@mlx-bun/inference 0.0.0-stage → 1.0.0

package/src/models/vision/qwen3vl.ts

@@ -0,0 +1,566 @@
+// Qwen3-VL vision tower (Qwen3.8's `vision_tower`) — faithful port
+// of mlx-vlm qwen3_vl/vision.py at the qwen3_5 configuration: depth 27,
+// hidden 1152, heads 16 (head_dim 72), patch 16, temporal 2, merge 2,
+// intermediate 4304 with gelu_pytorch_tanh, learned 48×48 pos-embed grid
+// bilinearly interpolated, 2D half/half rotary (dim 36) over full head_dim,
+// pre-LN blocks with biased LayerNorms, merger LayerNorm→fc1→GELU(erf)→fc2
+// to out_hidden 5120. deepstack is EMPTY for qwen3_5 and not ported.
+//
+// Weights: the artifact ships the tower as a bf16 sidecar
+// (optiq/optiq_vision.safetensors, 333 tensors, MLX conv layout). The
+// Conv3d patch embed has kernel == stride, so it is computed as one matmul
+// against the weight flattened to the preprocessor's (c, t, py, px) row
+// order — same values, GEMM accumulation (gate: tower output vs the mlx-vlm
+// capture; e2e tokens are the decisive gate, the SigLIP precedent).
+//
+// Input rows arrive MERGE-BLOCK-MAJOR from qwen3vl-preprocess.ts; rope
+// h/w positions and the pos-embed permutation are built in the same order.
+
+import { ptr, read } from "bun:ffi";
+import { existsSync, readFileSync } from "node:fs";
+import { tensorFingerprint } from "../../artifacts/fingerprint";
+import { MlxArray, cpuStream } from "@mlx-bun/mlx/array";
+import { C, Dtype } from "@mlx-bun/mlx/ffi";
+import * as ops from "@mlx-bun/mlx/ops";
+import type { Qwen3VLPreprocessed } from "../../input/vision/qwen3vl-preprocess";
+import {
+  QWEN3VL_MERGE_SIZE,
+  QWEN3VL_PATCH_SIZE,
+  QWEN3VL_TEMPORAL_PATCH_SIZE,
+} from "../../input/vision/qwen3vl-preprocess";
+
+const cstr = (s: string) => Buffer.from(s + "\0", "utf8");
+
+const DEPTH = 27;
+const HIDDEN = 1152;
+const HEADS = 16;
+const HEAD_DIM = HIDDEN / HEADS; // 72
+const ROT_DIM = HEAD_DIM / 2; // 36 (VisionRotaryEmbedding dim)
+const INTERMEDIATE = 4304;
+const OUT_HIDDEN = 5120;
+const POS_GRID = 48; // sqrt(num_position_embeddings 2304)
+const ROPE_THETA = 10000;
+
+function disposing(old: MlxArray, next: MlxArray): MlxArray {
+  old.dispose();
+  return next;
+}
+
+interface Block {
+  norm1W: MlxArray; norm1B: MlxArray;
+  norm2W: MlxArray; norm2B: MlxArray;
+  qkvW: MlxArray; qkvB: MlxArray;
+  projW: MlxArray; projB: MlxArray;
+  fc1W: MlxArray; fc1B: MlxArray;
+  fc2W: MlxArray; fc2B: MlxArray;
+}
+
+export class Qwen3VLVisionTower {
+  #identity: string | undefined;
+  readonly #identityWeights: Array<[string, MlxArray]>;
+
+  /** Computed once, under native preparation. Exact weight content prevents
+   * cross-artifact SSD reuse even when model aliases and shapes are identical. */
+  get cacheIdentity(): string {
+    if (this.#identity) return this.#identity;
+    return this.#identity = tensorFingerprint(this.#identityWeights);
+  }
+  #patchW: MlxArray; // [1536, 1152] — flattened conv, already transposed
+  #patchB: MlxArray;
+  #posEmbed: MlxArray; // [2304, 1152]
+  #blocks: Block[];
+  #mergerNormW: MlxArray; #mergerNormB: MlxArray;
+  #mergerFc1W: MlxArray; #mergerFc1B: MlxArray;
+  #mergerFc2W: MlxArray; #mergerFc2B: MlxArray;
+
+  private constructor(w: Map<string, MlxArray>) {
+    this.#identityWeights = [...w];
+    const t = (n: string): MlxArray => {
+      const a = w.get(`vision_tower.${n}`);
+      if (!a) throw new Error(`vision sidecar missing tensor vision_tower.${n}`);
+      return a;
+    };
+    // Conv3d weight [1152, 2, 16, 16, 3] (out, kT, kH, kW, in) — used with
+    // the REAL conv3d op (kernel == stride): a value-equivalent GEMM rounds
+    // differently in bf16 and the ulp seeds diverge through 27 blocks.
+    this.#patchW = t("patch_embed.proj.weight");
+    this.#patchB = t("patch_embed.proj.bias");
+    this.#posEmbed = t("pos_embed.weight");
+    this.#blocks = [];
+    for (let i = 0; i < DEPTH; i++) {
+      const p = `blocks.${i}`;
+      this.#blocks.push({
+        norm1W: t(`${p}.norm1.weight`), norm1B: t(`${p}.norm1.bias`),
+        norm2W: t(`${p}.norm2.weight`), norm2B: t(`${p}.norm2.bias`),
+        qkvW: t(`${p}.attn.qkv.weight`), qkvB: t(`${p}.attn.qkv.bias`),
+        projW: t(`${p}.attn.proj.weight`), projB: t(`${p}.attn.proj.bias`),
+        fc1W: t(`${p}.mlp.linear_fc1.weight`), fc1B: t(`${p}.mlp.linear_fc1.bias`),
+        fc2W: t(`${p}.mlp.linear_fc2.weight`), fc2B: t(`${p}.mlp.linear_fc2.bias`),
+      });
+    }
+    this.#mergerNormW = t("merger.norm.weight");
+    this.#mergerNormB = t("merger.norm.bias");
+    this.#mergerFc1W = t("merger.linear_fc1.weight");
+    this.#mergerFc1B = t("merger.linear_fc1.bias");
+    this.#mergerFc2W = t("merger.linear_fc2.weight");
+    this.#mergerFc2B = t("merger.linear_fc2.bias");
+  }
+
+  /** The 333 tensor names the tower models, by constructed name. */
+  static #tensorNames(): string[] {
+    const names: string[] = [
+      "vision_tower.patch_embed.proj.weight", "vision_tower.patch_embed.proj.bias",
+      "vision_tower.pos_embed.weight",
+      "vision_tower.merger.norm.weight", "vision_tower.merger.norm.bias",
+      "vision_tower.merger.linear_fc1.weight", "vision_tower.merger.linear_fc1.bias",
+      "vision_tower.merger.linear_fc2.weight", "vision_tower.merger.linear_fc2.bias",
+    ];
+    for (let i = 0; i < DEPTH; i++) {
+      for (const nm of [
+        "norm1.weight", "norm1.bias", "norm2.weight", "norm2.bias",
+        "attn.qkv.weight", "attn.qkv.bias", "attn.proj.weight", "attn.proj.bias",
+        "mlp.linear_fc1.weight", "mlp.linear_fc1.bias",
+        "mlp.linear_fc2.weight", "mlp.linear_fc2.bias",
+      ]) names.push(`vision_tower.blocks.${i}.${nm}`);
+    }
+    return names;
+  }
+
+  /** Checkpoint-generation naming for the tower. Pre-5.8 artifacts store
+   *  `vision_tower.*`; the transformers-5.8 export stores `model.visual.*`
+   *  (mlx-lm's text-only qwen3_5 drops BOTH prefixes in `Model.sanitize`, so
+   *  the mapping between them is ours). Verified 2026-09-01 across
+   *  Qwen3.8-27B-tqalloc-norot vs the staged pre-5.8 artifact: 331/333
+   *  tensors byte-identical under the rename, patch_embed.proj.weight
+   *  identical under the layout permutation below, and merger.linear_fc2
+   *  differs only because the staged artifact is rotation-folded. */
+  static #sourceName(canonical: string, prefix: string): string {
+    return prefix + canonical.slice("vision_tower.".length);
+  }
+
+  /** HF stores the Conv3d patch embed as (out, in, kT, kH, kW); MLX (and this
+   *  tower) want (out, kT, kH, kW, in). Proven a pure permutation on the
+   *  artifacts above — transpose(0,2,3,4,1) of the 5.8 tensor is bit-equal to
+   *  the pre-5.8 one. */
+  static #fixPatchEmbed(w: Map<string, MlxArray>): void {
+    const key = "vision_tower.patch_embed.proj.weight";
+    const a = w.get(key);
+    if (!a || a.shape.length !== 5) return;
+    if (a.shape[4] === 3) return; // already MLX layout
+    const t = ops.transposeAxes(a, [0, 2, 3, 4, 1]);
+    const c = ops.contiguous(t);
+    t.dispose();
+    a.dispose();
+    w.set(key, c);
+  }
+
+  /** Load one safetensors file's map and move `wanted` tensors into `out`
+   *  under their CANONICAL (`vision_tower.*`) names; `source` maps each
+   *  canonical name to the stored one. Missing names are skipped; the map is
+   *  always freed. */
+  static #takeFromFile(
+    path: string, wanted: string[], out: Map<string, MlxArray>,
+    source: (canonical: string) => string = (n) => n,
+  ): void {
+    const arrMap = new BigUint64Array([C.mlx_map_string_to_array_new()]);
+    const metaMap = new BigUint64Array([C.mlx_map_string_to_string_new()]);
+    const arrMapPtr = ptr(arrMap);
+    const metaMapPtr = ptr(metaMap);
+    const status = C.mlx_load_safetensors(arrMapPtr, metaMapPtr, ptr(cstr(path)), cpuStream);
+    C.mlx_map_string_to_string_free(read.u64(metaMapPtr, 0));
+    const handle = read.u64(arrMapPtr, 0);
+    if (status !== 0) {
+      // The out-param map was allocated regardless — free it on the error
+      // path too (2026-08-18 review; one-shot leak, but a leak).
+      C.mlx_map_string_to_array_free(handle);
+      throw new Error(`failed to load ${path}`);
+    }
+    try {
+      for (const name of wanted) {
+        if (out.has(name)) continue;
+        const slot = new BigUint64Array([C.mlx_array_new()]);
+        const slotPtr = ptr(slot);
+        if (C.mlx_map_string_to_array_get(slotPtr, handle, ptr(cstr(source(name)))) !== 0) {
+          C.mlx_array_free(read.u64(slotPtr, 0));
+          continue;
+        }
+        out.set(name, new MlxArray(read.u64(slotPtr, 0)));
+      }
+    } finally {
+      C.mlx_map_string_to_array_free(handle);
+    }
+  }
+
+  static load(modelDir: string): Qwen3VLVisionTower {
+    const names = Qwen3VLVisionTower.#tensorNames();
+    const weights = new Map<string, MlxArray>();
+    try {
+      const sidecar = `${modelDir}/optiq/optiq_vision.safetensors`;
+      if (existsSync(sidecar)) {
+        // OptiQ-convention sidecar (their artifacts + ours pre-2026-08-18).
+        Qwen3VLVisionTower.#takeFromFile(sidecar, names, weights);
+      } else {
+        // In-main fallback: the bf16 tower tensors live in the main
+        // safetensors shards (the mlx-vlm convention — one copy serves both
+        // ecosystems; sidecar duplication retired 2026-08-18). The 5.8-family
+        // export names them `model.visual.*` instead of `vision_tower.*`.
+        const idxPath = `${modelDir}/model.safetensors.index.json`;
+        let files: string[];
+        let prefixes = ["vision_tower.", "model.visual."];
+        if (existsSync(idxPath)) {
+          const idx = JSON.parse(readFileSync(idxPath, "utf8")) as
+            { weight_map: Record<string, string> };
+          prefixes = prefixes.filter((p) =>
+            Qwen3VLVisionTower.#sourceName(names[0]!, p) in idx.weight_map);
+          files = [...new Set(prefixes.flatMap((p) => names
+            .map((n) => idx.weight_map[Qwen3VLVisionTower.#sourceName(n, p)])
+            .filter((f): f is string => !!f)))];
+          if (files.length === 0)
+            throw new Error(`no vision tower tensors in ${modelDir} (index has none)`);
+        } else {
+          files = ["model.safetensors"];
+        }
+        for (const f of files)
+          for (const p of prefixes)
+            Qwen3VLVisionTower.#takeFromFile(
+              `${modelDir}/${f}`, names, weights,
+              (n) => Qwen3VLVisionTower.#sourceName(n, p),
+            );
+      }
+      for (const name of names)
+        if (!weights.has(name))
+          throw new Error(`qwen vision weights missing tensor ${name} (sidecar or in-main)`);
+      Qwen3VLVisionTower.#fixPatchEmbed(weights);
+    } catch (e) {
+      for (const [, a] of weights) a.dispose();
+      throw e;
+    }
+    return new Qwen3VLVisionTower(weights);
+  }
+
+  /** One vision block: pre-LN attention (+residual), pre-LN MLP (+residual).
+   *  Consumes `h`. */
+  #block(
+    h: MlxArray, blk: Block, cos: MlxArray, sin: MlxArray,
+    scale: number, n: number, segLen: number = n,
+  ): MlxArray {
+    const xn = ops.layerNorm(h, blk.norm1W, blk.norm1B, 1e-6);
+    const qkv = this.#linear(xn, blk.qkvW, blk.qkvB); // [n, 3456]
+    xn.dispose();
+    const qkv4 = disposing(qkv, ops.reshape(qkv, [n, 3, HEADS, HEAD_DIM]));
+    const [qs, ks, vs] = ops.split(qkv4, [1, 2], 1) as [MlxArray, MlxArray, MlxArray];
+    qkv4.dispose();
+    let q = disposing(qs, ops.reshape(qs, [n, HEADS, HEAD_DIM]));
+    let k = disposing(ks, ops.reshape(ks, [n, HEADS, HEAD_DIM]));
+    const v = disposing(vs, ops.reshape(vs, [n, HEADS, HEAD_DIM]));
+    q = disposing(q, this.#applyRope(q, cos, sin));
+    k = disposing(k, this.#applyRope(k, cos, sin));
+    // [1, HEADS, n, HEAD_DIM] full (non-causal) attention — one image =
+    // one cu_seqlens segment, so no mask. The reference's ensure_fused_sdpa
+    // ZERO-PADS head_dim 72 → 80 (nearest fused-kernel width), runs fused
+    // sdpa there, and slices back; the padded kernel's accumulation is the
+    // parity-relevant arithmetic, so replicate it exactly.
+    const PAD_DIM = 80;
+    const padTo = (x: MlxArray): MlxArray => {
+      const z = ops.zeros([n, HEADS, PAD_DIM - HEAD_DIM], Dtype.bfloat16);
+      const p = ops.concatAxis([x, z], -1);
+      z.dispose();
+      return disposing(x, p);
+    };
+    const lift = (x: MlxArray): MlxArray => {
+      const p = padTo(x); // consumes x
+      const tr = ops.transposeAxes(p, [1, 0, 2]);
+      p.dispose();
+      return disposing(tr, ops.expandDims(tr, 0));
+    };
+    const qT = lift(q);
+    const kT = lift(k);
+    const vT = lift(v);
+    // cu_seqlens split: one sdpa per temporal segment (single segment for
+    // images), outputs concatenated in order — the reference's exact walk.
+    let attnP: MlxArray;
+    if (segLen === n) {
+      attnP = ops.sdpa(qT, kT, vT, scale, "", null);
+    } else {
+      const outs: MlxArray[] = [];
+      for (let s = 0; s < n; s += segLen) {
+        const qs = qT.slice([0, 0, s, 0], [1, HEADS, s + segLen, PAD_DIM]);
+        const ks = kT.slice([0, 0, s, 0], [1, HEADS, s + segLen, PAD_DIM]);
+        const vsS = vT.slice([0, 0, s, 0], [1, HEADS, s + segLen, PAD_DIM]);
+        outs.push(ops.sdpa(qs, ks, vsS, scale, "", null));
+        qs.dispose(); ks.dispose(); vsS.dispose();
+      }
+      attnP = ops.concatAxis(outs, 2);
+      for (const o of outs) o.dispose();
+    }
+    qT.dispose(); kT.dispose(); vT.dispose();
+    const attn = disposing(
+      attnP, attnP.slice([0, 0, 0, 0], [1, HEADS, n, HEAD_DIM]),
+    );
+    const attnT = disposing(attn, ops.transposeAxes(attn, [0, 2, 1, 3]));
+    const merged = disposing(attnT, ops.reshape(attnT, [n, HIDDEN]));
+    const proj = disposing(merged, this.#linear(merged, blk.projW, blk.projB));
+    h = disposing(h, disposing(proj, ops.add(h, proj)));
+    const mn = ops.layerNorm(h, blk.norm2W, blk.norm2B, 1e-6);
+    const f1 = disposing(mn, this.#linear(mn, blk.fc1W, blk.fc1B));
+    const act = disposing(f1, ops.geluApprox(f1)); // gelu_pytorch_tanh
+    const f2 = disposing(act, this.#linear(act, blk.fc2W, blk.fc2B));
+    return disposing(h, disposing(f2, ops.add(h, f2)));
+  }
+
+  /** nn.Linear on bf16 — mx.addmm(bias, x, W^T), the reference's exact bias
+   *  path (a separate matmul+add rounds differently and seeded the ±1-ulp
+   *  block divergence found in the stage bisect). */
+  #linear(x: MlxArray, w: MlxArray, b: MlxArray): MlxArray {
+    const wT = ops.transposeAxes(w, [1, 0]);
+    const y = ops.addmm(b, x, wT);
+    wT.dispose();
+    return y;
+  }
+
+  /** Bilinearly interpolated learned pos embeds, in merge-block-major row
+   *  order — port of fast_pos_embed_interpolate + the merge permutation. */
+  #interpolatedPosEmbeds(gridH: number, gridW: number): MlxArray {
+    const n = gridH * gridW;
+    const idx = [new Int32Array(n), new Int32Array(n), new Int32Array(n), new Int32Array(n)];
+    const wts = [new Float32Array(n), new Float32Array(n), new Float32Array(n), new Float32Array(n)];
+    const lin = (count: number): Float32Array => {
+      // mx.linspace(0, POS_GRID-1, count) in f32.
+      const outArr = new Float32Array(count);
+      if (count === 1) { outArr[0] = 0; return outArr; }
+      const step = Math.fround((POS_GRID - 1) / (count - 1));
+      for (let i = 0; i < count; i++) outArr[i] = Math.fround(i * step);
+      return outArr;
+    };
+    const hIdx = lin(gridH);
+    const wIdx = lin(gridW);
+    for (let h = 0; h < gridH; h++) {
+      const hf = Math.trunc(hIdx[h]!);
+      const hc = Math.min(hf + 1, POS_GRID - 1);
+      const dh = Math.fround(hIdx[h]! - hf);
+      for (let x = 0; x < gridW; x++) {
+        const wf = Math.trunc(wIdx[x]!);
+        const wc = Math.min(wf + 1, POS_GRID - 1);
+        const dw = Math.fround(wIdx[x]! - wf);
+        const i = h * gridW + x;
+        idx[0]![i] = hf * POS_GRID + wf;
+        idx[1]![i] = hf * POS_GRID + wc;
+        idx[2]![i] = hc * POS_GRID + wf;
+        idx[3]![i] = hc * POS_GRID + wc;
+        wts[0]![i] = Math.fround((1 - dh) * (1 - dw));
+        wts[1]![i] = Math.fround((1 - dh) * dw);
+        wts[2]![i] = Math.fround(dh * (1 - dw));
+        wts[3]![i] = Math.fround(dh * dw);
+      }
+    }
+    // pos_embeds[k] = pos_embed[idx_k] * w_k (bf16, like the reference's
+    // weight tensor cast); summed pairwise in reference order.
+    let acc: MlxArray | null = null;
+    for (let k = 0; k < 4; k++) {
+      const ids = MlxArray.fromInt32(idx[k]!, [n]);
+      const rows = ops.takeAxis(this.#posEmbed, ids, 0); // [n, 1152] bf16
+      ids.dispose();
+      const wf = MlxArray.fromFloat32(wts[k]!, [n, 1]);
+      const wb = disposing(wf, wf.astype(Dtype.bfloat16));
+      const term = ops.mul(rows, wb);
+      rows.dispose();
+      wb.dispose();
+      acc = acc ? disposing(acc, disposing(term, ops.add(acc, term))) : term;
+    }
+    // Merge-block-major permutation (rows currently h-major patch order).
+    const ms = QWEN3VL_MERGE_SIZE;
+    const permIdx = new Int32Array(n);
+    let r = 0;
+    for (let h1 = 0; h1 < gridH / ms; h1++)
+      for (let w1 = 0; w1 < gridW / ms; w1++)
+        for (let mh = 0; mh < ms; mh++)
+          for (let mw = 0; mw < ms; mw++)
+            permIdx[r++] = (h1 * ms + mh) * gridW + (w1 * ms + mw);
+    const pIds = MlxArray.fromInt32(permIdx, [n]);
+    const out = ops.takeAxis(acc!, pIds, 0);
+    pIds.dispose();
+    acc!.dispose();
+    return out;
+  }
+
+  /** 2D rotary cos/sin [n, HEAD_DIM] f32 in merge-block-major order —
+   *  port of rot_pos_emb + apply_rotary_pos_emb_vision's tile(·,2). */
+  #ropeCosSin(gridH: number, gridW: number): [MlxArray, MlxArray] {
+    const n = gridH * gridW;
+    const half = ROT_DIM / 2; // 18 inv freqs
+    // inv_freq on-device, the reference's exact graph — Metal powf is not
+    // correctly rounded, so a JS Math.pow emulation differs by 1 ulp on some
+    // entries (seen as 2.4e-7 in the freq gate before this).
+    const invFreq = (() => {
+      const ar = ops.arange(0, ROT_DIM, 2, Dtype.float32);
+      const ex = disposing(ar, ops.mulScalar(ar, 1 / ROT_DIM));
+      const theta = MlxArray.fromFloat32(Float32Array.from([ROPE_THETA]), [1]);
+      const p = ops.pow(theta, ex);
+      theta.dispose();
+      ex.dispose();
+      const one = MlxArray.fromFloat32(Float32Array.from([1]), [1]);
+      const inv = disposing(p, ops.div(one, p));
+      one.dispose();
+      const v = inv.toFloat32();
+      inv.dispose();
+      return v;
+    })();
+    const ms = QWEN3VL_MERGE_SIZE;
+    const freqs = new Float32Array(n * ROT_DIM);
+    let r = 0;
+    for (let h1 = 0; h1 < gridH / ms; h1++) {
+      for (let w1 = 0; w1 < gridW / ms; w1++) {
+        for (let mh = 0; mh < ms; mh++) {
+          for (let mw = 0; mw < ms; mw++) {
+            const hPos = h1 * ms + mh;
+            const wPos = w1 * ms + mw;
+            const base = r * ROT_DIM;
+            for (let i = 0; i < half; i++) {
+              freqs[base + i] = Math.fround(hPos * invFreq[i]!);
+              freqs[base + half + i] = Math.fround(wPos * invFreq[i]!);
+            }
+            r++;
+          }
+        }
+      }
+    }
+    const f = MlxArray.fromFloat32(freqs, [n, ROT_DIM]);
+    const cosH = ops.cos(f);
+    const sinH = ops.sin(f);
+    f.dispose();
+    // tile(·, (1, 2)) over the last dim → [n, HEAD_DIM].
+    const cos = disposing(cosH, ops.concatAxis([cosH, cosH], -1));
+    const sin = disposing(sinH, ops.concatAxis([sinH, sinH], -1));
+    return [cos, sin];
+  }
+
+  /** (t·cos) + (rotate_half(t)·sin), f32 math, cast back to bf16. */
+  #applyRope(x: MlxArray, cos: MlxArray, sin: MlxArray): MlxArray {
+    // x [n, HEADS, HEAD_DIM]; cos/sin [n, 1, HEAD_DIM] broadcast over heads.
+    const half = HEAD_DIM / 2;
+    const n = x.shape[0]!;
+    const x1 = x.slice([0, 0, 0], [n, HEADS, half]);
+    const x2 = x.slice([0, 0, half], [n, HEADS, HEAD_DIM]);
+    const negX2 = ops.mulScalar(x2, -1);
+    x2.dispose();
+    const rot = disposing(negX2, ops.concatAxis([negX2, x1], -1));
+    x1.dispose();
+    const a = ops.mul(x, cos);
+    const b = disposing(rot, ops.mul(rot, sin));
+    const sum = ops.add(a, b);
+    a.dispose();
+    b.dispose();
+    return disposing(sum, sum.astype(Dtype.bfloat16));
+  }
+
+  /** Stage-instrumented encode for parity bisection: returns
+   *  named f32 copies of each pipeline stage alongside the final output. */
+  encodeStages(pp: Qwen3VLPreprocessed): Map<string, Float32Array> {
+    const stages = new Map<string, Float32Array>();
+    const out = this.encode(pp, (name, arr) => {
+      const f = arr.astype(Dtype.float32);
+      stages.set(name, f.toFloat32());
+      f.dispose();
+    });
+    const f = out.astype(Dtype.float32);
+    stages.set("merged", f.toFloat32());
+    f.dispose();
+    out.dispose();
+    return stages;
+  }
+
+  /** Parity-bisect helper: run ONLY the block chain (+ optional merger) on a
+   *  caller-provided hidden grid — used to separate patch-embed seed ulp
+   *  from block-math divergence in the stage gates. */
+  debugRunBlocks(
+    hF32: Float32Array, gridH: number, gridW: number, throughMerger = false,
+  ): Float32Array {
+    const n = gridH * gridW;
+    const hf = MlxArray.fromFloat32(hF32, [n, HIDDEN]);
+    let h = disposing(hf, hf.astype(Dtype.bfloat16));
+    const [cosF, sinF] = this.#ropeCosSin(gridH, gridW);
+    const cos = disposing(cosF, ops.expandDims(cosF, 1));
+    const sin = disposing(sinF, ops.expandDims(sinF, 1));
+    const scale = Math.pow(HEAD_DIM, -0.5);
+    for (const blk of this.#blocks) h = this.#block(h, blk, cos, sin, scale, n);
+    cos.dispose();
+    sin.dispose();
+    if (throughMerger) {
+      const mnorm = disposing(h, ops.layerNorm(h, this.#mergerNormW, this.#mergerNormB, 1e-6));
+      const grouped = disposing(mnorm, ops.reshape(mnorm, [n / 4, HIDDEN * 4]));
+      const f1 = disposing(grouped, this.#linear(grouped, this.#mergerFc1W, this.#mergerFc1B));
+      const act = disposing(f1, ops.geluPrecise(f1));
+      h = disposing(act, this.#linear(act, this.#mergerFc2W, this.#mergerFc2B));
+    }
+    const f = disposing(h, h.astype(Dtype.float32));
+    const out = f.toFloat32();
+    f.dispose();
+    return out;
+  }
+
+  /** Encode one preprocessed image → [imageTokens, 5120] bf16. */
+  encode(
+    pp: Qwen3VLPreprocessed,
+    tap?: (name: string, arr: MlxArray) => void,
+  ): MlxArray {
+    const [gridT, gridH, gridW] = pp.gridThw;
+    const n = pp.rows;
+    const pvF = MlxArray.fromFloat32(pp.pixelValues, [n, pp.cols]);
+    const pv = disposing(pvF, pvF.astype(Dtype.bfloat16));
+    // PatchEmbed: rows (c,t,py,px) → [N, tps, ps, ps, C] NDHWC → Conv3d with
+    // kernel == stride — the reference's exact reshape/moveaxis/conv graph.
+    const tps = QWEN3VL_TEMPORAL_PATCH_SIZE;
+    const ps = QWEN3VL_PATCH_SIZE;
+    const r5 = disposing(pv, ops.reshape(pv, [n, 3, tps, ps, ps]));
+    const tr = disposing(r5, ops.transposeAxes(r5, [0, 2, 3, 4, 1]));
+    const ndhwc = disposing(tr, ops.contiguous(tr));
+    const convOut = disposing(
+      ndhwc, ops.conv3d(ndhwc, this.#patchW, [tps, ps, ps]),
+    );
+    let h = disposing(convOut, ops.reshape(convOut, [n, HIDDEN]));
+    h = disposing(h, ops.add(h, this.#patchB));
+    tap?.("patch", h);
+    // Per-temporal-group pos/rope embeds are identical — tile the (h,w)
+    // block gridT times (the reference tiles coords/pos_embed per frame).
+    const tile = (a: MlxArray): MlxArray => {
+      if (gridT === 1) return a;
+      const parts = new Array(gridT).fill(a);
+      return disposing(a, ops.concatAxis(parts, 0));
+    };
+    const pos = tile(this.#interpolatedPosEmbeds(gridH, gridW));
+    tap?.("pos", pos);
+    h = disposing(h, disposing(pos, ops.add(h, pos)));
+
+    const [cosF0, sinF0] = this.#ropeCosSin(gridH, gridW);
+    const cosF = tile(cosF0);
+    const sinF = tile(sinF0);
+    const cos = disposing(cosF, ops.expandDims(cosF, 1)); // [n,1,72]
+    const sin = disposing(sinF, ops.expandDims(sinF, 1));
+
+    const scale = Math.pow(HEAD_DIM, -0.5);
+    // Attention never crosses temporal groups (cu_seqlens = h·w per frame
+    // pair): images run one segment, videos gridT segments.
+    const segLen = gridH * gridW;
+    let blkIdx = 0;
+    for (const blk of this.#blocks) {
+      h = this.#block(h, blk, cos, sin, scale, n, segLen);
+      tap?.(`blk${blkIdx}`, h);
+      blkIdx++;
+    }
+    tap?.("blocks", h);
+    cos.dispose();
+    sin.dispose();
+
+    // Merger: LayerNorm(1152) → reshape 4× → fc1 → GELU(erf) → fc2.
+    const mnorm = disposing(h, ops.layerNorm(h, this.#mergerNormW, this.#mergerNormB, 1e-6));
+    const grouped = disposing(mnorm, ops.reshape(mnorm, [n / 4, HIDDEN * 4]));
+    const f1 = disposing(grouped, this.#linear(grouped, this.#mergerFc1W, this.#mergerFc1B));
+    const act = disposing(f1, ops.geluPrecise(f1)); // reference nn.GELU()
+    return disposing(act, this.#linear(act, this.#mergerFc2W, this.#mergerFc2B));
+  }
+
+  dispose(): void {
+    // Sidecar arrays are process-pinned (house rule: mmap-backed weights are
+    // never handed a JS dtor); nothing to free eagerly.
+  }
+}