mlx-bun 0.0.13 → 0.3.0

package/src/spec/qwen-mtp-source.ts

@@ -0,0 +1,528 @@
+// Native Qwen3.8 MTP draft source — the Qwen-TRAINED multi-step prediction
+// head, split from the raw release's last shard and published as
+// mlx-community/Qwen3.8-27B-MTP-* (model_type "qwen3_5_mtp", block_size 3).
+//
+// Reference implementation (locate + port + gate, don't design):
+// mlx_vlm/speculative/drafters/qwen3_5_mtp/qwen3_5_mtp.py. Mechanism, per
+// step: h = fc(concat(rms_emb(embed(token)), rms_hidden(hidden))) → ONE
+// full-attention decoder layer (own KVCache, positions continue the target
+// sequence) → norm → the TARGET's lm_head → next draft token; recursive for
+// the block ("trained with multiple steps"). The drafter is NOT standalone:
+// it binds the target's embed_tokens + lm_head (mtp_use_dedicated_embeddings
+// false) and consumes the target's PRE-final-norm last-layer hidden (mlx-vlm
+// captures it with skip_final_norm=True) — carried here by the tapLayers
+// machinery, NOT the seam's post-norm anchorHidden.
+//
+// Row convention (predict-2-ahead): drafter KV row at position p is built
+// from (embed(token at position p+1), hidden at position p). Invariant
+// across rounds: after commit, the drafter offset equals the next pending
+// token's position minus zero — i.e. draft() always opens by building the
+// pending token's row from the TARGET's TRUE hidden at the emitted position
+// (held over from the verify tap). Computationally identical to mlx-vlm's
+// commit-time "seed" append, just performed at the top of the next round.
+//
+// Companion projections may be dense or quantized according to their metadata.
+// Published drafter norms are already sanitized to runtime layout; no +1.0 shift.
+
+import { MlxArray } from "../mlx/array";
+import * as ops from "../mlx/ops";
+import { materializeCopy } from "../mlx/materialize";
+import { toLogprobs } from "../sampler";
+import { loadModelConfig, type ModelConfig } from "../config";
+import { Weights } from "../weights";
+import { disposing, KVCache, QuantizedLinear, RMSNorm, type Mask } from "../model/gemma4-base";
+import { DenseLinear } from "../model/universal/modules";
+import type { DraftProvider, DraftSource, QwenMtpTarget } from "./source";
+import type { Cache } from "../model/gemma4";
+import { cloneKvCaches } from "../kv-store";
+import { cacheBytes } from "../prompt-cache";
+import { disposeResources } from "../engine/resources";
+import { SpeculativePrefixStore, type SpeculativePrefixState } from "./prefix-state";
+
+type Sampler = (logprobs: MlxArray, step: number) => MlxArray;
+
+const DRAFT_PREFILL_CHUNK = 2048;
+
+type MtpLinear = DenseLinear | QuantizedLinear;
+
+/** Projection tensors belong to Weights; dense transpose views belong to
+ *  the provider's resource stack. Quantized heads keep no dense copy. */
+function loadMtpLinear(
+  weights: Weights, path: string, config: ModelConfig, resources: DisposableStack,
+): MtpLinear {
+  if (weights.has(`${path}.scales`)) return QuantizedLinear.load(weights, path, config);
+  const layer = new DenseLinear(weights.tensor(`${path}.weight`), null);
+  resources.use(layer.wT);
+  return layer;
+}
+
+/** Qwen3Attention.forward with the companion's dense or quantized projections.
+ *  Attention operations follow src/model/qwen3_5.ts verbatim. */
+class MtpAttention {
+  readonly qProj: MtpLinear;
+  readonly kProj: MtpLinear;
+  readonly vProj: MtpLinear;
+  readonly oProj: MtpLinear;
+  readonly qNorm: RMSNorm;
+  readonly kNorm: RMSNorm;
+  readonly nHeads: number;
+  readonly nKvHeads: number;
+  readonly headDim: number;
+  readonly scale: number;
+  readonly ropeDims: number;
+  readonly ropeBase: number;
+
+  constructor(weights: Weights, config: ModelConfig, prefix: string, resources: DisposableStack) {
+    const t = config.text;
+    this.nHeads = t.numAttentionHeads;
+    this.nKvHeads = t.numKeyValueHeads;
+    this.headDim = t.headDim;
+    this.scale = Math.pow(this.headDim, -0.5);
+    this.ropeDims = Math.trunc(this.headDim * t.partialRotaryFactor);
+    this.ropeBase = t.ropeParameters.full_attention?.ropeTheta ?? 10000;
+    this.qProj = loadMtpLinear(weights, `${prefix}.q_proj`, config, resources);
+    this.kProj = loadMtpLinear(weights, `${prefix}.k_proj`, config, resources);
+    this.vProj = loadMtpLinear(weights, `${prefix}.v_proj`, config, resources);
+    this.oProj = loadMtpLinear(weights, `${prefix}.o_proj`, config, resources);
+    this.qNorm = new RMSNorm(weights.tensor(`${prefix}.q_norm.weight`), t.rmsNormEps);
+    this.kNorm = new RMSNorm(weights.tensor(`${prefix}.k_norm.weight`), t.rmsNormEps);
+  }
+
+  forward(x: MlxArray, mask: Mask, cache: KVCache): MlxArray {
+    const [B, L] = x.shape as [number, number, number];
+    const qp = this.qProj.forward(x);
+    const qpr = disposing(qp, ops.reshape(qp, [B, L, this.nHeads, this.headDim * 2]));
+    const [qHeads, gateHeads] = ops.split(qpr, [this.headDim], -1) as [MlxArray, MlxArray];
+    qpr.dispose();
+    const gate = disposing(gateHeads, ops.reshape(gateHeads, [B, L, this.nHeads * this.headDim]));
+
+    let k = this.kProj.forward(x);
+    let v = this.vProj.forward(x);
+
+    let q = this.qNorm.forward(qHeads);
+    qHeads.dispose();
+    q = disposing(q, ops.transposeAxes(q, [0, 2, 1, 3]));
+    k = disposing(k, ops.reshape(k, [B, L, this.nKvHeads, this.headDim]));
+    k = disposing(k, this.kNorm.forward(k));
+    k = disposing(k, ops.transposeAxes(k, [0, 2, 1, 3]));
+    v = disposing(v, ops.reshape(v, [B, L, this.nKvHeads, this.headDim]));
+    v = disposing(v, ops.transposeAxes(v, [0, 2, 1, 3]));
+
+    q = disposing(q, ops.rope(q, this.ropeDims, this.ropeBase, cache.offset, null));
+    k = disposing(k, ops.rope(k, this.ropeDims, this.ropeBase, cache.offset, null));
+
+    const [keys, values] = cache.updateAndFetch(k, v);
+    k.dispose();
+    v.dispose();
+    const attn = ops.sdpa(q, keys, values, this.scale, mask.mode, mask.arr);
+    keys.dispose();
+    values.dispose();
+    q.dispose();
+
+    const attnT = ops.transposeAxes(attn, [0, 2, 1, 3]);
+    attn.dispose();
+    const merged = ops.reshape(attnT, [B, L, -1]);
+    attnT.dispose();
+    const sig = ops.sigmoid(gate);
+    gate.dispose();
+    const gated = ops.mul(merged, sig);
+    merged.dispose();
+    sig.dispose();
+    const out = this.oProj.forward(gated);
+    gated.dispose();
+    return out;
+  }
+}
+
+/** The one MTP decoder block: fc-merge → attention → swiglu MLP → norm. */
+class MtpModule {
+  readonly fc: MtpLinear;
+  readonly preFcNormEmbedding: RMSNorm;
+  readonly preFcNormHidden: RMSNorm;
+  readonly attn: MtpAttention;
+  readonly mlpGate: MtpLinear;
+  readonly mlpUp: MtpLinear;
+  readonly mlpDown: MtpLinear;
+  readonly inputNorm: RMSNorm;
+  readonly postAttnNorm: RMSNorm;
+  readonly finalNorm: RMSNorm;
+
+  constructor(weights: Weights, config: ModelConfig, resources: DisposableStack) {
+    const eps = config.text.rmsNormEps;
+    this.fc = loadMtpLinear(weights, "fc", config, resources);
+    this.preFcNormEmbedding = new RMSNorm(weights.tensor("pre_fc_norm_embedding.weight"), eps);
+    this.preFcNormHidden = new RMSNorm(weights.tensor("pre_fc_norm_hidden.weight"), eps);
+    this.attn = new MtpAttention(weights, config, "layers.0.self_attn", resources);
+    this.mlpGate = loadMtpLinear(weights, "layers.0.mlp.gate_proj", config, resources);
+    this.mlpUp = loadMtpLinear(weights, "layers.0.mlp.up_proj", config, resources);
+    this.mlpDown = loadMtpLinear(weights, "layers.0.mlp.down_proj", config, resources);
+    this.inputNorm = new RMSNorm(weights.tensor("layers.0.input_layernorm.weight"), eps);
+    this.postAttnNorm = new RMSNorm(weights.tensor("layers.0.post_attention_layernorm.weight"), eps);
+    this.finalNorm = new RMSNorm(weights.tensor("norm.weight"), eps);
+  }
+
+  /** One block forward over [1,S,·]: token embeddings ([1,S,H], target
+   *  embed_tokens output) paired with hiddens ([1,S,H], target pre-final-norm
+   *  or the module's own chained output). Appends S rows to `cache`; returns
+   *  the module output [1,S,H] (post final norm — what the target lm_head
+   *  consumes AND what chains into the next step's `hidden`). */
+  forward(tokenEmbeds: MlxArray, hiddens: MlxArray, cache: KVCache): MlxArray {
+    const embNorm = this.preFcNormEmbedding.forward(tokenEmbeds);
+    const hidNorm = this.preFcNormHidden.forward(hiddens);
+    const joined = ops.concatAxis([embNorm, hidNorm], 2);
+    embNorm.dispose();
+    hidNorm.dispose();
+    const x = this.fc.forward(joined);
+    joined.dispose();
+
+    // Decoder layer (Qwen3Layer.forward shape).
+    const L = x.shape[1]!;
+    const mask = cache.makeMask(L, null);
+    const xn = this.inputNorm.forward(x);
+    const r = this.attn.forward(xn, mask, cache);
+    xn.dispose();
+    mask.arr?.dispose();
+    const h = ops.add(x, r);
+    x.dispose();
+    r.dispose();
+    const hn = this.postAttnNorm.forward(h);
+    const g = this.mlpGate.forward(hn);
+    const u = this.mlpUp.forward(hn);
+    hn.dispose();
+    const silu = ops.silu(g);
+    g.dispose();
+    const act = ops.mul(silu, u);
+    silu.dispose();
+    u.dispose();
+    const m = this.mlpDown.forward(act);
+    act.dispose();
+    const out = ops.add(h, m);
+    h.dispose();
+    m.dispose();
+    return disposing(out, this.finalNorm.forward(out));
+  }
+}
+
+class QwenMtpPrefixState implements SpeculativePrefixState {
+  constructor(readonly tokens: number[], readonly targetIdentity: object,
+    readonly namespace: string, readonly bytes: number, readonly target: Cache[],
+    public draft: KVCache | null, public hidden: MlxArray | null) {}
+
+  dispose(): void {
+    const resources = [...this.target.splice(0), this.draft, this.hidden];
+    this.draft = null;
+    this.hidden = null;
+    disposeResources(resources.filter((r) => r != null));
+  }
+}
+
+export class QwenMtpProvider implements DraftProvider {
+  readonly id: string;
+  readonly weightsBytes: number;
+  readonly #module: MtpModule;
+  readonly #config: ModelConfig;
+  readonly #resources: DisposableStack;
+  readonly #prefixStore = new SpeculativePrefixStore<QwenMtpPrefixState>();
+
+  private constructor(id: string, config: ModelConfig, weightsBytes: number, module: MtpModule, resources: DisposableStack) {
+    this.id = id;
+    this.#config = config;
+    this.#module = module;
+    this.weightsBytes = weightsBytes;
+    this.#resources = resources;
+  }
+
+  static async load(dir: string): Promise<QwenMtpProvider> {
+    const config = await loadModelConfig(dir);
+    if (config.modelType !== "qwen3_5_mtp")
+      throw new Error(`${dir}: not a qwen3_5_mtp drafter (model_type ${config.modelType})`);
+    const weights = await Weights.open(dir);
+    using resources = new DisposableStack();
+    resources.defer(() => weights.dispose());
+    const weightsBytes = [...weights.shards.files.values()]
+      .reduce((a, f) => a + f.mmap.size, 0);
+    const module = new MtpModule(weights, config, resources);
+    return new QwenMtpProvider(
+      dir.split("/").filter(Boolean).at(-1) ?? "qwen-mtp",
+      config, weightsBytes, module, resources.move(),
+    );
+  }
+
+  open(opts: Parameters<DraftProvider["open"]>[0]): DraftSource {
+    if (this.#resources.disposed)
+      throw new Error("qwen MTP provider is disposed");
+    const target = opts.target.qwenMtp;
+    if (!target)
+      throw new Error("qwen MTP drafting requires a qwen3_5-family target");
+    if (target.hiddenSize !== this.#config.text.hiddenSize) {
+      throw new Error(
+        `qwen MTP drafter hidden ${this.#config.text.hiddenSize} != target ` +
+        `${target.hiddenSize} — split from a different checkpoint?`,
+      );
+    }
+    return new QwenMtpSource(target, this.#module, opts.sampler,
+      { store: this.#prefixStore, identity: opts.target.identity });
+  }
+
+  dispose(): void {
+    // Release cached transpose views before their native weight maps. MLX
+    // retains buffers needed by outstanding GPU commands until completion.
+    disposeResources([this.#prefixStore, { dispose: () => this.#resources.dispose() }]);
+  }
+}
+
+export class QwenMtpSource implements DraftSource {
+  // Full-prompt target prefill (the bonus token exists before round 1, and
+  // the tap covers every prompt position — mlx-vlm's flow).
+  readonly prefillMode = "full" as const;
+  readonly weightsBytes = 0; // provider owns the drafter weights
+  /** Pre-final-norm tap: the LAST layer's output stream (index nLayers-1).
+   *  The seam's anchorHidden is post-final-norm and is deliberately unused. */
+  readonly tapLayers: number[];
+
+  readonly #target: QwenMtpTarget;
+  readonly #module: MtpModule;
+  readonly #sampler: Sampler;
+  #cache = new KVCache();
+  readonly prefix: DraftSource["prefix"];
+  #prefilledTokens = 0;
+  #hasDrafted = false;
+  /** Target pre-norm hidden at the position preceding the next pending
+   *  token: prefill's last tapped row, then each commit's vCtx row at the
+   *  emitted position. draft() consumes it to build the pending row. [1,1,H] */
+  #pendingTrueHidden: MlxArray | null = null;
+  #roundAppended = 0;
+  #closed = false;
+
+  constructor(target: QwenMtpTarget, module: MtpModule, sampler: Sampler,
+    prefix?: { store: SpeculativePrefixStore<QwenMtpPrefixState>; identity: object }) {
+    this.#target = target;
+    this.#module = module;
+    this.#sampler = sampler;
+    this.tapLayers = [target.layerCount - 1];
+    if (prefix) this.prefix = {
+      restore: (prompt, caches, namespace, maxBytes) => {
+        this.#checkOpen();
+        if (this.#prefilledTokens !== 0 || this.#hasDrafted)
+          throw new Error("Qwen MTP prefix restore requires a fresh source");
+        const entry = prefix.store.take(prompt, prefix.identity, namespace, maxBytes);
+        if (!entry) return 0;
+        try {
+          if (entry.target.length !== caches.length || !entry.draft || !entry.hidden ||
+            entry.draft.offset !== entry.tokens.length - 1)
+            throw new Error("invalid paired Qwen MTP prefix state");
+          disposeResources(caches);
+          caches.splice(0, caches.length, ...entry.target.splice(0));
+          this.#cache.dispose();
+          this.#cache = entry.draft;
+          entry.draft = null;
+          this.#pendingTrueHidden?.dispose();
+          this.#pendingTrueHidden = entry.hidden;
+          entry.hidden = null;
+          return this.#prefilledTokens = entry.tokens.length;
+        } finally { entry.dispose(); }
+      },
+      capture: (tokens, caches, namespace, maxBytes) => {
+        this.#checkOpen();
+        if (tokens.length !== this.#prefilledTokens || !this.#pendingTrueHidden ||
+          this.#cache.offset !== tokens.length - 1 || this.#roundAppended !== 0)
+          throw new Error("Qwen MTP snapshot requires an aligned prefill boundary");
+        const bytes = cacheBytes([...caches, this.#cache]) + this.#pendingTrueHidden.nbytes;
+        if (bytes > maxBytes) { prefix.store.dispose(); return; }
+        const retained: Cache[] = [];
+        let draft: KVCache | null = null, hidden: MlxArray | null = null;
+        try {
+          retained.push(...cloneKvCaches(caches));
+          draft = cloneKvCaches([this.#cache])[0] as KVCache;
+          hidden = this.#pendingTrueHidden.slice([0, 0, 0], this.#pendingTrueHidden.shape);
+          const entry = new QwenMtpPrefixState([...tokens], prefix.identity, namespace,
+            bytes, retained.splice(0), draft, hidden);
+          draft = null;
+          hidden = null;
+          prefix.store.put(entry, maxBytes);
+        } finally { disposeResources([...retained, draft, hidden].filter((r) => r != null)); }
+      },
+    };
+  }
+
+  /** Drafter prefill: rows for positions 0..L-2, keyed (token_{p+1}, h_p) —
+   *  the (pending, h_{L-1}) row is NOT built here (the pending token is
+   *  sampled after target prefill); its true hidden is kept for round 1. */
+  async prefill(promptIds: number[], ctxML?: MlxArray): Promise<void> {
+    this.#checkOpen();
+    if (!ctxML)
+      throw new Error("qwen MTP prefill requires the tapped pre-final-norm context");
+    try {
+      if (this.#hasDrafted)
+        throw new Error("qwen MTP prefill cannot follow drafting");
+      const L = promptIds.length;
+      const start = this.#prefilledTokens;
+      const H = ctxML.shape[2]!;
+      if (L <= start || ctxML.shape[1]! !== L - start)
+        throw new Error(`qwen MTP tap covered ${ctxML.shape[1]} of ${L - start} new prompt positions`);
+      if (start > 0) {
+        if (!this.#pendingTrueHidden || this.#cache.offset !== start - 1)
+          throw new Error("qwen MTP prefix state is not aligned");
+        this.#stepOne(promptIds[start]!, this.#pendingTrueHidden).dispose();
+      }
+      this.#pendingTrueHidden?.dispose();
+      // A slice would retain the complete prompt buffer in the saved prefix.
+      // Materialize this one row before handing off the prefill state.
+      using tail = ctxML.slice([0, L - start - 1, 0], [1, L - start, H]);
+      this.#pendingTrueHidden = materializeCopy(tail);
+      for (let pos = start; pos + 1 < L; pos += DRAFT_PREFILL_CHUNK) {
+        const n = Math.min(DRAFT_PREFILL_CHUNK, L - 1 - pos);
+        const shifted = promptIds.slice(pos + 1, pos + 1 + n);
+        const ids = ops.fromInt32(shifted, [1, n]);
+        const embeds = this.#target.embed(ids);
+        ids.dispose();
+        const hiddens = ctxML.slice([0, pos - start, 0], [1, pos + n - start, H]);
+        const out = this.#module.forward(embeds, hiddens, this.#cache);
+        embeds.dispose();
+        hiddens.dispose();
+        out.dispose(); // prefill outputs are not seeds; only KV matters here
+        // Bound the lazy KV graph at each chunk. Evaluating the unused output
+        // would also execute attention and the MLP, which prefill does not need.
+        ops.evalAll(this.#cache.state());
+      }
+      if (this.#cache.offset !== L - 1)
+        throw new Error(`qwen MTP prefill offset ${this.#cache.offset}, expected ${L - 1}`);
+      // Also covers a restored-prefix bridge with no subsequent full chunk.
+      ops.evalAll([this.#pendingTrueHidden, ...this.#cache.state()]);
+      this.#prefilledTokens = L;
+    } finally {
+      ctxML?.dispose(); // ownership per the seam contract
+    }
+  }
+
+  async draft(feed: number[], n: number, stepBase: number): Promise<number[]> {
+    this.#checkOpen();
+    this.#hasDrafted = true;
+    if (this.#roundAppended !== 0)
+      throw new Error("qwen MTP draft called before the prior round committed");
+    if (n <= 0) return [];
+    const pending = feed.at(-1);
+    if (!Number.isSafeInteger(pending) || pending! < 0)
+      throw new Error("qwen MTP requires a non-empty token feed");
+
+    const drafts: number[] = [];
+    let chained: MlxArray | null = null;
+    const startOffset = this.#cache.offset;
+    try {
+      // Build the pending token's row from the TRUE target hidden at the
+      // preceding position (prefill tail on round 1, verify tap afterwards).
+      if (!this.#pendingTrueHidden)
+        throw new Error("qwen MTP draft before prefill/commit");
+      const out = this.#stepOne(pending!, this.#pendingTrueHidden);
+      this.#pendingTrueHidden.dispose();
+      this.#pendingTrueHidden = null;
+      this.#roundAppended++;
+      chained = out;
+      drafts.push(this.#sample(out, stepBase));
+      while (drafts.length < n) {
+        const out = this.#stepOne(drafts.at(-1)!, chained!);
+        chained!.dispose();
+        chained = out;
+        this.#roundAppended++;
+        drafts.push(this.#sample(out, stepBase + drafts.length));
+      }
+      return drafts;
+    } catch (error) {
+      const appended = this.#cache.offset - startOffset;
+      if (appended > 0) this.#cache.trim(appended);
+      this.#roundAppended = 0;
+      throw error;
+    } finally {
+      chained?.dispose();
+    }
+  }
+
+  /** mlx-vlm accept semantics: keep drafted rows for accepted positions,
+   *  trim the rejected tail, then append the missing accepted row (all-accept
+   *  case) and the correction/bonus row using the TARGET's verified pre-norm
+   *  hiddens — harvesting the last output as the next round's seed. */
+  async commit(
+    d: number,
+    kAccept: number,
+    vCtxML?: MlxArray,
+    _verifiedHidden?: MlxArray,
+    acceptedTokens: readonly number[] = [],
+  ): Promise<void> {
+    this.#checkOpen();
+    try {
+      if (!vCtxML)
+        throw new Error("qwen MTP commit requires the tapped verify context");
+      if (kAccept > 0 && acceptedTokens.length !== kAccept)
+        throw new Error(
+          `qwen MTP received ${acceptedTokens.length} accepted tokens for k=${kAccept}`,
+        );
+      // Rows appended this round cover window tokens 0..d-1 (pending +
+      // drafts[0..d-2]); rows keyed by ACCEPTED window tokens (0..kAccept)
+      // survive — trim the rejected tail.
+      const trim = Math.max(this.#roundAppended - 1 - kAccept, 0);
+      if (trim > 0) this.#cache.trim(trim);
+
+      // All-accept case: window token d (= drafts[d-1]) has no drafted row —
+      // append it from the TRUE verify hidden (vCtx row d-1, the hidden at
+      // its preceding position). The correction/bonus row is NOT built here:
+      // the serve loop re-feeds that token in the next draft()'s `feed`, and
+      // draft() opens by building its row from #pendingTrueHidden below.
+      const H = vCtxML.shape[2]!;
+      if (kAccept === d && d > 0) {
+        const ids = ops.fromInt32([acceptedTokens[kAccept - 1]!], [1, 1]);
+        const embeds = this.#target.embed(ids);
+        ids.dispose();
+        const hidden = vCtxML.slice([0, d - 1, 0], [1, d, H]);
+        const out = this.#module.forward(embeds, hidden, this.#cache);
+        embeds.dispose();
+        hidden.dispose();
+        out.dispose();
+      }
+      // Next round's pending row input: the target's verified pre-norm
+      // hidden at the emitted position.
+      this.#pendingTrueHidden?.dispose();
+      this.#pendingTrueHidden = vCtxML.slice([0, kAccept, 0], [1, kAccept + 1, H]);
+    } finally {
+      vCtxML?.dispose(); // seam: commit takes ownership
+      this.#roundAppended = 0;
+    }
+  }
+
+  dispose(): void {
+    if (this.#closed) return;
+    this.#closed = true;
+    this.#cache.dispose();
+    this.#pendingTrueHidden?.dispose();
+  }
+
+  /** One module forward for (token, hidden) — appends one KV row. */
+  #stepOne(token: number, hidden: MlxArray): MlxArray {
+    using ids = ops.fromInt32([token], [1, 1]);
+    using embed = this.#target.embed(ids);
+    ids.dispose();
+    return this.#module.forward(embed, hidden, this.#cache);
+  }
+
+  /** Sample a draft token from the module output via the TARGET's lm head
+   *  and the request sampler (per-step RNG stream discipline). */
+  #sample(moduleOut: MlxArray, step: number): number {
+    using logits = this.#target.logitsFromHidden(moduleOut);
+    // Sampler contract is [1, V] (the main decode loop's shape). moduleOut
+    // is [1, 1, H] → logits [1, 1, V]; without this reshape any sampler
+    // that slices 2-D (top-k) throws "[slice] Invalid number of indices…
+    // dimension 3" — the serve-lane MTP 500 (chat defaults carry the
+    // model's top_k=20; the greedy bench harness never hit it).
+    const V = logits.shape[logits.shape.length - 1]!;
+    using flat = ops.reshape(logits, [1, V]);
+    logits.dispose();
+    using logprobs = toLogprobs(flat);
+    flat.dispose();
+    using tok = this.#sampler(logprobs, step);
+    logprobs.dispose();
+    return ops.itemUint32(tok);
+  }
+
+  #checkOpen(): void {
+    if (this.#closed) throw new Error("qwen MTP source is closed");
+  }
+}