mlx-native 0.6.2

Pure-Rust Metal GPU compute library for MLX-compatible inference on Apple Silicon
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
#include <metal_stdlib>
#include <metal_simdgroup_matrix>
using namespace metal;

// Wave 5b — chunk-parallel Gated DeltaNet inter-chunk state-recurrence kernel.
//
// Spec source (math): arXiv 2412.06464 §4 (Yang–Hatamizadeh; chunkwise
// parallelization of the gated delta rule).
// Spec source (algorithm structure): FLA's
// `chunk_gated_delta_rule_fwd_kernel_h_blockdim64` at
// /opt/vllm/vllm/model_executor/layers/fla/ops/chunk_delta_h.py:43-298.
//
// No FLA / Triton / CUDA / Metal code is copied. The math here is a
// re-derivation from the paper's recurrence (Eq. 5–6 + §4 chunkwise form);
// the algorithmic structure (load / dot / mask / accumulate per chunk)
// follows the FLA Triton kernel pattern but is open-coded for Metal.
//
// # Algorithm (per (batch b, head i_h, V-tile i_v))
//
//   bh := h0[b, i_h, BV-rows-of-tile, :]        # [BV, K] f32 in shared mem
//   for i_t in 0..NT:
//       store bh -> h_out[b, i_t, i_h, BV-rows, :]    # bf16 store
//       bv  := u[b, t_chunk, i_h, BV-rows]            # [BT, BV] f32
//             - w[b, t_chunk, i_h, :] @ bh^T          # [BT, K] @ [K, BV]
//       store bf16(bv) -> v_new[b, t_chunk, i_h, BV-rows]
//
//       last_t := (i_t+1)*BT - 1
//       g_last := g[b, last_t, i_h]                   # f32 scalar
//       g_blk  := g[b, t_start..t_end, i_h]           # f32 [BT]
//       bv     := bf16-roundtrip(bv) * exp(g_last - g_blk)[:, None]
//       bh     := bh * exp(g_last)
//
//       bh += bv.T @ k[b, t_chunk, i_h/group_ratio, :]  # outer accumulate
//
//   final_state[b, i_h, BV-rows, :] := bh             # f32 store
//
// # Wave 5b.2 iter 1 — simdgroup_matrix MMA on BOTH matmul sections
//
//   Both section 2b (`bv = u - w @ bh^T`) and section 2f (`bh += bv^T @ k`)
//   use `simdgroup_matrix<float, 8, 8>` 8×8 hardware MMA on Apple Silicon
//   AMX/Tensor units. Per-thread MAC loops in the prior implementation
//   underused the simdgroup matrix engines.
//
//   Tile partitioning (BT=64, BV=32, K=128, TG_THREADS=128 = 4 simdgroups):
//
//   Section 2b: `bv[BT, BV] = u - w @ bh^T`
//     * Each simdgroup `s` owns BT rows [s*16, s*16+16) (2 row-tiles of 8)
//       × all 4 BV col-tiles → 8 accumulator fragments per simdgroup.
//     * K-reduction: 16 8×8 K-tiles. Per K-tile: load 2 w-frags (bf16 from
//       device, K-row-stride = w_t_stride) + 4 bh-frags (f32 from
//       threadgroup, transposed via `simdgroup_load(..., true)`); 8 MMAs.
//     * After loop, extract per-thread fragment elements and compute
//       `local_v[c] = u_val - bv_acc[c]` cell-wise (16 cells/thread).
//
//   Section 2f: `bh[BV, K] += bv_stage @ k`
//     * Each simdgroup `s` owns BV rows [s*8, s*8+8) (1 row-tile of 8)
//       × all 16 K col-tiles → 16 accumulator fragments per simdgroup.
//     * Initial accumulator load: bh frags from threadgroup memory
//       (already gated by exp_g_last in section 2d).
//     * BT-reduction: 8 8×8 BT-tiles. Per BT-tile: load 1 bv-frag (bf16
//       from threadgroup, layout `[BV, BT]`) + 16 k-frags (bf16 from
//       device, K-row-stride = k_t_stride); 16 MMAs.
//     * After loop, store 16 bh frags back to threadgroup `bh`.
//
//   Per-thread fragment-element mapping (BaseMMAFrag<T, 8, 8>::get_coord):
//     qid = lane_id / 4
//     fm  = (qid & 4) + (lane_id / 2) % 4
//     fn  = (qid & 2) * 2 + (lane_id % 2) * 2
//   Each lane owns 2 elements per 8×8 frag at row=fm, cols=fn,fn+1.
//
//   For section 2b's per-thread `local_v[c]` (c ∈ 0..16):
//     i = c / 8                  (row-tile within sg, 0..1)
//     j = (c % 8) / 2            (col-tile, 0..3)
//     e = c % 2                  (lane sub-element, 0..1)
//     bt = sgid*16 + i*8 + fm
//     bv_idx = j*8 + fn + e
//
// # Numerical precision
//
//   Inputs / stored intermediates: bf16. Accumulators: f32. The bh state
//   stays in f32 in shared memory across the T-loop (matches FLA's policy
//   at chunk_delta_h.py:85 — `b_h1 = tl.zeros([BV, 64], dtype=tl.float32)`).
//
//   The 8×8 MMA reduction is deterministic across runs; the lane-element
//   ordering within an 8×8 frag is fixed by Apple Silicon's hardware
//   tensor unit. This gives bit-equivalent or strictly better numerical
//   error vs the per-thread serial sum (FLA-line-255 oracle still passes
//   at 0.0e0).
//
// # bf16 round-trip on bv (FLA parity, post-gate placement)
//
//   FLA places the only bf16 round-trip on b_v at line 255
//   (`b_v = b_v.to(k.dtype.element_ty)`) — AFTER the gate multiply at
//   line 213, BEFORE the outer-update dot at line 261. The cast must
//   sit between gate and dot, not before either.
//
//   The kernel mirrors that ordering exactly:
//     1. local_v = u - w @ bh^T              (f32, MMA in 2b)
//     2. local_v *= exp(g_last - g_blk)      (FLA :213, gate in f32)
//     3. bh      *= exp(g_last)              (FLA :215, gate in f32)
//     4. bv_stage[*] = bf16(local_v)         (FLA :255, post-gate cast)
//     5. bh += bv_stage @ b_k                (FLA :261, outer dot reads
//                                             bf16-rounded values, MMA in 2f)
//
//   Wave5b.1 iter1.5 corrected this from an earlier ordering that
//   bf16-rounded BEFORE the gate; that diverged from FLA by ~6e-4 on
//   final_state and was caught by the FLA-line-255 oracle in
//   tests/fixtures/gated_delta_net_chunk_oracle.py.
//
// # Threadgroup memory layout (post-iter-1 simdgroup_matrix MMA)
//
//   We allocate (BV*K * 4) + (BV*BT * 2) bytes = 16 KB + 4 KB = 20 KB at
//   threadgroup(0). The first BV*K floats are `bh` (the running f32 state
//   tile); the next BV*BT bfloats are `bv_stage` (used to publish per-thread
//   bv values to all threads for the outer-update). bv_stage layout is
//   `[BV, BT]` (transposed from the prior `[BT, BV]` f32 layout — this
//   saves 4 KB threadgroup memory and lets section 2f's MMA load
//   `simdgroup_matrix<bfloat>` frags directly without transpose).
//   20 KB < 32 KB M5 Max max threadgroup memory.
//
// # Memory layouts (innermost-first)
//
//   k:    [B, T, Hg, K]      bf16  — `K` innermost
//   w:    [B, T, H,  K]      bf16  — `K` innermost
//   u:    [B, T, H,  V]      bf16  — `V` innermost
//   g:    [B, T, H]          f32   — `H` innermost
//   h0:   [B, H, V, K]       f32   — `K` innermost (matches FLA p_h0 layout)
//   h_out:[B, NT, H, V, K]   bf16  — `K` innermost
//   v_new:[B, T, H, V]       bf16  — `V` innermost (same as u)
//   final:[B, H, V, K]       f32   — `K` innermost (same as h0)
//
// # Threading
//
//   Grid: (NV_TILES = V/BV, H, B)
//   Threadgroup: TG_THREADS = 128 (4 simdgroups × 32 lanes), flat 1D.
//
// # Buffer bindings (matching the host)
//
//   buffer(0): k           bf16
//   buffer(1): w           bf16
//   buffer(2): u           bf16
//   buffer(3): g           f32
//   buffer(4): h0          f32
//   buffer(5): h_out       bf16
//   buffer(6): v_new       bf16
//   buffer(7): final_state f32
//   buffer(8): params      uint[8] = [B, T, Hg, H, K, V, BT, NT]

constant uint BV = 32u;          // V-tile width (per-threadgroup V slice)
constant uint TG_THREADS = 128u; // threadgroup size
constant uint NSG = 4u;          // simdgroups per threadgroup (= TG_THREADS / 32)

// bf16 round-trip — cast to bfloat then back to float to truncate
// f32 -> bf16 -> f32 (matches PyTorch `.to(bfloat16).float()`).
inline float bf16_round(float x) {
    return float(bfloat(x));
}

kernel void gated_delta_net_chunk_inter_state_bf16(
    device const bfloat *k         [[buffer(0)]],
    device const bfloat *w         [[buffer(1)]],
    device const bfloat *u         [[buffer(2)]],
    device const float  *g         [[buffer(3)]],
    device const float  *h0        [[buffer(4)]],
    device bfloat       *h_out     [[buffer(5)]],
    device bfloat       *v_new     [[buffer(6)]],
    device float        *final_state [[buffer(7)]],
    device const uint   *params    [[buffer(8)]],
    threadgroup float   *shared_mem [[threadgroup(0)]],
    uint3 tid3 [[thread_position_in_threadgroup]],
    uint3 tgid [[threadgroup_position_in_grid]],
    ushort sgid [[simdgroup_index_in_threadgroup]],
    ushort slid [[thread_index_in_simdgroup]]
) {
    const uint B   = params[0];
    const uint T   = params[1];
    const uint Hg  = params[2];
    const uint H   = params[3];
    const uint K   = params[4];
    const uint V   = params[5];
    const uint BT  = params[6];
    const uint NT  = params[7];

    const uint i_v = tgid.x; // V-tile index
    const uint i_h = tgid.y; // V-head
    const uint i_b = tgid.z; // batch
    const uint tid = tid3.x;

    if (i_b >= B || i_h >= H) return;
    if (i_v * BV >= V) return;

    const uint v_base       = i_v * BV;          // first V-row in tile
    const uint group_ratio  = H / Hg;
    const uint kh           = i_h / group_ratio; // GQA-mapped K-head

    // Strides (in elements).
    const uint k_t_stride   = Hg * K;
    const uint k_seq_stride = T * k_t_stride;
    const uint w_t_stride   = H * K;
    const uint w_seq_stride = T * w_t_stride;
    const uint u_t_stride   = H * V;
    const uint u_seq_stride = T * u_t_stride;
    const uint g_t_stride   = H;
    const uint g_seq_stride = T * g_t_stride;
    const uint state_head_stride = V * K;
    const uint state_seq_stride  = H * state_head_stride;
    const uint h_chunk_stride    = H * state_head_stride;
    const uint h_seq_stride      = NT * h_chunk_stride;

    // Threadgroup-memory partition (20 KB total — host allocates exactly
    // this much; see gated_delta_net_chunk.rs `dispatch_*`).
    //   shared_mem[0          .. BV*K)            = bh          [BV, K]   f32 (16 KB)
    //   shared_mem[BV*K bytes .. BV*K+BV*BT bytes)= bv_stage_bf [BV, BT]  bf16 (4 KB)
    threadgroup float  *bh          = shared_mem;
    threadgroup bfloat *bv_stage_bf =
        (threadgroup bfloat *)(shared_mem + (BV * K));

    // simdgroup-frag lane coordinates (BaseMMAFrag<T, 8, 8>::get_coord).
    const ushort qid = slid / 4;
    const ushort fm  = (qid & 4) + ((slid / 2) % 4);
    const ushort fn  = (qid & 2) * 2 + (slid % 2) * 2;

    // ===================================================================
    // 1. Load h0 -> bh.
    // ===================================================================
    {
        const uint h0_base = i_b * state_seq_stride + i_h * state_head_stride;
        for (uint flat = tid; flat < BV * K; flat += TG_THREADS) {
            const uint vv = flat / K;
            const uint kk = flat - vv * K;
            const uint vidx = v_base + vv;
            float val = 0.0f;
            if (vidx < V) {
                val = h0[h0_base + vidx * K + kk];
            }
            bh[vv * K + kk] = val;
        }
    }
    threadgroup_barrier(mem_flags::mem_threadgroup);

    // Per-thread accumulator for bv during phase 2b — populated from MMA
    // fragment elements after the K-reduction MMA loop.
    //
    // CELLS_PER_THREAD = 16: 2 row-tiles × 4 col-tiles × 2 elements.
    // c-mapping: c = i*8 + j*2 + e where i = row-tile (0..1), j = col-tile
    // (0..3), e = lane sub-element (0..1).
    //   bt_local = i*8 + fm                (within sg)
    //   bv_local = j*8 + fn + e
    constexpr uint CELLS_PER_THREAD = 16u;
    thread float local_v[CELLS_PER_THREAD];

    // ===================================================================
    // 2. Per-chunk loop.
    // ===================================================================
    for (uint i_t = 0; i_t < NT; ++i_t) {
        const uint t_start = i_t * BT;
        const uint last_t  = t_start + BT - 1;

        // -----------------------------------------------------------
        // 2a. Snapshot bh -> h_out.
        // -----------------------------------------------------------
        const uint h_out_base =
            i_b * h_seq_stride + i_t * h_chunk_stride + i_h * state_head_stride;
        for (uint flat = tid; flat < BV * K; flat += TG_THREADS) {
            const uint vv = flat / K;
            const uint kk = flat - vv * K;
            const uint vidx = v_base + vv;
            if (vidx < V) {
                h_out[h_out_base + vidx * K + kk] = bfloat(bh[vv * K + kk]);
            }
        }
        // No barrier needed — bh is read-only here.

        // -----------------------------------------------------------
        // 2b. Compute bv = u - w @ bh^T  via simdgroup_matrix<float, 8, 8>
        //     8×8 MMA. Each simdgroup owns BT rows [sgid*16, sgid*16+16)
        //     (2 row-tiles) × all 4 BV col-tiles = 8 fragments.
        // -----------------------------------------------------------
        const uint w_base = i_b * w_seq_stride + (t_start * w_t_stride) + i_h * K;
        const uint u_base = i_b * u_seq_stride + (t_start * u_t_stride) + i_h * V;

        // 8 accumulator frags per simdgroup, indexed [i*4 + j].
        simdgroup_matrix<float, 8, 8> bv_acc[8];
        for (uint t = 0; t < 8; ++t) {
            bv_acc[t] = simdgroup_matrix<float, 8, 8>(0);
        }

        // K-reduction: 16 K-tiles of 8.
        const uint num_k_tiles = K / 8u;
        for (uint kt = 0; kt < num_k_tiles; ++kt) {
            const uint k_origin = kt * 8u;

            simdgroup_barrier(mem_flags::mem_none);

            // Load 2 w fragments — sgid's BT row-tiles (2 of them).
            //   w[B,T,H,K] row-major K-innermost; bt-row stride = w_t_stride.
            //   Frag i covers BT rows [sgid*16 + i*8, sgid*16 + i*8 + 8).
            simdgroup_matrix<bfloat, 8, 8> w_frag[2];
            const device bfloat *w_ptr_base =
                w + w_base + sgid * 16u * w_t_stride + k_origin;
            simdgroup_load(w_frag[0], w_ptr_base + 0u * 8u * w_t_stride,
                           w_t_stride, 0, false);
            simdgroup_load(w_frag[1], w_ptr_base + 1u * 8u * w_t_stride,
                           w_t_stride, 0, false);

            // Load 4 bh^T fragments — all 4 BV col-tiles.
            //   bh layout is [BV, K] row-major (K-innermost, row stride = K).
            //   bh^T frag at (k_origin..k_origin+8, bv_origin..bv_origin+8)
            //   reads bh[bv_origin+j][k_origin+i] which is bh + bv_origin*K
            //   + k_origin with row stride K and transpose=true.
            simdgroup_matrix<float, 8, 8> bh_frag_T[4];
            for (uint j = 0; j < 4; ++j) {
                const uint bv_origin = j * 8u;
                threadgroup const float *bh_ptr =
                    bh + bv_origin * K + k_origin;
                simdgroup_load(bh_frag_T[j], bh_ptr, K, 0, true);
            }

            simdgroup_barrier(mem_flags::mem_none);

            // 8 MMAs: bv_acc[i*4+j] += w_frag[i] @ bh_frag_T[j].
            for (uint i = 0; i < 2; ++i) {
                for (uint j = 0; j < 4; ++j) {
                    simdgroup_multiply_accumulate(
                        bv_acc[i * 4 + j], w_frag[i], bh_frag_T[j],
                        bv_acc[i * 4 + j]);
                }
            }
        }

        // Extract per-thread fragment elements into local_v and subtract u.
        // c-mapping: c = i*8 + j*2 + e (see CELLS_PER_THREAD comment).
        for (uint i = 0; i < 2; ++i) {
            for (uint j = 0; j < 4; ++j) {
                thread auto &elems = bv_acc[i * 4 + j].thread_elements();
                const uint bt = sgid * 16u + i * 8u + fm;
                const uint bv0 = j * 8u + fn;
                // u_val for the two elements at (bt, bv0) and (bt, bv0+1).
                float u0 = 0.0f, u1 = 0.0f;
                const uint vidx0 = v_base + bv0;
                const uint vidx1 = v_base + bv0 + 1u;
                if (vidx0 < V) {
                    u0 = float(u[u_base + bt * u_t_stride + vidx0]);
                }
                if (vidx1 < V) {
                    u1 = float(u[u_base + bt * u_t_stride + vidx1]);
                }
                const uint c0 = i * 8u + j * 2u + 0u;
                const uint c1 = i * 8u + j * 2u + 1u;
                local_v[c0] = u0 - float(elems[0]);
                local_v[c1] = u1 - float(elems[1]);
            }
        }

        // -----------------------------------------------------------
        // 2c. Store bf16(bv) -> v_new (BEFORE the gate multiply, matching
        //     FLA's tl.store at line 199-203 — the stored v_new is
        //     u - w @ h^T without the gate scaling).
        // -----------------------------------------------------------
        const uint vn_base = i_b * u_seq_stride + (t_start * u_t_stride) + i_h * V;
        for (uint i = 0; i < 2; ++i) {
            for (uint j = 0; j < 4; ++j) {
                const uint bt = sgid * 16u + i * 8u + fm;
                const uint bv0 = j * 8u + fn;
                const uint c0 = i * 8u + j * 2u + 0u;
                const uint c1 = i * 8u + j * 2u + 1u;
                const uint vidx0 = v_base + bv0;
                const uint vidx1 = v_base + bv0 + 1u;
                if (vidx0 < V) {
                    v_new[vn_base + bt * u_t_stride + vidx0] = bfloat(local_v[c0]);
                }
                if (vidx1 < V) {
                    v_new[vn_base + bt * u_t_stride + vidx1] = bfloat(local_v[c1]);
                }
            }
        }

        // -----------------------------------------------------------
        // 2d. Gate multiplies (FLA chunk_delta_h.py:213, :215).
        //     local_v[c] *= exp(g_last - g_blk[bt])   (f32, pre-bf16-round)
        //     bh[*]      *= exp(g_last)
        // -----------------------------------------------------------
        const uint g_base = i_b * g_seq_stride + i_h;
        const float g_last = g[g_base + last_t * g_t_stride];
        const float exp_g_last = metal::exp(g_last);

        // Per-thread gate of local_v. Each thread holds 16 cells across
        // 2 distinct BT rows (i = 0, 1) — so two distinct g_t values.
        for (uint i = 0; i < 2; ++i) {
            const uint bt = sgid * 16u + i * 8u + fm;
            const float g_t = g[g_base + (t_start + bt) * g_t_stride];
            const float scale = metal::exp(g_last - g_t);
            for (uint j = 0; j < 4; ++j) {
                const uint c0 = i * 8u + j * 2u + 0u;
                const uint c1 = i * 8u + j * 2u + 1u;
                local_v[c0] *= scale;
                local_v[c1] *= scale;
            }
        }

        // Scale bh by exp_g_last in place (cooperative across threads).
        for (uint flat = tid; flat < BV * K; flat += TG_THREADS) {
            bh[flat] *= exp_g_last;
        }
        threadgroup_barrier(mem_flags::mem_threadgroup);

        // -----------------------------------------------------------
        // 2e. Publish bf16-rounded local_v to bv_stage_bf (FLA :255).
        //     Layout `[BV, BT]` (transposed): bv_stage_bf[bv * BT + bt].
        //     This lets section 2f's MMA load bf16 frags directly without
        //     a transpose flag.
        // -----------------------------------------------------------
        for (uint i = 0; i < 2; ++i) {
            for (uint j = 0; j < 4; ++j) {
                const uint bt = sgid * 16u + i * 8u + fm;
                const uint bv0 = j * 8u + fn;
                const uint c0 = i * 8u + j * 2u + 0u;
                const uint c1 = i * 8u + j * 2u + 1u;
                bv_stage_bf[bv0 * BT + bt]      = bfloat(local_v[c0]);
                bv_stage_bf[(bv0 + 1u) * BT + bt] = bfloat(local_v[c1]);
            }
        }
        threadgroup_barrier(mem_flags::mem_threadgroup);

        // -----------------------------------------------------------
        // 2f. Outer-update: bh += bv_stage_bf @ k  (MMA).
        //     Each simdgroup `s` owns BV rows [s*8, s*8+8) (1 row-tile)
        //     × all 16 K col-tiles → 16 accumulator fragments per sg.
        //     Initial accumulator = (already-gated) bh frags loaded from
        //     threadgroup memory.
        // -----------------------------------------------------------
        const uint k_chunk_base =
            i_b * k_seq_stride + t_start * k_t_stride + kh * K;

        // Load 16 bh accumulator frags (BV row-tile = sgid, all K col-tiles).
        //   Tile count is hard-coded for K=128 (= 16 8-element K-tiles). The
        //   compile-time loop bound is LOAD-BEARING — Metal's MMA scheduler
        //   needs unrollable loops for the simdgroup_matrix tile sequence;
        //   making the bound runtime via `K/8` collapses inter_state from
        //   1.08 ms to 3.40 ms (3.15× regression, measured 2026-04-27).
        //   So MAX_K is narrowed to 128 in src/ops/gated_delta_net_chunk.rs;
        //   K=192 dispatch is rejected by validate(). To support K=192/256
        //   later, port FLA's b_h1..b_h4 bank-split (chunk_delta_h.py:215-221)
        //   which keeps each kernel's K-tile count compile-time-known per bank.
        simdgroup_matrix<float, 8, 8> bh_acc[16];
        {
            threadgroup const float *bh_row_ptr = bh + sgid * 8u * K;
            for (uint kj = 0; kj < 16; ++kj) {
                simdgroup_load(bh_acc[kj], bh_row_ptr + kj * 8u, K, 0, false);
            }
        }

        // BT-reduction: 8 BT-tiles of 8.
        const uint num_bt_tiles = BT / 8u;
        for (uint bj = 0; bj < num_bt_tiles; ++bj) {
            const uint bt_origin = bj * 8u;

            simdgroup_barrier(mem_flags::mem_none);

            // Load 1 bv frag — sgid's BV row-tile, 8 BT cols.
            //   bv_stage_bf layout `[BV, BT]` row-major (BT innermost,
            //   row stride = BT). Frag covers BV rows [sgid*8, sgid*8+8)
            //   × BT cols [bt_origin, bt_origin+8).
            simdgroup_matrix<bfloat, 8, 8> bv_frag;
            simdgroup_load(bv_frag,
                           bv_stage_bf + sgid * 8u * BT + bt_origin,
                           BT, 0, false);

            // Load 16 k frags — 8 BT rows × 8 K cols each.
            //   k[B,T,Hg,K] row-major K-innermost; bt-row stride = k_t_stride.
            simdgroup_matrix<bfloat, 8, 8> k_frag[16];
            const device bfloat *k_row_base =
                k + k_chunk_base + bt_origin * k_t_stride;
            for (uint kj = 0; kj < 16; ++kj) {
                simdgroup_load(k_frag[kj], k_row_base + kj * 8u,
                               k_t_stride, 0, false);
            }

            simdgroup_barrier(mem_flags::mem_none);

            // 16 MMAs: bh_acc[kj] += bv_frag @ k_frag[kj].
            for (uint kj = 0; kj < 16; ++kj) {
                simdgroup_multiply_accumulate(
                    bh_acc[kj], bv_frag, k_frag[kj], bh_acc[kj]);
            }
        }

        // Store bh frags back to threadgroup memory.
        {
            threadgroup float *bh_row_ptr = bh + sgid * 8u * K;
            for (uint kj = 0; kj < 16; ++kj) {
                simdgroup_store(bh_acc[kj], bh_row_ptr + kj * 8u, K, 0, false);
            }
        }
        threadgroup_barrier(mem_flags::mem_threadgroup);
    }

    // ===================================================================
    // 3. Store final state.
    // ===================================================================
    {
        const uint fs_base = i_b * state_seq_stride + i_h * state_head_stride;
        for (uint flat = tid; flat < BV * K; flat += TG_THREADS) {
            const uint vv = flat / K;
            const uint kk = flat - vv * K;
            const uint vidx = v_base + vv;
            if (vidx < V) {
                final_state[fs_base + vidx * K + kk] = bh[vv * K + kk];
            }
        }
    }
}