mlx-bun 0.0.7 → 0.0.9

package/src/model/gemma4.ts

@@ -295,7 +295,14 @@ class Attention {
     } else if (
       ta === "flash" && flashSupported(q) && (mask.mode === "causal" || mask.mode === "array")
     ) {
-      // Training: ops.sdpa's dK vjp is wrong, so use the validated flash kernel.
+      // Training, OPT-IN flash path (MLX_BUN_TRAIN_ATTN=flash) for O(L)-memory
+      // long context. NOTE the default (ops.sdpa, below) is the CORRECT and
+      // verified route: its dQ/dK/dV vjps are exact (finite-difference checked;
+      // it's what mlx-lm's tuner differentiates). The two historical bugs were
+      // in THIS flash kernel (spurious dK transpose, divergent dQ barrier),
+      // both fixed and gated by tests/flash-attention.test.ts vs the ops.sdpa
+      // oracle — see flash-attention.ts's routing comment. Flash also remains
+      // Gemma-guarded in the trainer (e4b >=2K SIGTRAP not yet re-validated).
       // Full layers → window 0 (pure causal); sliding layers → their window.
       const window = this.isSliding ? this.windowSize : 0;
       attn = flashAttention(q, shared.keys, shared.values, 1.0, true, window);
@@ -663,6 +670,13 @@ export class Gemma4Model {
   /** Set by the trainer to enable gradient checkpointing in forwardLayers
    *  (null = off; the non-checkpointed path is untouched). */
   gradCkpt: GradCheckpointCtx | null = null;
+  /** Optional per-forward hidden-state tap for DSpark's DFlash H_ctx (paper
+   *  Eq 2). null = off → byte-identical forward. `layers`: residual-output
+   *  indices to capture (0..nLayers-1 = post-layer residual; nLayers =
+   *  post-finalNorm). `pos`: if set, capture only [1,1,H] at that seq position
+   *  (live decode); else the full [1,L,H] (offline regen). The caller owns and
+   *  disposes the captured arrays. Additive/read-only → parity untouched. */
+  hiddenTap: { layers: Set<number>; pos?: number; captured: Map<number, MlxArray> } | null = null;
   // e2b/e4b per-layer input machinery
   readonly perLayerEmbed: QuantizedEmbedding | null;
   readonly perLayerModelProjection: QuantizedLinear | null;
@@ -820,6 +834,20 @@ export class Gemma4Model {
     return out;
   }
 
+  /** Store layer `i`'s residual stream `h` ([1,L,H]) if the hidden tap requests
+   *  it. Copies (astype bf16, optional slice) so the loop's `h.dispose()` can't
+   *  free the capture. No-op and no graph ops when the tap is null → the forward
+   *  is byte-identical to an untapped run. (DSpark DFlash H_ctx.) */
+  protected captureLayer(i: number, h: MlxArray): void {
+    const tap = this.hiddenTap;
+    if (!tap || !tap.layers.has(i)) return;
+    const H = h.shape[2]!;
+    const src = tap.pos !== undefined ? h.slice([0, tap.pos, 0], [1, tap.pos + 1, H]) : h;
+    const copy = src.astype(this.embed.scales.dtype); // bf16, matches regen dump dtype
+    if (src !== h) src.dispose();
+    tap.captured.set(i, copy);
+  }
+
   /** Consumes h. */
   protected forwardLayers(
     h0: MlxArray, cache: Cache[], bidir: MlxArray | null, ids: MlxArray | null,
@@ -878,6 +906,7 @@ export class Gemma4Model {
       pls?.dispose();
       h.dispose();
       h = next;
+      this.captureLayer(i, h); // DFlash H_ctx tap (no-op unless hiddenTap set)
       intermediates[i] = shared;
     }
     perLayer?.dispose();
@@ -895,7 +924,9 @@ export class Gemma4Model {
     }
     for (const m of masks.values()) m.arr?.dispose();
 
-    return disposing(h, this.finalNorm.forward(h));
+    h = disposing(h, this.finalNorm.forward(h));
+    this.captureLayer(this.layers.length, h); // post-finalNorm sentinel (index = nLayers)
+    return h;
   }
 
   // --- Segmented-backward support (docs/design/segmented-backward-training.md

package/src/model/minicpm5.ts

@@ -4,6 +4,7 @@
 import type { ModelConfig } from "../config";
 import type { Weights } from "../weights";
 import { MlxArray } from "../mlx/array";
+import { Dtype } from "../mlx/ffi";
 import * as ops from "../mlx/ops";
 import {
   argmaxLastPosition,
@@ -16,9 +17,56 @@ import {
   QuantizedLinear,
   quantizedSdpa,
   RMSNorm,
+  isCompiledTrace,
   type Cache,
   type Mask,
 } from "./gemma4-base";
+import {
+  fusedSwiglu,
+  fusedSwigluEnabled,
+  fusedSwigluSupported,
+  swigluMMin,
+  swigluUnderCap,
+} from "./fused-swiglu-kernel";
+import { steelLinear } from "./steel-linear-kernel";
+import { fusedMlp } from "./fused-mlp-kernel";
+
+/** Break the graph at a boundary: contiguous → eval → fresh graph-free leaf — the
+ *  exact primitive the segmented TRAINER uses so MLX never plans across boundaries
+ *  and prior-segment activations are freed (copy of train/segmented.ts detachLeaf). */
+function detachLeaf(a: MlxArray): MlxArray {
+  const c = ops.contiguous(a);
+  c.eval();
+  const leaf = MlxArray.fromBytesCopy(c.rawBytes(), c.shape, c.dtype);
+  c.dispose();
+  a.dispose();
+  return leaf;
+}
+
+/** A quantized projection as OUR steel dispatch when the controlled path is live
+ *  (no adapter, affine, clean tiling, M>1 so it's not a decode GEMV); otherwise
+ *  the normal linear. x is [B,L,H] or [M,H]. */
+function ourLinear(lin: QuantizedLinear, x: MlxArray): MlxArray {
+  const H = x.shape[x.shape.length - 1]!;
+  const M = x.shape.reduce((a, b) => a * b, 1) / H;
+  if (
+    !fusedSwigluEnabled() || isCompiledTrace() || M < swigluMMin() ||
+    lin.adapters != null || lin.biases == null ||
+    (x.dtype !== Dtype.bfloat16 && x.dtype !== Dtype.float16) ||
+    H % 32 !== 0 || lin.spec.groupSize % 32 !== 0
+  ) {
+    return lin.forward(x);
+  }
+  const x2 = x.shape.length === 2 ? x : ops.reshape(x, [M, H]);
+  const y = steelLinear(x2, lin.w, lin.scales, lin.biases, lin.spec);
+  if (x2 !== x) x2.dispose();
+  return x.shape.length === 2
+    ? y
+    : disposing(y, ops.reshape(y, [...x.shape.slice(0, -1), lin.w.shape[0]!]));
+}
+
+/** How often the inference fused path materializes the residual stream (layers). */
+const EVAL_EVERY = Math.max(1, Number(process.env.MLX_BUN_SWIGLU_EVAL_EVERY) || 1);
 import { flashAttention, getTrainingAttn, flashSupported } from "./flash-attention";
 
 // Shared prompt-prefix plan (lever 7). When set, the attention ropes the
@@ -156,7 +204,40 @@ class LlamaMLP {
     this.down = QuantizedLinear.load(weights, `${prefix}.down_proj`, config);
   }
 
-  forward(x: MlxArray): MlxArray {
+  /** MLP. When `residual` (the layer input h) is given AND the fused path applies,
+   *  the kernel computes h + mlp(x) directly — the residual is ABSORBED, so nothing
+   *  stays live across the kernel for MLX's planner to alias (it just runs our one
+   *  dispatch). Otherwise returns mlp(x) and the caller adds the residual. */
+  forward(x: MlxArray, residual?: MlxArray): MlxArray {
+    const H = x.shape[x.shape.length - 1]!;
+    const M = x.shape.reduce((a, b) => a * b, 1) / H;
+    if (
+      residual != null &&
+      fusedSwigluEnabled() && !isCompiledTrace() && M >= swigluMMin() && swigluUnderCap() &&
+      this.gate.adapters == null && this.up.adapters == null && this.down.adapters == null &&
+      this.down.biases != null && this.down.spec.groupSize % 32 === 0 &&
+      fusedSwigluSupported(x, this.gate.spec, this.up.spec, this.gate.biases, this.up.biases)
+    ) {
+      // The WHOLE MLP + residual in ONE kernel — gate+up+silu+down then +h, `hidden`
+      // never materialized, residual absorbed. MLX runs one dispatch; nothing for
+      // its planner to alias → chains across layers with no boundary.
+      const Hin = this.gate.scales.shape[1]! * this.gate.spec.groupSize;
+      const Ho = this.down.w.shape[0]!;
+      const x2 = x.shape.length === 2 ? x : ops.reshape(x, [M, Hin]);
+      const h2 = residual.shape.length === 2 ? residual : ops.reshape(residual, [M, Ho]);
+      const out2 = fusedMlp(
+        x2,
+        this.gate.w, this.gate.scales, this.gate.biases!, this.gate.spec,
+        this.up.w, this.up.scales, this.up.biases!, this.up.spec,
+        this.down.w, this.down.scales, this.down.biases!, this.down.spec,
+        h2,
+      );
+      if (x2 !== x) x2.dispose();
+      if (h2 !== residual) h2.dispose();
+      return x.shape.length === 2
+        ? out2
+        : disposing(out2, ops.reshape(out2, [...x.shape.slice(0, -1), Ho]));
+    }
     const gate = this.gate.forward(x);
     const up = this.up.forward(x);
     const sig = ops.sigmoid(gate);
@@ -166,9 +247,10 @@ class LlamaMLP {
     const hidden = ops.mul(silu, up);
     silu.dispose();
     up.dispose();
-    const out = this.down.forward(hidden);
+    const m = this.down.forward(hidden);
     hidden.dispose();
-    return out;
+    if (residual == null) return m;
+    return disposing(m, ops.add(residual, m));
   }
 }
 
@@ -192,11 +274,10 @@ class LlamaLayer {
     const h = ops.add(x, a);
     a.dispose();
     const hn = this.postAttnNorm.forward(h);
-    const m = this.mlp.forward(hn);
+    // Pass the residual h INTO the MLP so the fused kernel absorbs it (out = h + mlp).
+    const out = this.mlp.forward(hn, h);
     hn.dispose();
-    const out = ops.add(h, m);
     h.dispose();
-    m.dispose();
     return out;
   }
 }
@@ -256,7 +337,15 @@ export class MiniCPM5Model {
   }
 
   protected forwardLayers(h0: MlxArray, cache: Cache[]): MlxArray {
-    const h = this.runLayerRange(h0, 0, this.layers.length, cache);
+    // When the fused SwiGLU path is live, WE own the eval schedule: materialize
+    // each layer's output so the custom kernel sits in a shallow per-layer graph
+    // rather than a 24-deep lazy graph whose buffer planner aliases the kernel's
+    // output (the deep-graph corruption). Training/segmented-backward call
+    // runLayerRange directly and never set this, so autograd is untouched.
+    const h = this.runLayerRange(
+      h0, 0, this.layers.length, cache,
+      fusedSwigluEnabled() && process.env.MLX_BUN_SWIGLU_NOEVAL !== "1",
+    );
     h0.dispose(); // runLayerRange keeps its input alive; we own h0 here
     return disposing(h, this.finalNorm.forward(h));
   }
@@ -269,7 +358,7 @@ export class MiniCPM5Model {
    *  segmented-backward training path (docs/design/segmented-backward-training.md).
    *  MiniCPM5 is plain full-attention with no KV-sharing, so each layer reads
    *  its own stateless `cache[i]` and a single causal mask. */
-  runLayerRange(h: MlxArray, aIdx: number, bIdx: number, cache: Cache[]): MlxArray {
+  runLayerRange(h: MlxArray, aIdx: number, bIdx: number, cache: Cache[], evalPerLayer = false): MlxArray {
     if (aIdx >= bIdx) return h;
     const L = h.shape[1]!;
     const mask = cache[aIdx]!.makeMask(L, null);
@@ -278,6 +367,15 @@ export class MiniCPM5Model {
       const next = this.layers[i]!.forward(cur, mask, cache[i]!);
       if (i > aIdx) cur.dispose(); // keep the input boundary leaf; drop interiors
       cur = next;
+      // Per-layer boundary (segmented-trainer pattern): materialize each layer's
+      // output so MLX never plans across layers — our single-dispatch fused MLP is
+      // the only custom kernel it runs per layer. evalAll (GPU sync, no host copy)
+      // suffices for one kernel/boundary; MLX_BUN_SWIGLU_DETACH=1 forces the
+      // stronger graph-breaking detachLeaf copy.
+      if (evalPerLayer) {
+        if (process.env.MLX_BUN_SWIGLU_DETACH === "1") cur = detachLeaf(cur);
+        else ops.evalAll([cur]);
+      }
     }
     mask.arr?.dispose();
     return cur;

package/src/model/qwen3.ts

@@ -0,0 +1,276 @@
+// Concrete model graph for plain Qwen3 (Qwen3ForCausalLM).
+// Port target: mlx_lm.models.qwen3. A standard dense decoder — GQA softmax
+// attention with per-head q/k RMSNorm (applied over head_dim BEFORE RoPE),
+// full-head RoPE (theta 1e6), SwiGLU MLP, RMSNorm, tied embeddings.
+//
+// Primary use here is as a TEXT-EMBEDDING backbone (mlx-community Qwen3-Embedding-*):
+// `embedPooled` returns the last-token hidden, L2-normalized — the pooling
+// Qwen3-Embedding was trained with (an <|endoftext|> token is appended by the
+// caller; its hidden is the sentence vector). The full causal-LM surface
+// (forward/generate via the tied head) is wired too so it conforms to RuntimeModel.
+
+import type { ModelConfig } from "../config";
+import type { Weights } from "../weights";
+import { MlxArray } from "../mlx/array";
+import * as ops from "../mlx/ops";
+import {
+  argmaxLastPosition,
+  disposing,
+  KVCache,
+  LoraState,
+  QuantizedEmbedding,
+  QuantizedLinear,
+  RMSNorm,
+  type Cache,
+  type Mask,
+} from "./gemma4-base";
+
+class Qwen3Attention {
+  readonly qProj: QuantizedLinear;
+  readonly kProj: QuantizedLinear;
+  readonly vProj: QuantizedLinear;
+  readonly oProj: QuantizedLinear;
+  readonly qNorm: RMSNorm;
+  readonly kNorm: RMSNorm;
+  readonly nHeads: number;
+  readonly nKvHeads: number;
+  readonly headDim: number;
+  readonly scale: number;
+  readonly ropeBase: number;
+
+  constructor(weights: Weights, config: ModelConfig, prefix: string) {
+    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.ropeBase = t.ropeParameters.full_attention?.ropeTheta ?? 1000000;
+    this.qProj = QuantizedLinear.load(weights, `${prefix}.q_proj`, config);
+    this.kProj = QuantizedLinear.load(weights, `${prefix}.k_proj`, config);
+    this.vProj = QuantizedLinear.load(weights, `${prefix}.v_proj`, config);
+    this.oProj = QuantizedLinear.load(weights, `${prefix}.o_proj`, config);
+    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: Cache): MlxArray {
+    const [B, L] = x.shape as [number, number, number];
+    let q = this.qProj.forward(x);
+    let k = this.kProj.forward(x);
+    let v = this.vProj.forward(x);
+
+    // Reshape to heads, then q/k RMSNorm over head_dim BEFORE transpose+RoPE
+    // (mlx-lm qwen3 reference order).
+    q = disposing(q, ops.reshape(q, [B, L, this.nHeads, this.headDim]));
+    k = disposing(k, ops.reshape(k, [B, L, this.nKvHeads, this.headDim]));
+    v = disposing(v, ops.reshape(v, [B, L, this.nKvHeads, this.headDim]));
+    q = disposing(q, this.qNorm.forward(q));
+    k = disposing(k, this.kNorm.forward(k));
+
+    q = disposing(q, ops.transposeAxes(q, [0, 2, 1, 3]));
+    k = disposing(k, ops.transposeAxes(k, [0, 2, 1, 3]));
+    v = disposing(v, ops.transposeAxes(v, [0, 2, 1, 3]));
+
+    q = disposing(q, ops.rope(q, this.headDim, this.ropeBase, cache.offset, null));
+    k = disposing(k, ops.rope(k, this.headDim, 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 out = this.oProj.forward(merged);
+    merged.dispose();
+    return out;
+  }
+}
+
+class Qwen3MLP {
+  readonly gate: QuantizedLinear;
+  readonly up: QuantizedLinear;
+  readonly down: QuantizedLinear;
+
+  constructor(weights: Weights, config: ModelConfig, prefix: string) {
+    this.gate = QuantizedLinear.load(weights, `${prefix}.gate_proj`, config);
+    this.up = QuantizedLinear.load(weights, `${prefix}.up_proj`, config);
+    this.down = QuantizedLinear.load(weights, `${prefix}.down_proj`, config);
+  }
+
+  forward(x: MlxArray): MlxArray {
+    const gate = this.gate.forward(x);
+    const up = this.up.forward(x);
+    const sig = ops.sigmoid(gate);
+    const silu = ops.mul(gate, sig);
+    gate.dispose();
+    sig.dispose();
+    const hidden = ops.mul(silu, up);
+    silu.dispose();
+    up.dispose();
+    const out = this.down.forward(hidden);
+    hidden.dispose();
+    return out;
+  }
+}
+
+class Qwen3Layer {
+  readonly attn: Qwen3Attention;
+  readonly mlp: Qwen3MLP;
+  readonly inputNorm: RMSNorm;
+  readonly postAttnNorm: RMSNorm;
+
+  constructor(weights: Weights, config: ModelConfig, prefix: string) {
+    this.attn = new Qwen3Attention(weights, config, `${prefix}.self_attn`);
+    this.mlp = new Qwen3MLP(weights, config, `${prefix}.mlp`);
+    this.inputNorm = new RMSNorm(weights.tensor(`${prefix}.input_layernorm.weight`), config.text.rmsNormEps);
+    this.postAttnNorm = new RMSNorm(weights.tensor(`${prefix}.post_attention_layernorm.weight`), config.text.rmsNormEps);
+  }
+
+  forward(x: MlxArray, mask: Mask, cache: Cache): MlxArray {
+    const xn = this.inputNorm.forward(x);
+    const a = this.attn.forward(xn, mask, cache);
+    xn.dispose();
+    const h = ops.add(x, a);
+    a.dispose();
+    const hn = this.postAttnNorm.forward(h);
+    const m = this.mlp.forward(hn);
+    hn.dispose();
+    const out = ops.add(h, m);
+    h.dispose();
+    m.dispose();
+    return out;
+  }
+}
+
+export class Qwen3Model {
+  readonly config: ModelConfig;
+  readonly weightsBytes: number;
+  readonly prefixBase = "model";
+  readonly loraState = new LoraState();
+  readonly embed: QuantizedEmbedding;
+  readonly layers: Qwen3Layer[];
+  readonly finalNorm: RMSNorm;
+  /** Separate lm_head when NOT tied; otherwise the tied embedding's asLinear. */
+  readonly lmHead: QuantizedLinear | null;
+
+  constructor(weights: Weights, config: ModelConfig) {
+    this.config = config;
+    this.weightsBytes = [...weights.shards.files.values()]
+      .reduce((a, f) => a + f.mmap.size, 0);
+    this.embed = QuantizedEmbedding.load(weights, "model.embed_tokens", config);
+    this.layers = Array.from(
+      { length: config.text.numHiddenLayers },
+      (_, i) => new Qwen3Layer(weights, config, `model.layers.${i}`),
+    );
+    this.finalNorm = new RMSNorm(weights.tensor("model.norm.weight"), config.text.rmsNormEps);
+    this.lmHead = config.text.tieWordEmbeddings
+      ? null
+      : QuantizedLinear.load(weights, "lm_head", config);
+  }
+
+  loraTargets(): Map<string, QuantizedLinear> {
+    const out = new Map<string, QuantizedLinear>();
+    for (let i = 0; i < this.layers.length; i++) {
+      const l = this.layers[i]!;
+      const p = `model.layers.${i}`;
+      out.set(`${p}.self_attn.q_proj`, l.attn.qProj);
+      out.set(`${p}.self_attn.k_proj`, l.attn.kProj);
+      out.set(`${p}.self_attn.v_proj`, l.attn.vProj);
+      out.set(`${p}.self_attn.o_proj`, l.attn.oProj);
+      out.set(`${p}.mlp.gate_proj`, l.mlp.gate);
+      out.set(`${p}.mlp.up_proj`, l.mlp.up);
+      out.set(`${p}.mlp.down_proj`, l.mlp.down);
+    }
+    return out;
+  }
+
+  makeCache(): Cache[] {
+    return this.layers.map(() => new KVCache());
+  }
+
+  forwardHidden(ids: MlxArray, cache: Cache[]): MlxArray {
+    const h = this.embed.encode(ids);
+    return this.forwardLayers(h, cache);
+  }
+
+  forwardEmbeddings(_embeds: MlxArray, _cache: Cache[], _bidir: MlxArray | null): MlxArray {
+    throw new Error("qwen3 input-embedding path is not supported");
+  }
+
+  protected forwardLayers(h0: MlxArray, cache: Cache[]): MlxArray {
+    const L = h0.shape[1]!;
+    const mask = cache[0]!.makeMask(L, null);
+    let cur = h0;
+    for (let i = 0; i < this.layers.length; i++) {
+      const next = this.layers[i]!.forward(cur, mask, cache[i]!);
+      cur.dispose();
+      cur = next;
+    }
+    mask.arr?.dispose();
+    return disposing(cur, this.finalNorm.forward(cur));
+  }
+
+  /** Last-token pooled, L2-normalized sentence embedding for one sequence.
+   *  ids [1, L] → vector [1, hidden]. The caller appends the EOS/eod token
+   *  (Qwen3-Embedding's pooling token) to `ids`. */
+  embedPooled(ids: MlxArray): MlxArray {
+    const cache = this.makeCache();
+    try {
+      const h = this.forwardHidden(ids, cache); // [1, L, hidden]
+      const [B, L, H] = h.shape as [number, number, number];
+      const last = h.slice([0, L - 1, 0], [B, L, H]); // [1, 1, hidden]
+      h.dispose();
+      const vec = disposing(last, ops.reshape(last, [B, H])); // [1, hidden]
+      // L2 normalize: x / max(||x||_2, eps)
+      const sq = ops.square(vec);
+      const ss = disposing(sq, ops.sumAxis(sq, -1, true));
+      const norm = disposing(ss, ops.sqrt(ss));
+      const out = ops.div(vec, norm);
+      vec.dispose();
+      norm.dispose();
+      return out;
+    } finally {
+      for (const c of cache) c.dispose();
+    }
+  }
+
+  logitsFromHidden(h: MlxArray): MlxArray {
+    return this.lmHead ? this.lmHead.forward(h) : this.embed.asLinear(h);
+  }
+
+  forward(tokens: number[] | MlxArray, cache: Cache[]): MlxArray {
+    const ids = Array.isArray(tokens)
+      ? ops.fromInt32(tokens, [1, tokens.length])
+      : tokens;
+    const h = this.forwardHidden(ids, cache);
+    if (Array.isArray(tokens)) ids.dispose();
+    const logits = this.logitsFromHidden(h);
+    h.dispose();
+    return logits;
+  }
+
+  generate(promptTokens: number[], maxTokens: number, eosIds: number[] = []): number[] {
+    const cache = this.makeCache();
+    const out: number[] = [];
+    try {
+      let tokens = promptTokens;
+      for (let step = 0; step < maxTokens; step++) {
+        const logits = this.forward(tokens, cache);
+        const next = argmaxLastPosition(logits);
+        logits.dispose();
+        if (eosIds.includes(next)) break;
+        out.push(next);
+        tokens = [next];
+      }
+    } finally {
+      for (const c of cache) c.dispose();
+    }
+    return out;
+  }
+}

package/src/model/steel-linear-kernel.ts

@@ -0,0 +1,96 @@
+// Plain quantized linear y = x @ Wᵀ as OUR dispatch, composed from the steel
+// building blocks already in src/train/steel-qmm-header.ts (BlockMMA + Block/
+// QuantizedBlockLoader — MLX's verbatim GEMM, the same code qmm_t_impl wraps).
+// This is the projection primitive for the controlled forward (q/k/v/o/down):
+// we own the dispatch and the eval boundary instead of emitting an ops.* node
+// into MLX's lazy graph. N (out) and H (in/K) are baked as template literals;
+// M comes from the x_shape helper. Output bit-matches the BlockMMA store epilogue
+// (TransformNone → cast to T), the same store qmm_t_impl uses.
+
+import { MlxArray } from "../mlx/array";
+import { MetalKernel } from "../mlx/metal-kernel";
+import { STEEL_QMM_HEADER } from "../train/steel-qmm-header";
+import * as ops from "../mlx/ops";
+import type { QuantSpec } from "../mlx/ops";
+
+const BM = 32, BN = 32, BK = 32, WM = 2, WN = 2;
+const TGN = WM * WN * 32;
+
+// One BM×BN output tile per threadgroup (grid.y = N-tiles, grid.z = M-tiles), the
+// K=H reduction looped inside — the qmm_t_impl body, dims via templateInts.
+const SOURCE = String.raw`
+  const uint M=x_shape[0]; constexpr int N=N_T, H=H_T;
+  constexpr int gs=GS_T, bts=BITS_T;
+  constexpr int pf=get_pack_factor<bts,8>(), bpp=get_bytes_per_pack<bts>();
+  constexpr int BKp=BK_T+16/sizeof(T);
+  threadgroup T Xs[BM_T*BKp];
+  threadgroup T Ws[BN_T*BKp];
+  using mma_t=mlx::steel::BlockMMA<T,T,BM_T,BN_T,BK_T,WM_T,WN_T,false,true,BKp,BKp>;
+  using lx_t=mlx::steel::BlockLoader<T,BM_T,BK_T,BKp,1,WM_T*WN_T*SIMD_SIZE>;
+  using lw_t=QuantizedBlockLoader<T,BN_T,BK_T,BKp,1,WM_T*WN_T*SIMD_SIZE,gs,bts>;
+  const uint sg=simdgroup_index_in_threadgroup, sl=thread_index_in_simdgroup;
+  const int Hh=(int)H, K_w=Hh*bpp/pf, K_g=Hh/gs;
+  const int m0=(int)threadgroup_position_in_grid.z*BM_T;
+  const int n0=(int)threadgroup_position_in_grid.y*BN_T;
+  const short num_els=(short)min(BM_T,(int)M-m0);
+  const short num_outs=(short)min(BN_T,(int)N-n0);
+
+  mma_t mma_op(sg,sl);
+  lx_t  lx(x + (int64_t)m0*Hh, Hh, Xs, sg, sl);
+  lw_t  lw((const device uint8_t*)w + (int64_t)n0*K_w, scales + (int64_t)n0*K_g, biases + (int64_t)n0*K_g, Hh, Ws, sg, sl);
+
+  for (int k=0; k<Hh; k+=BK_T){
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    if (num_els<BM_T) lx.load_safe(short2(BK_T,num_els)); else lx.load_unsafe();
+    if (num_outs<BN_T) lw.load_safe(short2(BK_T,num_outs)); else lw.load_unsafe();
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    mma_op.mma(Xs,Ws);
+    lx.next(); lw.next();
+  }
+  threadgroup_barrier(mem_flags::mem_threadgroup);
+  // Store straight to device y[m0.., n0..] via the BlockMMA epilogue (float accum
+  // → T, the SAME rounding MLX's quantized_matmul applies).
+  device T* yp = y + (int64_t)m0*(int64_t)N + n0;
+  if (num_els<BM_T || num_outs<BN_T) mma_op.store_result_safe(yp, N, short2(num_outs,num_els));
+  else mma_op.store_result(yp, N);
+`;
+
+const _kernels = new Map<string, MetalKernel>();
+function getKernel(gs: number, bits: number): MetalKernel {
+  const key = `${gs}:${bits}`;
+  let k = _kernels.get(key);
+  if (!k) {
+    k = new MetalKernel({
+      name: "mlx_bun_steel_linear",
+      inputNames: ["x", "w", "scales", "biases"],
+      outputNames: ["y"],
+      source: SOURCE,
+      header:
+        `${STEEL_QMM_HEADER}\n` +
+        `#define GS_T ${gs}\n#define BITS_T ${bits}\n` +
+        `#define BM_T ${BM}\n#define BN_T ${BN}\n#define BK_T ${BK}\n` +
+        `#define WM_T ${WM}\n#define WN_T ${WN}\n#define TGN_T ${TGN}\n`,
+      ensureRowContiguous: true,
+    });
+    _kernels.set(key, k);
+  }
+  return k;
+}
+
+/** y = x @ Wᵀ, quantized, as our own dispatch. `x` is [M, H] (bf16/f16); `w` is
+ *  the packed weight [N, H*bits/32] with per-group scales/biases [N, H/groupSize].
+ *  Returns [M, N] in x.dtype. H % BK == 0 and groupSize % BK == 0 required. */
+export function steelLinear(
+  x: MlxArray, w: MlxArray, scales: MlxArray, biases: MlxArray, spec: QuantSpec,
+): MlxArray {
+  const M = x.shape[0]!, H = x.shape[1]!, N = w.shape[0]!;
+  const [y] = getKernel(spec.groupSize, spec.bits).apply([x, w, scales, biases], {
+    outputShapeFn: (ins) => [{ shape: [ins[0]!.shape[0]!, N], dtype: x.dtype }],
+    grid: (ins) => [TGN, Math.ceil(N / BN), Math.ceil(ins[0]!.shape[0]! / BM)],
+    threadGroup: [TGN, 1, 1],
+    templateInts: { N_T: N, H_T: H },
+    templateDtypes: { T: x.dtype },
+  });
+  void M;
+  return y!;
+}