut state_on_host = true;
    for (li, l) in layers.iter().enumerate() {
        if li >= layer_cap {
            break;
        }
        // The device path never ticked the profiler, so every per-token
        // number it printed described the two host-path tokens at the start
        // of a run — the ones that also pay for the upload. Ticking here is
        // what makes the chain's encode-and-wait split a per-token figure at
        // all.
        if prof::on() {
            prof::note_layer(li);
        }
        if chain && on_dev[li] {
            // Hash layers used to break the run in two: their forced expert
            // list changes per token, went through the (tag, len) upload
            // pool, and every layer of a submission shared one buffer. The
            // list has a per-layer slot now, so they chain like the rest.
            run.push(li);
            // CMF_DSV4_CHAIN_MAX=N caps a run's length. Diagnostic, not a
            // tuning knob: length-1 runs put ONE layer per submission, which
            // separates "the layer frame is wrong" from "layers in one
            // encoder contaminate each other" in a single ppl run.
            if run.len() >= chain_max() || dspark_wants(li) {
                let need_qn = run[0] == 0 || !on_dev[run[0] - 1];
                let captured = *run.last().unwrap();
                if !dsv4_chain_run(
                    layers,
                    &run,
                    cfg,
                    g,
                    st,
                    token_id,
                    &mut folded,
                    Some(state),
                    1,
                    &[],
                    need_qn,
                    pool,
                ) {
                    return false;
                }
                state_on_host = true;
                verify_fp("walk", st.pos, captured, state);
                dspark_note(captured, state, cfg);
                run.clear();
            }
            continue;
        }
        if chain && !run.is_empty() {
            // The very next layer is on the host, so bring its state back in
            // the chain's existing readback. Reading it in a second submit
            // below cost one fence per token on the release's 42+1 split.
            if !dsv4_chain_run(
                layers,
                &run,
                cfg,
                g,
                st,
                token_id,
                &mut folded,
                Some(state),
                1,
                &[],
                run[0] == 0 || !on_dev[run[0] - 1],
                pool,
            ) {
                return false;
            }
            state_on_host = true;
            let last = *run.last().unwrap();
            verify_fp("walk", st.pos, last, state);
            dspark_note(last, state, cfg);
        }
        run.clear();
        if partial_dev[li] && chain1_on() {
            if let Some(home) = dsv4_chain1_layer(
                state,
                &mut folded,
                layers,
                l,
                cfg,
                st,
                token_id,
                li,
                freqs_of(l),
                pool,
                state_on_host,
            ) {
                // chain1 advances the canonical host mirrors and rewrites
                // the device window/compressed cache from them, but it does
                // not go through dsv4_chain_run (the usual owner of these
                // arithmetic counters).  A speculative token-axis pass that
                // takes over on the next token must start from the same
                // extents, otherwise its very first partial layer attends to
                // an empty/stale prefix.
                st.dev_filled[li] = (st.window[li].len() / cfg.head_dim).min(cfg.window);
                st.dev_n_comp[li] = st.compressed[li].len() / cfg.head_dim;
                st.dev_n_ix[li] = l.indexer.as_ref().map_or(0, |ix| {
                    let ih = ix.weights_proj.rows();
                    let idim = ix.wq_b.rows() / ih.max(1);
                    st.index_kv[li].len() / idim.max(1)
                });
                state_on_host = home;
                if verify_fp_on(st.pos) {
                    if !state_on_host && !crate::gpu_wgpu::dsv4_state_read(state) {
                        return false;
                    }
                    state_on_host = true;
                    verify_fp("walk", st.pos, li, state);
                }
                // The draft captures THIS layer's state — which lives on
                // the card when no cold came home. Noting the stale host
                // array fed the draft garbage: 1265 drafted, 0 accepted.
                if dspark_wants(li) {
                    if !state_on_host && crate::gpu_wgpu::dsv4_state_read(state) {
                        state_on_host = true;
                    }
                    if state_on_host {
                        dspark_note(li, state, cfg);
                    }
                }
                continue;
            }
        }
        if partial_dev[li] && partial_walk_on() {
            // Attention and the resident expert subset stay on the card. The
            // router still sees every expert and returns only the winners
            // that did not fit; those are completed on the CPU and their
            // exact linear contribution is added back to device state.
            let Some(home) = dsv4_partial_layer(
                state,
                &mut folded,
                layers,
                l,
                cfg,
                st,
                token_id,
                li,
                freqs_of(l),
                pool,
                state_on_host,
            ) else {
                return false;
            };
            state_on_host = home;
            if home {
                dspark_note(li, state, cfg);
            }
            continue;
        }
        if !on_dev[li] {
            if !state_on_host && !crate::gpu_wgpu::dsv4_state_read(state) {
                return false;
            }
            state_on_host = true;
            let freqs = freqs_of(l);
            hc_block(
                state,
                &l.hc_attn_fn,
                &l.hc_attn_scale,
                &l.hc_attn_base,
                &l.attn_norm,
                cfg,
                scratch,
                pool,
                |f, o| attention_step(f, l, cfg, st, li, freqs, pool, None, o),
            );
            hc_block(
                state,
                &l.hc_ffn_fn,
                &l.hc_ffn_scale,
                &l.hc_ffn_base,
                &l.ffn_norm,
                cfg,
                scratch,
                pool,
                // The layer the card had no room for. Its experts are
                // reached one matvec at a time and the probe sends each to
                // the device — right per op, and a fence per op: this one
                // layer is why a token that submits ONCE for 42 layers
                // submits 13 times. CMF_DSV4_HOST_CPU_MOE=1 keeps them on
                // the host instead, trading arithmetic for round trips.
                |f, o| {
                    if host_cpu_moe() {
                        crate::gpu::cpu_scope(|| moe_step(f, l, cfg, token_id, li, pool, o))
                    } else {
                        moe_step(f, l, cfg, token_id, li, pool, o)
                    }
                },
            );
            // Only a following DEVICE layer needs the fold/hc slots and an
            // uploaded state. Consecutive host layers consume `state`
            // directly, and a final host layer is already exactly where the
            // head needs it — uploading then reading it back was pure sync.
            if layers.get(li + 1).is_some() && on_dev.get(li + 1).copied().unwrap_or(false) {
                let n = &layers[li + 1];
                let (f, p2, c2) = hc_fold_norm(
                    state,
                    &n.hc_attn_fn,
                    &n.hc_attn_scale,
                    &n.hc_attn_base,
                    &n.attn_norm,
                    cfg,
                    pool,
                );
                folded = f;
                if !crate::gpu_wgpu::dsv4_hc_write(&p2, &c2) {
                    return false;
                }
                if !crate::gpu_wgpu::dsv4_state_write(state) {
                    return false;
                }
            }
            dspark_note(li, state, cfg);
            continue;
        }
        let mut prep = AttnPrep::default();
        let _tp = prof::on().then(std::time::Instant::now);
        attention_step(
            &folded,
            l,
            cfg,
            st,
            li,
            freqs_of(l),
            pool,
            Some(&mut prep),
            &mut sink_out,
        );
        if let Some(t) = _tp {
            prof::PREP_NS.fetch_add(
                t.elapsed().as_nanos() as u64,
                std::sync::atomic::Ordering::Relaxed,
            );
        }
        // The caches the frame will read.
        let hd = cfg.head_dim;
        let n_comp = st.compressed[li].len() / hd;
        let cap = (cfg.window + n_comp.next_power_of_two().max(64)) * hd;
        let kv_id = st.kv_id;
        let _tc = prof::on().then(std::time::Instant::now);
        if !crate::gpu_wgpu::dsv4_cache_write(kv_id, li, 0, &st.window[li], cap)
            || (n_comp > 0
                && !crate::gpu_wgpu::dsv4_cache_write(
                    kv_id,
                    li,
                    cfg.window * hd,
                    &st.compressed[li],
                    cap,
                ))
        {
            return false;
        }
        if let Some(t) = _tc {
            prof::CACHEW_NS.fetch_add(
                t.elapsed().as_nanos() as u64,
                std::sync::atomic::Ordering::Relaxed,
            );
        }
        let idx32: Vec<u32> = prep
            .idxs
            .iter()
            .map(|&p| {
                if p < prep.win_len {
                    p as u32
                } else {
                    (cfg.window + (p - prep.win_len)) as u32
                }
            })
            .collect();
        let Some(pk) = pack_for(l, cfg, li) else {
            return false;
        };
        let (Some(wq_a), Some(wq_b), Some(wo_a), Some(wo_b)) = (
            l.wq_a.model_idx(),
            l.wq_b.model_idx(),
            l.wo_a.model_idx(),
            l.wo_b.model_idx(),
        ) else {
            return false;
        };
        let Some(model) = l.experts.first().and_then(|e| e.w1.model_arc()) else {
            return false;
        };
        let forced: Option<Vec<usize>> = l.tid2eid.as_ref().and_then(|tbl| {
            let v: Vec<usize> = if pk.needs_remap() {
                hash_route(tbl, cfg.vocab, cfg.top_k, token_id)
            } else {
                hash_route(tbl, cfg.vocab, cfg.top_k, token_id)
                    .into_iter()
                    .map(|gi| pk.to_slot[gi])
                    .collect()
            };
            if v.contains(&usize::MAX) {
                None
            } else {
                Some(v)
            }
        });
        if l.tid2eid.is_some() && forced.is_none() {
            return false;
        }
        let nxt = layers.get(li + 1);
        let w = crate::gpu_wgpu::Dsv4LayerW {
            attn: crate::gpu_wgpu::Dsv4AttnW {
                wq_a,
                wq_b,
                wo_a,
                wo_b,
                q_norm: &l.q_norm,
                sink: &l.attn_sink,
            },
            moe: crate::gpu_wgpu::Dsv4MoeW {
                router: &[],
                experts: &pk.tensors,
                logits: &[],
                // The PACK's bias, whose address outlives the process: the
                // frame's const cache is keyed on it, and a per-layer Vec
                // here handed every layer the first layer's — the exact
                // transient-Vec trap the const_buf war story describes,
                // reintroduced by this session and caught because the OFF
                // baseline moved.
                bias: pk.bias.as_deref(),
                mask: pk.mask.as_deref(),
                forced: forced.as_deref(),
                remap: pk.needs_remap().then_some(pk.remap.as_slice()),
                global: None,
                has_shared: true,
                shared_weight: 1.0,
                preweighted: false,
                qwen_softmax: false,
            },
            hc_ffn_fn: &l.hc_ffn_fn,
            hc_ffn_scale: &l.hc_ffn_scale,
            hc_ffn_base: &l.hc_ffn_base,
            hc_next_fn: nxt.map(|n| n.hc_attn_fn.as_slice()),
            hc_next_scale: nxt.map_or(&l.hc_attn_scale, |n| &n.hc_attn_scale),
            hc_next_base: nxt.map_or(&l.hc_attn_base, |n| n.hc_attn_base.as_slice()),
            ffn_norm: &l.ffn_norm,
            next_norm: nxt.map_or(&l.attn_norm, |n| n.attn_norm.as_slice()),
            next_q_norm: nxt.map_or(&l.q_norm, |n| n.q_norm.as_slice()),
            next_wq_a: nxt.and_then(|n| n.wq_a.model_idx()),
            router: &pk.router,
        };
        let geom = crate::gpu_wgpu::Dsv4LayerGeom {
            attn: crate::gpu_wgpu::Dsv4AttnGeom {
                dim,
                nh: cfg.n_heads,
                hd,
                rd: cfg.rope_head_dim,
                q_lora: cfg.q_lora_rank,
                o_lora: cfg.o_lora_rank,
                o_groups: cfg.o_groups,
                eps: cfg.norm_eps,
                scale: (hd as f32).powf(-0.5),
            },
            moe: crate::gpu_wgpu::Dsv4MoeGeom {
                hidden: dim,
                inter: cfg.moe_inter,
                top_k: cfg.top_k,
                route_scale: cfg.route_scale,
                swiglu_limit: cfg.swiglu_limit,
                gu_q2: l.experts.first().is_some_and(|e| {
                    e.w1.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP)
                }),
            },
            hc: cfg.hc_mult,
            hc_eps: cfg.hc_eps,
            sinkhorn_iters: cfg.hc_sinkhorn_iters,
        };
        let mut next = vec![0.0f32; dim];
        if !crate::gpu_wgpu::dsv4_layer_frame(
            &model,
            &w,
            geom,
            kv_id,
            li,
            Some(&prep.qr),
            &idx32,
            freqs_of(l),
            st.pos,
            &mut next,
            None,
            None,
            &mut Vec::new(),
        ) {
            return false;
        }
        state_on_host = false;
        folded = next;
        dspark_note(li, state, cfg);
    }
    let mut state_home = false;
    if chain {
        if !run.is_empty() {
            // The token's LAST run brings the state back with it. Only the
            // last: an earlier run's state is one the layers after it still
            // change.
            let need_qn = run[0] == 0 || !on_dev[run[0] - 1];
            let last_on_dev = *on_dev.last().unwrap_or(&false);
            let carry = last_on_dev && run.last() == Some(&(layers.len() - 1));
            let ok = if carry {
                let r = dsv4_chain_run(
                    layers,
                    &run,
                    cfg,
                    g,
                    st,
                    token_id,
                    &mut folded,
                    Some(state),
                    1,
                    &[],
                    need_qn,
                    pool,
                );
                state_home = r;
                state_on_host = r;
                if r {
                    dspark_note(*run.last().unwrap(), state, cfg);
                }
                r
            } else {
                let r = dsv4_chain_run(
                    layers,
                    &run,
                    cfg,
                    g,
                    st,
                    token_id,
                    &mut folded,
                    None,
                    1,
                    &[],
                    need_qn,
                    pool,
                );
                if r {
                    state_on_host = false;
                }
                r
            };
            if !ok {
                return false;
            }
        }
        if st.dev_set.is_empty() {
            st.dev_set = active_dev.clone();
            st.partial_set = partial_dev.clone();
            // The set is committed, so the card must keep it. Eviction by
            // score is right while the set is still being chosen and wrong
            // afterwards: an evicted layer drops off the card while its
            // caches stay there, and the loop then refuses the whole fast
            // path rather than read state from two sides.
            let mut idxs = Vec::new();
            for (li, l) in layers.iter().enumerate() {
                if !active_dev.get(li).copied().unwrap_or(false) {
                    continue;
                }
                for t in [&l.wq_a, &l.wq_b, &l.wkv, &l.wo_a, &l.wo_b, &l.gate] {
                    idxs.extend(t.model_idx());
                }
                if let Some(pk) = pack_for(l, cfg, li) {
                    for &(a, b, c) in &pk.tensors {
                        idxs.extend([a, b, c]);
                    }
                }
            }
            // Why a HOST layer stayed on the host, said in numbers. Its MoE
            // can still run on the card with a partial pack — `moe_frame` has
            // the remap and hands cold picks back — so the interesting figure
            // is how many experts it got. Zero means the upload order never
            // reached it; a few hundred means the readiness gate refused. The
            // two have different fixes and reading the code cannot tell them
            // apart.
            for (li, l) in layers.iter().enumerate() {
                if active_dev.get(li).copied().unwrap_or(false) {
                    continue;
                }
                let packed = pack_for(l, cfg, li).map_or(0, |p| p.globals.len());
                tracing::info!(
                    "слой {li} на хосте: упаковано {packed} экспертов из {}",
                    cfg.n_routed_experts
                );
            }
            let pinned = layers
                .iter()
                .find_map(|l| l.experts.first().and_then(|e| e.w1.model_arc()))
                .map_or(0, |m| crate::gpu_wgpu::pin_weights(&m, &idxs));
            tracing::info!(
                "закреплено на карте: {pinned} тензоров {} слоёв",
                on_dev.iter().filter(|&&x| x).count()
            );
        }
    }
    if state_home || state_on_host {
        return true;
    }
    crate::gpu_wgpu::dsv4_state_read(state)
}

/// Run a layer whose attention skeleton fits but only a subset of its MoE
/// experts does. This path is selected from the live VRAM budget, never from
/// a layer number. It is exact: routing spans all experts and cold winners
/// are folded back into the hyper-connection state before the next layer.
#[cfg(feature = "gpu")]
#[allow(clippy::too_many_arguments)]
fn dsv4_partial_layer(
    state: &mut [f32],
    folded: &mut Vec<f32>,
    layers: &[Dsv4Layer],
    l: &Dsv4Layer,
    cfg: &Dsv4Cfg,
    st: &mut Dsv4State,
    token_id: u32,
    li: usize,
    freqs: &[f32],
    pool: Option<&crate::pool::Pool>,
    state_on_host: bool,
) -> Option<bool> {
    let dim = cfg.dim;
    // The self-poisoning this walk was parked for: its frames read the
    // pooled post/comb/state slots, and whatever layer ran a frame LAST —
    // on this token or the previous one — left its own there. The walk
    // now seeds its OWN slots from the state it holds at entry, and is
    // immune to the neighbours. The state is home whenever the previous
    // layer exited through this walk or the host branch; a device exit
    // (full-layer frame) leaves it on the card, where the slots are
    // already this token's — nothing to reseed then.
    if state_on_host {
        let (f, post, comb) = hc_fold_norm(
            state,
            &l.hc_attn_fn,
            &l.hc_attn_scale,
            &l.hc_attn_base,
            &l.attn_norm,
            cfg,
            pool,
        );
        *folded = f;
        if !crate::gpu_wgpu::dsv4_hc_write(&post, &comb)
            || !crate::gpu_wgpu::dsv4_state_write(state)
        {
            return None;
        }
    }
    let mut prep = AttnPrep::default();
    let mut sink = vec![0.0f32; dim];
    attention_step(
        folded,
        l,
        cfg,
        st,
        li,
        freqs,
        pool,
        Some(&mut prep),
        &mut sink,
    );
    let hd = cfg.head_dim;
    let n_comp = st.compressed[li].len() / hd;
    let cap = (cfg.window + n_comp.next_power_of_two().max(64)) * hd;
    if !crate::gpu_wgpu::dsv4_cache_write(st.kv_id, li, 0, &st.window[li], cap)
        || (n_comp > 0
            && !crate::gpu_wgpu::dsv4_cache_write(
                st.kv_id,
                li,
                cfg.window * hd,
                &st.compressed[li],
                cap,
            ))
    {
        return None;
    }
    let a_tail = crate::gpu_wgpu::Dsv4HcTail {
        fn_: &l.hc_ffn_fn,
        scale: &l.hc_ffn_scale,
        base: &l.hc_ffn_base,
        norm: &l.ffn_norm,
        hc: cfg.hc_mult,
        sinkhorn_iters: cfg.hc_sinkhorn_iters,
        hc_eps: cfg.hc_eps,
        eps: cfg.norm_eps,
    };
    let scale = (cfg.head_dim as f32).powf(-0.5);
    // Optional FreeToken-style split point. Reading the normed FFN input
    // here adds one fence between attention and MoE, but it also lets the
    // host predict and execute cold experts while the resident experts are
    // running on the GPU. Keep the old no-readback path as the default: on
    // a warm/full pack the extra fence has nothing to hide and only hurts.
    let mut ffn_input = if cpu_overlap_on() {
        vec![0.0f32; dim]
    } else {
        Vec::new()
    };
    if !attn_frame(
        l,
        cfg,
        st,
        li,
        folded,
        &prep.qr,
        &prep.idxs,
        freqs,
        st.pos,
        prep.win_len,
        scale,
        Some(&a_tail),
        &mut ffn_input,
    ) {
        return None;
    }
    let nxt = layers.get(li + 1);
    let forced = l
        .tid2eid
        .as_ref()
        .map(|tbl| hash_route(tbl, cfg.vocab, cfg.top_k, token_id));
    let mut next = vec![0.0f32; dim];
    let (cold_sum, cold_count) = moe_frame(
        &ffn_input,
        l,
        cfg,
        li,
        &[],
        forced.as_deref(),
        pool,
        Some(&a_tail),
        // Do not pre-fold the next layer yet. That fold reuses the canonical
        // `post` slot; a cold correction still needs THIS layer's post. Once
        // the corrected state is home, the exact next fold is cheap on the
        // host and seeds either another partial frame or the next full run.
        None,
        &mut next,
    )?;
    // The resident contribution has already been expanded on the device. If
    // there were cold winners, add `post[j] * cold_sum` and retrieve the
    // corrected state in that submission; otherwise a plain readback is
    // enough. This state handoff is what makes partial layers composable at
    // arbitrary positions, not just at the tail of one checkpoint.
    let state_ok = if cold_count == 0 {
        crate::gpu_wgpu::dsv4_state_read(state)
    } else {
        crate::gpu_wgpu::dsv4_state_add_cold(&cold_sum, cfg.hc_mult, state)
    };
    if !state_ok {
        return None;
    }
    if let Some(n) = nxt {
        let (f, post, comb) = hc_fold_norm(
            state,
            &n.hc_attn_fn,
            &n.hc_attn_scale,
            &n.hc_attn_base,
            &n.attn_norm,
            cfg,
            pool,
        );
        *folded = f;
        if !crate::gpu_wgpu::dsv4_hc_write(&post, &comb)
            || !crate::gpu_wgpu::dsv4_state_write(state)
        {
            return None;
        }
    }
    // NB: the CALLER notes this layer for the draft's ring — a note here
    // as well double-counts the capture and fails `dspark_take`'s
    // completeness check (seen 4 of 3, measured), which reads exactly like
    // the starvation it was meant to fix.
    Some(true)
}

/// `CMF_DSV4_CHAIN1=1`: a partial layer runs as ONE submission — attention,
/// folds and the subset MoE in a single frame, the state staying on the
/// card when every winner was resident (the common case once the slots
/// warm). Cold winners pay the walk's exact correction from the preserved
/// post. Enabled by default after parity and long-run measurements on RTX
/// 5090, RTX PRO 6000 and A40; `CMF_DSV4_CHAIN1=0` keeps the old bisect.
#[cfg(feature = "gpu")]
fn chain1_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| {
        std::env::var("CMF_DSV4_CHAIN1")
            .map(|v| v != "0")
            .unwrap_or(true)
    })
}

/// `CMF_DSV4_CPU_OVERLAP=1`: on the two-frame partial walk, read the exact
/// normalized MoE input after attention and use it to overlap cold CPU
/// experts with the resident GPU frame. This is deliberately independent
/// from `CMF_DSV4_PARTIAL_WALK`: it is a measured alternative to chain-of-one,
/// not a new default for full packs.
#[cfg(feature = "gpu")]
fn cpu_overlap_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| std::env::var("CMF_DSV4_CPU_OVERLAP").is_ok_and(|v| v != "0"))
}

#[cfg(feature = "gpu")]
#[allow(clippy::too_many_arguments)]
fn dsv4_chain1_layer(
    state: &mut [f32],
    folded: &mut Vec<f32>,
    layers: &[Dsv4Layer],
    l: &Dsv4Layer,
    cfg: &Dsv4Cfg,
    st: &mut Dsv4State,
    token_id: u32,
    li: usize,
    freqs: &[f32],
    pool: Option<&crate::pool::Pool>,
    state_on_host: bool,
) -> Option<bool> {
    let dim = cfg.dim;
    let pk = pack_for(l, cfg, li)?;
    if pk.route_complete() {
        return None;
    }
    let model = l.experts.first().and_then(|e| e.w1.model_arc())?;
    // The same entry self-seed the repaired walk uses: the frame reads the
    // pooled post/comb/state slots, and this layer's own are the only ones
    // it may trust.
    if state_on_host {
        let (f, post, comb) = hc_fold_norm(
            state,
            &l.hc_attn_fn,
            &l.hc_attn_scale,
            &l.hc_attn_base,
            &l.attn_norm,
            cfg,
            pool,
        );
        *folded = f;
        if !crate::gpu_wgpu::dsv4_hc_write(&post, &comb)
            || !crate::gpu_wgpu::dsv4_state_write(state)
        {
            return None;
        }
    }
    let mut prep = AttnPrep::default();
    let mut sink = vec![0.0f32; dim];
    attention_step(
        folded,
        l,
        cfg,
        st,
        li,
        freqs,
        pool,
        Some(&mut prep),
        &mut sink,
    );
    let hd = cfg.head_dim;
    let n_comp = st.compressed[li].len() / hd;
    let cap = (cfg.window + n_comp.next_power_of_two().max(64)) * hd;
    let kv_id = st.kv_id;
    if !crate::gpu_wgpu::dsv4_cache_write(kv_id, li, 0, &st.window[li], cap)
        || (n_comp > 0
            && !crate::gpu_wgpu::dsv4_cache_write(
                kv_id,
                li,
                cfg.window * hd,
                &st.compressed[li],
                cap,
            ))
    {
        return None;
    }
    let idx32: Vec<u32> = prep
        .idxs
        .iter()
        .map(|&p| {
            if p < prep.win_len {
                p as u32
            } else {
                (cfg.window + (p - prep.win_len)) as u32
            }
        })
        .collect();
    let (Some(wq_a), Some(wq_b), Some(wo_a), Some(wo_b)) = (
        l.wq_a.model_idx(),
        l.wq_b.model_idx(),
        l.wo_a.model_idx(),
        l.wo_b.model_idx(),
    ) else {
        return None;
    };
    // Under the subset contract the forced list stays GLOBAL: the remap
    // either finds each hash winner a slot or returns it cold.
    let forced: Option<Vec<usize>> = l
        .tid2eid
        .as_ref()
        .map(|tbl| hash_route(tbl, cfg.vocab, cfg.top_k, token_id));
    let global_remap = pk
        .global
        .as_ref()
        .map(|gl| gl.pool.remap(gl.layer, cfg.n_routed_experts));
    let dynv = pk.dynslots.lock().unwrap();
    let live_remap = global_remap.as_deref().unwrap_or(dynv.remap.as_slice());
    let nxt = layers.get(li + 1);
    let w = crate::gpu_wgpu::Dsv4LayerW {
        attn: crate::gpu_wgpu::Dsv4AttnW {
            wq_a,
            wq_b,
            wo_a,
            wo_b,
            q_norm: &l.q_norm,
            sink: &l.attn_sink,
        },
        moe: crate::gpu_wgpu::Dsv4MoeW {
            router: &[],
            experts: &pk.tensors,
            logits: &[],
            // GLOBAL bias under the subset contract — the ranking spans
            // every expert, so a packed-order bias would misalign it.
            bias: pk.bias.as_deref(),
            mask: pk.mask.as_deref(),
            forced: forced.as_deref(),
            remap: Some(live_remap),
            global: pk.global.as_ref().map(|gl| crate::gpu_wgpu::Dsv4GlobalMoe {
                pool_uid: gl.pool.uid,
                shared_slot: gl.shared_slot,
                segment_slots: gl.pool.segment_slots as u32,
            }),
            has_shared: true,
            shared_weight: 1.0,
            preweighted: false,
            qwen_softmax: false,
        },
        hc_ffn_fn: &l.hc_ffn_fn,
        hc_ffn_scale: &l.hc_ffn_scale,
        hc_ffn_base: &l.hc_ffn_base,
        hc_next_fn: nxt.map(|n| n.hc_attn_fn.as_slice()),
        hc_next_scale: nxt.map_or(&l.hc_attn_scale, |n| &n.hc_attn_scale),
        hc_next_base: nxt.map_or(&l.hc_attn_base, |n| n.hc_attn_base.as_slice()),
        ffn_norm: &l.ffn_norm,
        next_norm: nxt.map_or(&l.attn_norm, |n| n.attn_norm.as_slice()),
        next_q_norm: nxt.map_or(&l.q_norm, |n| n.q_norm.as_slice()),
        next_wq_a: nxt.and_then(|n| n.wq_a.model_idx()),
        router: &pk.router,
    };
    let geom = crate::gpu_wgpu::Dsv4LayerGeom {
        attn: crate::gpu_wgpu::Dsv4AttnGeom {
            dim,
            nh: cfg.n_heads,
            hd,
            rd: cfg.rope_head_dim,
            q_lora: cfg.q_lora_rank,
            o_lora: cfg.o_lora_rank,
            o_groups: cfg.o_groups,
            eps: cfg.norm_eps,
            scale: (hd as f32).powf(-0.5),
        },
        moe: crate::gpu_wgpu::Dsv4MoeGeom {
            hidden: dim,
            inter: cfg.moe_inter,
            top_k: cfg.top_k,
            route_scale: cfg.route_scale,
            swiglu_limit: cfg.swiglu_limit,
            gu_q2: l
                .experts
                .first()
                .is_some_and(|e| e.w1.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP)),
        },
        hc: cfg.hc_mult,
        hc_eps: cfg.hc_eps,
        sinkhorn_iters: cfg.hc_sinkhorn_iters,
    };
    let mut next = vec![0.0f32; dim];
    let mut cold: Vec<(usize, f32)> = Vec::new();
    let mut routed: Vec<(usize, f32)> = Vec::new();
    let mut cold_x: Vec<f32> = Vec::new();
    let need_routes = route_stats_on() || pk.global.is_some();
    if !crate::gpu_wgpu::dsv4_layer_frame(
        &model,
        &w,
        geom,
        kv_id,
        li,
        Some(&prep.qr),
        &idx32,
        freqs,
        st.pos,
        &mut next,
        Some(&mut cold),
        need_routes.then_some(&mut routed),
        &mut cold_x,
    ) {
        return None;
    }
    if route_stats_on() {
        record_route(li, layers.len(), cfg.n_routed_experts, &routed);
    }
    drop(dynv);
    // The cold/readback slot already carries the device's real winners.
    // Feed all of them back to the allocator: this both installs misses and
    // refreshes hit ages, without a host-side router prediction or another
    // fence. A prediction disagreement therefore remains impossible here.
    if let Some(gl) = pk.global.as_ref() {
        let picks: Vec<usize> = routed.iter().map(|&(gi, _)| gi).collect();
        gl.pool
            .ensure_picks(&model, gl.layer, &picks, &l.experts, cfg.top_k, 1);
    }
    if cold.is_empty() {
        // Every winner was resident: the state stays on the card and the
        // frame's own next-fold is exact. This is the single-submission
        // path the whole function exists for.
        *folded = next;
        return Some(false);
    }
    // Cold winners: complete on the frame's own normed input, correct the
    // device state from the preserved post, and bring it home.
    if cold_x.len() < dim {
        return None;
    }
    let mut cold_sum = vec![0.0f32; dim];
    {
        let results: Vec<std::sync::Mutex<Vec<f32>>> = cold
            .iter()
            .map(|_| std::sync::Mutex::new(Vec::new()))
            .collect();
        let (cold_ref, results_ref, x_ref) = (&cold, &results, &cold_x[..dim]);
        std::thread::scope(|sc| {
            for i in 0..cold_ref.len() {
                let (gi, wt) = cold_ref[i];
                let Some(exp) = l.experts.get(gi) else {
                    continue;
                };
                let r = &results_ref[i];
                sc.spawn(move || {
                    let mut a = vec![0.0f32; cfg.dim];
                    crate::gpu::cpu_scope(|| run_expert(x_ref, exp, cfg, wt, None, &mut a));
                    *r.lock().unwrap() = a;
                });
            }
        });
        for r in &results {
            let a = r.lock().unwrap();
            for (o, v) in cold_sum.iter_mut().zip(a.iter()) {
                *o += v;
            }
        }
    }
    if !crate::gpu_wgpu::dsv4_state_add_cold_preserved(&cold_sum, cfg.hc_mult, state, kv_id, li) {
        return None;
    }
    // Reactive refill: the winners the slots did not hold are the likeliest
    // winners of the NEXT token — pull them in now, LRU-evicting.
    if pk.global.is_none() {
        let mut dynv = pk.dynslots.lock().unwrap();
        dynv.clock += 1;
        let clock = dynv.clock;
        for &(gi, _) in &cold {
            if gi >= dynv.remap.len() || dynv.remap[gi] != u32::MAX {
                continue;
            }
            let victim = (0..dynv.owner.len())
                .filter(|&sl| dynv.last[sl] != clock)
                .min_by_key(|&sl| dynv.last[sl]);
            let Some(victim) = victim else { break };
            let Some(exp) = l.experts.get(gi) else {
                continue;
            };
            let t3 = (|| {
                Some((
                    exp.w1.model_idx()?,
                    exp.w3.model_idx()?,
                    exp.w2.model_idx()?,
                ))
            })();
            let Some(t3) = t3 else { continue };
            let gu_q2 = exp.w1.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP);
            let pack_first = pk.tensors.first().map(|t| t.0).unwrap_or(usize::MAX);
            if !crate::gpu_wgpu::dsv4_slot_fill(
                &model,
                pack_first,
                victim,
                gi,
                t3,
                cfg.moe_inter,
                cfg.dim,
                gu_q2,
            ) {
                break;
            }
            let old = dynv.owner[victim] as usize;
            if old < dynv.remap.len() {
                dynv.remap[old] = u32::MAX;
            }
            dynv.remap[gi] = victim as u32;
            dynv.owner[victim] = gi as u32;
            dynv.last[victim] = clock;
            dynv.mutated = true;
        }
    }
    if let Some(n) = nxt {
        let (f, post, comb) = hc_fold_norm(
            state,
            &n.hc_attn_fn,
            &n.hc_attn_scale,
            &n.hc_attn_base,
            &n.attn_norm,
            cfg,
            pool,
        );
        *folded = f;
        if !crate::gpu_wgpu::dsv4_hc_write(&post, &comb) {
            return None;
        }
    }
    Some(true)
}

#[cfg(feature = "gpu")]
fn chain_max() -> usize {
    static N: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
    *N.get_or_init(|| {
        std::env::var("CMF_DSV4_CHAIN_MAX")
            .ok()
            .and_then(|v| v.parse().ok())
            .unwrap_or(usize::MAX)
    })
}

/// `CMF_DSV4_CHAIN=1`: put a run of consecutive device-capable layers in ONE
/// submission. Off by default until it has been measured on a real card.
#[cfg(feature = "gpu")]
/// `CMF_DSV4_HOST_CPU_MOE=1`: a layer that fell off the card runs its MoE on
/// the host WITHOUT the per-op device route — one fence a token instead of
/// one a matvec. Whether that wins is a measurement.
/// `CMF_DSV4_PARTIAL_WALK=1`: the fused device walk of a partial layer.
/// OFF until its self-poisoning is repaired: its attention frame reads the
/// pooled slots its own MoE frame rewrote on the previous token, so every
/// token after the first attends over leftovers — the drafts it captures
/// от такого состояния never match the verify (acceptance 0, measured).
/// The host branch walks these layers correctly; the pack stays resident
/// for the verify tail.
fn partial_walk_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| std::env::var("CMF_DSV4_PARTIAL_WALK").is_ok_and(|v| v != "0"))
}

fn host_cpu_moe() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| std::env::var("CMF_DSV4_HOST_CPU_MOE").is_ok_and(|v| v != "0"))
}

fn chain_enabled() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| {
        std::env::var("CMF_DSV4_CHAIN")
            .map(|v| v != "0")
            .unwrap_or(true)
    })
}

/// Encode a maximal run of consecutive device-capable layers and submit it
/// ONCE. Every layer in the run builds its own attention inputs on the card,
/// so nothing comes back between them — that is the whole saving.
///
/// The run's state belongs to the device from here on: `st.window`,
/// `st.compressed` and the compressor streams for these layers are stale on
/// the host afterwards, and only the counts in `st.dev_*` are kept. A layer
/// that has ever been in a run must therefore never be handed to the CPU
/// path again, which `dev_owned` records.
#[cfg(feature = "gpu")]
#[allow(clippy::too_many_arguments)]
fn dsv4_chain_run(
    layers: &[Dsv4Layer],
    run: &[usize],
    cfg: &Dsv4Cfg,
    g: &Dsv4Globals,
    st: &mut Dsv4State,
    token_id: u32,
    // In AND out: the run reads the fold it starts from and MUST leave the
    // fold it produced, because whatever follows — a host layer, or the next
    // run after a cap — seeds from this. Passing it read-only left every
    // later segment starting from a stale fold: exact with one unbroken run,
    // release-scale garbage the moment anything splits the chain.
    folded: &mut Vec<f32>,
    // When present, the hyper-connection state rides home in the run's own
    // submission instead of costing a second fence afterwards. Only the
    // token's LAST run passes it — an earlier one would read a state the
    // layers after it still change.
    state_out: Option<&mut [f32]>,
    // How many consecutive tokens this run carries. One is decode; more is a
    // prompt chunk or a speculative verify, which are the same shape of work.
    batch: usize,
    // Their ids, needed only when `batch > 1`: a hash layer forces its expert
    // list from the token's id, so the batch needs one list per token and the
    // single `token_id` above cannot supply them.
    batch_ids: &[u32],
    // Whether the device's qn buffer is stale: true at layer zero and after
    // a host layer. When the previous layer was chained, its frame's tail
    // already left THIS layer's LoRA vector on the card, and recomputing it
    // here was a full wq_a matvec on the CPU per run — at CHAIN_MAX=1 that
    // is one per LAYER, which is how a 43-fence path measured slower than
    // an 86-fence one.
    need_qn: bool,
    pool: Option<&crate::pool::Pool>,
) -> bool {
    if run.is_empty() {
        return true;
    }
    let (dim, hd) = (cfg.dim, cfg.head_dim);
    let first = run[0];
    let Some(model) = layers[first].experts.first().and_then(|e| e.w1.model_arc()) else {
        return false;
    };
    // Batch callers seed every token's fold and qn in its own slot. Seeding
    // the legacy shared slot here is not merely redundant: `folded` carries
    // only the eventual LAST output and is empty before the batch runs.
    if batch <= 1 && need_qn {
        let mut qn0 = vec![0.0f32; cfg.q_lora_rank];
        layers[first].wq_a.matvec(folded, &mut qn0, pool);
        rms_weighted(&mut qn0, &layers[first].q_norm, cfg.norm_eps);
        if !crate::gpu_wgpu::dsv4_chain_seed(folded, &qn0) {
            return false;
        }
    } else if batch <= 1 && !crate::gpu_wgpu::dsv4_chain_seed_fold(folded) {
        return false;
    }

    // Held apart from the borrowing structs below, which point into them.
    let mut packs = Vec::with_capacity(run.len());
    let mut forceds: Vec<Option<Vec<usize>>> = Vec::with_capacity(run.len());
    for &li in run {
        let Some(pk) = pack_for(&layers[li], cfg, li) else {
            return false;
        };
        // A chain cannot complete a cold pick between dependent layers, so
        // every expert OPEN in the route must be resident and the pack must
        // not have mutated. A masked-complete or hot-reordered pack carries
        // its immutable global-to-slot remap into the frame.
        // The verify batch reached here with cap-limited partial packs
        // and accepted 0 of 625 drafts — wrong experts, plausible sums.
        if !pk.route_complete() || pk.is_mutated() {
            return false;
        }
        let forced: Option<Vec<usize>> = layers[li].tid2eid.as_ref().and_then(|tbl| {
            let v: Vec<usize> = if pk.needs_remap() {
                hash_route(tbl, cfg.vocab, cfg.top_k, token_id)
            } else {
                hash_route(tbl, cfg.vocab, cfg.top_k, token_id)
                    .into_iter()
                    .map(|gi| pk.to_slot[gi])
                    .collect()
            };
            if v.contains(&usize::MAX) {
                None
            } else {
                Some(v)
            }
        });
        if layers[li].tid2eid.is_some() && forced.is_none() {
            return false;
        }
        forceds.push(forced);
        packs.push(pk);
    }

    let mut items = Vec::with_capacity(run.len());
    let mut freqs = Vec::with_capacity(run.len());
    for (i, &li) in run.iter().enumerate() {
        let l = &layers[li];
        let (Some(wq_a), Some(wq_b), Some(wo_a), Some(wo_b), Some(wkv)) = (
            l.wq_a.model_idx(),
            l.wq_b.model_idx(),
            l.wo_a.model_idx(),
            l.wo_b.model_idx(),
            l.wkv.model_idx(),
        ) else {
            return false;
        };
        let comp = match &l.compressor {
            None => None,
            Some(cp) => {
                let (Some(a), Some(b)) = (cp.wkv.model_idx(), cp.wgate.model_idx()) else {
                    return false;
                };
                Some((
                    crate::gpu_wgpu::Dsv4CompW {
                        wkv: a,
                        wgate: b,
                        norm: &cp.norm,
                        ape: &cp.ape,
                    },
                    crate::gpu_wgpu::Dsv4CompGeom {
                        width: cp.wkv.rows(),
                        hidden: dim,
                        ratio: cp.ratio,
                        overlap: cp.overlap,
                        rope_dim: cfg.rope_head_dim,
                        eps: cfg.norm_eps,
                    },
                ))
            }
        };
        let ix = match &l.indexer {
            None => None,
            Some(ixr) => {
                let cp = &ixr.compressor;
                let (Some(a), Some(b), Some(qb), Some(wp)) = (
                    cp.wkv.model_idx(),
                    cp.wgate.model_idx(),
                    ixr.wq_b.model_idx(),
                    ixr.weights_proj.model_idx(),
                ) else {
                    return false;
                };
                let ih = ixr.weights_proj.rows();
                Some((
                    crate::gpu_wgpu::Dsv4CompW {
                        wkv: a,
                        wgate: b,
                        norm: &cp.norm,
                        ape: &cp.ape,
                    },
                    crate::gpu_wgpu::Dsv4CompGeom {
                        width: cp.wkv.rows(),
                        hidden: dim,
                        ratio: cp.ratio,
                        overlap: cp.overlap,
                        rope_dim: cfg.rope_head_dim,
                        eps: cfg.norm_eps,
                    },
                    crate::gpu_wgpu::Dsv4IxW {
                        wq_b: qb,
                        weights_proj: wp,
                    },
                    crate::gpu_wgpu::Dsv4IxGeom {
                        ih,
                        idim: ixr.wq_b.rows() / ih.max(1),
                        q_lora: cfg.q_lora_rank,
                        hidden: dim,
                        rope_dim: cfg.rope_head_dim,
                        eps: cfg.norm_eps,
                        top_k: cfg.index_topk,
                        window: cfg.window,
                    },
                ))
            }
        };
        // The cache has to be big enough BEFORE the frame appends into it:
        // a chained layer never calls dsv4_cache_write, which is what used
        // to create and grow it.
        let ew_c0 = l.compressor.as_ref().map_or(0, |cp| {
            if cp.overlap {
                cp.wkv.rows() / 2
            } else {
                cp.wkv.rows()
            }
        });
        let comp_extra = l
            .compressor
            .as_ref()
            .map_or(0, |cp| batch.max(1).div_ceil(cp.ratio.max(1)));
        let need = cfg.window * hd
            + (st.dev_n_comp[li] + comp_extra + 1) * ew_c0.max(1)
            + (batch.max(1) + 1) * hd;
        if !crate::gpu_wgpu::dsv4_cache_ensure(st.kv_id, li, need.next_power_of_two()) {
            return false;
        }
        let ew_c = comp.as_ref().map_or(
            0,
            |(_, cg)| {
                if cg.overlap { cg.width / 2 } else { cg.width }
            },
        );
        let ew_i = ix.as_ref().map_or(
            0,
            |(_, cg, _, _)| {
                if cg.overlap { cg.width / 2 } else { cg.width }
            },
        );
        let prep = crate::gpu_wgpu::Dsv4Prep {
            wkv,
            kv_norm: &l.kv_norm,
            comp,
            ix,
            filled: st.dev_filled[li],
            window: cfg.window,
            n_comp: st.dev_n_comp[li],
            n_ix: st.dev_n_ix[li],
            comp_dst_off: cfg.window * hd + st.dev_n_comp[li] * ew_c,
            ix_dst_off: st.dev_n_ix[li] * ew_i,
            idx_cap: cfg.window
                + if l.indexer.is_some() {
                    cfg.index_topk
                } else {
                    st.dev_n_comp[li] + comp_extra + 1
                },
        };
        let nxt = layers.get(li + 1);
        let w = crate::gpu_wgpu::Dsv4LayerW {
            attn: crate::gpu_wgpu::Dsv4AttnW {
                wq_a,
                wq_b,
                wo_a,
                wo_b,
                q_norm: &l.q_norm,
                sink: &l.attn_sink,
            },
            moe: crate::gpu_wgpu::Dsv4MoeW {
                router: &packs[i].router,
                experts: &packs[i].tensors,
                logits: &[],
                // The PACK's slice, not a per-run Vec: the address stability
                // is the whole point (see Pack::bias).
                bias: packs[i].bias.as_deref(),
                mask: packs[i].mask.as_deref(),
                forced: forceds[i].as_deref(),
                remap: packs[i].needs_remap().then_some(packs[i].remap.as_slice()),
                global: None,
                has_shared: true,
                shared_weight: 1.0,
                preweighted: false,
                qwen_softmax: false,
            },
            hc_ffn_fn: &l.hc_ffn_fn,
            hc_ffn_scale: &l.hc_ffn_scale,
            hc_ffn_base: &l.hc_ffn_base,
            hc_next_fn: nxt.map(|n| n.hc_attn_fn.as_slice()),
            hc_next_scale: nxt.map_or(&l.hc_attn_scale, |n| &n.hc_attn_scale),
            hc_next_base: nxt.map_or(&l.hc_attn_base, |n| n.hc_attn_base.as_slice()),
            ffn_norm: &l.ffn_norm,
            next_norm: nxt.map_or(&l.attn_norm, |n| n.attn_norm.as_slice()),
            next_q_norm: nxt.map_or(&l.q_norm, |n| n.q_norm.as_slice()),
            next_wq_a: nxt.and_then(|n| n.wq_a.model_idx()),
            router: &packs[i].router,
        };
        let geom = crate::gpu_wgpu::Dsv4LayerGeom {
            attn: crate::gpu_wgpu::Dsv4AttnGeom {
                dim,
                nh: cfg.n_heads,
                hd,
                rd: cfg.rope_head_dim,
                q_lora: cfg.q_lora_rank,
                o_lora: cfg.o_lora_rank,
                o_groups: cfg.o_groups,
                eps: cfg.norm_eps,
                scale: (hd as f32).powf(-0.5),
            },
            moe: crate::gpu_wgpu::Dsv4MoeGeom {
                hidden: dim,
                inter: cfg.moe_inter,
                top_k: cfg.top_k,
                route_scale: cfg.route_scale,
                swiglu_limit: cfg.swiglu_limit,
                gu_q2: l.experts.first().is_some_and(|e| {
                    e.w1.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP)
                }),
            },
            hc: cfg.hc_mult,
            hc_eps: cfg.hc_eps,
            sinkhorn_iters: cfg.hc_sinkhorn_iters,
        };
        freqs.push(if l.compressor.is_some() {
            g.inv_freq_compress.as_slice()
        } else {
            g.inv_freq_window.as_slice()
        });
        items.push((w, geom, prep));
    }

    let mut out = vec![0.0f32; dim * batch.max(1)];
    if batch > 1 {
        // A batch keeps its own state per token. When a host tail follows,
        // all of those states ride home beside the folds in the same fence.
        // One forced row per token: same layers, the hash rows re-derived
        // from each token's own id.
        let mut forced_pt: Vec<Vec<Option<Vec<usize>>>> = Vec::with_capacity(batch);
        for t in 0..batch {
            let id = batch_ids.get(t).copied().unwrap_or(token_id);
            let mut row = Vec::with_capacity(run.len());
            for (i, &li) in run.iter().enumerate() {
                row.push(layers[li].tid2eid.as_ref().and_then(|tbl| {
                    let v: Vec<usize> = if packs[i].needs_remap() {
                        hash_route(tbl, cfg.vocab, cfg.top_k, id)
                    } else {
                        hash_route(tbl, cfg.vocab, cfg.top_k, id)
                            .into_iter()
                            .map(|gi| packs[i].to_slot[gi])
                            .collect()
                    };
                    if v.contains(&usize::MAX) {
                        None
                    } else {
                        Some(v)
                    }
                }));
                if layers[li].tid2eid.is_some() && row[i].is_none() {
                    return false;
                }
            }
            forced_pt.push(row);
        }
        if !crate::gpu_wgpu::dsv4_chain_batch(
            &model,
            &items,
            st.kv_id,
            first,
            &freqs,
            st.pos,
            batch,
            Some(&forced_pt),
            &mut out,
            state_out,
        ) {
            return false;
        }
        // The caller wants the LAST token's fold: it is the one whose logits
        // continue the sequence.
        *folded = out[(batch - 1) * dim..batch * dim].to_vec();
    } else {
        if !crate::gpu_wgpu::dsv4_layer_chain(
            &model, &items, st.kv_id, first, &freqs, st.pos, &mut out, state_out,
        ) {
            return false;
        }
        *folded = out;
    }
    // The device advanced these; the host keeps only the arithmetic. A batch
    // advanced them once per token, in order, so the host replays the same
    // rule that many times rather than inventing a closed form for it.
    for (i, &li) in run.iter().enumerate() {
        for t in 0..batch.max(1) {
            let pos = st.pos + t;
            st.dev_filled[li] = (st.dev_filled[li] + 1).min(cfg.window);
            if let Some((_, cg, ..)) = items[i].2.ix.as_ref() {
                if (pos + 1) % cg.ratio == 0 {
                    st.dev_n_ix[li] += 1;
                }
            }
            if let Some((_, cg)) = items[i].2.comp.as_ref() {
                if (pos + 1) % cg.ratio == 0 {
                    st.dev_n_comp[li] += 1;
                }
            }
        }
    }
    st.dev_owned = true;
    true
}

/// `CMF_DSV4_HC_DEV=0` puts the hyper-connections back on the host.
#[cfg(feature = "gpu")]
fn hc_on_device() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| {
        // OPT-IN. On the release checkpoint this path reads 3.234 against
        // the CPU's 3.282 — divergent — and the speed is unchanged, so there
        // is no trade to weigh: it must not be the default until it is
        // exact. The toy's near-agreement (129.787 vs 129.792) hid a real
        // fault the release exposes.
        std::env::var("CMF_DSV4_HC_DEV").is_ok_and(|v| v != "0") && crate::gpu::backend_available()
    })
}

/// The two-frame path with the hyper-connections on the card.
///
/// The host still prepares each layer's attention inputs — the compressor,
/// the indexer and the window, which are exact there — but it no longer
/// folds, Sinkhorns or norms, and it no longer carries the MoE half's input
/// between the halves: the attention frame leaves it on the device and the
/// MoE frame reads it from there. One readback a layer instead of two, and
/// 19 ms of host arithmetic a token gone.
#[cfg(feature = "gpu")]
#[allow(clippy::too_many_arguments)]
fn dsv4_two_frame_loop(
    state: &mut [f32],
    layers: &[Dsv4Layer],
    g: &Dsv4Globals,
    cfg: &Dsv4Cfg,
    st: &mut Dsv4State,
    token_id: u32,
    inv_freq: &[f32],
    pool: Option<&crate::pool::Pool>,
    scratch: &mut HcScratch,
) -> bool {
    let dim = cfg.dim;
    let freqs_of = |l: &Dsv4Layer| -> &[f32] {
        let f = if l.compressor.is_some() {
            &g.inv_freq_compress
        } else {
            &g.inv_freq_window
        };
        if f.is_empty() { inv_freq } else { f.as_slice() }
    };
    // Layer zero's fold has no frame before it, exactly as in the layer path.
    let (mut folded, post0, comb0) = hc_fold_norm(
        state,
        &layers[0].hc_attn_fn,
        &layers[0].hc_attn_scale,
        &layers[0].hc_attn_base,
        &layers[0].attn_norm,
        cfg,
        pool,
    );
    if !crate::gpu_wgpu::dsv4_state_write(state) || !crate::gpu_wgpu::dsv4_hc_write(&post0, &comb0)
    {
        return false;
    }
    // PRE-FLIGHT, before the first byte of state moves: a mid-loop refusal
    // would hand the token back to the ordinary loop AFTER these caches
    // advanced, and the second advance is not a slow answer but a wrong one.
    // The same discipline the layer loop states in the same words.
    let mut on_dev = vec![false; layers.len()];
    for (li, l) in layers.iter().enumerate() {
        let Some(pk) = pack_for(l, cfg, li) else {
            return false;
        };
        let Some(model) = l.experts.first().and_then(|e| e.w1.model_arc()) else {
            return false;
        };
        let gu_q2 = l
            .experts
            .first()
            .is_some_and(|e| e.w1.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP));
        let attn_ok = [
            l.wq_a.model_idx(),
            l.wq_b.model_idx(),
            l.wo_a.model_idx(),
            l.wo_b.model_idx(),
        ]
        .into_iter()
        .flatten()
        .all(|i| crate::gpu_wgpu::dsv4_weight_ready(&model, i));
        on_dev[li] = attn_ok
            && pk.route_complete()
            && crate::gpu_wgpu::dsv4_experts_ready(
                &model,
                &pk.tensors,
                cfg.moe_inter,
                dim,
                gu_q2,
                l.experts.first().is_some_and(|e| {
                    e.w2.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP)
                }),
            );
    }
    if !on_dev.iter().any(|&x| x) {
        return false;
    }
    let mut sink = vec![0.0f32; dim];
    for (li, l) in layers.iter().enumerate() {
        // A layer the card cannot hold runs on the host WHOLE, with the
        // state fetched and put back around it — the mixed ownership the
        // layer loop already proved out.
        if !on_dev[li] {
            if !crate::gpu_wgpu::dsv4_state_read(state) {
                return false;
            }
            let freqs = freqs_of(l);
            hc_block(
                state,
                &l.hc_attn_fn,
                &l.hc_attn_scale,
                &l.hc_attn_base,
                &l.attn_norm,
                cfg,
                scratch,
                pool,
                |f, o| attention_step(f, l, cfg, st, li, freqs, pool, None, o),
            );
            hc_block(
                state,
                &l.hc_ffn_fn,
                &l.hc_ffn_scale,
                &l.hc_ffn_base,
                &l.ffn_norm,
                cfg,
                scratch,
                pool,
                |f, o| moe_step(f, l, cfg, token_id, li, pool, o),
            );
            let nref = layers.get(li + 1).unwrap_or(l);
            let (f, p2, c2) = hc_fold_norm(
                state,
                &nref.hc_attn_fn,
                &nref.hc_attn_scale,
                &nref.hc_attn_base,
                &nref.attn_norm,
                cfg,
                pool,
            );
            folded = f;
            if !crate::gpu_wgpu::dsv4_hc_write(&p2, &c2)
                || !crate::gpu_wgpu::dsv4_state_write(state)
            {
                return false;
            }
            continue;
        }
        // The host's half: the caches and the attended list, untouched.
        let mut prep = AttnPrep::default();
        attention_step(
            &folded,
            l,
            cfg,
            st,
            li,
            freqs_of(l),
            pool,
            Some(&mut prep),
            &mut sink,
        );
        let hd = cfg.head_dim;
        let n_comp = st.compressed[li].len() / hd;
        let cap = (cfg.window + n_comp.next_power_of_two().max(64)) * hd;
        if !crate::gpu_wgpu::dsv4_cache_write(st.kv_id, li, 0, &st.window[li], cap)
            || (n_comp > 0
                && !crate::gpu_wgpu::dsv4_cache_write(
                    st.kv_id,
                    li,
                    cfg.window * hd,
                    &st.compressed[li],
                    cap,
                ))
        {
            return false;
        }
        let _idx32: Vec<u32> = prep
            .idxs
            .iter()
            .map(|&p| {
                if p < prep.win_len {
                    p as u32
                } else {
                    (cfg.window + (p - prep.win_len)) as u32
                }
            })
            .collect();
        let nxt = layers.get(li + 1);
        let a_tail = crate::gpu_wgpu::Dsv4HcTail {
            fn_: &l.hc_ffn_fn,
            scale: &l.hc_ffn_scale,
            base: &l.hc_ffn_base,
            norm: &l.ffn_norm,
            hc: cfg.hc_mult,
            sinkhorn_iters: cfg.hc_sinkhorn_iters,
            hc_eps: cfg.hc_eps,
            eps: cfg.norm_eps,
        };
        let scale = (cfg.head_dim as f32).powf(-0.5);
        if !attn_frame(
            l,
            cfg,
            st,
            li,
            &folded,
            &prep.qr,
            &prep.idxs,
            freqs_of(l),
            st.pos,
            prep.win_len,
            scale,
            Some(&a_tail),
            &mut [],
        ) {
            return false;
        }
        let m_tail = nxt.map(|n| crate::gpu_wgpu::Dsv4HcTail {
            fn_: &n.hc_attn_fn,
            scale: &n.hc_attn_scale,
            base: &n.hc_attn_base,
            norm: &n.attn_norm,
            hc: cfg.hc_mult,
            sinkhorn_iters: cfg.hc_sinkhorn_iters,
            hc_eps: cfg.hc_eps,
            eps: cfg.norm_eps,
        });
        let mut next = vec![0.0f32; dim];
        let pair = m_tail
            .as_ref()
            .zip(nxt)
            .map(|(t, n)| (t, n.attn_norm.as_slice()));
        let forced = l
            .tid2eid
            .as_ref()
            .map(|tbl| hash_route(tbl, cfg.vocab, cfg.top_k, token_id));
        if moe_frame(
            &[],
            l,
            cfg,
            li,
            &[],
            forced.as_deref(),
            pool,
            Some(&a_tail),
            pair,
            &mut next,
        )
        .is_none()
        {
            return false;
        }
        folded = next;
    }
    let _ = scratch;
    crate::gpu_wgpu::dsv4_state_read(state)
}

/// The host half of one hyper-connection block: mixes, Sinkhorn, fold, norm.
/// The device does this for every layer but the first, whose state it has not
/// seen yet.
#[cfg(feature = "gpu")]
#[allow(clippy::too_many_arguments)]
fn hc_fold_norm(
    state: &[f32],
    hc_fn: &[f32],
    hc_scale: &[f32; 3],
    hc_base: &[f32],
    norm_w: &[f32],
    cfg: &Dsv4Cfg,
    pool: Option<&crate::pool::Pool>,
) -> (Vec<f32>, Vec<f32>, Vec<f32>) {
    let (hc, dim) = (cfg.hc_mult, cfg.dim);
    let mix_hc = (2 + hc) * hc;
    let mut mixes = vec![0.0f32; mix_hc];
    hc_mixes(state, hc_fn, mix_hc, cfg.norm_eps, pool, &mut mixes);
    let mut pre = vec![0.0f32; hc];
    let mut post = vec![0.0f32; hc];
    let mut comb = vec![0.0f32; hc * hc];
    hc_split_sinkhorn(
        &mixes,
        hc_scale,
        hc_base,
        hc,
        cfg.hc_sinkhorn_iters,
        cfg.hc_eps,
        &mut pre,
        &mut post,
        &mut comb,
    );
    let mut folded = vec![0.0f32; dim];
    hc_fold(state, &pre, hc, dim, &mut folded);
    rms_weighted(&mut folded, norm_w, cfg.norm_eps);
    // post and comb travel with the fold: the frame's opening expand needs
    // exactly those, and they are not recoverable from the state alone.
    (folded, post, comb)
}

/// `CMF_DSV4_GPU_LAYER=1`: one submission per layer instead of two, with the
/// hyper-connection glue and the router on the device.
///
/// CORRECT — perplexity 5.211 against the CPU's 5.211 on the release, 128.576
/// against 128.576 on the toy — and SLOWER on this hardware: 6.0 tok/s where
/// the two-frame path gets 9.3. The reason is not the frame, it is the
/// all-or-nothing granularity underneath it. A layer whose experts miss VRAM
/// runs entirely on the host, attention included (6.5 ms a call against 0.9),
/// and with 100 GB of experts against a 98 GB card a fifth of the layers
/// miss. The two-frame path only loses the MoE half of those layers.
///
/// So the barrier it saves is real and the fallback it forces costs more. The
/// fix is the granularity: pack the experts that FIT, route over all of them
/// anyway, and run the few cold picks of a token on the host — per EXPERT,
/// not per layer. Then no layer ever leaves the device and this frame wins by
/// the 15 ms a token it was built to save.
#[cfg(feature = "gpu")]
fn gpu_layer_enabled() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| {
        std::env::var("CMF_DSV4_GPU_LAYER")
            .map(|v| v != "0")
            .unwrap_or(true)
            && crate::gpu::backend_available()
    })
}

/// The packed expert set of one layer: which globals made it in, and their
/// directory indices in packing order with the shared expert last. Built once
/// — the mask does not change during a run — and keyed by layer.
#[cfg(feature = "gpu")]
struct PackDyn {
    remap: Vec<u32>,
    owner: Vec<u32>,
    last: Vec<u64>,
    clock: u64,
    /// Per-expert recent-use tally (halved every 64 tokens): the q* rule.
    /// A slot upload only pays for itself when the expert is REUSED —
    /// FreeToken's split — so a fetch needs `seen >= CMF_DSV4_FETCH_MIN_SEEN`
    /// prior recent picks; a first-timer stays a cold pick and the CPU
    /// reads it at the shelf. The automatic default comes from a small
    /// one-time host→device bandwidth probe; an explicit env still wins.
    seen: Vec<u16>,
    /// Set on the first slot refill. The chain and the batch verify hand
    /// the device `remap: None` and trust the banks to still hold the
    /// BUILD-TIME packing — a mutated pack must never be claimed by them.
    mutated: bool,
}

/// One logical cache line in the model-wide expert pool. `layer` is the
/// first expert tensor's directory index rather than the ordinal: trunk and
/// MTP stages both use small ordinal numbers, while this identity is unique
/// for the lifetime of the mapped model.
#[cfg(feature = "gpu")]
#[derive(Clone, Copy)]
struct GlobalOwner {
    layer: usize,
    expert: usize,
    pinned: bool,
}

#[cfg(feature = "gpu")]
struct GlobalPoolDyn {
    slot_for: std::collections::HashMap<(usize, usize), u32>,
    owner: Vec<Option<GlobalOwner>>,
    last: Vec<u64>,
    seen: std::collections::HashMap<(usize, usize), u16>,
    occupancy: std::collections::HashMap<usize, usize>,
    clock: u64,
}

/// FreeToken's unified `(layer, expert) -> slot` table, with one deliberate
/// addition learned from this engine's routing traces: each trunk layer gets
/// a small protected floor. A plain global LRU under a deterministic
/// layer-by-layer sweep measured 4.7% hits, while equal per-layer LRUs hit
/// about 70%. The floor prevents cyclic scan eviction; every slot above it is
/// still borrowed and evicted globally, so skewed layers use otherwise idle
/// capacity.
#[cfg(feature = "gpu")]
struct GlobalPool {
    uid: u64,
    capacity: usize,
    segment_slots: usize,
    floor: usize,
    state: std::sync::Mutex<GlobalPoolDyn>,
}

#[cfg(feature = "gpu")]
impl GlobalPool {
    fn remap(&self, layer: usize, n: usize) -> Vec<u32> {
        let st = self.state.lock().unwrap();
        debug_assert_eq!(self.capacity, st.owner.len());
        (0..n)
            .map(|e| st.slot_for.get(&(layer, e)).copied().unwrap_or(u32::MAX))
            .collect()
    }

    fn seed_layer(
        &self,
        model: &std::sync::Arc<cortiq_core::CmfModel>,
        layer: usize,
        shared: (usize, usize, usize),
        _routed: &[(usize, (usize, usize, usize))],
    ) -> Option<u32> {
        let mut st = self.state.lock().unwrap();
        let shared_key = (layer, usize::MAX);
        let shared_slot = if let Some(&slot) = st.slot_for.get(&shared_key) {
            slot
        } else {
            let slot = st.owner.iter().position(Option::is_none)?;
            if !crate::gpu_wgpu::dsv4_global_slot_fill(model, slot, shared) {
                return None;
            }
            st.owner[slot] = Some(GlobalOwner {
                layer,
                expert: usize::MAX,
                pinned: true,
            });
            st.slot_for.insert(shared_key, slot as u32);
            *st.occupancy.entry(layer).or_insert(0) += 1;
            slot as u32
        };
        // Do not fill the remaining 40+ GiB by expert id. Route locality is
        // checkpoint- and prompt-dependent; the old eager seed spent a
        // minute uploading rows that the first token immediately replaced.
        // Empty slots are populated by the real top-k before its dispatch.
        Some(shared_slot)
    }

    fn ensure_picks(
        &self,
        model: &std::sync::Arc<cortiq_core::CmfModel>,
        layer: usize,
        picks: &[usize],
        experts: &[Dsv4Expert],
        quota: usize,
        min_seen: u16,
    ) -> Vec<u32> {
        let mut st = self.state.lock().unwrap();
        st.clock = st.clock.saturating_add(1);
        let clock = st.clock;
        if clock % 64 == 0 {
            for v in st.seen.values_mut() {
                *v >>= 1;
            }
        }
        for &expert in picks {
            let key = (layer, expert);
            let seen = st.seen.entry(key).or_insert(0);
            *seen = seen.saturating_add(1);
            if let Some(&slot) = st.slot_for.get(&key) {
                st.last[slot as usize] = clock;
            }
        }
        let mut fetched = 0usize;
        for &expert in picks {
            if fetched >= quota {
                break;
            }
            let key = (layer, expert);
            if st.slot_for.contains_key(&key) || st.seen.get(&key).copied().unwrap_or(0) < min_seen
            {
                continue;
            }
            let Some(exp) = experts.get(expert) else {
                continue;
            };
            let Some(tensors) = (|| {
                Some((
                    exp.w1.model_idx()?,
                    exp.w3.model_idx()?,
                    exp.w2.model_idx()?,
                ))
            })() else {
                continue;
            };
            let empty = st.owner.iter().position(Option::is_none);
            // First evict from a layer that currently borrows above its
            // floor. Do not evict any expert required by this very dispatch.
            let over_floor = |o: GlobalOwner, occ: &std::collections::HashMap<usize, usize>| {
                occ.get(&o.layer).copied().unwrap_or(0) > self.floor
            };
            let eligible =
                |o: GlobalOwner| !o.pinned && !(o.layer == layer && picks.contains(&o.expert));
            let victim = empty
                .or_else(|| {
                    st.owner
                        .iter()
                        .enumerate()
                        .filter_map(|(slot, &o)| {
                            o.filter(|&x| eligible(x) && over_floor(x, &st.occupancy))
                                .map(|_| slot)
                        })
                        .min_by_key(|&slot| st.last[slot])
                })
                .or_else(|| {
                    // If every layer sits exactly at its floor, replace within
                    // the requesting layer. This is per-layer LRU behaviour and
                    // cannot trigger the cyclic global scan collapse.
                    st.owner
                        .iter()
                        .enumerate()
                        .filter_map(|(slot, &o)| {
                            o.filter(|&x| eligible(x) && x.layer == layer).map(|_| slot)
                        })
                        .min_by_key(|&slot| st.last[slot])
                })
                .or_else(|| {
                    // A new/MTP layer has no protected share yet. Let it borrow
                    // the globally oldest unpinned line; trunk floors are a
                    // locality guarantee, not a permanent admission ban.
                    st.owner
                        .iter()
                        .enumerate()
                        .filter_map(|(slot, &o)| o.filter(|&x| eligible(x)).map(|_| slot))
                        .min_by_key(|&slot| st.last[slot])
                });
            let Some(victim) = victim else { break };
            if !crate::gpu_wgpu::dsv4_global_slot_fill(model, victim, tensors) {
                break;
            }
            if let Some(old) = st.owner[victim] {
                st.slot_for.remove(&(old.layer, old.expert));
                if let Some(n) = st.occupancy.get_mut(&old.layer) {
                    *n = n.saturating_sub(1);
                }
            }
            st.owner[victim] = Some(GlobalOwner {
                layer,
                expert,
                pinned: false,
            });
            st.slot_for.insert(key, victim as u32);
            *st.occupancy.entry(layer).or_insert(0) += 1;
            st.last[victim] = clock;
            fetched += 1;
        }
        drop(st);
        self.remap(layer, experts.len())
    }

    /// Reserve a small immutable slice of the unified arena for the draft.
    /// The draft graph cannot pause between its dependent stages to repair a
    /// slot that the exact trunk evicted, so its bounded resident subset is
    /// pinned.  These are still the SAME physical banks: no second expert
    /// allocation and no duplicate upload cache are created.
    fn pin_picks(
        &self,
        model: &std::sync::Arc<cortiq_core::CmfModel>,
        layer: usize,
        picks: &[usize],
        experts: &[Dsv4Expert],
    ) -> Vec<u32> {
        let remap = self.ensure_picks(model, layer, picks, experts, picks.len(), 1);
        let mut st = self.state.lock().unwrap();
        for &expert in picks {
            if let Some(&slot) = st.slot_for.get(&(layer, expert)) {
                if let Some(owner) = st.owner[slot as usize].as_mut() {
                    owner.pinned = true;
                }
            }
        }
        remap
    }
}

#[cfg(feature = "gpu")]
#[derive(Clone)]
struct GlobalPack {
    pool: std::sync::Arc<GlobalPool>,
    layer: usize,
    shared_slot: u32,
}

#[cfg(feature = "gpu")]
impl Pack {
    /// True once any slot was refilled away from the build-time packing.
    fn is_mutated(&self) -> bool {
        self.global.is_some() || self.dynslots.lock().unwrap().mutated
    }

    /// Complete means complete for the route the model will actually take.
    /// A task mask can close most of the 256 rows; packing every OPEN row is
    /// then a full device layer, not a partial layer with 200 imaginary cold
    /// experts. Hash layers deliberately carry no mask because their forced
    /// rows remain the exact checkpoint contract.
    fn route_complete(&self) -> bool {
        if self.global.is_some() {
            return false;
        }
        let need = self
            .mask
            .as_deref()
            .map_or(self.remap.len(), |m| m.iter().filter(|&&x| x != 0).count());
        self.globals.len() >= need
    }

    /// A global-to-slot table is needed for a masked set and for a complete
    /// pack whose hot-first order is not identity. Without it a full but
    /// reordered pack silently runs the right router index on the wrong bank.
    fn needs_remap(&self) -> bool {
        self.global.is_some()
            || self.mask.is_some()
            || self
                .remap
                .iter()
                .enumerate()
                .any(|(i, &slot)| slot != i as u32)
    }
}

/// The packed expert set of one layer: which globals made it in, and their
/// directory indices in packing order with the shared expert last. Keyed by
/// layer; the STATIC fields are built once, the dynamic slot state evolves.
#[cfg(feature = "gpu")]
struct Pack {
    /// The router as dense f32, expanded once. It is 4 MB a layer against a
    /// 112 GB model, it lives as long as the process — so the address-keyed
    /// device cache is sound for it, unlike anything built per call.
    router: Vec<f32>,
    /// global expert id -> packed slot, `usize::MAX` for the ones left out.
    to_slot: Vec<usize>,
    /// The same, as the u32 table the router reads.
    remap: Vec<u32>,
    /// packed order, globals only (shared is not in here).
    globals: Vec<usize>,
    tensors: Vec<(usize, usize, usize)>,
    /// Global 0/1 route mask consumed by the GPU router. Stable storage is
    /// part of the pack because the device constant cache keys by address.
    /// None on exact/hash routing.
    mask: Option<Vec<u32>>,
    /// FreeToken-style dynamic slots: the packed subset FOLLOWS the router
    /// instead of staying whatever load-time frequency guessed. `remap` here
    /// is the LIVE table (the immutable `remap` above is the initial state
    /// and stays only as the build artifact); `owner[slot]` is the global
    /// expert id occupying the slot; `last[slot]`/`clock` drive LRU. The
    /// device bank buffers accept `write_buffer` at slot offsets, and the
    /// frame re-uploads the remap every call — so a refill is two queue
    /// writes and no cache invalidation anywhere.
    dynslots: std::sync::Mutex<PackDyn>,
    /// The noaux_tc bias in GLOBAL order, kept here because it is the same
    /// every token and the pack lives as long as the process. Global order is
    /// required by masked/remapped routing; its stable address lets many
    /// layers share one submission. A bias uploaded through
    /// the per-call pool is written by every layer of a run BEFORE the run's
    /// single submit — queue writes do not interleave with passes — so every
    /// layer routed with the LAST layer's bias. On the release every scored
    /// layer carries one, which is the 50.280.
    bias: Option<Vec<f32>>,
    /// Present only on the descriptor-indexed Q4TP path. `remap` above is
    /// merely the build-time snapshot there; every frame takes a fresh map
    /// from this common allocator after its refills/evictions.
    global: Option<GlobalPack>,
}

#[cfg(feature = "gpu")]
/// Candidate order for a budget-limited pack: hottest expert first, by the
/// measured tally `CMF_DSV4_PACK_FREQ` points at (`layer<TAB>expert<TAB>count`
/// lines). None when the variable is unset, the file is unreadable, or the
/// tally has nothing for this layer — the caller keeps id order then. Ties
/// and untallied experts follow in id order, so the choice is deterministic.
fn pack_freq_order(li: usize, n: usize) -> Option<Vec<usize>> {
    use std::collections::HashMap;
    use std::sync::OnceLock;
    static FREQ: OnceLock<Option<HashMap<(usize, usize), u64>>> = OnceLock::new();
    let map = FREQ
        .get_or_init(|| {
            let path = std::env::var("CMF_DSV4_PACK_FREQ").ok()?;
            let text = match std::fs::read_to_string(&path) {
                Ok(t) => t,
                Err(e) => {
                    eprintln!("CMF_DSV4_PACK_FREQ={path} не читается ({e}) — порядок по id");
                    return None;
                }
            };
            let mut m = HashMap::new();
            for line in text.lines() {
                let mut it = line.split('\t');
                if let (Some(l), Some(e), Some(c)) = (it.next(), it.next(), it.next()) {
                    if let (Ok(l), Ok(e), Ok(c)) =
                        (l.trim().parse(), e.trim().parse(), c.trim().parse::<u64>())
                    {
                        *m.entry((l, e)).or_insert(0) += c;
                    }
                }
            }
            Some(m)
        })
        .as_ref()?;
    if !(0..n).any(|e| map.contains_key(&(li, e))) {
        return None;
    }
    let mut idx: Vec<usize> = (0..n).collect();
    idx.sort_by_key(|&e| {
        (
            std::cmp::Reverse(map.get(&(li, e)).copied().unwrap_or(0)),
            e,
        )
    });
    Some(idx)
}

#[cfg(feature = "gpu")]
fn global_pool_for(
    model: &std::sync::Arc<cortiq_core::CmfModel>,
    cfg: &Dsv4Cfg,
    gu_q2: bool,
    dn_q2: bool,
) -> Option<std::sync::Arc<GlobalPool>> {
    use std::collections::HashMap;
    use std::sync::{Arc, Mutex, OnceLock};
    if dn_q2
        || !crate::gpu_wgpu::dsv4_global_moe_supported()
        || std::env::var("CMF_MOE_MASK").is_ok()
    {
        return None;
    }
    static POOLS: OnceLock<Mutex<HashMap<u64, Arc<GlobalPool>>>> = OnceLock::new();
    let pools = POOLS.get_or_init(|| Mutex::new(HashMap::new()));
    if let Some(p) = pools.lock().unwrap().get(&model.uid()).cloned() {
        return Some(p);
    }
    let requested = crate::gpu_wgpu::dsv4_experts_fit(cfg.moe_inter, cfg.dim, gu_q2, false);
    let (capacity, segment_slots) =
        crate::gpu_wgpu::dsv4_global_moe_create(model, requested, cfg.moe_inter, cfg.dim, gu_q2)?;
    let layers = model.header.arch.num_layers.max(1);
    let p = Arc::new(GlobalPool {
        uid: model.uid(),
        capacity,
        segment_slots,
        // At least shared + one routed line remain protected per layer when
        // geometry permits it. Larger cards naturally raise the floor.
        floor: (capacity / layers).max(2),
        state: Mutex::new(GlobalPoolDyn {
            slot_for: HashMap::new(),
            owner: vec![None; capacity],
            last: vec![0; capacity],
            seen: HashMap::new(),
            occupancy: HashMap::new(),
            clock: 0,
        }),
    });
    pools.lock().unwrap().insert(model.uid(), p.clone());
    Some(p)
}

#[cfg(feature = "gpu")]
fn pack_for(l: &Dsv4Layer, cfg: &Dsv4Cfg, li: usize) -> Option<std::sync::Arc<Pack>> {
    use std::collections::HashMap;
    use std::sync::{Arc, Mutex, OnceLock};
    static CACHE: OnceLock<Mutex<HashMap<(u64, usize, usize), Option<Arc<Pack>>>>> =
        OnceLock::new();
    let cache = CACHE.get_or_init(|| Mutex::new(HashMap::new()));
    // Keyed by the layer's IDENTITY, not its ordinal. The draft's three
    // stages are layers too and they number 0, 1, 2 — under an ordinal key
    // they would be handed the trunk's first three packs: another layer's
    // router, another layer's tensor indices, another layer's bias. The gate
    // tensor is what actually distinguishes them.
    let model_uid = l
        .experts
        .first()
        .and_then(|e| e.w1.model_arc())
        .map_or(0, |m| m.uid());
    // Dense f32 routers need not have a directory handle, so the gate index
    // alone can be `None` for every layer. Pair the ordinal with the first
    // expert's mapped identity; model UID keeps long-lived multi-model
    // servers separate, while the expert index distinguishes trunk and MTP
    // layers that reuse ordinal 0/1/2.
    let first_expert = l
        .experts
        .first()
        .and_then(|e| e.w1.model_idx())
        .unwrap_or(usize::MAX);
    let key = (model_uid, li, first_expert);
    if let Some(v) = cache.lock().unwrap().get(&key) {
        return v.clone();
    }
    // `CMF_DSV4_PACK_MAX_LI=N` — do not pack layers above N at all. A layer
    // with no pack stays wholly host-owned, which is what both the batched
    // prefill and a speculative verify need of the tail: a device-owned
    // partial layer can join neither the batch (incomplete pack) nor the
    // causal host tail (its caches live on the card). This also carves the
    // VRAM the tail would have taken for the draft's own pack.
    if let Ok(v) = std::env::var("CMF_DSV4_PACK_MAX_LI") {
        if v.parse::<usize>().is_ok_and(|max| li > max) {
            cache.lock().unwrap().insert(key, None);
            return None;
        }
    }
    let build = || -> Option<Arc<Pack>> {
        let mut to_slot = vec![usize::MAX; cfg.n_routed_experts];
        let mut globals = Vec::new();
        let mut tensors = Vec::new();
        let idx3 = |e: &Dsv4Expert| -> Option<(usize, usize, usize)> {
            Some((e.w1.model_idx()?, e.w3.model_idx()?, e.w2.model_idx()?))
        };
        let route_mask: Option<Vec<u32>> = if l.tid2eid.is_none() {
            l.mask
                .as_deref()
                .map(|m| m.iter().map(|&open| u32::from(open)).collect())
        } else {
            None
        };
        // How many experts the card still has room for, minus one for the
        // shared expert, which always rides. Everything past that stays on the
        // host and is reached through the remap — the router still ranges over
        // all of them, so this costs speed and not a single bit of quality.
        let gu_q2 = l
            .experts
            .first()
            .is_some_and(|e| e.w1.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP));
        let dn_q2 = l
            .experts
            .first()
            .is_some_and(|e| e.w2.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP));
        // Exact default on Q4TP: one physical/logical cache for every
        // `(layer, expert)` pair. Explicit mask experiments keep the old
        // local path so this branch never combines two independent changes
        // to model semantics.
        if route_mask.is_none() {
            if let Some(model) = l.experts.first().and_then(|e| e.w1.model_arc()) {
                if let Some(gp) = global_pool_for(&model, cfg, gu_q2, dn_q2) {
                    let layer_key = first_expert;
                    let order = pack_freq_order(li, l.experts.len())
                        .unwrap_or_else(|| (0..l.experts.len()).collect());
                    let routed: Vec<_> = order
                        .into_iter()
                        .filter_map(|gi| Some((gi, idx3(&l.experts[gi])?)))
                        .collect();
                    let shared = idx3(&l.shared)?;
                    let shared_slot = gp.seed_layer(&model, layer_key, shared, &routed)?;
                    let remap = gp.remap(layer_key, cfg.n_routed_experts);
                    let to_slot: Vec<usize> = remap
                        .iter()
                        .map(|&s| {
                            if s == u32::MAX {
                                usize::MAX
                            } else {
                                s as usize
                            }
                        })
                        .collect();
                    let globals: Vec<usize> = remap
                        .iter()
                        .enumerate()
                        .filter_map(|(e, &s)| (s != u32::MAX).then_some(e))
                        .collect();
                    let (rows, cols) = (l.gate.rows(), l.gate.cols());
                    let mut router = vec![0.0f32; rows * cols];
                    for r in 0..rows {
                        l.gate.row_f32(r, &mut router[r * cols..(r + 1) * cols]);
                    }
                    return Some(Arc::new(Pack {
                        bias: l.gate_bias.clone(),
                        mask: None,
                        router,
                        to_slot,
                        remap: remap.clone(),
                        globals,
                        // Physical tensors live in the global GPU bank map,
                        // not in a second layer-local concatenation.
                        tensors: Vec::new(),
                        dynslots: std::sync::Mutex::new(PackDyn {
                            remap,
                            owner: Vec::new(),
                            last: Vec::new(),
                            clock: 0,
                            mutated: true,
                            seen: vec![0; cfg.n_routed_experts],
                        }),
                        global: Some(GlobalPack {
                            pool: gp,
                            layer: layer_key,
                            shared_slot,
                        }),
                    }));
                }
            }
        }
        // Pack what fits and leave the rest to the host. The router still
        // ranges over every expert; a missing winner is returned as a cold
        // pick and completed on the CPU. This is deliberately budget-driven,
        // not layer-driven: the same model scales from a small card (more
        // partial/host layers) to a large one (all experts resident) without
        // a checkpoint-specific cutoff.
        // `CMF_DSV4_PACK_MAX=N` caps the packing directly, so a toy can
        // reproduce the subset path without needing a card that runs out.
        if let Some(n) = std::env::var("CMF_DSV4_PACK_MAX")
            .ok()
            .and_then(|v| v.parse::<usize>().ok())
        {
            let mut to_slot = vec![usize::MAX; cfg.n_routed_experts];
            let mut globals = Vec::new();
            let mut tensors = Vec::new();
            for (gi, e) in l.experts.iter().enumerate() {
                if route_mask
                    .as_deref()
                    .is_some_and(|m| m.get(gi).copied().unwrap_or(1) == 0)
                {
                    continue;
                }
                if globals.len() >= n {
                    break;
                }
                to_slot[gi] = globals.len();
                globals.push(gi);
                tensors.push(idx3(e)?);
            }
            tensors.push(idx3(&l.shared)?);
            let (rows, cols) = (l.gate.rows(), l.gate.cols());
            let mut router = vec![0.0f32; rows * cols];
            for r in 0..rows {
                l.gate.row_f32(r, &mut router[r * cols..(r + 1) * cols]);
            }
            let remap: Vec<u32> = to_slot
                .iter()
                .map(|&sl| {
                    if sl == usize::MAX {
                        u32::MAX
                    } else {
                        sl as u32
                    }
                })
                .collect();
            return Some(Arc::new(Pack {
                bias: l.gate_bias.clone(),
                mask: route_mask.clone(),
                router,
                to_slot,
                dynslots: std::sync::Mutex::new(PackDyn {
                    remap: remap.clone(),
                    owner: globals.iter().map(|&g| g as u32).collect(),
                    last: vec![0; globals.len()],
                    clock: 0,
                    mutated: false,
                    seen: vec![0; cfg.n_routed_experts],
                }),
                remap,
                globals,
                tensors,
                global: None,
            }));
        }
        let dn_q2_fit = l
            .experts
            .first()
            .is_some_and(|e| e.w2.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP));
        let room = crate::gpu_wgpu::dsv4_experts_fit(cfg.moe_inter, cfg.dim, gu_q2, dn_q2_fit)
            .saturating_sub(1);
        // A greedy pack starves the tail: the first layers take the whole
        // expert budget and the tail falls off the device chain. Divide the
        // room still available by the layers still to pack. This depends on
        // format and VRAM geometry, not a card name: a small card gives every
        // layer a useful partial pack; when the whole model fits the quotient
        // naturally reaches all experts. Hash-routed head layers stay whole
        // whenever possible because their checkpoint table names exact rows.
        //
        // CMF_DSV4_PACK_LAYER_CAP=N remains an override; 0 restores the old
        // greedy packing for a performance bisect.
        let remaining_layers = l
            .experts
            .first()
            .and_then(|e| e.w1.model_arc())
            .map(|m| m.header.arch.num_layers.saturating_sub(li).max(1))
            .unwrap_or(1);
        let auto_cap = if l.mask.is_some() {
            // The mask is already the layer-specific cap. Its total mass was
            // counted by dspark_reserve_note before packing, so an additional
            // equal-per-layer cap only turns naturally uneven masked layers
            // partial and makes batched verification reject every draft.
            room
        } else if l.tid2eid.is_some() && room >= cfg.n_routed_experts {
            cfg.n_routed_experts
        } else {
            room.div_ceil(remaining_layers).max(1)
        };
        let room = match std::env::var("CMF_DSV4_PACK_LAYER_CAP")
            .ok()
            .and_then(|v| v.parse::<usize>().ok())
        {
            Some(0) => room,
            Some(cap) => room.min(cap),
            None => room.min(auto_cap),
        };
        // When the budget packs a SUBSET, which subset matters: a partial
        // layer completes its cold picks from the host, so every resident
        // expert that the routing actually reaches is host work saved.
        // `CMF_DSV4_PACK_FREQ` names a measured tally
        // (`CMF_DSV4_TRUNK_PICK_DUMP` wrote it) and reorders the candidates
        // hottest-first; layers absent from the tally keep id order. The
        // router still ranges over every expert either way — residency
        // choice changes speed, never a bit of the answer.
        let order =
            pack_freq_order(li, l.experts.len()).unwrap_or_else(|| (0..l.experts.len()).collect());
        for gi in order {
            let e = &l.experts[gi];
            if l.mask
                .as_deref()
                .is_some_and(|m| !m.get(gi).copied().unwrap_or(true))
            {
                continue;
            }
            if globals.len() >= room {
                break;
            }
            to_slot[gi] = globals.len();
            globals.push(gi);
            match idx3(e) {
                Some(t) => tensors.push(t),
                None => {
                    if std::env::var("CMF_DSV4_FRAME_DEBUG").is_ok() {
                        eprintln!("слой {li}: эксперт {gi} без индексов в каталоге");
                    }
                    return None;
                }
            }
        }
        if globals.is_empty() {
            // Two very different causes, and blaming the mask for the other
            // one sent a reader looking for a mask that was never set: an
            // actual empty mask, or a VRAM budget with no room left for even
            // one expert (`room` is 0, which is what a nearly-full card does
            // to the last layers).
            if room == 0 {
                static SAID_ZERO: std::sync::atomic::AtomicBool =
                    std::sync::atomic::AtomicBool::new(false);
                if !SAID_ZERO.swap(true, std::sync::atomic::Ordering::Relaxed) {
                    tracing::warn!(
                        "начиная со слоя {li}, в бюджете VRAM не осталось места даже под одного \
                         эксперта — остальные веса остаются mmap-backed и читаются по требованию"
                    );
                }
            } else {
                tracing::warn!("слой {li}: маска не оставила ни одного эксперта");
            }
            return None;
        }
        tensors.push(idx3(&l.shared)?); // shared rides last, as the kernels expect
        let (rows, cols) = (l.gate.rows(), l.gate.cols());
        let mut router = vec![0.0f32; rows * cols];
        for r in 0..rows {
            l.gate.row_f32(r, &mut router[r * cols..(r + 1) * cols]);
        }
        let remap: Vec<u32> = to_slot
            .iter()
            .map(|&sl| {
                if sl == usize::MAX {
                    u32::MAX
                } else {
                    sl as u32
                }
            })
            .collect();
        Some(Arc::new(Pack {
            bias: l.gate_bias.clone(),
            mask: route_mask,
            router,
            to_slot,
            dynslots: std::sync::Mutex::new(PackDyn {
                remap: remap.clone(),
                owner: globals.iter().map(|&g| g as u32).collect(),
                last: vec![0; globals.len()],
                clock: 0,
                mutated: false,
                seen: vec![0; cfg.n_routed_experts],
            }),
            remap,
            globals,
            tensors,
            global: None,
        }))
    };
    let v = build();
    cache.lock().unwrap().insert(key, v.clone());
    v
}

/// The whole MoE block in one submission, experts resident (default on;
/// `CMF_DSV4_GPU_MOE2=0` restores the host path). Returns false having
/// changed nothing if it cannot — a missing pack, a refused budget — so the
/// caller's CPU path stays correct to run. The early divergence this frame
/// once carried (0.44 relative, perplexity 5.162 vs 5.211) was the partial
/// -capture and hidden-seed defects, fixed since: perplexity gold 4.578 is
/// bit-exact against the CPU on every budget from 64 to 96.5 GB.
#[cfg(feature = "gpu")]
fn moe_frame(
    hidden: &[f32],
    l: &Dsv4Layer,
    cfg: &Dsv4Cfg,
    li: usize,
    logits: &[f32],
    forced: Option<&[usize]>,
    pool: Option<&crate::pool::Pool>,
    // The state handover: expand always when the device owns the state,
    // fold only when there is a next layer.
    hc_cur: Option<&crate::gpu_wgpu::Dsv4HcTail>,
    hc_next: Option<(&crate::gpu_wgpu::Dsv4HcTail, &[f32])>,
    out: &mut [f32],
) -> Option<(Vec<f32>, usize)> {
    macro_rules! no {
        ($($t:tt)*) => {{
            if std::env::var("CMF_DSV4_FRAME_DEBUG").is_ok() {
                eprintln!("кадр MoE отклонён: {}", format_args!($($t)*));
            }
            return None;
        }};
    }
    let Some(pk) = pack_for(l, cfg, li) else {
        no!("слой {li}: упаковка экспертов не построена");
    };
    // The router is a small f32 tensor and is usually NOT mapped; the handle
    // has to come from something that is.
    let Some(model) = l.experts.first().and_then(|e| e.w1.model_arc()) else {
        no!("слой {li}: эксперты не отображены из файла");
    };
    let subset = !pk.route_complete();
    let needs_remap = pk.needs_remap();
    // Dynamic slots (the FreeToken move): predict this token's winners on
    // the host and pull the missing ones into LRU slots BEFORE the frame
    // runs — up to CMF_DSV4_FETCH_MAX experts a layer a token. The device
    // still routes for real, so a wrong prediction costs one unused fill
    // and never a wrong number: an unmapped winner comes back as a cold
    // pick and the CPU completes it, exactly as before.
    fn fetch_quota() -> usize {
        static Q: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
        *Q.get_or_init(|| {
            std::env::var("CMF_DSV4_FETCH_MAX")
                .ok()
                .and_then(|v| v.parse().ok())
                .unwrap_or_else(|| crate::gpu_wgpu::dsv4_fetch_defaults().0)
        })
    }
    fn fetch_min_seen() -> u16 {
        static M: std::sync::OnceLock<u16> = std::sync::OnceLock::new();
        *M.get_or_init(|| {
            std::env::var("CMF_DSV4_FETCH_MIN_SEEN")
                .ok()
                .and_then(|v| v.parse().ok())
                .unwrap_or_else(|| crate::gpu_wgpu::dsv4_fetch_defaults().1)
        })
    }
    let mut dynv = pk.dynslots.lock().unwrap();
    // The winners, from the same logits the device will rank — used by the
    // slot refill below AND by the FreeToken-style overlap: the picks that
    // will NOT be resident are computed on the CPU while the device frame
    // runs, instead of serially after its wait.
    let mut pidx = Vec::new();
    let mut pwt = Vec::new();
    // A chained caller may leave the FFN input only on the device. The
    // overlap path intentionally reads it back; when it does, score a HOST
    // prediction without changing the device's routing source. The GPU still
    // recomputes and ranks its own logits when `logits` is empty, so a rounding
    // disagreement can waste an early CPU result but cannot change a token.
    let mut predicted_logits = Vec::new();
    let prediction_logits: &[f32] = if subset && logits.is_empty() && !hidden.is_empty() {
        predicted_logits.resize(cfg.n_routed_experts, 0.0);
        l.gate.matvec(hidden, &mut predicted_logits, pool);
        &predicted_logits
    } else {
        logits
    };
    if subset && !prediction_logits.is_empty() {
        route(
            prediction_logits,
            l.gate_bias.as_deref(),
            cfg.top_k,
            cfg.route_scale,
            forced,
            l.mask.as_deref(),
            &mut pidx,
            &mut pwt,
        );
    }
    let global_remap = pk.global.as_ref().map(|gl| {
        gl.pool.ensure_picks(
            &model, gl.layer, &pidx, &l.experts,
            // A global demand cache has empty lines during warmup and each
            // CPU cold completion costs far more than one expert upload.
            // Materialise every predicted winner immediately, FreeToken
            // style; device routing still returns any prediction drift as an
            // exact cold pick.
            cfg.top_k, 1,
        )
    });
    if subset && fetch_quota() > 0 && !pidx.is_empty() && !dynv.owner.is_empty() {
        dynv.clock += 1;
        let clock = dynv.clock;
        if clock % 64 == 0 {
            for v in dynv.seen.iter_mut() {
                *v >>= 1;
            }
        }
        for &pick in &pidx {
            dynv.seen[pick] = dynv.seen[pick].saturating_add(1);
            let sl = dynv.remap[pick];
            if sl != u32::MAX {
                dynv.last[sl as usize] = clock;
            }
        }
        let mut fetched = 0usize;
        for &pick in &pidx {
            if fetched >= fetch_quota() {
                break;
            }
            if dynv.remap[pick] != u32::MAX {
                continue;
            }
            if dynv.seen[pick] < fetch_min_seen() {
                continue; // one-shot so far: the CPU reads it at the shelf
            }
            // Victim: the LRU slot among those this token does not need.
            let victim = (0..dynv.owner.len())
                .filter(|&sl| dynv.last[sl] != clock)
                .min_by_key(|&sl| dynv.last[sl]);
            let Some(victim) = victim else { break };
            let Some(exp) = l.experts.get(pick) else {
                continue;
            };
            let t3 = (|| {
                Some((
                    exp.w1.model_idx()?,
                    exp.w3.model_idx()?,
                    exp.w2.model_idx()?,
                ))
            })();
            let Some(t3) = t3 else { continue };
            let gu_q2 = exp.w1.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP);
            let pack_first = pk.tensors.first().map(|t| t.0).unwrap_or(usize::MAX);
            if !crate::gpu_wgpu::dsv4_slot_fill(
                &model,
                pack_first,
                victim,
                pick,
                t3,
                cfg.moe_inter,
                cfg.dim,
                gu_q2,
            ) {
                break;
            }
            let old = dynv.owner[victim] as usize;
            if old < dynv.remap.len() {
                dynv.remap[old] = u32::MAX;
            }
            dynv.remap[pick] = victim as u32;
            dynv.owner[victim] = pick as u32;
            dynv.last[victim] = clock;
            dynv.mutated = true;
            fetched += 1;
        }
    }
    // With a complete pack the forced row is translated to packed numbering.
    // With a subset it stays global: the router's remap either finds its slot
    // or returns the forced expert as a cold pick, exactly like a scored one.
    let fpack: Option<Vec<usize>> = match forced {
        Some(f) if needs_remap => Some(f.to_vec()),
        Some(f) => {
            let v: Vec<usize> = f.iter().map(|&g| pk.to_slot[g]).collect();
            if v.contains(&usize::MAX) {
                no!("слой {li}: хеш-слой называет эксперта вне упаковки");
            }
            Some(v)
        }
        None => None,
    };
    // Routing ranges over EVERY expert; the remap turns a winner into a slot
    // or marks it cold. Nothing is masked, so nothing is lost.
    // Empty logits are the device-scored case: the frame computes them from
    // pk.router, whose rows are already in global order, so there is nothing
    // to reorder — and indexing an empty slice is how this line greeted the
    // first engaged run.
    let lg: Vec<f32> = if logits.is_empty() || needs_remap {
        logits.to_vec()
    } else {
        pk.globals.iter().map(|&g| logits[g]).collect()
    };
    let live_remap = global_remap.as_deref().unwrap_or(dynv.remap.as_slice());
    let w = crate::gpu_wgpu::Dsv4MoeW {
        router: &pk.router,
        experts: &pk.tensors,
        logits: &lg,
        bias: pk.bias.as_deref(),
        mask: pk.mask.as_deref(),
        forced: fpack.as_deref(),
        remap: needs_remap.then_some(live_remap),
        global: pk.global.as_ref().map(|gl| crate::gpu_wgpu::Dsv4GlobalMoe {
            pool_uid: gl.pool.uid,
            shared_slot: gl.shared_slot,
            segment_slots: gl.pool.segment_slots as u32,
        }),
        has_shared: true,
        shared_weight: 1.0,
        preweighted: false,
        qwen_softmax: false,
    };
    let g = crate::gpu_wgpu::Dsv4MoeGeom {
        hidden: cfg.dim,
        inter: cfg.moe_inter,
        top_k: cfg.top_k,
        route_scale: cfg.route_scale,
        swiglu_limit: cfg.swiglu_limit,
        gu_q2: l
            .experts
            .first()
            .is_some_and(|e| e.w1.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP)),
    };
    let mut cold = Vec::new();
    let mut cold_x = Vec::new();
    // The FreeToken overlap: the predicted winners that will NOT be
    // resident are computed on the CPU WHILE the device frame runs,
    // instead of serially after its wait. Unweighted (weight 1) — the
    // device's own cold weights scale the result at the merge, so a
    // routing drift between the host's ranking and the card's costs one
    // wasted thread, never a wrong number. Only the per-layer path has
    // the input on the host (`hidden` non-empty); the chain keeps its
    // own economy.
    let overlap: Vec<usize> = if !hidden.is_empty() {
        pidx.iter()
            .copied()
            .filter(|&pick| live_remap.get(pick).copied().unwrap_or(u32::MAX) == u32::MAX)
            .collect()
    } else {
        Vec::new()
    };
    let mut early: std::collections::HashMap<usize, Vec<f32>> = std::collections::HashMap::new();
    let frame_ok = std::thread::scope(|sc| {
        let handles: Vec<_> = overlap
            .iter()
            .filter_map(|&gi| l.experts.get(gi).map(|exp| (gi, exp)))
            .map(|(gi, exp)| {
                sc.spawn(move || {
                    let mut a = vec![0.0f32; cfg.dim];
                    crate::gpu::cpu_scope(|| run_expert(hidden, exp, cfg, 1.0, None, &mut a));
                    (gi, a)
                })
            })
            .collect();
        let ok = crate::gpu_wgpu::dsv4_moe_frame(
            &model,
            &w,
            g,
            hidden,
            &mut cold,
            &mut cold_x,
            hc_cur,
            hc_next,
            out,
        );
        for h in handles {
            if let Ok((gi, a)) = h.join() {
                early.insert(gi, a);
            }
        }
        ok
    });
    if !frame_ok {
        return None;
    }
    // The picks the card had no room for, finished here and added in. Their
    // weights already carry the top-k normalisation the device applied.
    if std::env::var("CMF_DSV4_MOE_CHECK").is_ok() {
        let csum: f32 = cold.iter().map(|c| c.1).sum();
        eprintln!(
            "[холодные] слой {li}: вернулось {} из {} | сумма холодных {csum:.4} | \
             route_scale {:.4} | {:?}",
            cold.len(),
            cfg.top_k,
            cfg.route_scale,
            &cold[..cold.len().min(3)]
        );
    }
    let mut acc = vec![0.0f32; cfg.dim];
    let mut cold_sum = vec![0.0f32; cfg.dim];
    let cold_input = if hidden.is_empty() {
        cold_x.as_slice()
    } else {
        hidden
    };
    // Cold means out-of-core by contract. The tensors remain mmap-backed:
    // missing pages are faulted from the CMF file and the OS may evict
    // them again under RAM pressure. Do not let the generic matvec probe
    // turn this into an unbounded second GPU cache behind the packer's
    // back.
    //
    // The unit of parallelism is the EXPERT, not the row: a 2048-row
    // matvec split across 380 workers is five rows per worker — all
    // dispatch, no arithmetic. One worker per cold expert, whole matvecs
    // inside (inner pool None), was the difference between ~7 ms and ~1 ms
    // per cold expert on the 384-core stand. cpu_scope is thread-local, so
    // it sits INSIDE the worker closure.
    if !early.is_empty() {
        // The overlap already computed (most of) the cold picks; scale by
        // the DEVICE's weight and add in cold order — the same order the
        // serial path used, so parity holds. A cold pick the prediction
        // missed (ranking drift) is computed inline, cpu_scope'd.
        for &(gi, wt) in &cold {
            if let Some(a) = early.get(&gi) {
                for ((o, sum), v) in out.iter_mut().zip(&mut cold_sum).zip(a.iter()) {
                    *o += v * wt;
                    *sum += v * wt;
                }
                continue;
            }
            let Some(exp) = l.experts.get(gi) else {
                continue;
            };
            crate::gpu::cpu_scope(|| run_expert(cold_input, exp, cfg, wt, pool, &mut acc));
            for ((o, sum), a) in out.iter_mut().zip(&mut cold_sum).zip(&acc) {
                *o += a;
                *sum += a;
            }
        }
        return Some((cold_sum, cold.len()));
    }
    match pool {
        Some(p) if cold.len() > 1 => {
            let results: Vec<std::sync::Mutex<Vec<f32>>> = cold
                .iter()
                .map(|_| std::sync::Mutex::new(Vec::new()))
                .collect();
            let (cold_ref, results_ref) = (&cold, &results);
            p.run_rows(cold.len(), &move |cs, ce| {
                for i in cs..ce {
                    let (gi, wt) = cold_ref[i];
                    let Some(exp) = l.experts.get(gi) else {
                        continue;
                    };
                    let mut a = vec![0.0f32; cfg.dim];
                    crate::gpu::cpu_scope(|| run_expert(cold_input, exp, cfg, wt, None, &mut a));
                    *results_ref[i].lock().unwrap() = a;
                }
            });
            // Serial reduce in cold order — the accumulation order the
            // scalar path had, so parity holds bit for bit.
            for r in &results {
                let a = r.lock().unwrap();
                if a.is_empty() {
                    continue;
                }
                for ((o, sum), v) in out.iter_mut().zip(&mut cold_sum).zip(a.iter()) {
                    *o += v;
                    *sum += v;
                }
            }
        }
        _ => {
            for &(gi, wt) in &cold {
                let Some(exp) = l.experts.get(gi) else {
                    continue;
                };
                crate::gpu::cpu_scope(|| run_expert(cold_input, exp, cfg, wt, pool, &mut acc));
                for ((o, sum), a) in out.iter_mut().zip(&mut cold_sum).zip(&acc) {
                    *o += a;
                    *sum += a;
                }
            }
        }
    }
    Some((cold_sum, cold.len()))
}

/// How much of each layer's compressed cache already sits on the card. ONE
/// map: a reader and a writer with a `static` each are two maps, and the
/// reader would never see a thing the writer put down.
/// The reallocation counter as of the last successful tail write. Any change
/// means some buffer was rebuilt and every tail count is stale.
#[cfg(feature = "gpu")]
fn last_grew(now: u64) -> u64 {
    use std::sync::atomic::{AtomicU64, Ordering};
    static SEEN: AtomicU64 = AtomicU64::new(0);
    let was = SEEN.load(Ordering::Relaxed);
    if was != now {
        SEEN.store(now, Ordering::Relaxed);
        compressed_map().lock().unwrap().clear();
        return u64::MAX; // force a full write this round
    }
    now
}

#[cfg(feature = "gpu")]
fn compressed_map() -> &'static std::sync::Mutex<std::collections::HashMap<(u64, usize), usize>> {
    use std::collections::HashMap;
    use std::sync::{Mutex, OnceLock};
    static W: OnceLock<Mutex<HashMap<(u64, usize), usize>>> = OnceLock::new();
    W.get_or_init(|| Mutex::new(HashMap::new()))
}

#[cfg(feature = "gpu")]
fn compressed_written(kv_id: u64, li: usize) -> usize {
    compressed_map()
        .lock()
        .unwrap()
        .get(&(kv_id, li))
        .copied()
        .unwrap_or(0)
}

/// Reset to zero whenever a write fails or the buffer grows — a grown buffer
/// keeps none of its contents.
#[cfg(feature = "gpu")]
fn note_compressed(kv_id: u64, li: usize, n: usize) {
    compressed_map().lock().unwrap().insert((kv_id, li), n);
}

#[cfg(feature = "gpu")]
fn gpu_moe2_enabled() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| {
        std::env::var("CMF_DSV4_GPU_MOE2")
            .map(|v| v != "0")
            .unwrap_or(true)
            && crate::gpu::backend_available()
    })
}

pub fn moe_step(
    hidden: &[f32],
    l: &Dsv4Layer,
    cfg: &Dsv4Cfg,
    token_id: u32,
    // Layer index — only used to bucket routing statistics.
    li: usize,
    pool: Option<&crate::pool::Pool>,
    out: &mut [f32],
) {
    let _t0 = prof::on().then(std::time::Instant::now);
    let _guard = scopeguard_moe(_t0, li);
    let mut logits = vec![0.0f32; cfg.n_routed_experts];
    l.gate.matvec(hidden, &mut logits, pool);
    let (mut idx, mut w) = (Vec::new(), Vec::new());
    route(
        &logits,
        l.gate_bias.as_deref(),
        cfg.top_k,
        cfg.route_scale,
        l.tid2eid
            .as_ref()
            .map(|tbl| hash_route(tbl, cfg.vocab, cfg.top_k, token_id))
            .as_deref(),
        l.mask.as_deref(),
        &mut idx,
        &mut w,
    );
    if route_stats_on() {
        let routed: Vec<(usize, f32)> = idx.iter().copied().zip(w.iter().copied()).collect();
        record_route(li, 0, cfg.n_routed_experts, &routed);
    }
    // Same trace the generic MoE path writes (`CMF_MOE_TRACE`): one
    // `layer:e1,e2,…` line per routed token. The first arena run on this
    // architecture measured a 4.5% hit rate — random level for the arena's
    // size — and only a per-token trace can say whether that is the
    // router's true entropy or the cache structure destroying locality.
    crate::pipeline::moe_trace_at(li as i32, &idx);
    // The whole block on the device, in one submission, or nothing. Routing
    // happens there too — the logits above are what it starts from, so the
    // CPU's own choice is discarded rather than second-guessed.
    #[cfg(feature = "gpu")]
    if gpu_moe2_enabled() && crate::gpu::enabled_here() {
        let forced = l
            .tid2eid
            .as_ref()
            .map(|tbl| hash_route(tbl, cfg.vocab, cfg.top_k, token_id));
        if moe_frame(
            hidden,
            l,
            cfg,
            li,
            &logits,
            forced.as_deref(),
            pool,
            None,
            None,
            out,
        )
        .is_some()
        {
            // CMF_DSV4_MOE_CHECK=1 recomputes the same block on the CPU and
            // reports where they part. A wrong MoE does not fail — it answers
            // differently — and the toy agreed bit for bit while the release
            // did not, so the difference lives in something the toy has no
            // instance of. Only a per-layer number will say which.
            if std::env::var("CMF_DSV4_MOE_CHECK").is_ok() {
                let mut want = vec![0.0f32; out.len()];
                let mut acc = vec![0.0f32; cfg.dim];
                // This is a CPU oracle. Without cpu_scope the generic
                // quantized matvec probe uploaded a second copy of every
                // selected expert, so the diagnostic itself exhausted VRAM
                // after otherwise-correct global-pool layers.
                crate::gpu::cpu_scope(|| {
                    for (e, &ei) in idx.iter().enumerate() {
                        let Some(exp) = l.experts.get(ei) else {
                            continue;
                        };
                        run_expert(
                            hidden,
                            exp,
                            cfg,
                            w.get(e).copied().unwrap_or(0.0),
                            pool,
                            &mut acc,
                        );
                        for (o, a) in want.iter_mut().zip(&acc) {
                            *o += a;
                        }
                    }
                    run_expert(hidden, &l.shared, cfg, 1.0, pool, &mut acc);
                    for (o, a) in want.iter_mut().zip(&acc) {
                        *o += a;
                    }
                });
                let num: f32 = want
                    .iter()
                    .zip(out.iter())
                    .map(|(a, b)| (a - b) * (a - b))
                    .sum();
                let den: f32 = want.iter().map(|a| a * a).sum::<f32>().max(1e-20);
                let rel = (num / den).sqrt();
                if rel > 1e-3 {
                    let packed = pack_for(l, cfg, li).map_or(0, |p| p.globals.len());
                    eprintln!(
                        "[кадр MoE] слой {li}: расхождение {rel:.3e} | выбрано {} | \
                         упаковано {packed} из {} | хеш={} | смещение={}",
                        idx.len(),
                        cfg.n_routed_experts,
                        l.tid2eid.is_some(),
                        l.gate_bias.is_some()
                    );
                }
            }
            return;
        }
    }
    // Cheap tally for the batching question: how many DISTINCT experts a
    // group of tokens reaches. If five tokens want thirty different experts,
    // a batched MoE reads thirty weights and amortises nothing — which is
    // the difference between a speculative verify that pays for itself and
    // one that does not. Disarmed it costs one thread-local read.
    PICK_TALLY.with(|t| {
        if let Some(v) = t.borrow_mut().as_mut() {
            v.push((li, idx.to_vec()));
        }
    });
    if dump_path().is_some() {
        PICKED.with(|p| {
            let mut p = p.borrow_mut();
            if p.len() <= li {
                p.resize(li + 1, Vec::new());
            }
            p[li] = idx.clone();
        });
    }
    // One submission for the whole block — the chosen experts plus the
    // shared one. Per-expert dispatches are what made MoE slow elsewhere,
    // and the device keeps the weights across tokens, so the cost is the
    // arithmetic rather than the traffic. A refusal (missing kernel, mixed
    // layouts, weights that do not fit the budget) falls to the CPU whole,
    // never half.
    // CORRECT but SLOWER, so off by default. Parity holds on real weights
    // (perplexity 6.808 → 6.839 at 64 tokens, 5.102 → 5.146 at 200), and an
    // honest alternating A/B says 1.0 tok/s against the CPU's 2.2. The first
    // measurement claimed the opposite — 0.7 → 2.0 — because the CPU arm ran
    // first and paged in 158 GB for the GPU arm to inherit.
    //
    // The cost is not arithmetic, it is round trips: this submits and reads
    // back once per layer, forty-three times a token, and a discrete card
    // charges milliseconds for each. Fixing it means one submission per
    // token — the whole-token graph — not a faster kernel.
    //
    // `CMF_DSV4_GPU_MOE=1` opts in.
    fn gpu_moe_on() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| std::env::var("CMF_DSV4_GPU_MOE").is_ok_and(|v| v != "0"))
    }
    if gpu_moe_on() && crate::gpu::enabled_here() {
        let mut jobs = Vec::with_capacity(idx.len() + 1);
        let mut model_ref = None;
        let mut ok = true;
        for (e, &ei) in idx.iter().enumerate() {
            let Some(exp) = l.experts.get(ei) else {
                continue;
            };
            ok &= crate::pipeline::moe_push_job_parts(
                &exp.w1,
                &exp.w3,
                &exp.w2,
                hidden,
                w.get(e).copied().unwrap_or(0.0),
                cfg.swiglu_limit,
                &mut jobs,
                &mut model_ref,
            )
            .is_some();
        }
        ok &= crate::pipeline::moe_push_job_parts(
            &l.shared.w1,
            &l.shared.w3,
            &l.shared.w2,
            hidden,
            1.0,
            cfg.swiglu_limit,
            &mut jobs,
            &mut model_ref,
        )
        .is_some();
        if ok {
            if let Some(m) = model_ref.as_ref() {
                if crate::gpu::moe_block(m, &jobs, out) {
                    // CMF_DSV4_GPU_CHECK=1 recomputes the same block on the
                    // CPU and reports the divergence. A GPU MoE that is wrong
                    // does not fail — it answers differently — so the only way
                    // to know is to ask both.
                    if std::env::var("CMF_DSV4_GPU_CHECK").is_ok() {
                        let mut want = vec![0.0f32; out.len()];
                        let mut acc = vec![0.0f32; cfg.dim];
                        for (e, &ei) in idx.iter().enumerate() {
                            let Some(exp) = l.experts.get(ei) else {
                                continue;
                            };
                            run_expert(
                                hidden,
                                exp,
                                cfg,
                                w.get(e).copied().unwrap_or(0.0),
                                pool,
                                &mut acc,
                            );
                            for (o, a) in want.iter_mut().zip(&acc) {
                                *o += a;
                            }
                        }
                        run_expert(hidden, &l.shared, cfg, 1.0, pool, &mut acc);
                        for (o, a) in want.iter_mut().zip(&acc) {
                            *o += a;
                        }
                        let num: f32 = want
                            .iter()
                            .zip(out.iter())
                            .map(|(a, b)| (a - b) * (a - b))
                            .sum();
                        let den: f32 = want.iter().map(|a| a * a).sum::<f32>().max(1e-20);
                        eprintln!(
                            "[dsv4-gpu] слой {li}: расхождение {:.3e} | |CPU|={:.5} |GPU|={:.5} | экспертов {}",
                            (num / den).sqrt(),
                            den.sqrt(),
                            out.iter().map(|x| x * x).sum::<f32>().sqrt(),
                            jobs.len()
                        );
                    }
                    return;
                }
            }
        }
    }
    out.fill(0.0);
    // Layers the packer had no room for land here whole. Same shape as the
    // frame's cold completion: one worker per expert (the shared one rides
    // as an extra job), whole matvecs inside, cpu_scope INSIDE the worker —
    // on the main thread it would gate nothing, and the generic matvec
    // would upload every expert to the card tensor by tensor, which is
    // exactly the per-token PCIe churn this path exists to avoid.
    let jobs: Vec<(Option<usize>, f32)> = idx
        .iter()
        .enumerate()
        .map(|(e, &ei)| (Some(ei), w.get(e).copied().unwrap_or(0.0)))
        .chain(std::iter::once((None, 1.0)))
        .collect();
    match pool {
        Some(p) if jobs.len() > 1 => {
            let results: Vec<std::sync::Mutex<Vec<f32>>> = jobs
                .iter()
                .map(|_| std::sync::Mutex::new(Vec::new()))
                .collect();
            let (jobs_ref, results_ref) = (&jobs, &results);
            p.run_rows(jobs.len(), &move |cs, ce| {
                for i in cs..ce {
                    let (ei, wt) = jobs_ref[i];
                    let exp = match ei {
                        Some(ei) => match l.experts.get(ei) {
                            Some(x) => x,
                            None => continue,
                        },
                        None => &l.shared,
                    };
                    let mut a = vec![0.0f32; cfg.dim];
                    crate::gpu::cpu_scope(|| run_expert(hidden, exp, cfg, wt, None, &mut a));
                    *results_ref[i].lock().unwrap() = a;
                }
            });
            for r in &results {
                let a = r.lock().unwrap();
                for (o, v) in out.iter_mut().zip(a.iter()) {
                    *o += v;
                }
            }
        }
        _ => {
            let mut acc = vec![0.0f32; cfg.dim];
            for &(ei, wt) in &jobs {
                let exp = match ei {
                    Some(ei) => match l.experts.get(ei) {
                        Some(x) => x,
                        None => continue,
                    },
                    None => &l.shared,
                };
                crate::gpu::cpu_scope(|| run_expert(hidden, exp, cfg, wt, pool, &mut acc));
                for (o, a) in out.iter_mut().zip(&acc) {
                    *o += a;
                }
            }
        }
    }
}

/// The routed and shared experts both come through here, so the clamp and
/// the weight folding have exactly one implementation — `expert_swiglu`.
fn run_expert(
    x: &[f32],
    e: &Dsv4Expert,
    cfg: &Dsv4Cfg,
    weight: f32,
    pool: Option<&crate::pool::Pool>,
    out: &mut [f32],
) {
    expert_swiglu(
        x,
        &|src, dst| e.w1.matvec(src, dst, pool),
        &|src, dst| e.w3.matvec(src, dst, pool),
        &|src, dst| e.w2.matvec(src, dst, pool),
        cfg.moe_inter,
        weight,
        cfg.swiglu_limit,
        out,
    );
}

/// The same expert computation for several inputs, streaming each selected
/// weight once. Used by DSpark's trained five-position block: running five
/// ordinary `moe_step`s rereads the shared expert five times and every
/// coincident routed expert once per position.
fn moe_step_block(
    xs: &[f32],
    b: usize,
    l: &Dsv4Layer,
    cfg: &Dsv4Cfg,
    token_ids: &[u32],
    tally_layer: usize,
    pool: Option<&crate::pool::Pool>,
    out: &mut [f32],
) {
    let (dim, inter) = (cfg.dim, cfg.moe_inter);
    debug_assert_eq!(xs.len(), b * dim);
    debug_assert_eq!(out.len(), b * dim);
    out.fill(0.0);

    let mut logits = vec![0.0f32; b * cfg.n_routed_experts];
    l.gate.matmat(xs, b, &mut logits, pool);
    let mut picks: Vec<Vec<usize>> = Vec::with_capacity(b);
    let mut weights: Vec<Vec<f32>> = Vec::with_capacity(b);
    for bi in 0..b {
        let mut idx = Vec::new();
        let mut wt = Vec::new();
        let forced = l.tid2eid.as_ref().map(|tbl| {
            hash_route(
                tbl,
                cfg.vocab,
                cfg.top_k,
                token_ids.get(bi).copied().unwrap_or(0),
            )
        });
        route(
            &logits[bi * cfg.n_routed_experts..(bi + 1) * cfg.n_routed_experts],
            l.gate_bias.as_deref(),
            cfg.top_k,
            cfg.route_scale,
            forced.as_deref(),
            l.mask.as_deref(),
            &mut idx,
            &mut wt,
        );
        PICK_TALLY.with(|t| {
            if let Some(v) = t.borrow_mut().as_mut() {
                v.push((tally_layer, idx.clone()));
            }
        });
        picks.push(idx);
        weights.push(wt);
    }

    // Preserve the scalar path's accumulation order by keeping every routed
    // slot separate; grouping below changes only when a weight is read.
    let mut routed = vec![0.0f32; b * cfg.top_k * dim];
    // Group the token slots by expert first — the list is also the unit of
    // parallelism below.
    let mut active: Vec<(usize, Vec<(usize, usize, f32)>)> = Vec::new();
    for ei in 0..l.experts.len() {
        let mut jobs = Vec::new();
        for bi in 0..b {
            for (slot, &picked) in picks[bi].iter().enumerate() {
                if picked == ei {
                    jobs.push((bi, slot, weights[bi][slot]));
                }
            }
        }
        if !jobs.is_empty() {
            active.push((ei, jobs));
        }
    }
    // One expert's forward, single-threaded, returning the scaled down
    // projections in job order.
    let expert_fwd =
        |ei: usize, jobs: &[(usize, usize, f32)], inner: Option<&crate::pool::Pool>| -> Vec<f32> {
            let e = &l.experts[ei];
            let n = jobs.len();
            let mut xj = vec![0.0f32; n * dim];
            for (j, &(bi, _, _)) in jobs.iter().enumerate() {
                xj[j * dim..(j + 1) * dim].copy_from_slice(&xs[bi * dim..(bi + 1) * dim]);
            }
            let mut gate = vec![0.0f32; n * inter];
            let mut up = vec![0.0f32; n * inter];
            e.w1.matmat(&xj, n, &mut gate, inner);
            e.w3.matmat(&xj, n, &mut up, inner);
            for (j, &(_, _, wt)) in jobs.iter().enumerate() {
                let (gj, uj) = (
                    &mut gate[j * inter..(j + 1) * inter],
                    &mut up[j * inter..(j + 1) * inter],
                );
                if cfg.swiglu_limit > 0.0 {
                    for u in uj.iter_mut() {
                        *u = u.clamp(-cfg.swiglu_limit, cfg.swiglu_limit);
                    }
                    for g in gj.iter_mut() {
                        *g = g.min(cfg.swiglu_limit);
                    }
                }
                for (g, &u) in gj.iter_mut().zip(uj.iter()) {
                    *g = (*g / (1.0 + (-*g).exp())) * u * wt;
                }
            }
            let mut down = vec![0.0f32; n * dim];
            e.w2.matmat(&gate, n, &mut down, inner);
            down
        };
    // ~370 non-resident experts a token used to run this loop ONE AFTER
    // ANOTHER: a 2048-row matvec cannot occupy a big pool, and the loop
    // serialised the only real parallelism there is — across experts.
    // Measured on a 384-core host with DeepSeek-V4-Flash: ~3.3 s/token
    // flat across every fetch-side improvement, because the wall was
    // here. Parallel across experts, each single-threaded and pinned to
    // the CPU on ITS OWN worker (cpu_scope is thread-local, so it must
    // be entered inside the closure, not around the pool call — the
    // documented trap).
    match pool {
        Some(p) if active.len() > 1 => {
            let results: Vec<std::sync::Mutex<Vec<f32>>> = active
                .iter()
                .map(|_| std::sync::Mutex::new(Vec::new()))
                .collect();
            let active_ref = &active;
            let results_ref = &results;
            let fwd = &expert_fwd;
            p.run_rows(active_ref.len(), &move |s, e| {
                for i in s..e {
                    let (ei, jobs) = &active_ref[i];
                    let d = crate::gpu::cpu_scope(|| fwd(*ei, jobs, None));
                    *results_ref[i].lock().unwrap() = d;
                }
            });
            for (i, (_, jobs)) in active.iter().enumerate() {
                let down = results[i].lock().unwrap();
                for (j, &(bi, slot, _)) in jobs.iter().enumerate() {
                    routed[(bi * cfg.top_k + slot) * dim..(bi * cfg.top_k + slot + 1) * dim]
                        .copy_from_slice(&down[j * dim..(j + 1) * dim]);
                }
            }
        }
        _ => {
            for (ei, jobs) in &active {
                let down = expert_fwd(*ei, jobs, pool);
                for (j, &(bi, slot, _)) in jobs.iter().enumerate() {
                    routed[(bi * cfg.top_k + slot) * dim..(bi * cfg.top_k + slot + 1) * dim]
                        .copy_from_slice(&down[j * dim..(j + 1) * dim]);
                }
            }
        }
    }

    // Shared expert: all positions always use it, so this is the highest
    // certainty weight-sharing win in the block.
    let mut sg = vec![0.0f32; b * inter];
    let mut su = vec![0.0f32; b * inter];
    l.shared.w1.matmat(xs, b, &mut sg, pool);
    l.shared.w3.matmat(xs, b, &mut su, pool);
    for bi in 0..b {
        let (gj, uj) = (
            &mut sg[bi * inter..(bi + 1) * inter],
            &mut su[bi * inter..(bi + 1) * inter],
        );
        if cfg.swiglu_limit > 0.0 {
            for u in uj.iter_mut() {
                *u = u.clamp(-cfg.swiglu_limit, cfg.swiglu_limit);
            }
            for g in gj.iter_mut() {
                *g = g.min(cfg.swiglu_limit);
            }
        }
        for (g, &u) in gj.iter_mut().zip(uj.iter()) {
            *g = (*g / (1.0 + (-*g).exp())) * u;
        }
    }
    let mut shared = vec![0.0f32; b * dim];
    l.shared.w2.matmat(&sg, b, &mut shared, pool);

    for bi in 0..b {
        let dst = &mut out[bi * dim..(bi + 1) * dim];
        for slot in 0..picks[bi].len() {
            let src = &routed[(bi * cfg.top_k + slot) * dim..(bi * cfg.top_k + slot + 1) * dim];
            for (o, &v) in dst.iter_mut().zip(src) {
                *o += v;
            }
        }
        for (o, &v) in dst.iter_mut().zip(&shared[bi * dim..(bi + 1) * dim]) {
            *o += v;
        }
    }
}

/// Complete only the routed experts absent from a partial GPU pack.  Jobs
/// are grouped by expert so coincident speculative positions stream that
/// expert's weights once; per-token slot accumulation remains in route order,
/// matching the ordinary exact cold-correction path.
fn cold_step_block(
    xs: &[f32],
    b: usize,
    l: &Dsv4Layer,
    cfg: &Dsv4Cfg,
    cold: &[Vec<(usize, f32)>],
    pool: Option<&crate::pool::Pool>,
    out: &mut [f32],
) {
    let (dim, inter) = (cfg.dim, cfg.moe_inter);
    debug_assert_eq!(xs.len(), b * dim);
    debug_assert_eq!(cold.len(), b);
    debug_assert_eq!(out.len(), b * dim);
    out.fill(0.0);
    let slots = cold.iter().map(Vec::len).max().unwrap_or(0);
    if slots == 0 {
        return;
    }
    // The ordinary partial walk evaluates every cold winner with matvec.
    // The grouped arm streams a coincident expert once for the whole verify
    // block; release-scale fingerprints and row-zero logits were identical,
    // and it moved the fixed A40 bench 1.9 → 2.3 tok/s.  Keep the scalar arm
    // as a parity escape hatch for a new quant layout/adapter.
    let grouped = std::env::var("CMF_DSV4_COLD_MATMAT")
        .map(|v| v != "0")
        .unwrap_or(true);
    if !grouped {
        let jobs: Vec<(usize, usize, usize, f32)> = cold
            .iter()
            .enumerate()
            .flat_map(|(bi, row)| {
                row.iter()
                    .enumerate()
                    .map(move |(slot, &(ei, wt))| (bi, slot, ei, wt))
            })
            .collect();
        let results: Vec<std::sync::Mutex<Vec<f32>>> = jobs
            .iter()
            .map(|_| std::sync::Mutex::new(Vec::new()))
            .collect();
        match pool {
            Some(p) if jobs.len() > 1 => {
                let (jobs_ref, results_ref) = (&jobs, &results);
                p.run_rows(jobs.len(), &move |s, e| {
                    for i in s..e {
                        let (_, _, ei, wt) = jobs_ref[i];
                        let Some(exp) = l.experts.get(ei) else {
                            continue;
                        };
                        let bi = jobs_ref[i].0;
                        let mut a = vec![0.0f32; dim];
                        crate::gpu::cpu_scope(|| {
                            run_expert(&xs[bi * dim..(bi + 1) * dim], exp, cfg, wt, None, &mut a)
                        });
                        *results_ref[i].lock().unwrap() = a;
                    }
                });
            }
            _ => {
                for (i, &(bi, _, ei, wt)) in jobs.iter().enumerate() {
                    let Some(exp) = l.experts.get(ei) else {
                        continue;
                    };
                    let mut a = vec![0.0f32; dim];
                    crate::gpu::cpu_scope(|| {
                        run_expert(&xs[bi * dim..(bi + 1) * dim], exp, cfg, wt, pool, &mut a)
                    });
                    *results[i].lock().unwrap() = a;
                }
            }
        }
        // Reduce in each token's routing order, exactly like
        // dsv4_chain1_layer. The jobs vector was built in that order.
        for (i, &(bi, _, _, _)) in jobs.iter().enumerate() {
            let a = results[i].lock().unwrap();
            for (o, &v) in out[bi * dim..(bi + 1) * dim].iter_mut().zip(a.iter()) {
                *o += v;
            }
        }
        return;
    }
    let mut active: Vec<(usize, Vec<(usize, usize, f32)>)> = Vec::new();
    for ei in 0..l.experts.len() {
        let mut jobs = Vec::new();
        for bi in 0..b {
            for (slot, &(picked, wt)) in cold[bi].iter().enumerate() {
                if picked == ei {
                    jobs.push((bi, slot, wt));
                }
            }
        }
        if !jobs.is_empty() {
            active.push((ei, jobs));
        }
    }
    let expert_fwd =
        |ei: usize, jobs: &[(usize, usize, f32)], inner: Option<&crate::pool::Pool>| -> Vec<f32> {
            let e = &l.experts[ei];
            let n = jobs.len();
            let mut xj = vec![0.0f32; n * dim];
            for (j, &(bi, _, _)) in jobs.iter().enumerate() {
                xj[j * dim..(j + 1) * dim].copy_from_slice(&xs[bi * dim..(bi + 1) * dim]);
            }
            let mut gate = vec![0.0f32; n * inter];
            let mut up = vec![0.0f32; n * inter];
            e.w1.matmat(&xj, n, &mut gate, inner);
            e.w3.matmat(&xj, n, &mut up, inner);
            for (j, &(_, _, wt)) in jobs.iter().enumerate() {
                let (gj, uj) = (
                    &mut gate[j * inter..(j + 1) * inter],
                    &mut up[j * inter..(j + 1) * inter],
                );
                if cfg.swiglu_limit > 0.0 {
                    for u in uj.iter_mut() {
                        *u = u.clamp(-cfg.swiglu_limit, cfg.swiglu_limit);
                    }
                    for g in gj.iter_mut() {
                        *g = g.min(cfg.swiglu_limit);
                    }
                }
                for (g, &u) in gj.iter_mut().zip(uj.iter()) {
                    *g = (*g / (1.0 + (-*g).exp())) * u * wt;
                }
            }
            let mut down = vec![0.0f32; n * dim];
            e.w2.matmat(&gate, n, &mut down, inner);
            down
        };
    let mut routed = vec![0.0f32; b * slots * dim];
    match pool {
        Some(p) if active.len() > 1 => {
            let results: Vec<std::sync::Mutex<Vec<f32>>> = active
                .iter()
                .map(|_| std::sync::Mutex::new(Vec::new()))
                .collect();
            let active_ref = &active;
            let results_ref = &results;
            let fwd = &expert_fwd;
            p.run_rows(active.len(), &move |s, e| {
                for i in s..e {
                    let (ei, jobs) = &active_ref[i];
                    let d = crate::gpu::cpu_scope(|| fwd(*ei, jobs, None));
                    *results_ref[i].lock().unwrap() = d;
                }
            });
            for (i, (_, jobs)) in active.iter().enumerate() {
                let down = results[i].lock().unwrap();
                for (j, &(bi, slot, _)) in jobs.iter().enumerate() {
                    routed[(bi * slots + slot) * dim..(bi * slots + slot + 1) * dim]
                        .copy_from_slice(&down[j * dim..(j + 1) * dim]);
                }
            }
        }
        _ => {
            for (ei, jobs) in &active {
                let down = expert_fwd(*ei, jobs, pool);
                for (j, &(bi, slot, _)) in jobs.iter().enumerate() {
                    routed[(bi * slots + slot) * dim..(bi * slots + slot + 1) * dim]
                        .copy_from_slice(&down[j * dim..(j + 1) * dim]);
                }
            }
        }
    }
    for bi in 0..b {
        let dst = &mut out[bi * dim..(bi + 1) * dim];
        for slot in 0..cold[bi].len() {
            let src = &routed[(bi * slots + slot) * dim..(bi * slots + slot + 1) * dim];
            for (o, &v) in dst.iter_mut().zip(src) {
                *o += v;
            }
        }
    }
}

/// Feed the exact route of the speculative batch's guaranteed-accepted first
/// token back into the same FreeToken-style live slots ordinary decode uses.
/// Rejected draft suffixes must not train or pollute the LRU, so the caller
/// deliberately passes only row zero.
#[cfg(feature = "gpu")]
fn refill_route_slots(l: &Dsv4Layer, cfg: &Dsv4Cfg, pk: &Pack, picks: &[usize]) {
    if picks.is_empty() || pk.route_complete() {
        return;
    }
    let quota = {
        static Q: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
        *Q.get_or_init(|| {
            std::env::var("CMF_DSV4_FETCH_MAX")
                .ok()
                .and_then(|v| v.parse().ok())
                .unwrap_or_else(|| crate::gpu_wgpu::dsv4_fetch_defaults().0)
        })
    };
    if quota == 0 {
        return;
    }
    let min_seen = {
        static M: std::sync::OnceLock<u16> = std::sync::OnceLock::new();
        *M.get_or_init(|| {
            std::env::var("CMF_DSV4_FETCH_MIN_SEEN")
                .ok()
                .and_then(|v| v.parse().ok())
                .unwrap_or_else(|| crate::gpu_wgpu::dsv4_fetch_defaults().1)
        })
    };
    let Some(model) = l.experts.first().and_then(|e| e.w1.model_arc()) else {
        return;
    };
    let mut dynv = pk.dynslots.lock().unwrap();
    if dynv.owner.is_empty() {
        return;
    }
    dynv.clock += 1;
    let clock = dynv.clock;
    if clock % 64 == 0 {
        for seen in &mut dynv.seen {
            *seen >>= 1;
        }
    }
    for &pick in picks {
        if pick >= dynv.seen.len() {
            continue;
        }
        dynv.seen[pick] = dynv.seen[pick].saturating_add(1);
        let slot = dynv.remap[pick];
        if slot != u32::MAX {
            dynv.last[slot as usize] = clock;
        }
    }
    let pack_first = pk.tensors.first().map(|t| t.0).unwrap_or(usize::MAX);
    let mut fetched = 0usize;
    for &pick in picks {
        if fetched >= quota || pick >= dynv.remap.len() {
            break;
        }
        if dynv.remap[pick] != u32::MAX || dynv.seen[pick] < min_seen {
            continue;
        }
        let victim = (0..dynv.owner.len())
            .filter(|&slot| dynv.last[slot] != clock)
            .min_by_key(|&slot| dynv.last[slot]);
        let Some(victim) = victim else { break };
        let Some(exp) = l.experts.get(pick) else {
            continue;
        };
        let tensors = (|| {
            Some((
                exp.w1.model_idx()?,
                exp.w3.model_idx()?,
                exp.w2.model_idx()?,
            ))
        })();
        let Some(tensors) = tensors else { continue };
        let gu_q2 = exp.w1.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP);
        if !crate::gpu_wgpu::dsv4_slot_fill(
            &model,
            pack_first,
            victim,
            pick,
            tensors,
            cfg.moe_inter,
            cfg.dim,
            gu_q2,
        ) {
            break;
        }
        let old = dynv.owner[victim] as usize;
        if old < dynv.remap.len() {
            dynv.remap[old] = u32::MAX;
        }
        dynv.remap[pick] = victim as u32;
        dynv.owner[victim] = pick as u32;
        dynv.last[victim] = clock;
        dynv.mutated = true;
        fetched += 1;
    }
}

/// Grouped output projection for a block. `wo_a` cannot use a plain matmat
/// because each group sees a different attention slice; reading a quantized
/// row once and applying it to every block position gives the same dot order
/// without rereading/dequantizing that row B times.
fn o_project_block(
    attn: &[f32],
    b: usize,
    wo_a: &crate::qtensor::QTensor,
    wo_b: &crate::qtensor::QTensor,
    groups: usize,
    lora: usize,
    pool: Option<&crate::pool::Pool>,
    out: &mut [f32],
) {
    let attn_len = attn.len() / b;
    let per_group = attn_len / groups;
    let rows = groups * lora;
    let mut mid = vec![0.0f32; b * rows];
    let mid_addr = crate::pool::SendMut::new(mid.as_mut_ptr());
    let run = |start: usize, end: usize| {
        let mut wr = vec![0.0f32; wo_a.cols()];
        for r in start..end {
            wo_a.row_f32(r, &mut wr);
            let group = r / lora;
            for bi in 0..b {
                let x = &attn
                    [bi * attn_len + group * per_group..bi * attn_len + (group + 1) * per_group];
                let v = wr.iter().zip(x).map(|(w, x)| w * x).sum();
                unsafe { *mid_addr.at(bi * rows + r) = v };
            }
        }
    };
    match pool {
        Some(p) if rows >= 256 => p.run_rows(rows, &run),
        _ => run(0, rows),
    }
    wo_b.matmat(&mid, b, out, pool);
}

/// `CMF_DSV4_TRACE=1` prints the hidden state's RMS after each half-block and
/// the logits' shape at the end. A 300B model that decodes nonsense gives no
/// other handle: this says whether the state grew, collapsed or went
/// non-finite, and at which layer — before anyone reaches for a debugger on a
/// hundred-gigabyte file.
fn no_compressed() -> bool {
    static OFF: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *OFF.get_or_init(|| std::env::var("CMF_DSV4_NO_COMPRESSED").is_ok_and(|v| v != "0"))
}

fn trace_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| std::env::var("CMF_DSV4_TRACE").is_ok_and(|v| v != "0"))
}

fn rms_of(v: &[f32]) -> f32 {
    if v.is_empty() {
        return 0.0;
    }
    (v.iter().map(|x| x * x).sum::<f32>() / v.len() as f32).sqrt()
}

#[cfg(feature = "gpu")]
fn verify_fp_on(pos: usize) -> bool {
    static POS: std::sync::OnceLock<Option<usize>> = std::sync::OnceLock::new();
    let want = *POS.get_or_init(|| {
        std::env::var("CMF_DSV4_FP_POS")
            .ok()
            .and_then(|v| v.parse().ok())
    });
    want == Some(pos)
}

#[cfg(feature = "gpu")]
fn verify_fp(tag: &str, pos: usize, li: usize, state: &[f32]) {
    if !verify_fp_on(pos) {
        return;
    }
    let mut fp = 0xcbf29ce484222325u64;
    for &x in state {
        fp ^= x.to_bits() as u64;
        fp = fp.wrapping_mul(0x100000001b3);
    }
    eprintln!(
        "[dsv4-fp] {tag} pos={pos} li={li} fp={fp:016x} rms={:.7} head={:?}",
        rms_of(state),
        &state[..4.min(state.len())]
    );
}

/// `CMF_DSV4_DUMP=<path>` appends one JSON line per token: the embedding, the
/// hyper-connection state after every layer, the folded-and-normed head input
/// and the logits. It exists to be diffed against the reference forward on
/// the same weights — the numerical parity this port has never had, which at
/// toy scale is a few thousand floats and entirely tractable.
thread_local! {
    /// The attention body's input and output per layer, interleaved.
    static BODY: std::cell::RefCell<Vec<String>> = const { std::cell::RefCell::new(Vec::new()) };
    /// Experts chosen per layer for the token being decoded — the dump needs
    /// them, because two implementations that pick DIFFERENT experts diverge
    /// hugely for a reason that is not a bug in either.
    static PICKED: std::cell::RefCell<Vec<Vec<usize>>> =
        const { std::cell::RefCell::new(Vec::new()) };
    /// (layer, chosen experts) in call order, when armed.
    static PICK_TALLY: std::cell::RefCell<Option<Vec<(usize, Vec<usize>)>>> =
        const { std::cell::RefCell::new(None) };
}

/// Start recording expert picks. Idempotent; the previous tally is dropped.
pub fn pick_tally_arm() {
    PICK_TALLY.with(|t| *t.borrow_mut() = Some(Vec::new()));
}

/// Take what was recorded and stop recording.
pub fn pick_tally_take() -> Vec<(usize, Vec<usize>)> {
    PICK_TALLY.with(|t| t.borrow_mut().take().unwrap_or_default())
}

/// How many distinct experts a set of per-token pick lists reaches, and how
/// many picks it makes. The ratio is what a batched MoE can hope to save.
pub fn tally_unique(picks: &[(usize, Vec<usize>)]) -> (usize, usize) {
    // Keyed by (layer, expert). Expert 17 of layer 3 and expert 17 of layer 4
    // are different weights, and counting them as one understated the traffic
    // a batch has to read — badly for the draft, whose three stages each have
    // their own 256.
    let mut seen = std::collections::HashSet::new();
    let mut total = 0;
    for (li, v) in picks {
        total += v.len();
        for &e in v {
            seen.insert((*li, e));
        }
    }
    (seen.len(), total)
}

fn dump_path() -> Option<&'static str> {
    static P: std::sync::OnceLock<Option<String>> = std::sync::OnceLock::new();
    P.get_or_init(|| std::env::var("CMF_DSV4_DUMP").ok())
        .as_deref()
}

fn dump_line(json: &str) {
    if let Some(p) = dump_path() {
        use std::io::Write as _;
        if let Ok(mut f) = std::fs::OpenOptions::new()
            .create(true)
            .append(true)
            .open(p)
        {
            let _ = writeln!(f, "{json}");
        }
    }
}

fn vec_json(v: &[f32]) -> String {
    let mut s = String::with_capacity(v.len() * 9);
    s.push('[');
    for (i, x) in v.iter().enumerate() {
        if i > 0 {
            s.push(',');
        }
        s.push_str(&format!("{x:.6e}"));
    }
    s.push(']');
    s
}

/// One token through the whole stack.
///
/// The hidden state is `hc_mult` copies of a `dim`-vector from the very
/// first line to the very last: the embedding is replicated, every layer
/// folds/expands around its two halves, and only `hc_head_fold` collapses
/// it before the output norm and the head. There is no point in this
/// function where an ordinary residual would fit.
#[allow(clippy::too_many_arguments)]
/// A chunk of prompt tokens. Stage one of the batched prefill (see
/// docs/DSV4_PREFILL.md): the walk itself, with the head skipped for every
/// token but the last.
///
/// Prefill costs `len × per-token` today, and on a 2500-token prompt that is
/// a minute and a half before the first word. The stages that follow batch
/// the weight reads — which is where the nine-fold gap to the bandwidth
/// floor lives — but this one is the scaffolding they hang on, and it
/// already stops computing 129 280 logits for tokens nobody asks about.
#[allow(clippy::too_many_arguments)]
/// `CMF_DSV4_BATCH=N` — how many prompt tokens go through the card in one
/// submission. 1 keeps the walk. The chunk still bounds it: a batch never
/// spans two chunks, so cancellation stays as responsive as it was.
fn batch_prefill() -> usize {
    static N: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
    *N.get_or_init(|| {
        std::env::var("CMF_DSV4_BATCH")
            .ok()
            .and_then(|v| v.parse::<usize>().ok())
            .filter(|&n| (1..=32).contains(&n))
            // A partial expert pack used to make the device-chain batch
            // decline, so the conservative default was one.  The exact host
            // batch below now handles that geometry: routing still spans all
            // experts and every cold winner is computed from the mmap-backed
            // checkpoint.  Eight amortises expert reads without making a long
            // prompt's cancellation granularity coarse.
            .unwrap_or(8)
    })
}

/// The prompt as batches instead of a walk, when every layer will take one.
///
/// Refuses before touching any state, never half way: the caller's fallback
/// is the per-token walk, and a batch that advanced the caches and then gave
/// up would have them advanced twice. So everything that can decline is asked
/// first, and after the first dispatch the only outcomes are success and a
/// hard failure.
///
/// Hash layers are the one shape it cannot take: their expert list is forced
/// by the TOKEN's id and the layer description carries one list, not one per
/// token. The release has three of them (0, 1, 2); a file without them
/// batches the whole stack.
#[allow(clippy::too_many_arguments)]
fn forward_chunk_batched(
    g: &Dsv4Globals,
    layers: &[Dsv4Layer],
    cfg: &Dsv4Cfg,
    st: &mut Dsv4State,
    ids: &[u32],
    pos0: usize,
    inv_freq: &[f32],
    pool: Option<&crate::pool::Pool>,
    logits: &mut Vec<f32>,
    want_logits: bool,
) -> bool {
    #[cfg(not(feature = "gpu"))]
    {
        let _ = (
            g,
            layers,
            cfg,
            st,
            ids,
            pos0,
            inv_freq,
            pool,
            logits,
            want_logits,
        );
        false
    }
    #[cfg(feature = "gpu")]
    {
        let b = ids.len();
        // Complete packs form the fused device prefix.  Partial packs belong
        // to the exact causal tail: that tail routes over all experts and
        // completes cold winners, so gpu_end == 0 is a useful (and common on
        // smaller cards) batch rather than a reason to walk token by token.
        let gpu_end = st
            .dev_set
            .iter()
            .enumerate()
            .position(|(li, &on)| {
                !on || pack_for(&layers[li], cfg, li).is_none_or(|p| !p.route_complete())
            })
            .unwrap_or(st.dev_set.len());
        let why = if b < 2 {
            "токенов меньше двух"
        } else if !chain_enabled() {
            "цепочка выключена"
        } else if !st.dev_owned {
            "карта ещё не владеет состоянием"
        } else if st.dev_set.len() != layers.len() {
            "набор слоёв ещё не зафиксирован"
        } else if st.dev_set[gpu_end.min(st.dev_set.len())..]
            .iter()
            .enumerate()
            .any(|(i, &on)| on && !st.partial_set.get(gpu_end + i).copied().unwrap_or(false))
        {
            "слои на карте не образуют префикс"
        } else {
            ""
        };
        if !why.is_empty() {
            // This is an expected preflight refusal: the first prompt chunk
            // establishes device ownership and a one-token tail is too small
            // to batch.  The caller immediately takes the exact walk, so a
            // warning here looked like a generation failure when nothing had
            // failed. Keep the reason available only to the diagnostic gate.
            tracing::debug!("dsv4: пакет preflight — {why}");
            if std::env::var("CMF_DSV4_FRAME_DEBUG").is_ok() {
                eprintln!("dsv4: пакет preflight — {why}");
            }
            return false;
        }
        let (hc, dim) = (cfg.hc_mult, cfg.dim);
        let mut emb = vec![0.0f32; dim];
        let mut states = vec![0.0f32; b * hc * dim];
        for (t, &id) in ids.iter().enumerate() {
            let mut state = vec![0.0f32; hc * dim];
            g.embed.row_f32(id as usize, &mut emb);
            for j in 0..hc {
                state[j * dim..(j + 1) * dim].copy_from_slice(&emb);
            }
            let (folded, post0, comb0) = hc_fold_norm(
                &state,
                &layers[0].hc_attn_fn,
                &layers[0].hc_attn_scale,
                &layers[0].hc_attn_base,
                &layers[0].attn_norm,
                cfg,
                pool,
            );
            let mut qn0 = vec![0.0f32; layers[0].wq_a.rows()];
            layers[0].wq_a.matvec(&folded, &mut qn0, pool);
            rms_weighted(&mut qn0, &layers[0].q_norm, cfg.norm_eps);
            if !crate::gpu_wgpu::dsv4_state_write_t(&state, t)
                || !crate::gpu_wgpu::dsv4_hc_write_t(&post0, &comb0, t)
                || !crate::gpu_wgpu::dsv4_chain_seed_t(&folded, &qn0, t)
                || !crate::gpu_wgpu::dsv4_chain_seed_bt(t, b, &state, &post0, &comb0, &folded, &qn0)
            {
                return false;
            }
            states[t * hc * dim..(t + 1) * hc * dim].copy_from_slice(&state);
        }
        let run: Vec<usize> = (0..gpu_end).collect();
        let mut folded = Vec::new();
        st.pos = pos0;
        if gpu_end > 0 {
            if !dsv4_chain_run(
                layers,
                &run,
                cfg,
                g,
                st,
                *ids.last().unwrap(),
                &mut folded,
                Some(&mut states),
                b,
                ids,
                true,
                pool,
            ) {
                return false;
            }
        }
        // Finish the trailing host layers in causal token order. Their KV
        // caches are host-owned, while the device prefix advanced its own
        // caches inside the one submission above. On the release this loop
        // is exactly layer 42; keeping it general makes smaller VRAM budgets
        // correct as long as the resident layers remain one prefix.
        let mut scratch = HcScratch::new(cfg);
        host_tail_walk_batch(
            g,
            layers,
            cfg,
            st,
            gpu_end,
            &mut states,
            ids,
            pos0,
            b,
            inv_freq,
            &mut scratch,
            pool,
            None,
        );
        st.pos = pos0 + b;
        // Said once. A gate that compares a batched prompt against a walked
        // one proves nothing if the batch quietly declined — the numbers match
        // because the same code produced both. This line is what tells the
        // two apart.
        {
            static SAID: std::sync::Once = std::sync::Once::new();
            SAID.call_once(|| tracing::warn!("dsv4: префилл пакетами по {b}"));
        }
        // Only the last token's logits are read; the rest of the chunk exists
        // to fill the caches. The head consumes the hyper-connection state,
        // not the chain's intermediate fold — skipping this final learned
        // fold used to make a full-device batch fast and wrong.
        if want_logits {
            let last = &states[(b - 1) * hc * dim..b * hc * dim];
            let mut h = vec![0.0f32; dim];
            hc_head_fold(
                last,
                &g.hc_head_fn,
                g.hc_head_scale,
                &g.hc_head_base,
                cfg,
                pool,
                &mut h,
            );
            rms_weighted(&mut h, &g.norm, cfg.norm_eps);
            logits.resize(cfg.vocab, 0.0);
            g.head.matvec(&h, logits, pool);
        } else {
            logits.clear();
        }
        true
    }
}

/// Everything a speculative verify must be able to put back.
///
/// Device caches roll back by restore-then-replay: the shadow puts the
/// window rings and compressor streams where they were BEFORE the pass, and
/// the replay re-appends the accepted tokens' state from the hidden inputs
/// the pass retained. Append-only regions roll back by count. Host-owned
/// tail layers roll back by clone-and-rewalk.
#[cfg(feature = "gpu")]
pub struct Dsv4SpecTxn {
    pos0: usize,
    batch: usize,
    pub(crate) gpu_end: usize,
    dev_filled: Vec<usize>,
    dev_n_comp: Vec<usize>,
    dev_n_ix: Vec<usize>,
    host: Vec<(usize, HostLayerSnap)>,
    /// Per host layer, per verified token: the layer's state right after
    /// that token's attention — what a rollback restores INSTEAD of
    /// re-walking the tail it already walked (the values are identical;
    /// only the side effects were ever needed).
    host_steps: Vec<(usize, Vec<HostLayerSnap>)>,
    /// Every token's hyper-connection state as it left the device prefix,
    /// BEFORE the host tail walked (and mutated) anything: the rewalk's
    /// input, and the head's.
    pub states: Vec<f32>,
    shadow: Option<crate::gpu_wgpu::Dsv4SpecShadow>,
}

#[cfg(feature = "gpu")]
struct HostLayerSnap {
    window: Vec<f32>,
    compressed: Vec<f32>,
    index_kv: Vec<f32>,
    pending_kv: Vec<f32>,
    pending_score: Vec<f32>,
    prev_kv: Vec<f32>,
    prev_score: Vec<f32>,
    pending_ix_kv: Vec<f32>,
    pending_ix_score: Vec<f32>,
    prev_ix_kv: Vec<f32>,
    prev_ix_score: Vec<f32>,
}

#[cfg(feature = "gpu")]
fn host_snap(st: &Dsv4State, li: usize) -> HostLayerSnap {
    HostLayerSnap {
        window: st.window[li].clone(),
        compressed: st.compressed[li].clone(),
        index_kv: st.index_kv[li].clone(),
        pending_kv: st.pending_kv[li].clone(),
        pending_score: st.pending_score[li].clone(),
        prev_kv: st.prev_kv[li].clone(),
        prev_score: st.prev_score[li].clone(),
        pending_ix_kv: st.pending_ix_kv[li].clone(),
        pending_ix_score: st.pending_ix_score[li].clone(),
        prev_ix_kv: st.prev_ix_kv[li].clone(),
        prev_ix_score: st.prev_ix_score[li].clone(),
    }
}

#[cfg(feature = "gpu")]
fn host_restore(st: &mut Dsv4State, li: usize, s: &HostLayerSnap) {
    st.window[li] = s.window.clone();
    st.compressed[li] = s.compressed.clone();
    st.index_kv[li] = s.index_kv.clone();
    st.pending_kv[li] = s.pending_kv.clone();
    st.pending_score[li] = s.pending_score.clone();
    st.prev_kv[li] = s.prev_kv.clone();
    st.prev_score[li] = s.prev_score.clone();
    st.pending_ix_kv[li] = s.pending_ix_kv.clone();
    st.pending_ix_score[li] = s.pending_ix_score.clone();
    st.prev_ix_kv[li] = s.prev_ix_kv.clone();
    st.prev_ix_score[li] = s.prev_ix_score.clone();
}

/// One host-tail walk of token `t`'s state through layers `gpu_end..`,
/// mutating `state` in place and the layers' host caches. Exactly the loop
/// the batch runs, factored so the verify can re-run it for accepted tokens.
#[cfg(feature = "gpu")]
#[allow(clippy::too_many_arguments)]
fn host_tail_walk(
    g: &Dsv4Globals,
    layers: &[Dsv4Layer],
    cfg: &Dsv4Cfg,
    st: &mut Dsv4State,
    gpu_end: usize,
    state: &mut [f32],
    token_id: u32,
    pos: usize,
    inv_freq: &[f32],
    scratch: &mut HcScratch,
    pool: Option<&crate::pool::Pool>,
) {
    st.pos = pos;
    for (li, l) in layers.iter().enumerate().skip(gpu_end) {
        let freqs = if l.compressor.is_some() {
            &g.inv_freq_compress
        } else {
            &g.inv_freq_window
        };
        let freqs = if freqs.is_empty() {
            inv_freq
        } else {
            freqs.as_slice()
        };
        hc_block(
            state,
            &l.hc_attn_fn,
            &l.hc_attn_scale,
            &l.hc_attn_base,
            &l.attn_norm,
            cfg,
            scratch,
            pool,
            |f, o| attention_step(f, l, cfg, st, li, freqs, pool, None, o),
        );
        hc_block(
            state,
            &l.hc_ffn_fn,
            &l.hc_ffn_scale,
            &l.hc_ffn_base,
            &l.ffn_norm,
            cfg,
            scratch,
            pool,
            |f, o| {
                if host_cpu_moe() {
                    crate::gpu::cpu_scope(|| moe_step(f, l, cfg, token_id, li, pool, o))
                } else {
                    moe_step(f, l, cfg, token_id, li, pool, o)
                }
            },
        );
        dspark_note(li, state, cfg);
    }
}

/// The host tail for a whole batch: attention stays causal per token (its
/// window mutates), the MoE half runs through the block-grouped path — the
/// same accumulation order as the position walk, which the block tests pin
/// bit for bit. This is the verify's tail; the single-token paths keep
/// `hc_block`.
#[cfg(feature = "gpu")]
#[allow(clippy::too_many_arguments)]
fn host_tail_walk_batch(
    g: &Dsv4Globals,
    layers: &[Dsv4Layer],
    cfg: &Dsv4Cfg,
    st: &mut Dsv4State,
    gpu_end: usize,
    states: &mut [f32],
    ids: &[u32],
    pos0: usize,
    b: usize,
    inv_freq: &[f32],
    scratch: &mut HcScratch,
    pool: Option<&crate::pool::Pool>,
    mut steps: Option<&mut Vec<(usize, Vec<HostLayerSnap>)>>,
) {
    let (hc, dim) = (cfg.hc_mult, cfg.dim);
    let mix_hc = (2 + hc) * hc;
    let mut folds = vec![0.0f32; b * dim];
    let mut mo = vec![0.0f32; b * dim];
    let mut posts = vec![0.0f32; b * hc];
    let mut combs = vec![0.0f32; b * hc * hc];
    let mut resid = vec![0.0f32; b * hc * dim];
    let spec_time = {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| std::env::var("CMF_DSV4_SPEC_TIME").is_ok_and(|v| v != "0"))
    };
    for (li, l) in layers.iter().enumerate().skip(gpu_end) {
        let t_attn = std::time::Instant::now();
        let freqs = if l.compressor.is_some() {
            &g.inv_freq_compress
        } else {
            &g.inv_freq_window
        };
        let freqs = if freqs.is_empty() {
            inv_freq
        } else {
            freqs.as_slice()
        };
        for t in 0..b {
            st.pos = pos0 + t;
            let state = &mut states[t * hc * dim..(t + 1) * hc * dim];
            hc_block(
                state,
                &l.hc_attn_fn,
                &l.hc_attn_scale,
                &l.hc_attn_base,
                &l.attn_norm,
                cfg,
                scratch,
                pool,
                |f, o| attention_step(f, l, cfg, st, li, freqs, pool, None, o),
            );
            if let Some(steps) = steps.as_mut() {
                match steps.iter_mut().find(|(l, _)| *l == li) {
                    Some((_, v)) => v.push(host_snap(st, li)),
                    None => steps.push((li, vec![host_snap(st, li)])),
                }
            }
        }
        let t_glue = std::time::Instant::now();
        for t in 0..b {
            let state = &states[t * hc * dim..(t + 1) * hc * dim];
            hc_mixes(
                state,
                &l.hc_ffn_fn,
                mix_hc,
                cfg.norm_eps,
                pool,
                &mut scratch.mixes,
            );
            hc_split_sinkhorn(
                &scratch.mixes,
                &l.hc_ffn_scale,
                &l.hc_ffn_base,
                hc,
                cfg.hc_sinkhorn_iters,
                cfg.hc_eps,
                &mut scratch.pre,
                &mut posts[t * hc..(t + 1) * hc],
                &mut combs[t * hc * hc..(t + 1) * hc * hc],
            );
            let fold = &mut folds[t * dim..(t + 1) * dim];
            hc_fold(state, &scratch.pre, hc, dim, fold);
            let ms = fold.iter().map(|v| v * v).sum::<f32>() / dim as f32;
            let inv = 1.0 / (ms + cfg.norm_eps).sqrt();
            for (v, w) in fold.iter_mut().zip(&l.ffn_norm) {
                *v = *v * inv * w;
            }
            resid[t * hc * dim..(t + 1) * hc * dim]
                .copy_from_slice(&states[t * hc * dim..(t + 1) * hc * dim]);
        }
        let t_moe = std::time::Instant::now();
        // A tail layer with a device expert pack (partial or full) runs its
        // hot winners on the card per token and completes the cold ones on
        // the host — the same exact split the partial walk uses. Default on
        // (measured: the tail fell 27.4 → 18.2 ms of the verify round);
        // `CMF_DSV4_TAIL_PACK=0` restores the batched host block.
        let tail_pack = {
            static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
            *ON.get_or_init(|| {
                std::env::var("CMF_DSV4_TAIL_PACK")
                    .map(|v| v != "0")
                    .unwrap_or(true)
            })
        };
        let mut packed_done = false;
        if tail_pack && pack_for(l, cfg, li).is_some() {
            packed_done = true;
            for t in 0..b {
                let f = &folds[t * dim..(t + 1) * dim];
                let forced = l.tid2eid.as_ref().map(|tbl| {
                    hash_route(tbl, cfg.vocab, cfg.top_k, ids.get(t).copied().unwrap_or(0))
                });
                let o = &mut mo[t * dim..(t + 1) * dim];
                match moe_frame(f, l, cfg, li, &[], forced.as_deref(), pool, None, None, o) {
                    Some((cold_sum, n)) => {
                        if n > 0 {
                            for (od, cd) in o.iter_mut().zip(cold_sum.iter()) {
                                *od += cd;
                            }
                        }
                    }
                    None => {
                        packed_done = false;
                        break;
                    }
                }
            }
        }
        if !packed_done {
            if host_cpu_moe() {
                crate::gpu::cpu_scope(|| moe_step_block(&folds, b, l, cfg, ids, li, pool, &mut mo));
            } else {
                moe_step_block(&folds, b, l, cfg, ids, li, pool, &mut mo);
            }
        }
        let t_exp = std::time::Instant::now();
        for t in 0..b {
            let state = &mut states[t * hc * dim..(t + 1) * hc * dim];
            hc_expand(
                &mo[t * dim..(t + 1) * dim],
                &resid[t * hc * dim..(t + 1) * hc * dim],
                &posts[t * hc..(t + 1) * hc],
                &combs[t * hc * hc..(t + 1) * hc * hc],
                hc,
                dim,
                state,
            );
            dspark_note(li, state, cfg);
        }
        if spec_time {
            eprintln!(
                "хвост слоя {li}: attn {:.1} мс, клей {:.1}, moe {:.1}, expand {:.1}",
                (t_glue - t_attn).as_secs_f64() * 1e3,
                (t_moe - t_glue).as_secs_f64() * 1e3,
                (t_exp - t_moe).as_secs_f64() * 1e3,
                t_exp.elapsed().as_secs_f64() * 1e3,
            );
        }
    }
}

/// One exact token-axis layer over a partial expert pack.  The device runs
/// attention, routing over all experts, the resident MoE rows and the
/// hyper-connection join once for the whole batch.  Cold winners are grouped
/// by expert on the host, corrected into the returned state, and only then is
/// the next dependent layer allowed to start.
#[cfg(feature = "gpu")]
#[allow(clippy::too_many_arguments)]
fn partial_layer_batch(
    g: &Dsv4Globals,
    layers: &[Dsv4Layer],
    cfg: &Dsv4Cfg,
    st: &mut Dsv4State,
    li: usize,
    states: &mut [f32],
    ids: &[u32],
    pos0: usize,
    b: usize,
    pool: Option<&crate::pool::Pool>,
) -> bool {
    let l = &layers[li];
    let (hc, dim, hd) = (cfg.hc_mult, cfg.dim, cfg.head_dim);
    if states.len() < b * hc * dim || ids.len() < b {
        return false;
    }
    let Some(pk) = pack_for(l, cfg, li) else {
        return false;
    };
    if pk.route_complete() {
        return false;
    }
    let Some(model) = l.experts.first().and_then(|e| e.w1.model_arc()) else {
        return false;
    };
    let (Some(wq_a), Some(wq_b), Some(wo_a), Some(wo_b), Some(wkv)) = (
        l.wq_a.model_idx(),
        l.wq_b.model_idx(),
        l.wo_a.model_idx(),
        l.wo_b.model_idx(),
        l.wkv.model_idx(),
    ) else {
        return false;
    };
    let comp = match &l.compressor {
        None => None,
        Some(cp) => {
            let (Some(a), Some(bx)) = (cp.wkv.model_idx(), cp.wgate.model_idx()) else {
                return false;
            };
            Some((
                crate::gpu_wgpu::Dsv4CompW {
                    wkv: a,
                    wgate: bx,
                    norm: &cp.norm,
                    ape: &cp.ape,
                },
                crate::gpu_wgpu::Dsv4CompGeom {
                    width: cp.wkv.rows(),
                    hidden: dim,
                    ratio: cp.ratio,
                    overlap: cp.overlap,
                    rope_dim: cfg.rope_head_dim,
                    eps: cfg.norm_eps,
                },
            ))
        }
    };
    let ix = match &l.indexer {
        None => None,
        Some(ixr) => {
            let cp = &ixr.compressor;
            let (Some(a), Some(bx), Some(qb), Some(wp)) = (
                cp.wkv.model_idx(),
                cp.wgate.model_idx(),
                ixr.wq_b.model_idx(),
                ixr.weights_proj.model_idx(),
            ) else {
                return false;
            };
            let ih = ixr.weights_proj.rows();
            Some((
                crate::gpu_wgpu::Dsv4CompW {
                    wkv: a,
                    wgate: bx,
                    norm: &cp.norm,
                    ape: &cp.ape,
                },
                crate::gpu_wgpu::Dsv4CompGeom {
                    width: cp.wkv.rows(),
                    hidden: dim,
                    ratio: cp.ratio,
                    overlap: cp.overlap,
                    rope_dim: cfg.rope_head_dim,
                    eps: cfg.norm_eps,
                },
                crate::gpu_wgpu::Dsv4IxW {
                    wq_b: qb,
                    weights_proj: wp,
                },
                crate::gpu_wgpu::Dsv4IxGeom {
                    ih,
                    idim: ixr.wq_b.rows() / ih.max(1),
                    q_lora: cfg.q_lora_rank,
                    hidden: dim,
                    rope_dim: cfg.rope_head_dim,
                    eps: cfg.norm_eps,
                    top_k: cfg.index_topk,
                    window: cfg.window,
                },
            ))
        }
    };
    let ew_c = comp.as_ref().map_or(
        0,
        |(_, cg)| {
            if cg.overlap { cg.width / 2 } else { cg.width }
        },
    );
    let ew_i = ix.as_ref().map_or(
        0,
        |(_, cg, _, _)| {
            if cg.overlap { cg.width / 2 } else { cg.width }
        },
    );
    let comp_extra = comp
        .as_ref()
        .map_or(0, |(_, cg)| b.div_ceil(cg.ratio.max(1)));
    let need = cfg.window * hd + (st.dev_n_comp[li] + comp_extra + 1) * ew_c.max(1) + (b + 1) * hd;
    if !crate::gpu_wgpu::dsv4_cache_ensure(st.kv_id, li, need.next_power_of_two()) {
        return false;
    }
    let base = crate::gpu_wgpu::Dsv4Prep {
        wkv,
        kv_norm: &l.kv_norm,
        comp,
        ix,
        filled: st.dev_filled[li],
        window: cfg.window,
        n_comp: st.dev_n_comp[li],
        n_ix: st.dev_n_ix[li],
        comp_dst_off: cfg.window * hd + st.dev_n_comp[li] * ew_c,
        ix_dst_off: st.dev_n_ix[li] * ew_i,
        idx_cap: cfg.window
            + if l.indexer.is_some() {
                cfg.index_topk
            } else {
                st.dev_n_comp[li] + comp_extra + 1
            },
    };
    let mut preps = Vec::with_capacity(b);
    for t in 0..b {
        let mut p = base.clone();
        p.filled = (base.filled + t).min(base.window);
        let advanced = |ratio: usize| -> usize {
            if ratio == 0 {
                0
            } else {
                (0..t).filter(|k| (pos0 + k + 1) % ratio == 0).count()
            }
        };
        if let Some((_, cg)) = base.comp.as_ref() {
            p.n_comp = base.n_comp + advanced(cg.ratio);
            p.comp_dst_off = base.comp_dst_off + (p.n_comp - base.n_comp) * ew_c;
        }
        if let Some((_, cg, _, _)) = base.ix.as_ref() {
            p.n_ix = base.n_ix + advanced(cg.ratio);
            p.ix_dst_off = base.ix_dst_off + (p.n_ix - base.n_ix) * ew_i;
        }
        preps.push(p);
    }

    // Every row enters with an exact host state because the previous partial
    // layer was corrected before returning.  Seed the token-axis slots and
    // the q-LoRA vector the batched attention consumes.
    for t in 0..b {
        let state = &states[t * hc * dim..(t + 1) * hc * dim];
        let (fold, post, comb) = hc_fold_norm(
            state,
            &l.hc_attn_fn,
            &l.hc_attn_scale,
            &l.hc_attn_base,
            &l.attn_norm,
            cfg,
            pool,
        );
        let mut qn = vec![0.0f32; cfg.q_lora_rank];
        l.wq_a.matvec(&fold, &mut qn, pool);
        rms_weighted(&mut qn, &l.q_norm, cfg.norm_eps);
        if !crate::gpu_wgpu::dsv4_chain_seed_bt(t, b, state, &post, &comb, &fold, &qn) {
            return false;
        }
    }
    let forced_rows: Vec<Option<Vec<usize>>> = ids
        .iter()
        .take(b)
        .map(|&id| {
            l.tid2eid
                .as_ref()
                .map(|tbl| hash_route(tbl, cfg.vocab, cfg.top_k, id))
        })
        .collect();
    let dynv = pk.dynslots.lock().unwrap();
    let w = crate::gpu_wgpu::Dsv4LayerW {
        attn: crate::gpu_wgpu::Dsv4AttnW {
            wq_a,
            wq_b,
            wo_a,
            wo_b,
            q_norm: &l.q_norm,
            sink: &l.attn_sink,
        },
        moe: crate::gpu_wgpu::Dsv4MoeW {
            router: &[],
            experts: &pk.tensors,
            logits: &[],
            bias: pk.bias.as_deref(),
            mask: pk.mask.as_deref(),
            forced: None,
            remap: Some(&dynv.remap),
            global: None,
            has_shared: true,
            shared_weight: 1.0,
            preweighted: false,
            qwen_softmax: false,
        },
        hc_ffn_fn: &l.hc_ffn_fn,
        hc_ffn_scale: &l.hc_ffn_scale,
        hc_ffn_base: &l.hc_ffn_base,
        // Stop after this layer's state.  The next-layer fold must see the
        // cold-corrected state, not the resident-only state on the card.
        hc_next_fn: None,
        hc_next_scale: &l.hc_attn_scale,
        hc_next_base: &l.hc_attn_base,
        ffn_norm: &l.ffn_norm,
        next_norm: &l.attn_norm,
        next_q_norm: &l.q_norm,
        next_wq_a: None,
        router: &pk.router,
    };
    let geom = crate::gpu_wgpu::Dsv4LayerGeom {
        attn: crate::gpu_wgpu::Dsv4AttnGeom {
            dim,
            nh: cfg.n_heads,
            hd,
            rd: cfg.rope_head_dim,
            q_lora: cfg.q_lora_rank,
            o_lora: cfg.o_lora_rank,
            o_groups: cfg.o_groups,
            eps: cfg.norm_eps,
            scale: (hd as f32).powf(-0.5),
        },
        moe: crate::gpu_wgpu::Dsv4MoeGeom {
            hidden: dim,
            inter: cfg.moe_inter,
            top_k: cfg.top_k,
            route_scale: cfg.route_scale,
            swiglu_limit: cfg.swiglu_limit,
            gu_q2: l
                .experts
                .first()
                .is_some_and(|e| e.w1.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP)),
        },
        hc,
        hc_eps: cfg.hc_eps,
        sinkhorn_iters: cfg.hc_sinkhorn_iters,
    };
    let freqs = if l.compressor.is_some() {
        g.inv_freq_compress.as_slice()
    } else {
        g.inv_freq_window.as_slice()
    };
    let Some(mut got) = crate::gpu_wgpu::dsv4_layer_batch_partial(
        &model,
        &w,
        geom,
        st.kv_id,
        li,
        b,
        &preps,
        Some(&forced_rows),
        freqs,
        pos0,
    ) else {
        return false;
    };
    drop(dynv);
    let mut cold_sum = vec![0.0f32; b * dim];
    cold_step_block(&got.cold_x, b, l, cfg, &got.cold, pool, &mut cold_sum);
    for t in 0..b {
        let state = &mut got.states[t * hc * dim..(t + 1) * hc * dim];
        let post = &got.posts[t * hc..(t + 1) * hc];
        let cold = &cold_sum[t * dim..(t + 1) * dim];
        for j in 0..hc {
            for d in 0..dim {
                state[j * dim + d] += post[j] * cold[d];
            }
        }
    }
    states[..b * hc * dim].copy_from_slice(&got.states);
    if !crate::gpu_wgpu::dsv4_spec_cap_write_host(li, b, hc * dim, &got.states) {
        return false;
    }
    if let Some(first_route) = got.routed.first() {
        refill_route_slots(l, cfg, &pk, first_route);
    }
    for t in 0..b {
        let pos = pos0 + t;
        st.dev_filled[li] = (st.dev_filled[li] + 1).min(cfg.window);
        if let Some((_, cg, ..)) = base.ix.as_ref() {
            if (pos + 1) % cg.ratio == 0 {
                st.dev_n_ix[li] += 1;
            }
        }
        if let Some((_, cg)) = base.comp.as_ref() {
            if (pos + 1) % cg.ratio == 0 {
                st.dev_n_comp[li] += 1;
                note_compressed(st.kv_id, li, st.dev_n_comp[li]);
            }
        }
    }
    true
}

/// A speculative verify pass: run `ids` (the committed next token followed
/// by draft proposals) at positions `pos0..pos0+B` through the trunk in one
/// batched submission, WITHOUT giving up the ability to roll back, and
/// return every position's greedy answer. The caller decides the accepted
/// prefix and calls [`dsv4_spec_finish`], which either keeps everything
/// (`accepted == B`) or restores-and-replays to the accepted length.
///
/// `logits_out` takes B rows of vocab logits, `argmax_out` their argmaxes.
#[cfg(feature = "gpu")]
#[allow(clippy::too_many_arguments)]
pub fn dsv4_verify_chunk(
    g: &Dsv4Globals,
    layers: &[Dsv4Layer],
    cfg: &Dsv4Cfg,
    st: &mut Dsv4State,
    ids: &[u32],
    pos0: usize,
    inv_freq: &[f32],
    pool: Option<&crate::pool::Pool>,
    cap_targets: &[usize],
    argmax_out: &mut Vec<u32>,
    logits_out: &mut Vec<f32>,
    walked_out: &mut Vec<f32>,
) -> Option<Dsv4SpecTxn> {
    let b = ids.len();
    // The complete prefix still runs as one fused chain.  Immediately after
    // it, contiguous PARTIAL packs can now stay on the device too: each one
    // is corrected with its cold experts before the next layer is seeded.
    let complete_end = st
        .dev_set
        .iter()
        .enumerate()
        .position(|(li, &on)| {
            !on || pack_for(&layers[li], cfg, li).is_none_or(|p| !p.route_complete())
        })
        .unwrap_or(st.dev_set.len());
    fn partial_batch_on() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| {
            std::env::var("CMF_DSV4_PARTIAL_BATCH")
                .map(|v| v != "0")
                .unwrap_or(true)
        })
    }
    let device_end = if partial_batch_on() {
        (complete_end..layers.len())
            .take_while(|&li| {
                st.dev_set.get(li).copied().unwrap_or(false)
                    && st.partial_set.get(li).copied().unwrap_or(false)
                    && pack_for(&layers[li], cfg, li).is_some_and(|p| !p.route_complete())
            })
            .last()
            .map_or(complete_end, |li| li + 1)
    } else {
        complete_end
    };
    // A PARTIAL layer after a host gap is allowed to walk in the host tail.
    // A FULL device layer there would violate the contiguous-prefix contract.
    let full_beyond = st.dev_set[device_end.min(st.dev_set.len())..]
        .iter()
        .enumerate()
        .any(|(i, &on)| on && !st.partial_set.get(device_end + i).copied().unwrap_or(false));
    // `CMF_DSV4_HOST_VERIFY=1` lets the verify run with NO device prefix:
    // every layer walks in the host tail, batched — which is where a
    // many-core host amortises the weight read and the unpack across the
    // draft (the whole point of a batched verify). Off, a partial layer 0
    // (dynamic-slot packs) silently priced the entire speculation at zero:
    // 625 drafted, 0 verified, all cost and no candidate.
    fn host_verify_on() -> bool {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| {
            std::env::var("CMF_DSV4_HOST_VERIFY")
                .map(|v| v != "0")
                // On a small card every layer can have a useful partial pack
                // while no layer has a complete one.  The exact batched tail
                // is specifically built for that shape; silently pricing
                // speculation at zero here defeated the automatic fast path.
                .unwrap_or(true)
        })
    }
    if b < 2
        || !chain_enabled()
        || !st.dev_owned
        || st.dev_set.len() != layers.len()
        || (device_end == 0 && !host_verify_on())
        || full_beyond
    {
        return None;
    }
    let (hc, dim, hd) = (cfg.hc_mult, cfg.dim, cfg.head_dim);
    // ── the transaction ──
    let metas: Vec<(usize, usize, usize, usize)> = (0..device_end)
        .map(|li| (li, hd, cfg.window, st.dev_filled[li]))
        .collect();
    let shadow = crate::gpu_wgpu::dsv4_spec_shadow(st.kv_id, &metas, b)?;
    let mut txn = Dsv4SpecTxn {
        pos0,
        batch: b,
        gpu_end: device_end,
        dev_filled: st.dev_filled.clone(),
        dev_n_comp: st.dev_n_comp.clone(),
        dev_n_ix: st.dev_n_ix.clone(),
        host: (device_end..layers.len())
            .map(|li| (li, host_snap(st, li)))
            .collect(),
        states: Vec::new(),
        host_steps: Vec::new(),
        shadow: Some(shadow),
    };
    // The capture targets that live on the device: photograph their states.
    let dev_caps: Vec<usize> = cap_targets
        .iter()
        .copied()
        .filter(|&t| t < device_end)
        .collect();
    crate::gpu_wgpu::dsv4_spec_retain_arm(device_end, &dev_caps);

    // ── seed and run the batch (the prefill batch's own shape) ──
    let mut emb = vec![0.0f32; dim];
    let mut states = vec![0.0f32; b * hc * dim];
    for (t, &id) in ids.iter().enumerate() {
        let mut state = vec![0.0f32; hc * dim];
        g.embed.row_f32(id as usize, &mut emb);
        for j in 0..hc {
            state[j * dim..(j + 1) * dim].copy_from_slice(&emb);
        }
        states[t * hc * dim..(t + 1) * hc * dim].copy_from_slice(&state);
        let (folded, post0, comb0) = hc_fold_norm(
            &state,
            &layers[0].hc_attn_fn,
            &layers[0].hc_attn_scale,
            &layers[0].hc_attn_base,
            &layers[0].attn_norm,
            cfg,
            pool,
        );
        let mut qn0 = vec![0.0f32; layers[0].wq_a.rows()];
        layers[0].wq_a.matvec(&folded, &mut qn0, pool);
        rms_weighted(&mut qn0, &layers[0].q_norm, cfg.norm_eps);
        if !crate::gpu_wgpu::dsv4_state_write_t(&state, t)
            || !crate::gpu_wgpu::dsv4_hc_write_t(&post0, &comb0, t)
            || !crate::gpu_wgpu::dsv4_chain_seed_t(&folded, &qn0, t)
            || !crate::gpu_wgpu::dsv4_chain_seed_bt(t, b, &state, &post0, &comb0, &folded, &qn0)
        {
            crate::gpu_wgpu::dsv4_spec_retain_arm(0, &[]);
            return None;
        }
    }
    let spec_time = {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| std::env::var("CMF_DSV4_SPEC_TIME").is_ok_and(|v| v != "0"))
    };
    let t0 = std::time::Instant::now();
    let mut folded = Vec::new();
    st.pos = pos0;
    // Diagnostic split: the fused prefix normally has one fence.  Splitting
    // only while fingerprinting tells us which complete layer first departs
    // from scalar decode; it must never become a user-facing tuning flag.
    let fp_split = verify_fp_on(pos0) && std::env::var("CMF_DSV4_FP_SPLIT").is_ok_and(|v| v != "0");
    let mut ok = true;
    if fp_split {
        for li in 0..complete_end {
            let one = [li];
            ok = dsv4_chain_run(
                layers,
                &one,
                cfg,
                g,
                st,
                *ids.last().unwrap(),
                &mut folded,
                Some(&mut states),
                b,
                ids,
                li == 0,
                pool,
            );
            if !ok {
                break;
            }
            verify_fp("verify", pos0, li, &states[..hc * dim]);
        }
    } else if complete_end > 0 {
        let run: Vec<usize> = (0..complete_end).collect();
        ok = dsv4_chain_run(
            layers,
            &run,
            cfg,
            g,
            st,
            *ids.last().unwrap(),
            &mut folded,
            Some(&mut states),
            b,
            ids,
            true,
            pool,
        );
        if ok {
            verify_fp("verify", pos0, complete_end - 1, &states[..hc * dim]);
        }
    }
    if ok {
        for li in complete_end..device_end {
            if !partial_layer_batch(g, layers, cfg, st, li, &mut states, ids, pos0, b, pool) {
                ok = false;
                break;
            }
            verify_fp("verify", pos0, li, &states[..hc * dim]);
        }
    }
    crate::gpu_wgpu::dsv4_spec_retain_arm(0, &[]);
    if !ok {
        // Nothing committed on the host; the device may hold half-appended
        // state, so put the snapshot back before declining.
        if let Some(sh) = txn.shadow.take() {
            let _ = crate::gpu_wgpu::dsv4_spec_restore(&sh);
        }
        st.dev_filled = txn.dev_filled;
        st.dev_n_comp = txn.dev_n_comp;
        st.dev_n_ix = txn.dev_n_ix;
        st.pos = pos0;
        return None;
    }
    txn.states = states.clone();
    let t_chain = t0.elapsed();
    if std::env::var("CMF_DSV4_FOLD_DBG").is_ok() {
        // Any indexer fold this window landed: read the entry back and
        // print a fingerprint, so the fused and per-token folds can be
        // held against each other on the release shapes.
        for li in 0..device_end {
            let Some(ixr) = &layers[li].indexer else {
                continue;
            };
            let ratio = ixr.compressor.ratio;
            for t in 0..b {
                if (pos0 + t + 1) % ratio == 0 {
                    let ew = {
                        let w = ixr.compressor.wkv.rows();
                        if ixr.compressor.overlap { w / 2 } else { w }
                    };
                    let idx_new = txn.dev_n_ix[li]
                        + (0..=t).filter(|k| (pos0 + k + 1) % ratio == 0).count()
                        - 1;
                    if let Some(v) =
                        crate::gpu_wgpu::dsv4_dbg_read_ix(st.kv_id, li, idx_new * ew, ew.min(8))
                    {
                        let sum: f32 = v.iter().sum();
                        eprintln!(
                            "[fold] li={li} pos={} entry={idx_new} head={:?} sum={sum:.6}",
                            pos0 + t,
                            &v[..4.min(v.len())]
                        );
                    }
                }
            }
        }
    }

    // ── host tail + every position's head ──
    let mut scratch = HcScratch::new(cfg);
    argmax_out.clear();
    logits_out.clear();
    logits_out.resize(b * cfg.vocab, 0.0);
    let mut head_in = vec![0.0f32; b * dim];
    let mut host_steps: Vec<(usize, Vec<HostLayerSnap>)> = Vec::new();
    host_tail_walk_batch(
        g,
        layers,
        cfg,
        st,
        device_end,
        &mut states,
        ids,
        pos0,
        b,
        inv_freq,
        &mut scratch,
        pool,
        Some(&mut host_steps),
    );
    txn.host_steps = host_steps;
    for t in 0..b {
        let state = &states[t * hc * dim..(t + 1) * hc * dim];
        let h = &mut head_in[t * dim..(t + 1) * dim];
        hc_head_fold(
            state,
            &g.hc_head_fn,
            g.hc_head_scale,
            &g.hc_head_base,
            cfg,
            pool,
            h,
        );
        rms_weighted(h, &g.norm, cfg.norm_eps);
    }
    // The experimental B-wide head uses a different reduction kernel from
    // ordinary decode.  On the release q4tp it changed row-zero argmax under
    // a force-reject transaction, so it is diagnostic-only until parity is
    // proven; speculative execution must inherit the canonical head exactly.
    let batch_head = std::env::var("CMF_DSV4_SPEC_BATCH_HEAD").is_ok_and(|v| v != "0");
    let head_gpu = batch_head
        && g.head.model_idx().is_some_and(|hi| {
            let model = layers[0].experts.first().and_then(|e| e.w1.model_arc());
            model.is_some_and(|m| {
                crate::gpu_wgpu::q4tp_matvec_batch_for_test(
                    &m, hi, &head_in, b, cfg.vocab, dim, logits_out,
                )
            })
        });
    for t in 0..b {
        if !head_gpu {
            let h = &head_in[t * dim..(t + 1) * dim];
            g.head
                .matvec(h, &mut logits_out[t * cfg.vocab..(t + 1) * cfg.vocab], pool);
        }
        let row = &logits_out[t * cfg.vocab..(t + 1) * cfg.vocab];
        let mut best = 0usize;
        for v in 1..cfg.vocab {
            if row[v] > row[best] {
                best = v;
            }
        }
        argmax_out.push(best as u32);
    }
    walked_out.clear();
    walked_out.extend_from_slice(&states);
    st.pos = pos0 + b;
    if spec_time {
        eprintln!(
            "verify: тень+сид+цепочка {:.1} мс, хвост+голова {:.1} мс",
            t_chain.as_secs_f64() * 1e3,
            (t0.elapsed() - t_chain).as_secs_f64() * 1e3,
        );
    }
    Some(txn)
}

/// Keep the accepted prefix of a verify pass and put everything else back.
///
/// `accepted` counts the FED tokens whose state stays (at least 1 — the
/// first fed token was already committed by the caller). With
/// `accepted == batch` this is free; otherwise the device restores its
/// snapshot and replays the accepted tokens' state appends, and the host
/// tail re-walks them.
#[cfg(feature = "gpu")]
pub fn dsv4_spec_finish(
    g: &Dsv4Globals,
    layers: &[Dsv4Layer],
    cfg: &Dsv4Cfg,
    st: &mut Dsv4State,
    mut txn: Dsv4SpecTxn,
    accepted: usize,
    ids: &[u32],
    inv_freq: &[f32],
    pool: Option<&crate::pool::Pool>,
) -> bool {
    macro_rules! sfail {
        ($($t:tt)*) => {{
            if std::env::var("CMF_DSV4_SPEC_DEBUG").is_ok() {
                eprintln!("spec_finish: {}", format_args!($($t)*));
            }
            return false;
        }};
    }
    let b = txn.batch;
    let k = accepted.min(b);
    let (hc, dim, hd) = (cfg.hc_mult, cfg.dim, cfg.head_dim);
    // The staged batch never slid the windows; land the accepted prefix now,
    // whatever k is.
    let win_metas: Vec<(usize, usize, usize, usize)> = (0..txn.gpu_end)
        .map(|li| (li, txn.dev_filled[li], cfg.window, hd))
        .collect();
    if !crate::gpu_wgpu::dsv4_spec_commit_windows(st.kv_id, &win_metas, b, k) {
        sfail!("коммит окон");
    }
    if k == b {
        // Every stream mutation was the walk's own kernels in walk order —
        // nothing to put back.
        return true;
    }
    // ── device: restore to the snapshot, then replay the accepted tokens ──
    let Some(sh) = txn.shadow.take() else {
        sfail!("нет тени")
    };
    if !crate::gpu_wgpu::dsv4_spec_restore(&sh) {
        sfail!("restore");
    }
    let Some(model) = layers[0].experts.first().and_then(|e| e.w1.model_arc()) else {
        sfail!("нет модели");
    };
    let mut plan: Vec<(usize, crate::gpu_wgpu::Dsv4Prep)> = Vec::new();
    let mut freqs_own: Vec<&[f32]> = Vec::new();
    for li in 0..txn.gpu_end {
        let l = &layers[li];
        let Some(wkv) = l.wkv.model_idx() else {
            sfail!("wkv слоя {li}")
        };
        let comp = match &l.compressor {
            None => None,
            Some(cp) => {
                let (Some(a), Some(bx)) = (cp.wkv.model_idx(), cp.wgate.model_idx()) else {
                    sfail!("компрессор слоя {li}");
                };
                Some((
                    crate::gpu_wgpu::Dsv4CompW {
                        wkv: a,
                        wgate: bx,
                        norm: &cp.norm,
                        ape: &cp.ape,
                    },
                    crate::gpu_wgpu::Dsv4CompGeom {
                        width: cp.wkv.rows(),
                        hidden: dim,
                        ratio: cp.ratio,
                        overlap: cp.overlap,
                        rope_dim: cfg.rope_head_dim,
                        eps: cfg.norm_eps,
                    },
                ))
            }
        };
        let ix = match &l.indexer {
            None => None,
            Some(ixr) => {
                let cp = &ixr.compressor;
                let (Some(a), Some(bx), Some(qb), Some(wp)) = (
                    cp.wkv.model_idx(),
                    cp.wgate.model_idx(),
                    ixr.wq_b.model_idx(),
                    ixr.weights_proj.model_idx(),
                ) else {
                    sfail!("индексер слоя {li}");
                };
                let ih = ixr.weights_proj.rows();
                Some((
                    crate::gpu_wgpu::Dsv4CompW {
                        wkv: a,
                        wgate: bx,
                        norm: &cp.norm,
                        ape: &cp.ape,
                    },
                    crate::gpu_wgpu::Dsv4CompGeom {
                        width: cp.wkv.rows(),
                        hidden: dim,
                        ratio: cp.ratio,
                        overlap: cp.overlap,
                        rope_dim: cfg.rope_head_dim,
                        eps: cfg.norm_eps,
                    },
                    crate::gpu_wgpu::Dsv4IxW {
                        wq_b: qb,
                        weights_proj: wp,
                    },
                    crate::gpu_wgpu::Dsv4IxGeom {
                        ih,
                        idim: ixr.wq_b.rows() / ih.max(1),
                        q_lora: cfg.q_lora_rank,
                        hidden: dim,
                        rope_dim: cfg.rope_head_dim,
                        eps: cfg.norm_eps,
                        top_k: cfg.index_topk,
                        window: cfg.window,
                    },
                ))
            }
        };
        let ew_c = comp.as_ref().map_or(
            0,
            |(_, cg)| {
                if cg.overlap { cg.width / 2 } else { cg.width }
            },
        );
        let ew_i = ix.as_ref().map_or(
            0,
            |(_, cg, _, _)| {
                if cg.overlap { cg.width / 2 } else { cg.width }
            },
        );
        let prep = crate::gpu_wgpu::Dsv4Prep {
            wkv,
            kv_norm: &l.kv_norm,
            comp,
            ix,
            filled: txn.dev_filled[li],
            window: cfg.window,
            n_comp: txn.dev_n_comp[li],
            n_ix: txn.dev_n_ix[li],
            comp_dst_off: cfg.window * hd + txn.dev_n_comp[li] * ew_c,
            ix_dst_off: txn.dev_n_ix[li] * ew_i,
            idx_cap: cfg.window
                + if l.indexer.is_some() {
                    cfg.index_topk
                } else {
                    0
                },
        };
        let fr = if l.compressor.is_some() {
            g.inv_freq_compress.as_slice()
        } else {
            g.inv_freq_window.as_slice()
        };
        freqs_own.push(if fr.is_empty() { inv_freq } else { fr });
        plan.push((li, prep));
    }
    if !crate::gpu_wgpu::dsv4_spec_replay(
        &model,
        &plan,
        st.kv_id,
        txn.pos0,
        b,
        k,
        &freqs_own,
        hd,
        dim,
        cfg.rope_head_dim,
        cfg.norm_eps,
        true,
    ) {
        sfail!("replay k={k}");
    }
    // ── host counts: the snapshot advanced by k tokens ──
    let advanced = |ratio: usize| -> usize {
        if ratio == 0 {
            return 0;
        }
        (0..k).filter(|t| (txn.pos0 + t + 1) % ratio == 0).count()
    };
    for li in 0..txn.gpu_end {
        let l = &layers[li];
        st.dev_filled[li] = (txn.dev_filled[li] + k).min(cfg.window);
        let ac = l.compressor.as_ref().map_or(0, |cp| advanced(cp.ratio));
        let ai = l
            .indexer
            .as_ref()
            .map_or(0, |ix| advanced(ix.compressor.ratio));
        st.dev_n_comp[li] = txn.dev_n_comp[li] + ac;
        st.dev_n_ix[li] = txn.dev_n_ix[li] + ai;
        note_compressed(st.kv_id, li, st.dev_n_comp[li]);
    }
    // ── host tail: the verify pass already walked these tokens; restore
    //    the per-token snapshot it took instead of walking them again. ──
    if k >= 1 && txn.host_steps.iter().all(|(_, v)| v.len() >= k) && !txn.host_steps.is_empty() {
        for (li, v) in &txn.host_steps {
            host_restore(st, *li, &v[k - 1]);
        }
    } else {
        for (li, snap) in &txn.host {
            host_restore(st, *li, snap);
        }
        let mut scratch = HcScratch::new(cfg);
        let mut states = txn.states.clone();
        host_tail_walk_batch(
            g,
            layers,
            cfg,
            st,
            txn.gpu_end,
            &mut states[..k * hc * dim],
            ids,
            txn.pos0,
            k,
            inv_freq,
            &mut scratch,
            pool,
            None,
        );
    }
    st.pos = txn.pos0 + k;
    true
}

pub fn forward_chunk(
    g: &Dsv4Globals,
    layers: &[Dsv4Layer],
    cfg: &Dsv4Cfg,
    st: &mut Dsv4State,
    ids: &[u32],
    pos0: usize,
    inv_freq: &[f32],
    pool: Option<&crate::pool::Pool>,
    logits: &mut Vec<f32>,
    want_logits: bool,
) {
    let bs = batch_prefill();
    if bs > 1 {
        // The first token walks, always. The batch will only run where every
        // layer has already proved it takes the card, and that proof is a
        // completed single-token run — with the whole prompt arriving as one
        // chunk there is otherwise no first run to give it, and the batch
        // declines for the entire prompt while a gate comparing it against
        // the walk reports agreement it never tested.
        let mut i = 0;
        if !st.dev_owned && !ids.is_empty() {
            st.pos = pos0;
            forward_token_inner(
                g,
                layers,
                cfg,
                st,
                ids[0],
                inv_freq,
                pool,
                logits,
                ids.len() == 1,
            );
            i = 1;
        }
        while i < ids.len() {
            let end = (i + bs).min(ids.len());
            st.pos = pos0 + i;
            if !forward_chunk_batched(
                g,
                layers,
                cfg,
                st,
                &ids[i..end],
                pos0 + i,
                inv_freq,
                pool,
                logits,
                want_logits && end == ids.len(),
            ) {
                break;
            }
            i = end;
        }
        if i == ids.len() {
            return;
        }
        // Refused before touching anything; the walk starts where it left off.
        for (k, &id) in ids.iter().enumerate().skip(i) {
            st.pos = pos0 + k;
            let last = want_logits && k + 1 == ids.len();
            forward_token_inner(g, layers, cfg, st, id, inv_freq, pool, logits, last);
        }
        return;
    }
    for (i, &id) in ids.iter().enumerate() {
        st.pos = pos0 + i;
        let last = want_logits && i + 1 == ids.len();
        forward_token_inner(g, layers, cfg, st, id, inv_freq, pool, logits, last);
    }
}

pub fn forward_token(
    g: &Dsv4Globals,
    layers: &[Dsv4Layer],
    cfg: &Dsv4Cfg,
    st: &mut Dsv4State,
    token_id: u32,
    inv_freq: &[f32],
    pool: Option<&crate::pool::Pool>,
    logits: &mut Vec<f32>,
) {
    forward_token_inner(g, layers, cfg, st, token_id, inv_freq, pool, logits, true);
}

#[allow(clippy::too_many_arguments)]
fn forward_token_inner(
    g: &Dsv4Globals,
    layers: &[Dsv4Layer],
    cfg: &Dsv4Cfg,
    st: &mut Dsv4State,
    token_id: u32,
    inv_freq: &[f32],
    pool: Option<&crate::pool::Pool>,
    logits: &mut Vec<f32>,
    // Prompt tokens other than the last one have their logits thrown away.
    want_logits: bool,
) {
    let _t_all = prof::on().then(std::time::Instant::now);
    let _all_guard = Charge(_t_all, &prof::ALL_NS);
    let (hc, dim) = (cfg.hc_mult, cfg.dim);

    // Embedding, replicated into the copies.
    let mut emb = vec![0.0f32; dim];
    g.embed.row_f32(token_id as usize, &mut emb);
    let mut state = vec![0.0f32; hc * dim];
    for j in 0..hc {
        state[j * dim..(j + 1) * dim].copy_from_slice(&emb);
    }

    let mut scratch = HcScratch::new(cfg);
    let mut dump: Vec<String> = Vec::new();
    if dump_path().is_some() {
        dump.push(format!("\"embed\":{}", vec_json(&emb)));
        PICKED.with(|p| p.borrow_mut().clear());
        BODY.with(|b| b.borrow_mut().clear());
        dump.push(",\"layers\":[".into());
    }
    if trace_on() {
        eprintln!(
            "[dsv4] tok={token_id} pos={} embed rms={:.5}",
            st.pos,
            rms_of(&emb)
        );
    }
    // ── one submission per layer, when the device will take it ──
    #[cfg(feature = "gpu")]
    let layer_frames = gpu_layer_enabled()
        && dsv4_layer_loop(
            &mut state,
            layers,
            g,
            cfg,
            st,
            token_id,
            inv_freq,
            pool,
            &mut scratch,
        );
    #[cfg(not(feature = "gpu"))]
    let layer_frames = false;

    // ── the fast two-frame path: hyper-connections on the card ──
    // Measured on the release, the fold, the Sinkhorn and the norms cost 19
    // ms of a 57 ms token on the host and hundredths of one on the device.
    // With both frames doing their own, the host carries nothing between a
    // layer's halves and the MoE half's input never leaves the card — one
    // readback a layer instead of two.
    #[cfg(feature = "gpu")]
    let hc_dev = hc_on_device()
        && !layer_frames
        && gpu_attn_enabled()
        && gpu_moe2_enabled()
        && dump_path().is_none();
    #[cfg(not(feature = "gpu"))]
    let hc_dev = false;
    // The device loop's verdict as a VALUE, not as a cfg-gated `if`. It used
    // to be the latter, with the CPU loop in the `else` arm — so a build
    // without the gpu feature compiled no layer loop at all and every token
    // passed through untouched. The window test said so ("sliding window
    // never filled") and only in the CPU-only build, which is the one
    // configuration the gate was not running.
    #[cfg(feature = "gpu")]
    let two_frame_done = hc_dev
        && dsv4_two_frame_loop(
            &mut state,
            layers,
            g,
            cfg,
            st,
            token_id,
            inv_freq,
            pool,
            &mut scratch,
        );
    #[cfg(not(feature = "gpu"))]
    let two_frame_done = false;
    if !two_frame_done {
        for (li, l) in layers.iter().enumerate() {
            if layer_frames {
                break;
            }
            // attention half
            hc_block(
                &mut state,
                &l.hc_attn_fn,
                &l.hc_attn_scale,
                &l.hc_attn_base,
                &l.attn_norm,
                cfg,
                &mut scratch,
                pool,
                |folded, out| {
                    if dump_path().is_some() {
                        // The body's own input and output, so the reference can be
                        // fed the port's input: then only the body can differ.
                        BODY.with(|b| b.borrow_mut().push(vec_json(folded)));
                    }
                    // The layer's kind decides its frequencies, not the model's.
                    let freqs = if l.compressor.is_some() {
                        &g.inv_freq_compress
                    } else {
                        &g.inv_freq_window
                    };
                    let freqs = if freqs.is_empty() {
                        inv_freq
                    } else {
                        freqs.as_slice()
                    };
                    attention_step(folded, l, cfg, st, li, freqs, pool, None, out);
                    if dump_path().is_some() {
                        BODY.with(|b| b.borrow_mut().push(vec_json(out)));
                    }
                },
            );
            if dump_path().is_some() {
                // After the attention half only — this is what separates an
                // attention discrepancy from an expert one.
                dump.push(format!(
                    "{}{}",
                    if li == 0 { "" } else { "," },
                    vec_json(&state)
                ));
            }
            // FFN half
            let _t_hc2 = prof::on().then(std::time::Instant::now);
            hc_block(
                &mut state,
                &l.hc_ffn_fn,
                &l.hc_ffn_scale,
                &l.hc_ffn_base,
                &l.ffn_norm,
                cfg,
                &mut scratch,
                pool,
                |folded, out| moe_step(folded, l, cfg, token_id, li, pool, out),
            );
            if let Some(t) = _t_hc2 {
                // The block's own time minus the expert step inside it — what the
                // fold, the norm and the expand cost on their own.
                prof::HC_NS.fetch_add(
                    t.elapsed().as_nanos() as u64,
                    std::sync::atomic::Ordering::Relaxed,
                );
            }
            if dump_path().is_some() {
                dump.push(format!(",{}", vec_json(&state)));
            }
            if trace_on() && (st.pos % 64 == 0 || st.pos == 199) {
                eprintln!(
                    "[dsv4]  кеши слоя {li}: окно={} сжатых={} индекс={} (ratio={:?})",
                    st.window[li].len() / cfg.head_dim.max(1),
                    st.compressed[li].len() / cfg.head_dim.max(1),
                    st.index_kv[li].len().max(1) / 128,
                    l.compressor.as_ref().map(|c| (c.ratio, c.overlap)),
                );
            }
            if trace_on() {
                let bad = state.iter().filter(|v| !v.is_finite()).count();
                eprintln!(
                    "[dsv4]  layer {li:>2}: rms={:.5}{}",
                    rms_of(&state),
                    if bad > 0 {
                        format!("  NON-FINITE x{bad}")
                    } else {
                        String::new()
                    }
                );
            }
            dspark_note(li, &state, cfg);
        }
    }
    st.pos += 1;

    // Collapse the copies, normalize, project to the vocabulary.
    let mut h = vec![0.0f32; dim];
    hc_head_fold(
        &state,
        &g.hc_head_fn,
        g.hc_head_scale,
        &g.hc_head_base,
        cfg,
        pool,
        &mut h,
    );
    if !want_logits {
        logits.clear();
        return;
    }
    let _t_head = prof::on().then(std::time::Instant::now);
    rms_weighted(&mut h, &g.norm, cfg.norm_eps);
    logits.clear();
    logits.resize(g.head.rows(), 0.0);
    g.head.matvec(&h, logits, pool);
    if let Some(t) = _t_head {
        prof::HEAD_NS.fetch_add(
            t.elapsed().as_nanos() as u64,
            std::sync::atomic::Ordering::Relaxed,
        );
    }
    if dump_path().is_some() {
        dump.push("]".into());
        let picked = PICKED.with(|p| {
            p.borrow()
                .iter()
                .map(|v| {
                    format!(
                        "[{}]",
                        v.iter()
                            .map(|e| e.to_string())
                            .collect::<Vec<_>>()
                            .join(",")
                    )
                })
                .collect::<Vec<_>>()
                .join(",")
        });
        dump.push(format!(",\"experts\":[{picked}]"));
        let body = BODY.with(|b| b.borrow().join(","));
        dump.push(format!(",\"attn_io\":[{body}]"));
        dump_line(&format!(
            "{{\"tok\":{token_id},\"pos\":{},{},\"head\":{},\"logits\":{}}}",
            st.pos - 1,
            dump.join(""),
            vec_json(&h),
            vec_json(logits)
        ));
    }
    if trace_on() {
        let (mut top, mut best) = (0usize, f32::NEG_INFINITY);
        for (i, &v) in logits.iter().enumerate() {
            if v > best {
                best = v;
                top = i;
            }
        }
        let lo = logits.iter().cloned().fold(f32::MAX, f32::min);
        eprintln!(
            "[dsv4]  head: rms={:.5} logits[{}..{:.3}] argmax={top}",
            rms_of(&h),
            format_args!("{lo:.3}"),
            best
        );
    }
}

/// Build the runtime weights from a converted `.cmf`.
///
/// Names are the converter's output (see `canon_name`'s deepseek_v4 arm):
/// attention keeps DeepSeek's own spelling under `self_attn.`, the MoE is
/// rewritten into the layout every other MoE here uses, and the hyper-
/// connection tensors ride under the layer prefix.
pub fn load(
    model: &std::sync::Arc<cortiq_core::CmfModel>,
    cfg: &Dsv4Cfg,
    n_layers: usize,
) -> Result<(Dsv4Globals, Vec<Dsv4Layer>), String> {
    let q = |name: &str| -> Result<crate::qtensor::QTensor, String> {
        crate::qtensor::QTensor::from_model(model, name)
    };
    // The small pieces — norms, the sink, ape, the hyper-connection
    // projections — are read as plain f32. They are not all 2-D (a norm is a
    // vector), so this cannot go through QTensor, which requires a matrix.
    let f = |name: &str| -> Result<Vec<f32>, String> {
        crate::loader::load_f32(model, name, &crate::loader::Overlay::None)
    };

    // Two frequency tables, chosen per layer by whether it compresses. The
    // release's compress_rope_theta (160 000) is not in config.json — it
    // lives in inference/config.json — so it is pinned here with the other
    // constants the header cannot carry.
    let rope_of = |base: f32, yarn: bool| -> Vec<f32> {
        if yarn {
            crate::attention::yarn_inv_freq(cfg.rope_head_dim, base, 16.0, 65536, 32.0, 1.0)
        } else {
            crate::attention::rope_inv_freq(cfg.rope_head_dim, base)
        }
    };
    let globals = Dsv4Globals {
        inv_freq_compress: rope_of(160_000.0, true),
        inv_freq_window: rope_of(10_000.0, false),
        embed: q("model.embed_tokens.weight")?,
        norm: f("model.norm.weight")?,
        head: q("lm_head.weight")?,
        hc_head_fn: f("model.hc_head_fn")?,
        hc_head_base: f("model.hc_head_base")?,
        hc_head_scale: *f("model.hc_head_scale")?
            .first()
            .ok_or("dsv4: empty hc_head_scale")?,
    };

    let mut layers = Vec::with_capacity(n_layers);
    for li in 0..n_layers {
        layers.push(load_layer(
            model,
            cfg,
            &format!("model.layers.{li}"),
            Scheme::Main,
        )?);
    }
    // The projection this loader exists to serve: with a RAM tier configured,
    // pin the MASKED expert set with one sequential sweep of the file at
    // streaming rate, before decode discovers it miss by miss in random
    // order. Experts outside a layer's mask are skipped; a layer without a
    // mask keeps all of its experts (the budget caps the sweep).
    #[cfg(feature = "gpu")]
    if crate::gpu_wgpu::host_banks_on() {
        // Host banks: one background sweep, layer by layer, oldest first.
        let sets: Vec<(usize, Vec<(usize, usize, usize)>, bool)> = layers
            .iter()
            .filter_map(|l| {
                let idx3 = |e: &Dsv4Expert| {
                    Some((e.w1.model_idx()?, e.w3.model_idx()?, e.w2.model_idx()?))
                };
                let mut v: Vec<_> = l.experts.iter().filter_map(idx3).collect();
                let first = v.first()?.0;
                v.push(idx3(&l.shared)?);
                let gu_q2 =
                    l.experts.first()?.w1.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP);
                Some((first, v, gu_q2))
            })
            .collect();
        let m2 = model.clone();
        let (inter, dim) = (cfg.moe_inter, cfg.dim);
        std::thread::spawn(move || {
            for (first, v, gu_q2) in sets {
                crate::gpu_wgpu::dsv4_host_bank_build(&m2, first, &v, inter, dim, gu_q2);
            }
            tracing::info!("host banks: built");
        });
    }
    #[cfg(feature = "gpu")]
    {
        let masks: Vec<Option<Vec<bool>>> = layers.iter().map(|l| l.mask.clone()).collect();
        crate::gpu_wgpu::prefetch_tier(model, &|name: &str| {
            let Some(i) = name.find(".experts.") else {
                return false;
            };
            let rest = &name[i + 9..];
            let e: usize = match rest[..rest.find('.').unwrap_or(rest.len())].parse() {
                Ok(v) => v,
                Err(_) => return false,
            };
            let li: usize = {
                let Some(j) = name.find("layers.") else {
                    return false;
                };
                let r = &name[j + 7..];
                match r[..r.find('.').unwrap_or(r.len())].parse() {
                    Ok(v) => v,
                    Err(_) => return false,
                }
            };
            match masks.get(li).and_then(|m| m.as_ref()) {
                Some(m) => m.get(e).copied().unwrap_or(false),
                None => true,
            }
        });
    }
    Ok((globals, layers))
}

/// Where a layer's tensors live in the file.
///
/// The MTP modules are the same layer as any other — attention, a
/// hyper-connection pair, a gated MoE over 256 experts — but the converter
/// wrote them under DeepSeek's internal names rather than the HF ones it used
/// for the trunk. Two schemes, one loader: a second copy would drift.
#[derive(Clone, Copy, PartialEq, Eq, Debug)]
pub enum Scheme {
    Main,
    Mtp,
}

impl Scheme {
    fn attn(self) -> &'static str {
        match self {
            Scheme::Main => "self_attn",
            Scheme::Mtp => "attn",
        }
    }
    fn attn_norm(self) -> &'static str {
        match self {
            Scheme::Main => "input_layernorm.weight",
            Scheme::Mtp => "attn_norm.weight",
        }
    }
    fn ffn_norm(self) -> &'static str {
        match self {
            Scheme::Main => "post_attention_layernorm.weight",
            Scheme::Mtp => "ffn_norm.weight",
        }
    }
    fn mlp(self) -> &'static str {
        match self {
            Scheme::Main => "mlp",
            Scheme::Mtp => "ffn",
        }
    }
    /// The router's per-expert bias. Absent on the trunk's hash layers, which
    /// is how they are recognised; always present on an MTP module.
    fn gate_bias(self) -> &'static str {
        match self {
            Scheme::Main => "expert_bias",
            Scheme::Mtp => "gate.bias",
        }
    }
    fn shared(self) -> &'static str {
        match self {
            Scheme::Main => "shared_expert",
            Scheme::Mtp => "shared_experts",
        }
    }
    /// gate, down, up — in that order, which is w1/w2/w3 upstream.
    fn w(self, i: u8) -> &'static str {
        match (self, i) {
            (Scheme::Main, 1) => "gate_proj.weight",
            (Scheme::Main, 2) => "down_proj.weight",
            (Scheme::Main, _) => "up_proj.weight",
            (Scheme::Mtp, 1) => "w1.weight",
            (Scheme::Mtp, 2) => "w2.weight",
            (Scheme::Mtp, _) => "w3.weight",
        }
    }
}

/// One layer, wherever it lives in the file.
pub fn load_layer(
    model: &std::sync::Arc<cortiq_core::CmfModel>,
    cfg: &Dsv4Cfg,
    p: &str,
    s: Scheme,
) -> Result<Dsv4Layer, String> {
    let q = |name: &str| -> Result<crate::qtensor::QTensor, String> {
        crate::qtensor::QTensor::from_model(model, name)
    };
    let f = |name: &str| -> Result<Vec<f32>, String> {
        crate::loader::load_f32(model, name, &crate::loader::Overlay::None)
    };
    let opt_f = |name: &str| -> Option<Vec<f32>> { f(name).ok() };
    let at = s.attn();
    let ml = s.mlp();
    {
        let scale3 = |name: &str| -> Result<[f32; 3], String> {
            let v = f(name)?;
            if v.len() < 3 {
                return Err(format!("{name}: expected 3 scales, got {}", v.len()));
            }
            Ok([v[0], v[1], v[2]])
        };
        // The compressor exists on every layer whose ratio is non-zero;
        // its presence in the file is the only signal we need.
        let compressor = match q(&format!("{p}.{at}.compressor.wkv.weight")) {
            Ok(wkv) => {
                let ape = f(&format!("{p}.{at}.compressor.ape"))?;
                // ape is [ratio, coff*head_dim]; coff is 2 when the windows
                // overlap, which the release does at ratio 4.
                let width = wkv.rows();
                let ratio = (ape.len() / width.max(1)).max(1);
                Some(Dsv4Compressor {
                    wkv,
                    wgate: q(&format!("{p}.{at}.compressor.wgate.weight"))?,
                    norm: f(&format!("{p}.{at}.compressor.norm.weight"))?,
                    ape,
                    ratio,
                    overlap: ratio == 4,
                })
            }
            Err(_) => None,
        };
        let indexer = match q(&format!("{p}.{at}.indexer.wq_b.weight")) {
            Ok(wq_b) => {
                let ape = f(&format!("{p}.{at}.indexer.compressor.ape"))?;
                let cwkv = q(&format!("{p}.{at}.indexer.compressor.wkv.weight"))?;
                let width = cwkv.rows();
                let ratio = (ape.len() / width.max(1)).max(1);
                Some(Dsv4Indexer {
                    wq_b,
                    weights_proj: q(&format!("{p}.{at}.indexer.weights_proj.weight"))?,
                    compressor: Dsv4Compressor {
                        wkv: cwkv,
                        wgate: q(&format!("{p}.{at}.indexer.compressor.wgate.weight"))?,
                        norm: f(&format!("{p}.{at}.indexer.compressor.norm.weight"))?,
                        ape,
                        ratio,
                        overlap: ratio == 4,
                    },
                })
            }
            Err(_) => None,
        };

        let mut experts = Vec::with_capacity(cfg.n_routed_experts);
        for e in 0..cfg.n_routed_experts {
            let ep = format!("{p}.{ml}.experts.{e}");
            experts.push(Dsv4Expert {
                w1: q(&format!("{ep}.{w}", w = s.w(1)))?,
                w2: q(&format!("{ep}.{w}", w = s.w(2)))?,
                w3: q(&format!("{ep}.{w}", w = s.w(3)))?,
            });
        }

        Ok(Dsv4Layer {
            attn_norm: f(&format!("{p}.{an}", an = s.attn_norm()))?,
            ffn_norm: f(&format!("{p}.{fnm}", fnm = s.ffn_norm()))?,
            wq_a: q(&format!("{p}.{at}.wq_a.weight"))?,
            q_norm: f(&format!("{p}.{at}.q_norm.weight"))?,
            wq_b: q(&format!("{p}.{at}.wq_b.weight"))?,
            wkv: q(&format!("{p}.{at}.wkv.weight"))?,
            kv_norm: f(&format!("{p}.{at}.kv_norm.weight"))?,
            wo_a: q(&format!("{p}.{at}.wo_a.weight"))?,
            wo_b: q(&format!("{p}.{at}.wo_b.weight"))?,
            attn_sink: f(&format!("{p}.{at}.attn_sink"))?,
            compressor,
            indexer,
            hc_attn_fn: f(&format!("{p}.hc_attn_fn"))?,
            hc_attn_base: f(&format!("{p}.hc_attn_base"))?,
            hc_attn_scale: scale3(&format!("{p}.hc_attn_scale"))?,
            hc_ffn_fn: f(&format!("{p}.hc_ffn_fn"))?,
            hc_ffn_base: f(&format!("{p}.hc_ffn_base"))?,
            hc_ffn_scale: scale3(&format!("{p}.hc_ffn_scale"))?,
            gate: q(&format!("{p}.{ml}.gate.weight"))?,
            // The bias is absent exactly on the hash layers, and the table
            // is present exactly there — the file itself says which is which.
            gate_bias: opt_f(&format!("{p}.{ml}.{b}", b = s.gate_bias())),
            tid2eid: opt_f(&format!("{p}.{ml}.tid2eid")),
            experts,
            mask: if model.tensor(&format!("{p}.{ml}.tid2eid")).is_some() {
                None
            } else {
                crate::loader::moe_task_mask(model, &format!("{p}."), cfg.n_routed_experts)
            },
            shared: Dsv4Expert {
                w1: q(&format!("{p}.{ml}.{sh}.{w}", sh = s.shared(), w = s.w(1)))?,
                w2: q(&format!("{p}.{ml}.{sh}.{w}", sh = s.shared(), w = s.w(2)))?,
                w3: q(&format!("{p}.{ml}.{sh}.{w}", sh = s.shared(), w = s.w(3)))?,
            },
        })
    }
}

/// One module of the speculation stack.
///
/// The release carries three, so the draft is three deep, and the last one
/// also holds a confidence head — the model scores its own proposals rather
/// than leaving acceptance to a threshold we would have to invent. Each
/// module is a full layer with its own 256 experts; what makes it an MTP
/// module rather than a 44th layer is `main_proj`, which folds the previous
/// hidden state into the next embedding before the layer runs.
pub struct Dsv4Mtp {
    pub layer: Dsv4Layer,
    /// Stage 0 only: the projection that turns the trunk's captured hidden
    /// states into the block's input. Later stages take the block from the
    /// stage before them, so they carry none.
    pub main_proj: Option<crate::qtensor::QTensor>,
    pub main_norm: Option<Vec<f32>>,
    /// Last module only: what turns a draft hidden state into logits.
    pub norm: Option<Vec<f32>>,
    pub hc_head_fn: Option<Vec<f32>>,
    pub hc_head_base: Option<Vec<f32>>,
    pub hc_head_scale: Option<f32>,
    pub confidence: Option<crate::qtensor::QTensor>,
    /// Last stage only: a rank-256 bigram table that biases the draft's
    /// logits, and whose embedding also feeds the confidence head. Cheap
    /// enough that the draft samples through it position by position while
    /// the network itself runs the whole block at once.
    pub markov_w1: Option<crate::qtensor::QTensor>,
    pub markov_w2: Option<crate::qtensor::QTensor>,
}

/// Load as much of the speculation stack as the file carries, up to
/// `max_depth`. Missing is not an error: a checkpoint without MTP simply
/// yields an empty stack, and the caller falls back to plain decoding.
pub fn load_mtp(
    model: &std::sync::Arc<cortiq_core::CmfModel>,
    cfg: &Dsv4Cfg,
    max_depth: usize,
) -> Vec<Dsv4Mtp> {
    let f = |name: &str| -> Option<Vec<f32>> {
        crate::loader::load_f32(model, name, &crate::loader::Overlay::None).ok()
    };
    let mut out = Vec::new();
    for d in 0..max_depth {
        let p = format!("model.mtp.{d}");
        // A stage is recognised by its attention, not by `main_proj`: only
        // stage 0 has that, and only the last has the head. Keying on either
        // end found one module of three.
        if model.tensor(&format!("{p}.attn.wq_a.weight")).is_none() {
            break;
        }
        let layer = match load_layer(model, cfg, &p, Scheme::Mtp) {
            Ok(l) => l,
            Err(e) => {
                eprintln!("MTP {d}: пропущен, {e}");
                break;
            }
        };
        out.push(Dsv4Mtp {
            layer,
            main_proj: crate::qtensor::QTensor::from_model(model, &format!("{p}.main_proj.weight"))
                .ok(),
            main_norm: f(&format!("{p}.main_norm.weight")),
            norm: f(&format!("{p}.norm.weight")),
            hc_head_fn: f(&format!("{p}.hc_head_fn")),
            hc_head_base: f(&format!("{p}.hc_head_base")),
            hc_head_scale: f(&format!("{p}.hc_head_scale")).and_then(|v| v.first().copied()),
            confidence: crate::qtensor::QTensor::from_model(
                model,
                &format!("{p}.confidence_head.proj.weight"),
            )
            .ok(),
            markov_w1: crate::qtensor::QTensor::from_model(
                model,
                &format!("{p}.markov_head.markov_w1.weight"),
            )
            .ok(),
            markov_w2: crate::qtensor::QTensor::from_model(
                model,
                &format!("{p}.markov_head.markov_w2.weight"),
            )
            .ok(),
        });
    }
    dspark_apply_mask(&mut out);
    if !out.is_empty() {
        let mp = out
            .iter()
            .find_map(|m| m.main_proj.as_ref())
            .map(|t| format!("[{}, {}]", t.rows(), t.cols()))
            .unwrap_or_else(|| "нет".into());
        eprintln!(
            "MTP: {} стади(я/и/й), main_proj {mp}, экспертов {}, \
             голова уверенности {}, марков {}",
            out.len(),
            out[0].layer.experts.len(),
            if out.iter().any(|m| m.confidence.is_some()) {
                "есть"
            } else {
                "нет"
            },
            if out.iter().any(|m| m.markov_w1.is_some()) {
                "есть"
            } else {
                "нет"
            },
        );
    }
    out
}

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

    // A whole model, small enough to reason about: 2 layers, 4 heads, 8
    // experts. Weights are deterministic and tiny, which is the point —
    // this test is about shapes, indexing and cache bookkeeping, the things
    // that a 138 GB file would surface only after an hour of loading.
    fn toy() -> (Dsv4Globals, Vec<Dsv4Layer>, Dsv4Cfg) {
        use crate::qtensor::QTensor;
        let cfg = Dsv4Cfg {
            dim: 32,
            n_heads: 4,
            head_dim: 8,
            rope_head_dim: 4,
            q_lora_rank: 16,
            o_lora_rank: 16,
            o_groups: 2,
            hc_mult: 4,
            hc_sinkhorn_iters: 20,
            hc_eps: 1e-6,
            norm_eps: 1e-6,
            n_routed_experts: 8,
            top_k: 2,
            moe_inter: 16,
            route_scale: 1.0,
            swiglu_limit: 10.0,
            window: 6,
            index_topk: 8,
            vocab: 24,
        };
        // Deterministic pseudo-random in a narrow band: big enough to move
        // the state, small enough that nothing saturates.
        let w = |n: usize, seed: usize| -> Vec<f32> {
            (0..n)
                .map(|i| (((i * 7 + seed * 13) % 101) as f32 / 101.0 - 0.5) * 0.3)
                .collect()
        };
        let t = |rows: usize, cols: usize, seed: usize| {
            QTensor::from_f32(w(rows * cols, seed), rows, cols)
        };
        let ones = |n: usize| vec![1.0f32; n];

        let (dim, hc) = (cfg.dim, cfg.hc_mult);
        // q is n_heads*head_dim wide and kv is one head wide; rope rides the
        // tail of each rather than widening anything.
        let q_width = cfg.n_heads * cfg.head_dim;
        let kv_width = cfg.head_dim;
        let o_per_group = q_width / cfg.o_groups;
        let mut layers = Vec::new();
        for li in 0..2 {
            let experts: Vec<Dsv4Expert> = (0..cfg.n_routed_experts)
                .map(|e| Dsv4Expert {
                    w1: t(cfg.moe_inter, dim, 40 + e + li * 8),
                    w2: t(dim, cfg.moe_inter, 60 + e + li * 8),
                    w3: t(cfg.moe_inter, dim, 80 + e + li * 8),
                })
                .collect();
            // Layer 0 is a hash layer (table-routed); layer 1 routes normally
            // and carries the compressor — both paths get exercised.
            layers.push(Dsv4Layer {
                attn_norm: ones(dim),
                ffn_norm: ones(dim),
                wq_a: t(cfg.q_lora_rank, dim, 1 + li),
                q_norm: ones(cfg.q_lora_rank),
                wq_b: t(q_width, cfg.q_lora_rank, 3 + li),
                wkv: t(kv_width, dim, 5 + li),
                kv_norm: ones(kv_width),
                wo_a: t(cfg.o_groups * cfg.o_lora_rank, o_per_group, 7 + li),
                wo_b: t(dim, cfg.o_groups * cfg.o_lora_rank, 9 + li),
                attn_sink: vec![0.1; cfg.n_heads],
                // Layer 1 carries the OVERLAPPING compressor, as the release
                // does at ratio 4: the projection is twice the entry width.
                compressor: if li == 1 {
                    Some(Dsv4Compressor {
                        wkv: t(2 * kv_width, dim, 11),
                        wgate: t(2 * kv_width, dim, 13),
                        norm: ones(kv_width),
                        ape: vec![0.01; 4 * 2 * kv_width],
                        ratio: 4,
                        overlap: true,
                    })
                } else {
                    None
                },
                indexer: if li == 1 {
                    Some(Dsv4Indexer {
                        wq_b: t(2 * 16, cfg.q_lora_rank, 41),
                        weights_proj: t(2, dim, 43),
                        compressor: Dsv4Compressor {
                            wkv: t(2 * 16, dim, 45),
                            wgate: t(2 * 16, dim, 47),
                            norm: ones(16),
                            ape: vec![0.01; 4 * 2 * 16],
                            ratio: 4,
                            overlap: true,
                        },
                    })
                } else {
                    None
                },
                hc_attn_fn: w((2 + hc) * hc * hc * dim, 15 + li),
                hc_attn_base: w((2 + hc) * hc, 17 + li),
                hc_attn_scale: [1.0, 1.0, 1.0],
                hc_ffn_fn: w((2 + hc) * hc * hc * dim, 19 + li),
                hc_ffn_base: w((2 + hc) * hc, 21 + li),
                hc_ffn_scale: [1.0, 1.0, 1.0],
                gate: t(cfg.n_routed_experts, dim, 23 + li),
                gate_bias: if li == 1 {
                    Some(vec![0.0; cfg.n_routed_experts])
                } else {
                    None
                },
                tid2eid: if li == 0 {
                    Some(
                        (0..cfg.vocab * cfg.top_k)
                            .map(|i| (i % cfg.n_routed_experts) as f32)
                            .collect(),
                    )
                } else {
                    None
                },
                experts,
                mask: None,
                shared: Dsv4Expert {
                    w1: t(cfg.moe_inter, dim, 25 + li),
                    w2: t(dim, cfg.moe_inter, 27 + li),
                    w3: t(cfg.moe_inter, dim, 29 + li),
                },
            });
        }
        let inv = |base: f32| -> Vec<f32> {
            (0..cfg.rope_head_dim / 2)
                .map(|i| 1.0 / base.powf(2.0 * i as f32 / cfg.rope_head_dim as f32))
                .collect()
        };
        let g = Dsv4Globals {
            inv_freq_compress: inv(160000.0),
            inv_freq_window: inv(10000.0),
            embed: t(cfg.vocab, dim, 31),
            norm: ones(dim),
            head: t(cfg.vocab, dim, 33),
            hc_head_fn: w(hc * hc * dim, 35),
            hc_head_base: w(hc, 37),
            hc_head_scale: 1.0,
        };
        (g, layers, cfg)
    }

    /// The whole stack, decoding a sequence. Every block is on the path:
    /// hyper-connections, the double-LoRA attention with its sink, the KV
    /// compressor firing on its ratio boundary, hash routing on one layer
    /// and score routing on the other.
    #[test]
    fn forward_token_decodes_a_sequence_without_falling_over() {
        let (g, layers, cfg) = toy();
        let mut st = Dsv4State::new(layers.len());
        let inv_freq: Vec<f32> = (0..cfg.rope_head_dim / 2)
            .map(|i| 1.0 / 10000f32.powf(2.0 * i as f32 / cfg.rope_head_dim as f32))
            .collect();
        let mut logits = Vec::new();

        // Ten tokens: more than twice the compressor's ratio, so the
        // compressed cache is written on a boundary and read afterwards.
        let mut first: Option<Vec<f32>> = None;
        for (step, tok) in [3u32, 7, 1, 9, 4, 2, 8, 5, 6, 0].into_iter().enumerate() {
            forward_token(
                &g,
                &layers,
                &cfg,
                &mut st,
                tok,
                &inv_freq,
                None,
                &mut logits,
            );
            assert_eq!(logits.len(), cfg.vocab, "step {step}: logit count");
            assert!(
                logits.iter().all(|v| v.is_finite()),
                "step {step}: non-finite logit — {logits:?}"
            );
            // A model that has collapsed returns the same distribution
            // regardless of input; that is the failure this catches.
            let spread = logits.iter().cloned().fold(f32::MIN, f32::max)
                - logits.iter().cloned().fold(f32::MAX, f32::min);
            assert!(spread > 1e-6, "step {step}: logits are flat ({spread})");
            if step == 0 {
                first = Some(logits.clone());
            }
            assert_eq!(st.pos, step + 1, "position bookkeeping");
        }

        // The cache has to have grown, and the compressor layer must have
        // emitted compressed entries (10 tokens / ratio 4 = 2 windows).
        assert!(!st.window[0].is_empty(), "sliding window never filled");
        // Ten tokens through a window of six: it must have slid, not grown.
        for (li, w) in st.window.iter().enumerate() {
            assert!(
                w.len() / cfg.head_dim <= cfg.window,
                "layer {li}: window holds {} positions, cap is {}",
                w.len() / cfg.head_dim,
                cfg.window
            );
        }
        assert!(
            !st.compressed[1].is_empty(),
            "compressor layer produced no compressed KV in 10 tokens"
        );
        // Ten tokens at ratio 4 fold twice, and the entries must be one head
        // wide — the overlapping projection is 2x that, so a width mistake
        // shows up here rather than as quiet nonsense.
        assert_eq!(
            st.compressed[1].len() / cfg.head_dim,
            2,
            "expected two folds in ten tokens at ratio 4"
        );
        assert!(
            !st.prev_kv[1].is_empty(),
            "the overlapping compressor never kept a previous window"
        );
        // Every layer that HAS an indexer must have filled the indexer's own
        // cache: it is what decides which compressed positions attention
        // reads, and an empty one silently discards the whole long-range
        // memory rather than failing.
        for (li, l) in layers.iter().enumerate() {
            if l.indexer.is_some() {
                assert!(
                    !st.index_kv[li].is_empty(),
                    "layer {li} has an indexer but its cache stayed empty"
                );
            }
        }

        // Context must matter: the same token at position 0 of a fresh state
        // and at the end of a filled one cannot give identical logits.
        let mut fresh = Dsv4State::new(layers.len());
        let mut relogits = Vec::new();
        forward_token(
            &g,
            &layers,
            &cfg,
            &mut fresh,
            3,
            &inv_freq,
            None,
            &mut relogits,
        );
        assert_eq!(
            relogits,
            first.unwrap(),
            "the same token from a fresh state must reproduce exactly"
        );
    }

    /// The reference clamps `up` on both sides but `gate` only from above.
    /// Getting that symmetric would quietly change every expert's output on
    /// the tokens that saturate, which is the hardest kind of bug to see.
    #[test]
    fn swiglu_limit_clamps_up_both_ways_and_gate_only_from_above() {
        let inter = 4;
        // gate = [-50, 50, 1, -1], up = [50, -50, 1, -1]
        let gate_src = [-50.0f32, 50.0, 1.0, -1.0];
        let up_src = [50.0f32, -50.0, 1.0, -1.0];
        let limit = 10.0f32;
        let mut got = vec![0.0f32; inter];
        expert_swiglu(
            &[0.0],
            &|_, d| d.copy_from_slice(&gate_src),
            &|_, d| d.copy_from_slice(&up_src),
            &|src, d| d.copy_from_slice(src),
            inter,
            1.0,
            limit,
            &mut got,
        );
        let silu = |g: f32| g / (1.0 + (-g).exp());
        // gate: only the +50 is cut, the -50 rides through silu untouched.
        let want = [
            silu(-50.0) * limit,
            silu(limit) * -limit,
            silu(1.0) * 1.0,
            -silu(-1.0),
        ];
        for (i, w) in want.iter().enumerate() {
            assert!(
                (got[i] - w).abs() < 1e-5,
                "lane {i}: got {} want {w}",
                got[i]
            );
        }
        // And with the clamp off nothing is touched.
        let mut raw = vec![0.0f32; inter];
        expert_swiglu(
            &[0.0],
            &|_, d| d.copy_from_slice(&gate_src),
            &|_, d| d.copy_from_slice(&up_src),
            &|src, d| d.copy_from_slice(src),
            inter,
            1.0,
            0.0,
            &mut raw,
        );
        assert!(
            (raw[1] - silu(50.0) * -50.0).abs() < 1e-3,
            "limit 0 must not clamp"
        );
    }

    /// The grouped projection writes its intermediate from several threads
    /// at once. Disjoint indices are the whole argument for that being safe,
    /// so the pooled result has to equal the serial one exactly — a race
    /// here would show up as occasional wrong tokens, not as a crash.
    #[test]
    fn grouped_projection_is_identical_with_and_without_a_pool() {
        let (groups, lora, per_group, dim) = (4usize, 128usize, 64usize, 32usize);
        let attn: Vec<f32> = (0..groups * per_group)
            .map(|i| ((i * 13) as f32 * 0.021).sin())
            .collect();
        let wo_a: Vec<f32> = (0..groups * lora * per_group)
            .map(|i| ((i * 7) as f32 * 0.011).cos())
            .collect();
        let wo_b: Vec<f32> = (0..dim * groups * lora)
            .map(|i| ((i * 5) as f32 * 0.009).sin())
            .collect();
        let row = |r: usize, x: &[f32], _sc: &mut [f32]| -> f32 {
            wo_a[r * per_group..(r + 1) * per_group]
                .iter()
                .zip(x)
                .map(|(a, b)| a * b)
                .sum()
        };
        let project = |mid: &[f32], dst: &mut [f32]| {
            for (d, o) in dst.iter_mut().enumerate() {
                *o = wo_b[d * mid.len()..(d + 1) * mid.len()]
                    .iter()
                    .zip(mid)
                    .map(|(a, b)| a * b)
                    .sum();
            }
        };

        let mut serial = vec![0.0f32; dim];
        o_project(
            &attn,
            &row,
            per_group,
            &project,
            groups,
            lora,
            None,
            &mut serial,
        );

        let pool = crate::pool::Pool::new(4);
        let mut pooled = vec![0.0f32; dim];
        o_project(
            &attn,
            &row,
            per_group,
            &project,
            groups,
            lora,
            Some(&pool),
            &mut pooled,
        );
        assert_eq!(serial, pooled, "the pooled projection diverged");
        assert!(
            serial.iter().any(|v| v.abs() > 1e-6),
            "test data is degenerate"
        );
    }

    #[test]
    fn block_grouped_projection_matches_position_walk() {
        let (_g, layers, cfg) = toy();
        let l = &layers[1];
        let b = 5;
        let attn_len = cfg.n_heads * cfg.head_dim;
        let attn: Vec<f32> = (0..b * attn_len)
            .map(|i| ((i * 17) as f32 * 0.013).sin())
            .collect();
        let mut walked = vec![0.0f32; b * cfg.dim];
        for bi in 0..b {
            o_project(
                &attn[bi * attn_len..(bi + 1) * attn_len],
                &|r, x, sc| l.wo_a.row_dot(r, x, sc),
                l.wo_a.cols(),
                &|mid, dst| l.wo_b.matvec(mid, dst, None),
                cfg.o_groups,
                cfg.o_lora_rank,
                None,
                &mut walked[bi * cfg.dim..(bi + 1) * cfg.dim],
            );
        }
        let mut batched = vec![0.0f32; b * cfg.dim];
        o_project_block(
            &attn,
            b,
            &l.wo_a,
            &l.wo_b,
            cfg.o_groups,
            cfg.o_lora_rank,
            None,
            &mut batched,
        );
        assert_eq!(batched, walked);
    }

    #[test]
    fn block_moe_matches_position_walk_in_route_order() {
        let (_g, layers, cfg) = toy();
        // The scored layer exercises repeated and distinct experts without
        // tying the result to a token-id table.
        let l = &layers[1];
        let b = 5;
        let xs: Vec<f32> = (0..b * cfg.dim)
            .map(|i| ((i * 11) as f32 * 0.019).cos())
            .collect();
        let ids = [1u32, 2, 3, 4, 5];
        let mut walked = vec![0.0f32; b * cfg.dim];
        for bi in 0..b {
            moe_step(
                &xs[bi * cfg.dim..(bi + 1) * cfg.dim],
                l,
                &cfg,
                ids[bi],
                1,
                None,
                &mut walked[bi * cfg.dim..(bi + 1) * cfg.dim],
            );
        }
        let mut batched = vec![0.0f32; b * cfg.dim];
        moe_step_block(&xs, b, l, &cfg, &ids, 1, None, &mut batched);
        assert_eq!(batched, walked);
    }

    /// The overlapping compressor folds 2*ratio slots, not ratio: the
    /// previous window contributes its first half of dimensions and the
    /// current one its second half. Treating it as a plain compressor makes
    /// the entry twice as wide as the cache expects, which lands the whole
    /// thing in the wrong store rather than raising anything.
    #[test]
    fn overlapping_compressor_folds_both_windows() {
        let (ratio, d) = (2usize, 3usize);
        // Current window: two tokens, 2*d wide each. Second half is what the
        // current window contributes.
        let cur_kv: Vec<f32> = vec![
            1.0, 1.0, 1.0, /*|*/ 10.0, 20.0, 30.0, // token 0
            2.0, 2.0, 2.0, /*|*/ 40.0, 50.0, 60.0, // token 1
        ];
        // Make the current window's second-half scores dominate everywhere.
        let cur_sc: Vec<f32> = vec![
            0.0, 0.0, 0.0, /*|*/ 0.0, 0.0, 100.0, //
            0.0, 0.0, 0.0, /*|*/ 100.0, 100.0, 0.0,
        ];
        // Previous window: its FIRST half is what it contributes.
        let prev_kv: Vec<f32> = vec![
            7.0, 8.0, 9.0, /*|*/ 0.0, 0.0, 0.0, //
            5.0, 6.0, 7.0, /*|*/ 0.0, 0.0, 0.0,
        ];
        let prev_sc = vec![0.0f32; ratio * 2 * d];

        let mut out = vec![0.0f32; d];
        compress_window_overlap(&prev_kv, &prev_sc, &cur_kv, &cur_sc, ratio, d, &mut out);
        // dim 0 and 1: token 1's second half wins (score 100)
        assert!((out[0] - 40.0).abs() < 1e-3, "dim0 = {}", out[0]);
        assert!((out[1] - 50.0).abs() < 1e-3, "dim1 = {}", out[1]);
        // dim 2: token 0's second half wins
        assert!((out[2] - 30.0).abs() < 1e-3, "dim2 = {}", out[2]);

        // With no previous window the fold still works and uses only the
        // current one — this is the very first window of a generation.
        let mut first = vec![0.0f32; d];
        compress_window_overlap(&[], &[], &cur_kv, &cur_sc, ratio, d, &mut first);
        assert!(
            first.iter().all(|v| v.is_finite()),
            "first window: {first:?}"
        );
        assert!((first[0] - 40.0).abs() < 1e-3, "first dim0 = {}", first[0]);

        // And a previous window with real scores does pull the result.
        let mut both = vec![0.0f32; d];
        let strong_prev = vec![100.0f32; ratio * 2 * d];
        compress_window_overlap(
            &prev_kv,
            &strong_prev,
            &cur_kv,
            &cur_sc,
            ratio,
            d,
            &mut both,
        );
        assert!(
            (both[0] - 40.0).abs() > 1.0,
            "a scored previous window must move the fold, got {}",
            both[0]
        );
    }

    /// Numerical parity with the reference. The vectors below come from
    /// running `kernel.py::hc_split_sinkhorn`'s own formula on a fixed
    /// input; matching them pins the exponent order, the eps placement and
    /// the off-by-one in the iteration count all at once — a property test
    /// alone would pass with any of those wrong.
    #[test]
    fn sinkhorn_matches_the_reference_numbers() {
        let hc = 4;
        let mixes: Vec<f32> = (0..24).map(|i| (i as f32 * 0.37).sin() * 3.0).collect();
        let base: Vec<f32> = (0..24).map(|i| (i as f32 * 0.11).cos()).collect();
        let (mut pre, mut post, mut comb) = (vec![0.0; hc], vec![0.0; hc], vec![0.0; hc * hc]);
        hc_split_sinkhorn(
            &mixes,
            &[1.0, 1.0, 1.0],
            &base,
            hc,
            20,
            1e-6,
            &mut pre,
            &mut post,
            &mut comb,
        );
        let want_pre = [0.7310596, 0.8888268, 0.9525191, 0.97424865];
        let want_post = [1.9600224, 1.9534285, 1.9201256, 1.8160983];
        let want_comb = [
            0.5996052,
            0.282_535_9,
            0.09218107,
            0.025676856,
            0.17564717,
            0.22228767,
            0.271_745_4,
            0.330_318_8,
            0.029528176,
            0.12206022,
            0.32619134,
            0.5222193,
            0.19521846,
            0.37311527,
            0.30988118,
            0.12178412,
        ];
        for (i, w) in want_pre.iter().enumerate() {
            assert!((pre[i] - w).abs() < 1e-5, "pre[{i}]: {} vs {w}", pre[i]);
        }
        for (i, w) in want_post.iter().enumerate() {
            assert!((post[i] - w).abs() < 1e-5, "post[{i}]: {} vs {w}", post[i]);
        }
        for (i, w) in want_comb.iter().enumerate() {
            assert!((comb[i] - w).abs() < 1e-4, "comb[{i}]: {} vs {w}", comb[i]);
        }
    }

    /// Sinkhorn's whole point is a doubly stochastic matrix: every row and
    /// every column sums to one. If the alternating normalization is wrong
    /// (or the loop count is off by one) the sums drift, and the residual
    /// mixing quietly gains or loses mass on every layer.
    #[test]
    fn sinkhorn_leaves_the_mixing_matrix_doubly_stochastic() {
        let hc = 4;
        let mix_hc = (2 + hc) * hc;
        // a deliberately lopsided projection
        let mixes: Vec<f32> = (0..mix_hc).map(|i| (i as f32 * 0.37).sin() * 3.0).collect();
        let base: Vec<f32> = (0..mix_hc).map(|i| (i as f32 * 0.11).cos()).collect();
        let (mut pre, mut post, mut comb) = (vec![0.0; hc], vec![0.0; hc], vec![0.0; hc * hc]);
        hc_split_sinkhorn(
            &mixes,
            &[1.0, 1.0, 1.0],
            &base,
            hc,
            20,
            1e-6,
            &mut pre,
            &mut post,
            &mut comb,
        );
        for j in 0..hc {
            let r: f32 = comb[j * hc..(j + 1) * hc].iter().sum();
            assert!((r - 1.0).abs() < 2e-3, "row {j} sums to {r}");
            let c: f32 = (0..hc).map(|k| comb[k * hc + j]).sum();
            assert!((c - 1.0).abs() < 2e-3, "col {j} sums to {c}");
        }
        // pre is a gate in (eps, 1+eps); post carries the factor 2
        assert!(pre.iter().all(|&v| v > 0.0 && v < 1.001));
        assert!(post.iter().all(|&v| (0.0..=2.0).contains(&v)));
    }

    /// Folding four copies and expanding them back must preserve a constant
    /// state exactly when the block contributes nothing: with post = 0 the
    /// expansion is a doubly stochastic mix of identical copies, i.e. itself.
    #[test]
    fn expand_of_identical_copies_is_a_fixed_point() {
        let (hc, dim) = (4usize, 3usize);
        let residual: Vec<f32> = std::iter::repeat_n([1.5f32, -2.0, 0.25], hc)
            .flatten()
            .collect();
        let comb = {
            // exactly doubly stochastic: uniform
            vec![0.25f32; hc * hc]
        };
        let post = vec![0.0f32; hc];
        let mut out = vec![0.0f32; hc * dim];
        hc_expand(&[0.0; 3], &residual, &post, &comb, hc, dim, &mut out);
        for (o, r) in out.iter().zip(&residual) {
            assert!((o - r).abs() < 1e-6, "{o} vs {r}");
        }
    }

    /// The bias must move the SELECTION without touching the weights: with a
    /// large bias on a low-scoring expert it gets picked, but its weight is
    /// still its own (small) score, renormalized.
    #[test]
    fn selection_bias_steers_the_choice_but_not_the_weights() {
        let scores = [3.0f32, 0.1, 2.0, 0.05];
        let bias = [0.0f32, 10.0, 0.0, 0.0];
        let (mut idx, mut w) = (Vec::new(), Vec::new());
        route(&scores, Some(&bias), 2, 1.5, None, None, &mut idx, &mut w);
        assert_eq!(idx[0], 1, "the biased expert must win selection");
        assert_eq!(idx[1], 0);
        // weights come from sqrt(softplus(score)) BEFORE the bias, so the
        // biased expert's share must be the smaller of the two
        assert!(w[0] < w[1], "biased expert kept its own (small) weight");
        let sum: f32 = w.iter().sum();
        assert!((sum - 1.5).abs() < 1e-5, "weights renormalize then scale");
    }

    /// The sink is an extra logit with no value: it must lower every
    /// weight without adding output. With a huge sink the head should
    /// attend to almost nothing.
    #[test]
    fn attention_sink_drains_weight_without_contributing_output() {
        let hd = 2;
        let q = [1.0f32, 0.0];
        let kv = [1.0f32, 0.0, 0.0, 1.0];
        let mut out = vec![0.0f32; hd];
        sparse_attend(&q, &kv, &[0, 1], f32::NEG_INFINITY, 1.0, hd, &mut out);
        let plain = out.clone();
        assert!(plain[0] > plain[1], "the aligned key must dominate");
        sparse_attend(&q, &kv, &[0, 1], 20.0, 1.0, hd, &mut out);
        assert!(
            out[0] < plain[0] * 0.01 && out[1] < plain[1] * 0.01,
            "a large sink must drain nearly all the mass: {out:?}"
        );
    }

    /// A masked slot must be ignored entirely — not folded in as a zero
    /// key, which would still add exp(0) to the denominator.
    #[test]
    fn masked_positions_leave_the_denominator_alone() {
        let hd = 2;
        let q = [1.0f32, 0.0];
        let kv = [1.0f32, 0.0, 0.0, 1.0];
        let (mut a, mut b) = (vec![0.0f32; hd], vec![0.0f32; hd]);
        sparse_attend(&q, &kv, &[0], f32::NEG_INFINITY, 1.0, hd, &mut a);
        sparse_attend(
            &q,
            &kv,
            &[0, usize::MAX],
            f32::NEG_INFINITY,
            1.0,
            hd,
            &mut b,
        );
        for (x, y) in a.iter().zip(&b) {
            assert!((x - y).abs() < 1e-6, "{x} vs {y}");
        }
    }

    /// Forward then inverse rotation is the identity — the property the
    /// output path depends on.
    #[test]
    fn rope_tail_inverts_itself() {
        let inv_freq = [1.0f32, 0.5];
        let orig = [9.0f32, 8.0, 1.0, 2.0, 3.0, 4.0];
        let mut v = orig;
        rope_tail(&mut v, &inv_freq, 7, 4, false);
        assert!(v[..2] == orig[..2], "the non-rope head must not move");
        assert!(v[2..] != orig[2..], "the tail must actually rotate");
        rope_tail(&mut v, &inv_freq, 7, 4, true);
        for (a, b) in v.iter().zip(&orig) {
            assert!((a - b).abs() < 1e-5, "{a} vs {b}");
        }
    }

    /// The window pooling is a softmax per DIMENSION over the ratio, with
    /// the position bias inside the exponent.
    #[test]
    fn compressor_pools_the_window_per_dimension() {
        let (ratio, width) = (2usize, 2usize);
        let kv = [1.0f32, 10.0, 3.0, 20.0];
        // dim 0: equal scores → mean; dim 1: second token wins by a mile
        let score = [0.0f32, 0.0, 0.0, 50.0];
        let ape = vec![0.0f32; ratio * width];
        let mut out = vec![0.0f32; width];
        compress_window(&kv, &score, &ape, ratio, width, &mut out);
        assert!(
            (out[0] - 2.0).abs() < 1e-5,
            "equal scores average: {}",
            out[0]
        );
        assert!(
            (out[1] - 20.0).abs() < 1e-3,
            "a dominant score wins: {}",
            out[1]
        );
    }

    /// A negative dot product must not drag a position down: the relu
    /// means heads abstain rather than veto.
    #[test]
    fn index_scores_relu_before_weighting() {
        let (nh, hd) = (2usize, 2usize);
        // head 0 aligns with position 0, head 1 anti-aligns with it
        let q = [1.0f32, 0.0, -1.0, 0.0];
        let kv = [1.0f32, 0.0, 0.0, 1.0];
        let w = [1.0f32, 1.0];
        let mut sc = Vec::new();
        index_scores(&q, &kv, &w, nh, hd, 2, 2, None, &mut sc);
        // without the relu the anti-aligned head would cancel head 0 to zero
        assert!(sc[0] > 0.9, "abstention, not veto: {:?}", sc);
    }

    #[test]
    fn index_scores_mask_the_future() {
        let (nh, hd) = (1usize, 2usize);
        let q = [1.0f32, 0.0];
        let kv = [1.0f32, 0.0, 1.0, 0.0, 1.0, 0.0];
        let w = [1.0f32];
        let mut sc = Vec::new();
        index_scores(&q, &kv, &w, nh, hd, 3, 2, None, &mut sc);
        assert!(sc[0].is_finite() && sc[1].is_finite());
        assert!(sc[2] == f32::NEG_INFINITY, "position 2 is in the future");
        let mut idx = Vec::new();
        top_k_positions(&sc, 3, &mut idx);
        assert_eq!(idx, vec![0, 1], "a masked slot never wins a slot");
    }

    #[test]
    fn top_k_is_deterministic_on_ties() {
        let sc = [1.0f32, 1.0, 1.0, 0.0];
        let mut idx = Vec::new();
        top_k_positions(&sc, 2, &mut idx);
        assert_eq!(idx, vec![0, 1], "ties resolve to the lower index");
    }

    /// The block cycle must leave the state's SHAPE intact (hc copies in,
    /// hc copies out) and must actually route the block's output back in:
    /// a block that writes a constant has to move every copy.
    #[test]
    fn hc_block_preserves_the_copy_structure_and_applies_the_block() {
        let cfg = Dsv4Cfg {
            dim: 4,
            n_heads: 1,
            head_dim: 4,
            rope_head_dim: 2,
            q_lora_rank: 4,
            o_lora_rank: 2,
            o_groups: 1,
            hc_mult: 4,
            hc_sinkhorn_iters: 20,
            hc_eps: 1e-6,
            norm_eps: 1e-6,
            n_routed_experts: 2,
            top_k: 1,
            moe_inter: 4,
            route_scale: 1.0,
            swiglu_limit: 10.0,
            window: 128,
            index_topk: 4,
            vocab: 8,
        };
        let (hc, dim) = (cfg.hc_mult, cfg.dim);
        let mix_hc = (2 + hc) * hc;
        let hc_fn: Vec<f32> = (0..mix_hc * hc * dim)
            .map(|i| ((i % 13) as f32 - 6.0) * 0.05)
            .collect();
        let hc_base: Vec<f32> = (0..mix_hc).map(|i| (i as f32 * 0.2).sin()).collect();
        let norm_w = vec![1.0f32; dim];
        let mut state: Vec<f32> = (0..hc * dim).map(|i| (i as f32 * 0.3).cos()).collect();
        let before = state.clone();
        let mut scratch = HcScratch::new(&cfg);
        hc_block(
            &mut state,
            &hc_fn,
            &[1.0, 1.0, 1.0],
            &hc_base,
            &norm_w,
            &cfg,
            &mut scratch,
            None,
            |_folded, out: &mut [f32]| out.iter_mut().for_each(|o| *o = 1.0),
        );
        assert_eq!(state.len(), before.len(), "copy structure must survive");
        assert!(state.iter().all(|v| v.is_finite()), "{state:?}");
        assert!(
            state.iter().zip(&before).any(|(a, b)| (a - b).abs() > 1e-4),
            "the block's output has to reach the state"
        );
    }

    #[test]
    fn hash_route_reads_the_table_row() {
        // vocab 3, top_k 2
        let table = [7.0f32, 9.0, 1.0, 2.0, 5.0, 6.0];
        assert_eq!(hash_route(&table, 3, 2, 0), vec![7, 9]);
        assert_eq!(hash_route(&table, 3, 2, 2), vec![5, 6]);
        // out-of-range ids clamp instead of panicking
        assert_eq!(hash_route(&table, 3, 2, 99), vec![5, 6]);
    }

    /// A task mask restricts SELECTION and nothing else: the weights still
    /// come from the pre-bias scores and still renormalize, now over what
    /// survives. Masking must never reroute — an expert the mask forbids has
    /// to be absent, not replaced by a neighbour with the wrong weight.
    #[test]
    fn a_task_mask_restricts_selection_and_renormalizes() {
        // Expert 3 scores highest, then 1, then 2, then 0.
        let scores = [0.1f32, 4.0, 1.0, 9.0];
        let (mut idx, mut w) = (Vec::new(), Vec::new());
        route(&scores, None, 2, 1.0, None, None, &mut idx, &mut w);
        assert_eq!(idx, vec![3, 1], "unmasked: the two best win");
        let sum: f32 = w.iter().sum();
        assert!((sum - 1.0).abs() < 1e-5, "weights must sum to route_scale");

        // Forbid the winner: the next two take its place and the weights
        // renormalize over them.
        let mask = [true, false, true, true];
        let (mut i2, mut w2) = (Vec::new(), Vec::new());
        route(&scores, None, 2, 1.0, None, Some(&mask), &mut i2, &mut w2);
        assert_eq!(i2, vec![3, 2], "masked expert must not be selected");
        let sum2: f32 = w2.iter().sum();
        assert!((sum2 - 1.0).abs() < 1e-5, "masked weights must renormalize");

        // A mask leaving fewer than top_k experts yields fewer, not garbage.
        let tight = [false, false, false, true];
        let (mut i3, mut w3) = (Vec::new(), Vec::new());
        route(&scores, None, 2, 1.0, None, Some(&tight), &mut i3, &mut w3);
        assert_eq!(i3, vec![3]);
        assert_eq!(w3.len(), 1);
    }

    /// On a hash layer the reference gathers the scores AT THE TABLE's
    /// experts. Choosing top-k first and swapping the indices afterwards
    /// leaves every weight attached to a different expert than the one it
    /// scales — silently, since both lists are the right length.
    #[test]
    fn hash_layers_weight_the_experts_the_table_names() {
        // Expert 3 scores highest, expert 0 lowest; the table names 0 and 1.
        let scores = [0.1f32, 0.4, 0.2, 5.0];
        let table = vec![0.0f32, 1.0];
        let idx_forced = hash_route(&table, 1, 2, 0);
        assert_eq!(idx_forced, vec![0, 1]);

        let (mut idx, mut w) = (Vec::new(), Vec::new());
        route(
            &scores,
            None,
            2,
            1.0,
            Some(&idx_forced),
            None,
            &mut idx,
            &mut w,
        );
        assert_eq!(idx, vec![0, 1], "the table must decide the experts");

        // The weights must be the table experts' own scores, normalized.
        let sp = |x: f32| (1.0 + x.exp()).ln().sqrt();
        let (s0, s1) = (sp(scores[0]), sp(scores[1]));
        let tot = s0 + s1;
        assert!(
            (w[0] - s0 / tot).abs() < 1e-6,
            "w[0]={} want {}",
            w[0],
            s0 / tot
        );
        assert!(
            (w[1] - s1 / tot).abs() < 1e-6,
            "w[1]={} want {}",
            w[1],
            s1 / tot
        );

        // And the top-k path is untouched: expert 3 still wins there.
        let (mut idx2, mut w2) = (Vec::new(), Vec::new());
        route(&scores, None, 2, 1.0, None, None, &mut idx2, &mut w2);
        assert_eq!(idx2[0], 3, "without a table the highest score still wins");
    }
}

// ══ DSpark: the block-parallel draft ══════════════════════════════════
//
// Not a classic MTP chain. One pass through the three stages produces the
// WHOLE block of `block_size` positions at once: position 0 carries the token
// the trunk just emitted, the rest carry a noise token, and every position
// attends to every other one — which is why the block cannot be measured a
// position at a time and pretend to be faithful. Depth comes from the block,
// not from the stage count.
//
// The stages' KV cache is built from the trunk's hidden state, not from the
// draft's own tokens: one entry per real position, `kv_norm(wkv(main_x))`,
// in a ring of `window`. The block's own keys and values are appended for
// the duration of the block and then discarded.

/// The noise token the block's unknown positions carry
/// (`dspark_noise_token_id`).
pub const DSPARK_NOISE_TOKEN: u32 = 128799;
/// `dspark_block_size` — the width of the draft block, and NOT a tuning knob.
///
/// All five positions attend to each other and the model was trained with
/// exactly four noise slots behind the real token, so a narrower block is a
/// different draft model, not a cheaper one. What the survival curve argues
/// for is verifying fewer of the five — see `dspark_verify_k` — which costs
/// less without changing what the draft computes.
pub fn dspark_block() -> usize {
    5
}

/// How many of the block's proposals the trunk actually checks.
///
/// Survival is [0.67, 0.50, 0.29, 0.08, 0.04]: positions four and five are
/// paid for on every verify and delivered on a twelfth of them. Three yields
/// 2.46 tokens a cycle against five's 2.58, for three fifths of the verify.
/// `CMF_DSPARK_VERIFY_K=N` sets it.
#[cfg(feature = "gpu")]
fn dspark_native_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| {
        std::env::var("CMF_DSPARK_NATIVE")
            .map(|v| v != "0")
            // The q4 checkpoint's native draft accepts far more proposals
            // than its upload-time q4→q2 recode. A q2 file stays q2 below.
            .unwrap_or(true)
    })
}

pub fn dspark_verify_k() -> usize {
    static K: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
    *K.get_or_init(|| {
        std::env::var("CMF_DSPARK_VERIFY_K")
            .ok()
            .and_then(|v| v.parse::<usize>().ok())
            .filter(|&n| (1..=DSPARK_BLOCK_MAX).contains(&n))
            .unwrap_or(DSPARK_BLOCK_MAX)
    })
}

/// The trained block width.
pub const DSPARK_BLOCK_MAX: usize = 5;

/// Per-sequence state of the draft: one KV ring per stage, and the trunk
/// hidden states the block's input is projected from.
pub struct DsparkState {
    /// `[stage][window * kv_width]`, written at `pos % window`.
    pub win: Vec<Vec<f32>>,
    /// How many real positions each ring holds, capped at `window`.
    pub filled: Vec<usize>,
    /// The trunk's captured hidden, `dim * n_targets`, refreshed every token.
    pub main_hidden: Vec<f32>,
    /// True once `main_hidden` holds this position's capture.
    pub have_hidden: bool,
}

impl DsparkState {
    pub fn new(stages: usize, cfg: &Dsv4Cfg, targets: usize) -> Self {
        Self {
            win: vec![Vec::new(); stages],
            filled: vec![0; stages],
            main_hidden: vec![0.0; cfg.dim * targets],
            have_hidden: false,
        }
    }
}

/// Which trunk layers the draft reads. Upstream names them explicitly
/// (`dspark_target_layer_ids`); the file says the same thing less directly —
/// `main_proj` has one `dim`-wide input block per captured layer — and the
/// release captures the last three. Deriving it from the weight keeps the
/// two from disagreeing.
pub fn dspark_targets(mtp: &[Dsv4Mtp], cfg: &Dsv4Cfg, n_layers: usize) -> Vec<usize> {
    let Some(mp) = mtp.iter().find_map(|m| m.main_proj.as_ref()) else {
        return Vec::new();
    };
    let n = (mp.cols() / cfg.dim.max(1)).clamp(1, n_layers);
    (n_layers - n..n_layers).collect()
}

thread_local! {
    /// The armed capture: which layers to take, and the buffer they fill.
    /// A thread-local rather than a parameter because the capture has to
    /// reach into the middle of a layer loop that eight call sites share,
    /// and threading an optional buffer through all of them to serve one
    /// diagnostic is a worse trade than this.
    static DSPARK_CAP: std::cell::RefCell<(Vec<usize>, Vec<f32>, usize)> =
        const { std::cell::RefCell::new((Vec::new(), Vec::new(), 0)) };
}

/// Arm the capture for the layers `targets`, in order.
pub fn dspark_arm(targets: &[usize], dim: usize) {
    DSPARK_CAP.with(|c| {
        let mut c = c.borrow_mut();
        c.0 = targets.to_vec();
        c.1 = vec![0.0; dim * targets.len()];
        c.2 = 0;
    });
}

/// Whether the armed MTP capture needs the state immediately after `li`.
/// The normal decode path keeps a full run in one submission; DSpark is the
/// only caller that needs an intermediate state to cross the device boundary.
fn dspark_wants(li: usize) -> bool {
    DSPARK_CAP.with(|c| c.borrow().0.contains(&li))
}

/// Called after every host layer. Free when nothing is armed.
pub fn dspark_note(li: usize, state: &[f32], cfg: &Dsv4Cfg) {
    DSPARK_CAP.with(|c| {
        let mut c = c.borrow_mut();
        if c.0.is_empty() {
            return;
        }
        if let Some(slot) = c.0.iter().position(|&t| t == li) {
            let (_, buf, seen) = &mut *c;
            dspark_capture(state, cfg, slot, buf);
            // Counted, not "was the last one" — under the device chain only
            // the layers left on the host call this, and taking the last
            // target as the signal would hand the draft a buffer whose other
            // slots still hold the previous token, or nothing at all.
            *seen = if slot == 0 { 1 } else { *seen + 1 };
            if std::env::var("CMF_DSPARK_CAP_DBG").is_ok() {
                eprintln!("[cap] note li={li} slot={slot} seen={}", *seen);
            }
        }
    });
}

/// Read one slot of the armed capture buffer as-is, complete or not. The
/// speculative verify fills the DEVICE targets from its own photographs and
/// only needs the host layers' slots from here — `dspark_take`'s
/// completeness contract would never be met on that path.
pub fn dspark_peek_slot(slot: usize, dim: usize, out: &mut [f32]) -> bool {
    DSPARK_CAP.with(|c| {
        let c = c.borrow();
        let lo = slot * dim;
        if c.1.len() < lo + dim {
            return false;
        }
        out[..dim].copy_from_slice(&c.1[lo..lo + dim]);
        true
    })
}

/// Move the capture out, if this token produced a complete one.
pub fn dspark_take(out: &mut Vec<f32>) -> bool {
    DSPARK_CAP.with(|c| {
        let mut c = c.borrow_mut();
        if c.0.is_empty() || c.2 != c.0.len() {
            if std::env::var("CMF_DSPARK_CAP_DBG").is_ok() {
                eprintln!("[cap] take FAIL armed={:?} seen={}", c.0, c.2);
            }
            return false;
        }
        out.clear();
        out.extend_from_slice(&c.1);
        c.2 = 0;
        true
    })
}

/// The trunk's contribution: the mean over the hyper-connection copies,
/// appended in target order. Costs one pass over `hc * dim` per captured
/// layer and nothing else.
pub fn dspark_capture(state: &[f32], cfg: &Dsv4Cfg, slot: usize, out: &mut [f32]) {
    let (hc, dim) = (cfg.hc_mult, cfg.dim);
    let dst = &mut out[slot * dim..(slot + 1) * dim];
    let inv = 1.0 / hc as f32;
    for d in 0..dim {
        let mut s = 0.0;
        for j in 0..hc {
            s += state[j * dim + d];
        }
        dst[d] = s * inv;
    }
}

/// `CMF_DSPARK_PICK_DUMP=path` — accumulate the draft's expert picks per
/// stage and periodically rewrite `path` with `stage<TAB>expert<TAB>count`
/// lines. Rewritten every 32 blocks rather than at exit, so a run that is
/// killed still leaves the tallies on disk.
pub fn dspark_freq_note(picks: &[(usize, Vec<usize>)]) {
    static FREQ: std::sync::Mutex<Option<(std::collections::HashMap<(usize, usize), u64>, u64)>> =
        std::sync::Mutex::new(None);
    let Ok(path) = std::env::var("CMF_DSPARK_PICK_DUMP") else {
        return;
    };
    let mut g = FREQ.lock().unwrap();
    let (map, blocks) = g.get_or_insert_with(|| (std::collections::HashMap::new(), 0));
    for (stage, idx) in picks {
        for &e in idx {
            *map.entry((*stage, e)).or_insert(0) += 1;
        }
    }
    *blocks += 1;
    if *blocks % 32 == 0 {
        let mut lines: Vec<_> = map.iter().collect();
        lines.sort();
        let body: String = lines
            .iter()
            .map(|((s, e), n)| format!("{s}\t{e}\t{n}\n"))
            .collect();
        let _ = std::fs::write(&path, body);
    }
}

/// `CMF_DSV4_TRUNK_PICK_DUMP=path` — the same tally for the TRUNK's layers:
/// `layer<TAB>expert<TAB>count`, rewritten every 32 tokens. The pick lists
/// come from the probe's own tally window, so only layers that route on the
/// host are counted — which is exactly the population a partial pack serves.
pub fn trunk_freq_note(picks: &[(usize, Vec<usize>)]) {
    static FREQ: std::sync::Mutex<Option<(std::collections::HashMap<(usize, usize), u64>, u64)>> =
        std::sync::Mutex::new(None);
    let Ok(path) = std::env::var("CMF_DSV4_TRUNK_PICK_DUMP") else {
        return;
    };
    let mut g = FREQ.lock().unwrap();
    let (map, blocks) = g.get_or_insert_with(|| (std::collections::HashMap::new(), 0));
    for (li, idx) in picks {
        for &e in idx {
            *map.entry((*li, e)).or_insert(0) += 1;
        }
    }
    *blocks += 1;
    if *blocks % 32 == 0 {
        let mut lines: Vec<_> = map.iter().collect();
        lines.sort();
        let body: String = lines
            .iter()
            .map(|((l, e), n)| format!("{l}\t{e}\t{n}\n"))
            .collect();
        let _ = std::fs::write(&path, body);
    }
}

/// `CMF_DSPARK_MASK=path` — restrict the draft's routed experts to an
/// explicit per-stage keep-set: line `d` of the file lists the expert ids
/// stage `d` may route to, comma-separated. Weights renormalize over what
/// remains (the `Dsv4Layer::mask` contract). The draft only proposes — the
/// trunk still verifies every token — so a thinner draft costs acceptance,
/// never correctness. This is the offline dial for sizing a resident
/// device pack before one exists.
fn dspark_apply_mask(out: &mut [Dsv4Mtp]) {
    let Ok(path) = std::env::var("CMF_DSPARK_MASK") else {
        return;
    };
    let Ok(text) = std::fs::read_to_string(&path) else {
        eprintln!("DSpark: CMF_DSPARK_MASK={path} не читается — маска не применена");
        return;
    };
    for (d, line) in text.lines().enumerate() {
        let Some(m) = out.get_mut(d) else { break };
        let n = m.layer.experts.len();
        let mut mask = vec![false; n];
        let mut kept = 0usize;
        for tok in line.split(',') {
            if let Ok(e) = tok.trim().parse::<usize>() {
                if e < n && !mask[e] {
                    mask[e] = true;
                    kept += 1;
                }
            }
        }
        if kept == 0 {
            continue;
        }
        eprintln!("DSpark: стадия {d} ограничена {kept}/{n} экспертами");
        m.layer.mask = Some(mask);
    }
}

/// The draft's device residency: which experts of each stage live on the
/// card, and how the device router reaches them.
///
/// The draft only proposes — the trunk verifies every token — so the pack
/// is free to keep a SUBSET of each stage's experts and mask the routing to
/// it: acceptance pays, correctness never does. The subset is chosen by
/// measured routing frequency (`CMF_DSPARK_PACK` names the tally file that
/// `CMF_DSPARK_PICK_DUMP` wrote; `CMF_DSPARK_RESIDENT` caps experts per
/// stage, default 48).
#[cfg(feature = "gpu")]
pub struct DsparkPack {
    pub stages: Vec<DsparkStagePack>,
    /// Gate/up requantized to q2tp at upload (the binary registered an
    /// encoder); the graph then dispatches the q2tp kernels.
    pub gu_q2: bool,
    /// The down planes too (native in the file, never requantized at
    /// upload); the graph dispatches the 2-bit down kernel.
    pub dn_q2: bool,
    /// Dequantized router and bias per stage, f32 — address-stable for the
    /// life of the pack, which is what the device's const cache needs.
    pub routers: Vec<Vec<f32>>,
    pub biases: Vec<Option<Vec<f32>>>,
}

#[cfg(feature = "gpu")]
pub struct DsparkStagePack {
    /// Selectable experts (true = resident).
    pub mask: Vec<bool>,
    /// Global expert id → pack slot; usize::MAX where cold.
    pub to_slot: Vec<usize>,
    /// The same two as the device consumes them — u32, address-stable for
    /// the pack's lifetime (the const cache keys on the pointer).
    pub mask_u32: Vec<u32>,
    pub map_u32: Vec<u32>,
    /// (gate, up, down) directory indices, pack order, shared LAST.
    pub tensors: Vec<(usize, usize, usize)>,
    pub n_resident: usize,
    /// A protected subset inside the trunk's model-wide physical banks.
    /// None keeps the legacy independent local pack for adapters without
    /// descriptor-indexed storage arrays and for q2tp.
    global: Option<GlobalPack>,
}

/// The q2tp encoder, registered by the binary that has one (the CLI's
/// converter owns the rung-search implementation and the engine must not
/// depend on the CLI). When present, the draft's gate/up experts are
/// requantized q4tp → q2tp AT UPLOAD — half the VRAM and the same kernels
/// the trunk's q2tp experts already use. Draft-only fidelity: acceptance
/// pays, correctness never does.
pub static DSPARK_Q2TP_ENCODE: std::sync::OnceLock<fn(&[f32], usize, usize) -> Vec<u8>> =
    std::sync::OnceLock::new();

/// `CMF_DSPARK_GPU=1` — the probe (and later the speculative loop) drafts
/// on the card instead of the CPU/disk tier.
#[cfg(feature = "gpu")]
pub fn dspark_gpu_on() -> bool {
    static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
    *ON.get_or_init(|| {
        std::env::var("CMF_DSPARK_GPU")
            .map(|v| v != "0")
            .unwrap_or(true)
    })
}

/// The pack, built once per process (the stand runs one model).
#[cfg(feature = "gpu")]
pub fn dspark_pack_get(mtp: &[Dsv4Mtp], cfg: &Dsv4Cfg) -> Option<&'static DsparkPack> {
    static P: std::sync::OnceLock<Option<Box<DsparkPack>>> = std::sync::OnceLock::new();
    P.get_or_init(|| dspark_pack_build(mtp, cfg).map(Box::new))
        .as_deref()
}

/// Build and upload the draft's pack. Returns `None` when the stack is
/// absent, the budget refuses, or a stage's weights are not where the
/// device path needs them — the caller falls back to the CPU draft.
/// Reserve the VRAM the speculative draft's device pack will take, so the
/// trunk's greedy packing leaves it room. Called at load, before any trunk
/// pack is built; a no-op when there is no MTP stack or speculation is off.
/// The estimate uses the draft's native dtypes — an upload-time re-encode
/// only shrinks it, which errs on the safe side of the physical ceiling.
///
/// A budget that cannot pack the trunk to the draft's capture layers gets NO
/// reservation.  Host-batch verify is exact there, but the measured A40 q4tp
/// result is 0.63 tok/s versus the faster ordinary exact walk: reserving the
/// draft shrinks every trunk layer and makes speculation a net loss.  The
/// threshold is geometric (nine tenths of the trunk's own expert bytes plus
/// the draft), never a card name.
#[cfg(feature = "gpu")]
pub fn dspark_reserve_note(mtp: &[Dsv4Mtp], cfg: &Dsv4Cfg, layers: &[Dsv4Layer]) {
    if mtp.is_empty() || std::env::var("CMF_DSV4_SPEC").is_ok_and(|v| v == "0") || !dspark_gpu_on()
    {
        return;
    }
    // `=1` is the diagnostic force path used to measure a configuration the
    // zero-knob geometric gate would reject.  Production auto-selection keeps
    // the gate below; forcing must reserve before the trunk is packed or the
    // late draft upload simply OOMs.
    let forced = std::env::var("CMF_DSV4_SPEC").is_ok_and(|v| v == "1");
    let dt = |q2: bool| {
        if q2 {
            cortiq_core::TensorDtype::Q2TiledP
        } else {
            cortiq_core::TensorDtype::Q4TiledP
        }
    };
    let gu_q2 = mtp[0]
        .layer
        .experts
        .first()
        .is_some_and(|e| e.w1.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP));
    let dn_q2 = mtp[0]
        .layer
        .experts
        .first()
        .is_some_and(|e| e.w2.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP));
    // A native-q4 draft can share the descriptor-indexed trunk arena and
    // therefore needs no second expert reservation.  Do not arm automatic
    // speculation here yet: a frequency pack says which experts are common,
    // not that the draft's proposals pay for a five-token exact verify.  The
    // explicit diagnostic gate can exercise the shared path without OOM;
    // zero-knob production remains the measured faster exact walk until an
    // acceptance profile has passed its own quality/speed gate.
    if !gu_q2
        && !dn_q2
        && std::env::var("CMF_MOE_MASK").is_err()
        && crate::gpu_wgpu::dsv4_global_moe_supported()
    {
        crate::gpu_wgpu::DRAFT_PACK_RESERVE.store(0, std::sync::atomic::Ordering::Relaxed);
        crate::gpu_wgpu::DRAFT_RESERVE.store(0, std::sync::atomic::Ordering::Relaxed);
        return;
    }
    let gu = cortiq_core::quant::expected_nbytes(dt(gu_q2), &[cfg.moe_inter, cfg.dim]).unwrap_or(0);
    let dn = cortiq_core::quant::expected_nbytes(dt(dn_q2), &[cfg.dim, cfg.moe_inter]).unwrap_or(0);
    let per = (2 * gu + dn) as u64;
    let n_res: usize = std::env::var("CMF_DSPARK_RESIDENT")
        .ok()
        .and_then(|v| v.parse().ok())
        // The default matches the measured acceptance plateau's low edge:
        // residency below it costs acceptance, above it only costs VRAM.
        .unwrap_or(40);
    // Routed residents per stage, plus each stage's shared expert.
    let bytes = per * (n_res * mtp.len() + mtp.len() + 1) as u64;
    // The trunk's own expert bytes, from the route it will actually serve.
    // A task-specialist mask changes the physical working set: counting all
    // 256 rows here made a compact, fully resident masked trunk look like the
    // 158 GB general model, so the zero-knob gate silently disabled the draft
    // that is responsible for the second half of its speedup.  Hash layers
    // deliberately have no mask and still count every checkpoint-named row.
    // Keep the production gate used by the zero-knob path: if nearly all of
    // the effective trunk plus the draft cannot fit, spend the whole budget
    // on trunk slots.
    let trunk: u64 = layers
        .iter()
        .map(|l| {
            let Some(e) = l.experts.first() else {
                return 0;
            };
            let gu = cortiq_core::quant::expected_nbytes(
                dt(e.w1.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP)),
                &[cfg.moe_inter, cfg.dim],
            )
            .unwrap_or(0);
            let dn = cortiq_core::quant::expected_nbytes(
                dt(e.w2.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP)),
                &[cfg.dim, cfg.moe_inter],
            )
            .unwrap_or(0);
            let routed = l
                .mask
                .as_deref()
                .map_or(l.experts.len(), |m| m.iter().filter(|&&open| open).count());
            ((2 * gu + dn) * (routed + 1)) as u64
        })
        .sum();
    if let Some(budget) = crate::gpu_wgpu::dsv4_vram_budget() {
        if !forced && budget < trunk / 10 * 9 + bytes {
            return;
        }
    }
    // The pack is not the draft's whole physical footprint.  Its three
    // attention skeletons, block-axis activations, captures and retained
    // verify states are ordinary wgpu allocations and therefore do not
    // appear in the resident-weight ledger.  Keeping only `bytes` here made
    // q4tp fit on paper and then panic the A40 driver while building DSpark.
    // A geometry-scaled workspace (bounded to 512..1024 MiB) is separate from
    // the expert reservation: dsv4_draft_fit must hand back only PACK bytes,
    // never turn scratch headroom into more resident experts.
    let mib = 1024 * 1024u64;
    let workspace = match crate::gpu_wgpu::dsv4_vram_budget() {
        // Smaller discrete heaps have less slack between the reported weight
        // ceiling and the driver's physical allocation ceiling.  One GiB is
        // still only ~2% of an A40 and is cheaper than an OOM/restart.
        Some(b) if b <= 64 * 1024 * mib => 1024 * mib,
        _ => {
            ((cfg.dim * cfg.hc_mult * DSPARK_BLOCK_MAX * 4096) as u64).clamp(512 * mib, 1024 * mib)
        }
    };
    crate::gpu_wgpu::DRAFT_PACK_RESERVE.store(bytes, std::sync::atomic::Ordering::Relaxed);
    crate::gpu_wgpu::DRAFT_RESERVE.store(
        bytes.saturating_add(workspace),
        std::sync::atomic::Ordering::Relaxed,
    );
}

#[cfg(not(feature = "gpu"))]
pub fn dspark_reserve_note(_mtp: &[Dsv4Mtp], _cfg: &Dsv4Cfg, _layers: &[Dsv4Layer]) {}

#[cfg(feature = "gpu")]
pub fn dspark_pack_build(mtp: &[Dsv4Mtp], cfg: &Dsv4Cfg) -> Option<DsparkPack> {
    if mtp.is_empty() {
        return None;
    }
    let model = mtp[0]
        .layer
        .experts
        .first()
        .and_then(|e| e.w1.model_arc())?;
    let native_q2 = mtp[0]
        .layer
        .experts
        .first()
        .is_some_and(|e| e.w1.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP));
    let dn_native = mtp[0]
        .layer
        .experts
        .first()
        .is_some_and(|e| e.w2.model_dtype() == Some(cortiq_core::TensorDtype::Q2TiledP));
    let global_share = !native_q2
        && !dn_native
        && std::env::var("CMF_MOE_MASK").is_err()
        && crate::gpu_wgpu::dsv4_global_moe_supported()
        && crate::gpu_wgpu::dsv4_global_moe_ready(&model);
    let n_res: usize = std::env::var("CMF_DSPARK_RESIDENT")
        .ok()
        .and_then(|v| v.parse().ok())
        .unwrap_or_else(|| {
            if global_share {
                // The measured acceptance plateau starts here. These
                // rows occupy existing arena slots rather than asking
                // the weight allocator for another multi-GiB pack.
                return 40;
            }
            // No knob: take what the card actually has left, whatever the
            // card is. The stages split the fit evenly after their shared
            // experts. Forty is the measured acceptance plateau: larger
            // packs still make the router and upload more rows without a
            // useful increase in accepted tokens (64 was slower on A40).
            let gu_q2 = native_q2 || (!dspark_native_on() && DSPARK_Q2TP_ENCODE.get().is_some());
            let room = crate::gpu_wgpu::dsv4_draft_fit(cfg.moe_inter, cfg.dim, gu_q2, dn_native);
            (room.saturating_sub(mtp.len() + 1) / mtp.len().max(1)).clamp(8, 40)
        });
    // Frequency tallies: lines of `stage<TAB>expert<TAB>count`. Named by
    // `CMF_DSPARK_PACK`, or found as `<model>.dspark.tsv` beside the model
    // file — ship the tally next to the checkpoint and no knob is needed.
    let mut freq: Vec<Vec<(u64, usize)>> = vec![Vec::new(); mtp.len()];
    let pack_path = std::env::var("CMF_DSPARK_PACK").ok().or_else(|| {
        let m = mtp[0].layer.experts.first()?.w1.model_arc()?;
        let mut s = m.path.as_os_str().to_os_string();
        s.push(".dspark.tsv");
        let p = std::path::PathBuf::from(s);
        p.exists().then(|| p.to_string_lossy().into_owned())
    });
    if let Some(path) = pack_path {
        if let Ok(text) = std::fs::read_to_string(&path) {
            for line in text.lines() {
                let mut it = line.split_whitespace();
                if let (Some(s), Some(e), Some(n)) = (it.next(), it.next(), it.next()) {
                    if let (Ok(s), Ok(e), Ok(n)) =
                        (s.parse::<usize>(), e.parse::<usize>(), n.parse::<u64>())
                    {
                        if s < freq.len() {
                            freq[s].push((n, e));
                        }
                    }
                }
            }
        }
    }
    let mut stages = Vec::with_capacity(mtp.len());
    let mut routers = Vec::with_capacity(mtp.len());
    let mut biases = Vec::with_capacity(mtp.len());
    for (si, m) in mtp.iter().enumerate() {
        let l = &m.layer;
        let n = l.experts.len();
        // Frequency order, then the untallied ids — a cold start still
        // packs SOMETHING deterministic.
        let mut order: Vec<usize> = {
            let mut f = freq[si].clone();
            f.sort_by(|a, b| b.0.cmp(&a.0));
            let mut seen = vec![false; n];
            let mut o: Vec<usize> = f
                .into_iter()
                .map(|(_, e)| e)
                .filter(|&e| {
                    if e < n && !seen[e] {
                        seen[e] = true;
                        true
                    } else {
                        false
                    }
                })
                .collect();
            o.extend((0..n).filter(|&e| !seen[e]));
            o
        };
        order.truncate(n_res.min(n));
        let (global, mask, to_slot, tensors) = if global_share {
            let pk = pack_for(l, cfg, model.header.arch.num_layers + si)?;
            let gl = pk.global.clone()?;
            let remap = gl.pool.pin_picks(&model, gl.layer, &order, &l.experts);
            let mut mask = vec![false; n];
            let mut to_slot = vec![usize::MAX; n];
            for e in 0..n {
                if remap.get(e).copied().unwrap_or(u32::MAX) != u32::MAX {
                    mask[e] = true;
                    to_slot[e] = remap[e] as usize;
                }
            }
            (Some(gl), mask, to_slot, Vec::new())
        } else {
            let mut mask = vec![false; n];
            let mut to_slot = vec![usize::MAX; n];
            let mut tensors = Vec::with_capacity(order.len() + 1);
            for (slot, &e) in order.iter().enumerate() {
                let ex = &l.experts[e];
                let (Some(w1), Some(w3), Some(w2)) =
                    (ex.w1.model_idx(), ex.w3.model_idx(), ex.w2.model_idx())
                else {
                    return None;
                };
                mask[e] = true;
                to_slot[e] = slot;
                tensors.push((w1, w3, w2));
            }
            let (Some(s1), Some(s3), Some(s2)) = (
                l.shared.w1.model_idx(),
                l.shared.w3.model_idx(),
                l.shared.w2.model_idx(),
            ) else {
                return None;
            };
            tensors.push((s1, s3, s2));
            (None, mask, to_slot, tensors)
        };
        // The router and bias, dequantized once.
        let mut router = vec![0.0f32; n * cfg.dim];
        for (r, row) in (0..n).zip(router.chunks_mut(cfg.dim)) {
            l.gate.row_f32(r, row);
        }
        routers.push(router);
        biases.push(l.gate_bias.clone());
        let mask_u32: Vec<u32> = mask.iter().map(|&m| m as u32).collect();
        let map_u32: Vec<u32> = to_slot
            .iter()
            .map(|&x| if x == usize::MAX { u32::MAX } else { x as u32 })
            .collect();
        stages.push(DsparkStagePack {
            mask,
            to_slot,
            mask_u32,
            map_u32,
            tensors,
            n_resident: order.len(),
            global,
        });
    }
    // ── upload: the small skeleton FIRST, the expert stacks after — the
    //    documented admission order (experts fill the card and the skeleton
    //    then misses). ──
    let mut skeleton = Vec::new();
    for m in mtp {
        let l = &m.layer;
        for t in [&l.wq_a, &l.wq_b, &l.wkv, &l.wo_a, &l.wo_b] {
            skeleton.push(t.model_idx()?);
        }
    }
    if let Some(mp) = mtp[0].main_proj.as_ref() {
        skeleton.push(mp.model_idx()?);
    }
    for &idx in &skeleton {
        if !crate::gpu_wgpu::dsv4_weight_ready(&model, idx) {
            eprintln!("DSpark: скелет драфта не влез в VRAM — GPU-черновик выключен");
            return None;
        }
    }
    // The dtype in the FILE decides: a properly converted CMF stores the
    // draft's gate/up as q2tp and uploads through the same path as the
    // trunk's 2-bit experts. The at-upload requant is only the fallback for
    // files published before the converter's q2tp profile covered the MTP
    // stack (and only when the binary registered an encoder).
    let gu_q2 = native_q2
        || (!crate::dsv4::dspark_native_on() && crate::dsv4::DSPARK_Q2TP_ENCODE.get().is_some());
    for (si, sp) in stages.iter().enumerate() {
        let ok = if sp.global.is_some() {
            true
        } else if native_q2 {
            crate::gpu_wgpu::dsv4_experts_ready(
                &model,
                &sp.tensors,
                cfg.moe_inter,
                cfg.dim,
                true,
                dn_native,
            )
        } else if gu_q2 {
            crate::gpu_wgpu::moe_expert_bufs_requant_gu(&model, &sp.tensors, cfg.moe_inter, cfg.dim)
                .is_some()
        } else {
            crate::gpu_wgpu::dsv4_experts_ready(
                &model,
                &sp.tensors,
                cfg.moe_inter,
                cfg.dim,
                false,
                false,
            )
        };
        if !ok {
            eprintln!(
                "DSpark: эксперты стадии {si} ({} + shared) не влезли в VRAM — GPU-черновик выключен",
                sp.n_resident
            );
            return None;
        }
    }
    let _ = crate::gpu_wgpu::pin_weights(&model, &skeleton);
    eprintln!(
        "DSpark: {} — {} стадии по {} экспертов + shared",
        if global_share {
            "защищённая доля единого пула"
        } else {
            "отдельный пак драфта на карте"
        },
        stages.len(),
        stages
            .iter()
            .map(|s| s.n_resident.to_string())
            .collect::<Vec<_>>()
            .join("/")
    );
    Some(DsparkPack {
        stages,
        gu_q2: if global_share { false } else { gu_q2 },
        dn_q2: dn_native,
        routers,
        biases,
    })
}

/// Append one real position's entry to every stage's KV ring, from the
/// trunk captures in `ds.main_hidden`. The draft does this for the position
/// it drafts at; a speculative decode also owes an entry for every accepted
/// position it never drafted from — a hole in the ring silently starves
/// later blocks of context, which reads as "acceptance decayed" and not as
/// a bug.
pub fn dspark_ring_append(
    g: &Dsv4Globals,
    mtp: &[Dsv4Mtp],
    cfg: &Dsv4Cfg,
    ds: &mut DsparkState,
    pos: usize,
    pool: Option<&crate::pool::Pool>,
) {
    let (dim, hd, rd) = (cfg.dim, cfg.head_dim, cfg.rope_head_dim);
    let inv_freq = &g.inv_freq_window;
    let Some(stage0) = mtp.first() else { return };
    let (Some(mp), Some(mn)) = (stage0.main_proj.as_ref(), stage0.main_norm.as_ref()) else {
        return;
    };
    let mut main_x = vec![0.0f32; dim];
    mp.matvec(&ds.main_hidden, &mut main_x, pool);
    rms_weighted(&mut main_x, mn, cfg.norm_eps);
    for (si, m) in mtp.iter().enumerate() {
        let kvw = m.layer.wkv.rows();
        if ds.win[si].len() < cfg.window * kvw {
            ds.win[si].resize(cfg.window * kvw, 0.0);
        }
        let mut kv = vec![0.0f32; kvw];
        m.layer.wkv.matvec(&main_x, &mut kv, pool);
        rms_weighted(&mut kv, &m.layer.kv_norm, cfg.norm_eps);
        rope_tail(&mut kv[kvw - hd..], inv_freq, pos, rd, false);
        let slot = pos % cfg.window;
        ds.win[si][slot * kvw..(slot + 1) * kvw].copy_from_slice(&kv);
        ds.filled[si] = (pos + 1).min(cfg.window);
    }
}

/// The draft block on the card: one submission for all three stages and
/// five positions, states home in one fence, the head on the host. The
/// markov bias is skipped (its per-position chain through the previous
/// PROPOSAL is the one part a single graph cannot batch) — compare against
/// the CPU draft under `CMF_DSPARK_NO_MARKOV=1`.
#[cfg(feature = "gpu")]
#[allow(clippy::too_many_arguments)]
pub fn dspark_draft_gpu(
    g: &Dsv4Globals,
    mtp: &[Dsv4Mtp],
    cfg: &Dsv4Cfg,
    ds: &mut DsparkState,
    pack: &DsparkPack,
    kv_id: u64,
    last_token: u32,
    pos: usize,
    pool: Option<&crate::pool::Pool>,
    out_conf: &mut Vec<f32>,
) -> Vec<u32> {
    let (hc, dim) = (cfg.hc_mult, cfg.dim);
    let block = dspark_block();
    let Some(model) = mtp[0].layer.experts.first().and_then(|e| e.w1.model_arc()) else {
        return Vec::new();
    };
    let (Some(mp), Some(mn)) = (mtp[0].main_proj.as_ref(), mtp[0].main_norm.as_ref()) else {
        return Vec::new();
    };
    let Some(mp_idx) = mp.model_idx() else {
        return Vec::new();
    };
    let mut stages = Vec::with_capacity(mtp.len());
    for (si, m) in mtp.iter().enumerate() {
        let l = &m.layer;
        let (Some(wq_a), Some(wq_b), Some(wo_a), Some(wo_b), Some(wkv)) = (
            l.wq_a.model_idx(),
            l.wq_b.model_idx(),
            l.wo_a.model_idx(),
            l.wo_b.model_idx(),
            l.wkv.model_idx(),
        ) else {
            return Vec::new();
        };
        let sp = &pack.stages[si];
        stages.push(crate::gpu_wgpu::DsparkStageW {
            wq_a,
            wq_b,
            wo_a,
            wo_b,
            wkv,
            q_norm: &l.q_norm,
            kv_norm: &l.kv_norm,
            attn_norm: &l.attn_norm,
            ffn_norm: &l.ffn_norm,
            sink: &l.attn_sink,
            hc_attn_fn: &l.hc_attn_fn,
            hc_attn_scale: &l.hc_attn_scale,
            hc_attn_base: &l.hc_attn_base,
            hc_ffn_fn: &l.hc_ffn_fn,
            hc_ffn_scale: &l.hc_ffn_scale,
            hc_ffn_base: &l.hc_ffn_base,
            router: &pack.routers[si],
            bias: pack.biases[si].as_deref(),
            experts: &sp.tensors,
            mask_u32: &sp.mask_u32,
            map_u32: &sp.map_u32,
            global: sp.global.as_ref().map(|gl| crate::gpu_wgpu::Dsv4GlobalMoe {
                pool_uid: gl.pool.uid,
                shared_slot: gl.shared_slot,
                segment_slots: gl.pool.segment_slots as u32,
            }),
        });
    }
    let geom = crate::gpu_wgpu::DsparkGeom {
        dim,
        hc,
        nh: cfg.n_heads,
        hd: cfg.head_dim,
        rd: cfg.rope_head_dim,
        q_lora: cfg.q_lora_rank,
        o_lora: cfg.o_lora_rank,
        o_groups: cfg.o_groups,
        inter: cfg.moe_inter,
        n_experts: cfg.n_routed_experts,
        top_k: cfg.top_k,
        window: cfg.window,
        eps: cfg.norm_eps,
        hc_eps: cfg.hc_eps,
        sinkhorn_iters: cfg.hc_sinkhorn_iters,
        route_scale: cfg.route_scale,
        swiglu_limit: cfg.swiglu_limit,
        scale: (cfg.head_dim as f32).powf(-0.5),
        gu_q2: pack.gu_q2,
        dn_q2: pack.dn_q2,
    };
    // ── seed states: the real token, then noise, replicated over copies ──
    let ids: Vec<u32> = (0..block)
        .map(|i| {
            if i == 0 {
                last_token
            } else {
                DSPARK_NOISE_TOKEN
            }
        })
        .collect();
    let mut states0 = vec![0.0f32; block * hc * dim];
    let mut emb = vec![0.0f32; dim];
    for (i, &id) in ids.iter().enumerate() {
        g.embed.row_f32(id as usize, &mut emb);
        for j in 0..hc {
            states0[(i * hc + j) * dim..(i * hc + j + 1) * dim].copy_from_slice(&emb);
        }
    }
    let dspark_time = {
        static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
        *ON.get_or_init(|| std::env::var("CMF_DSPARK_TIME").is_ok_and(|v| v != "0"))
    };
    let t0 = std::time::Instant::now();
    let filled = (pos + 1).min(cfg.window);
    let mut states = vec![0.0f32; block * hc * dim];
    if !crate::gpu_wgpu::dspark_graph(
        &model,
        &stages,
        geom,
        kv_id,
        mp_idx,
        mn,
        &ds.main_hidden,
        &states0,
        pos,
        filled,
        &g.inv_freq_window,
        block,
        &mut states,
    ) {
        return Vec::new();
    }
    for si in 0..mtp.len() {
        ds.filled[si] = filled;
    }
    let t_graph = t0.elapsed();

    // ── head, on the host: fold, norm, one B-wide matmat, argmax ──
    let last = &mtp[mtp.len() - 1];
    let (Some(hfn), Some(hbase), Some(hscale), Some(hnorm)) = (
        last.hc_head_fn.as_ref(),
        last.hc_head_base.as_ref(),
        last.hc_head_scale,
        last.norm.as_ref(),
    ) else {
        return Vec::new();
    };
    let mut head_in = vec![0.0f32; block * dim];
    let mut pre_norms = vec![vec![0.0f32; dim]; block];
    for i in 0..block {
        hc_head_fold(
            &states[i * hc * dim..(i + 1) * hc * dim],
            hfn,
            hscale,
            hbase,
            cfg,
            pool,
            &mut head_in[i * dim..(i + 1) * dim],
        );
        pre_norms[i].copy_from_slice(&head_in[i * dim..(i + 1) * dim]);
        rms_weighted(&mut head_in[i * dim..(i + 1) * dim], hnorm, cfg.norm_eps);
    }
    let t_fold = t0.elapsed();
    let mut logits = vec![0.0f32; block * cfg.vocab];
    // The B-axis q4tp kernel, one submission: `matmat` at B=5 falls to the
    // CPU tile path and measured 46 ms of a 60 ms draft.
    let head_gpu = g.head.model_idx().is_some_and(|hi| {
        crate::gpu_wgpu::q4tp_matvec_batch_for_test(
            &model,
            hi,
            &head_in,
            block,
            cfg.vocab,
            dim,
            &mut logits,
        )
    });
    if !head_gpu {
        g.head.matmat(&head_in, block, &mut logits, pool);
    }
    let t_head = t0.elapsed();
    // The markov bigram is not optional: without it acceptance fell 1.02 →
    // 0.42 on natural text. Its chain runs through the previous PROPOSAL,
    // so it stays position-by-position; the w2 matvec is big enough that
    // the QTensor route puts it on the card by itself.
    let mut proposals = Vec::with_capacity(block);
    out_conf.clear();
    let mut prev = last_token;
    let mut mk_embed = vec![0.0f32; last.markov_w1.as_ref().map_or(0, |t| t.cols())];
    let mut bias = vec![0.0f32; cfg.vocab];
    for i in 0..block {
        let row = &mut logits[i * cfg.vocab..(i + 1) * cfg.vocab];
        if let (Some(w1), Some(w2)) = (last.markov_w1.as_ref(), last.markov_w2.as_ref()) {
            w1.row_f32(prev as usize, &mut mk_embed);
            w2.matvec(&mk_embed, &mut bias, pool);
            for (a, b) in row.iter_mut().zip(&bias) {
                *a += *b;
            }
        }
        let mut best = 0usize;
        for v in 1..row.len() {
            if row[v] > row[best] {
                best = v;
            }
        }
        if let Some(cf) = last.confidence.as_ref() {
            let mut cat = pre_norms[i].clone();
            cat.extend_from_slice(&mk_embed);
            let mut sc = [0.0f32; 1];
            if cat.len() == cf.cols() {
                cf.matvec(&cat, &mut sc, pool);
            }
            out_conf.push(sc[0]);
        }
        proposals.push(best as u32);
        prev = best as u32;
    }
    if dspark_time {
        eprintln!(
            "DSpark GPU: граф {:.1} мс, фолды {:.1}, голова {:.1}, марков+argmax {:.1}",
            t_graph.as_secs_f64() * 1e3,
            (t_fold - t_graph).as_secs_f64() * 1e3,
            (t_head - t_fold).as_secs_f64() * 1e3,
            (t0.elapsed() - t_head).as_secs_f64() * 1e3,
        );
    }
    proposals
}

/// One draft: `DSPARK_BLOCK` proposed tokens and a confidence per position.
///
/// `pos` is the position of `last_token` — the block predicts `pos+1 ..
/// pos+BLOCK`. Returns the proposals in order; `out_conf` takes the
/// confidence head's score where the last stage carries one.
#[allow(clippy::too_many_arguments)]
pub fn dspark_draft(
    g: &Dsv4Globals,
    mtp: &[Dsv4Mtp],
    cfg: &Dsv4Cfg,
    ds: &mut DsparkState,
    last_token: u32,
    pos: usize,
    pool: Option<&crate::pool::Pool>,
    out_conf: &mut Vec<f32>,
) -> Vec<u32> {
    let (hc, dim, hd, rd) = (cfg.hc_mult, cfg.dim, cfg.head_dim, cfg.rope_head_dim);
    let block = dspark_block();
    let inv_freq = &g.inv_freq_window;

    // ── the block's input: main_norm(main_proj(captured hiddens)) ──
    let Some(stage0) = mtp.first() else {
        return Vec::new();
    };
    let (Some(_mp), Some(_mn)) = (stage0.main_proj.as_ref(), stage0.main_norm.as_ref()) else {
        return Vec::new();
    };
    dspark_ring_append(g, mtp, cfg, ds, pos, pool);

    // ── the block: the real token, then noise ──
    let ids: Vec<u32> = (0..block)
        .map(|i| {
            if i == 0 {
                last_token
            } else {
                DSPARK_NOISE_TOKEN
            }
        })
        .collect();
    let mut states = vec![vec![0.0f32; hc * dim]; block];
    let mut emb = vec![0.0f32; dim];
    for (i, &id) in ids.iter().enumerate() {
        g.embed.row_f32(id as usize, &mut emb);
        for j in 0..hc {
            states[i][j * dim..(j + 1) * dim].copy_from_slice(&emb);
        }
    }

    let mut scratch = HcScratch::new(cfg);
    for (si, m) in mtp.iter().enumerate() {
        let l = &m.layer;
        let kvw = l.wkv.rows();
        // ── attention half: fold every position first, because each one's
        //    keys are visible to all the others. ──
        let mut post = vec![vec![0.0f32; hc]; block];
        let mut comb = vec![vec![0.0f32; hc * hc]; block];
        let mut resid = vec![vec![0.0f32; hc * dim]; block];
        let mut folded = vec![vec![0.0f32; dim]; block];
        let mix_hc = (2 + hc) * hc;
        for i in 0..block {
            hc_mixes(
                &states[i],
                &l.hc_attn_fn,
                mix_hc,
                cfg.norm_eps,
                pool,
                &mut scratch.mixes,
            );
            hc_split_sinkhorn(
                &scratch.mixes,
                &l.hc_attn_scale,
                &l.hc_attn_base,
                hc,
                cfg.hc_sinkhorn_iters,
                cfg.hc_eps,
                &mut scratch.pre,
                &mut post[i],
                &mut comb[i],
            );
            hc_fold(&states[i], &scratch.pre, hc, dim, &mut folded[i]);
            rms_weighted(&mut folded[i], &l.attn_norm, cfg.norm_eps);
            resid[i].copy_from_slice(&states[i]);
        }
        // Keys and values of the block itself — kept for this block only.
        let folded_all: Vec<f32> = folded.iter().flatten().copied().collect();
        let mut blk_kv = vec![0.0f32; block * kvw];
        l.wkv.matmat(&folded_all, block, &mut blk_kv, pool);
        for i in 0..block {
            let dst = &mut blk_kv[i * kvw..(i + 1) * kvw];
            rms_weighted(dst, &l.kv_norm, cfg.norm_eps);
            rope_tail(&mut dst[kvw - hd..], inv_freq, pos + 1 + i, rd, false);
        }
        // The attended set: every cached real position, then the whole block.
        let win_len = ds.filled[si];
        let mut cache = Vec::with_capacity((win_len + block) * hd);
        for p in 0..win_len {
            let e = &ds.win[si][p * kvw..(p + 1) * kvw];
            cache.extend_from_slice(&e[kvw - hd..]);
        }
        for i in 0..block {
            let e = &blk_kv[i * kvw..(i + 1) * kvw];
            cache.extend_from_slice(&e[kvw - hd..]);
        }
        let idxs: Vec<usize> = (0..win_len + block).collect();
        let scale = (hd as f32).powf(-0.5);
        let qrank = l.wq_a.rows();
        let qdim = cfg.n_heads * hd;
        let mut qr = vec![0.0f32; block * qrank];
        l.wq_a.matmat(&folded_all, block, &mut qr, pool);
        for i in 0..block {
            rms_weighted(&mut qr[i * qrank..(i + 1) * qrank], &l.q_norm, cfg.norm_eps);
        }
        let mut q = vec![0.0f32; block * qdim];
        l.wq_b.matmat(&qr, block, &mut q, pool);
        let mut attn = vec![0.0f32; block * qdim];
        for i in 0..block {
            let qi = &mut q[i * qdim..(i + 1) * qdim];
            let ai = &mut attn[i * qdim..(i + 1) * qdim];
            let qpos = pos + 1 + i;
            for h in 0..cfg.n_heads {
                let head = &mut qi[h * hd..(h + 1) * hd];
                rms_inplace(head, cfg.norm_eps);
                rope_tail(head, inv_freq, qpos, rd, false);
            }
            for h in 0..cfg.n_heads {
                let qh = &qi[h * hd..(h + 1) * hd];
                let oh = &mut ai[h * hd..(h + 1) * hd];
                sparse_attend(qh, &cache, &idxs, l.attn_sink[h], scale, hd, oh);
                rope_tail(oh, inv_freq, qpos, rd, true);
            }
        }
        let mut blk_out = vec![0.0f32; block * dim];
        o_project_block(
            &attn,
            block,
            &l.wo_a,
            &l.wo_b,
            cfg.o_groups,
            cfg.o_lora_rank,
            pool,
            &mut blk_out,
        );
        for i in 0..block {
            let mut next = vec![0.0f32; hc * dim];
            hc_expand(
                &blk_out[i * dim..(i + 1) * dim],
                &resid[i],
                &post[i],
                &comb[i],
                hc,
                dim,
                &mut next,
            );
            states[i] = next;
        }
        // ── MoE: fold every position, group equal experts, then expand in
        //    the original per-position route order. ──
        let mut ffn_fold = vec![0.0f32; block * dim];
        let mut ffn_post = vec![vec![0.0f32; hc]; block];
        let mut ffn_comb = vec![vec![0.0f32; hc * hc]; block];
        let mut ffn_resid = vec![vec![0.0f32; hc * dim]; block];
        for i in 0..block {
            hc_mixes(
                &states[i],
                &l.hc_ffn_fn,
                mix_hc,
                cfg.norm_eps,
                pool,
                &mut scratch.mixes,
            );
            hc_split_sinkhorn(
                &scratch.mixes,
                &l.hc_ffn_scale,
                &l.hc_ffn_base,
                hc,
                cfg.hc_sinkhorn_iters,
                cfg.hc_eps,
                &mut scratch.pre,
                &mut ffn_post[i],
                &mut ffn_comb[i],
            );
            hc_fold(
                &states[i],
                &scratch.pre,
                hc,
                dim,
                &mut ffn_fold[i * dim..(i + 1) * dim],
            );
            rms_weighted(
                &mut ffn_fold[i * dim..(i + 1) * dim],
                &l.ffn_norm,
                cfg.norm_eps,
            );
            ffn_resid[i].copy_from_slice(&states[i]);
        }
        let mut moe_out = vec![0.0f32; block * dim];
        moe_step_block(&ffn_fold, block, l, cfg, &ids, si, pool, &mut moe_out);
        for i in 0..block {
            let mut next = vec![0.0f32; hc * dim];
            hc_expand(
                &moe_out[i * dim..(i + 1) * dim],
                &ffn_resid[i],
                &ffn_post[i],
                &ffn_comb[i],
                hc,
                dim,
                &mut next,
            );
            states[i] = next;
        }
    }

    // ── head: the last stage's fold, the trunk's own head ──
    let last = &mtp[mtp.len() - 1];
    let (Some(hfn), Some(hbase), Some(hscale), Some(hnorm)) = (
        last.hc_head_fn.as_ref(),
        last.hc_head_base.as_ref(),
        last.hc_head_scale,
        last.norm.as_ref(),
    ) else {
        return Vec::new();
    };
    let mut proposals = Vec::with_capacity(block);
    out_conf.clear();
    let mut prev = last_token;
    let mut head_in = vec![0.0f32; block * dim];
    let mut pre_norms = vec![vec![0.0f32; dim]; block];
    for i in 0..block {
        hc_head_fold(
            &states[i],
            hfn,
            hscale,
            hbase,
            cfg,
            pool,
            &mut head_in[i * dim..(i + 1) * dim],
        );
        pre_norms[i].copy_from_slice(&head_in[i * dim..(i + 1) * dim]);
        rms_weighted(&mut head_in[i * dim..(i + 1) * dim], hnorm, cfg.norm_eps);
    }
    let mut logits = vec![0.0f32; block * cfg.vocab];
    g.head.matmat(&head_in, block, &mut logits, pool);
    let mut mk_embed = vec![0.0f32; last.markov_w1.as_ref().map_or(0, |t| t.cols())];
    for i in 0..block {
        let logits_i = &mut logits[i * cfg.vocab..(i + 1) * cfg.vocab];
        // The markov head biases the logits from the PREVIOUS token — a
        // rank-256 bigram the draft samples through position by position,
        // while the network itself ran the whole block at once.
        // `CMF_DSPARK_NO_MARKOV=1` drops it: the bias is sequential through
        // the block (each position needs the previous PROPOSAL), which is
        // the one part of the draft a single device graph cannot batch — so
        // its acceptance value has to be known before it earns that
        // complexity.
        let no_markov = {
            static ON: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
            *ON.get_or_init(|| std::env::var("CMF_DSPARK_NO_MARKOV").is_ok_and(|v| v != "0"))
        };
        if no_markov {
            // Still feed the confidence head's embedding slot below.
            if let Some(w1) = last.markov_w1.as_ref() {
                w1.row_f32(prev as usize, &mut mk_embed);
            }
        } else if let (Some(w1), Some(w2)) = (last.markov_w1.as_ref(), last.markov_w2.as_ref()) {
            w1.row_f32(prev as usize, &mut mk_embed);
            let mut bias = vec![0.0f32; cfg.vocab];
            w2.matvec(&mk_embed, &mut bias, pool);
            for (a, b) in logits_i.iter_mut().zip(&bias) {
                *a += *b;
            }
        }
        let mut best = 0usize;
        for v in 1..logits_i.len() {
            if logits_i[v] > logits_i[best] {
                best = v;
            }
        }
        if let Some(cf) = last.confidence.as_ref() {
            let mut cat = pre_norms[i].clone();
            cat.extend_from_slice(&mk_embed);
            let mut s = [0.0f32; 1];
            if cat.len() == cf.cols() {
                cf.matvec(&cat, &mut s, pool);
            }
            out_conf.push(s[0]);
        }
        proposals.push(best as u32);
        prev = best as u32;
    }
    proposals
}