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
#include <metal_stdlib>
#include <metal_simdgroup_matrix>
using namespace metal;

// Wave 5b.1 iter 3 — chunk_fwd_o Metal kernel.
// Wave 5b.2 iter 2 — simdgroup_matrix<float, 8, 8> MMA on three matmul
// sections (q · h^T, q · k^T, b_A · v_new), mirroring the inter_state
// iter 1 recipe.
//
// Spec source:
// - FLA reference: `chunk_fwd_kernel_o` at
//   /opt/vllm/vllm/model_executor/layers/fla/ops/chunk_o.py:42-138
//
// No FLA / Triton / CUDA / Metal code is copied. The math here is a
// re-derivation from the FLA spec; the algorithmic structure (load q/k/h,
// dot-and-accumulate across K-tiles, gate, mask, bf16 cast, dot with v)
// follows the FLA Triton kernel pattern but is open-coded for Metal.
//
// # Algorithm (per (batch b, V-head i_h, chunk i_t, V-tile i_v))
//
//   kh         = i_h / (H / Hg)                           # GQA-mapped K-head
//   bo_acc     = zeros([BT, BV])                          # f32 in MMA frags
//   bA_acc     = zeros([BT, BT])                          # f32 in MMA frags
//
//   # K-reduction (FLA :93-113), FUSED — both sections share the same q-load.
//   for kt in 0..(K // 8):
//       bo_acc[BT, BV] += q[BT, kt..] @ trans(h[BV, kt..])      # FLA :111
//       bA_acc[BT, BT] += q[BT, kt..] @ trans(k[BT, kt..])      # FLA :113
//
//   # Gate (FLA :115-120; USE_G == True for GDN):
//   bo_acc[i, j] *= exp(g[t_start + i])
//   bA_acc[i, j] *= exp(g[t_start + i] - g[t_start + j])
//
//   # Causal+diag mask (FLA :122-125 — INCLUSIVE of diagonal, `>=`):
//   bA_acc[i, j] = (i >= j) ? bA_acc[i, j] : 0
//
//   # bf16 round-trip on bA_acc (FLA :137 — AFTER mask, BEFORE dot with v):
//   bA_stage[i, j] = bfloat(bA_acc[i, j])  # publish to threadgroup mem
//
//   # Closing dot — bf16 b_A × bf16 v_new → f32 acc (MMA), then scale + store:
//   for bt_tile in 0..(BT // 8):
//       bo_acc[BT, BV] += bA_stage[BT, bt_tile..] @ v_new[bt_tile.., BV]
//   o[BT, BV] = bfloat(bo_acc * scale)                    # FLA :137
//
// # Wave 5b.2 iter 2 — simdgroup_matrix MMA on THREE matmul sections
//
//   All three matmul sections (2A: q·h^T over K, 2B: q·k^T over K,
//   2C: bA_bf16·v_new over BT) use `simdgroup_matrix<float, 8, 8>` 8×8
//   hardware MMA on Apple Silicon AMX/Tensor units. Per-thread MAC loops
//   in iter 3 underused the simdgroup matrix engines; iter 1's lesson on
//   inter_state (17.4× speedup) transfers directly here, with the bonus
//   that 2A and 2B share the same q-frag per K-tile (single q load amortized
//   across 4 BV-col MMAs + 8 BT-col MMAs = 12 MMAs/load).
//
//   Tile partitioning (BT=64, BV=32, K=128, TG_THREADS=256 = 8 simdgroups):
//
//   Section 2A: `bo_acc[BT, BV] += q · h^T`
//     * Each simdgroup `s` owns BT row [s*8, s*8+8) (1 row-tile)
//       × all 4 BV col-tiles → 4 accumulator fragments per simdgroup.
//     * K-reduction: 16 8×8 K-tiles. Per K-tile: load 1 q-frag (bf16,
//       sgid's BT row, 8 K cols, K-row-stride = qk_t_stride) + 4 h-frags
//       (bf16, h is [V, K] row-major K-innermost; transpose-load via
//       `simdgroup_load(..., true)` produces h^T tile [K, BV])
//       — wait: simdgroup_load reads as [rows, cols] in source layout.
//       For h^T frag covering (k_origin..+8, bv_origin..+8) we read
//       h[bv_origin+j][k_origin+i] which is base + bv_origin*K + k_origin
//       with row stride K and `transpose=true`. 4 MMAs.
//
//   Section 2B: `bA_acc[BT, BT] += q · k^T` (FUSED into 2A's K-loop)
//     * Each simdgroup `s` owns BT row [s*8, s*8+8) (same as 2A)
//       × all 8 BT col-tiles → 8 accumulator fragments per simdgroup.
//     * Per K-tile (already inside 2A's loop, q-frag reused):
//       load 8 k^T-frags (bf16, k is [BT, K] row-major K-innermost,
//       transpose-load via `simdgroup_load(..., true)`); 8 MMAs.
//
//   Section 2C: `bo_acc += bA_bf16 · v_new`
//     * Each simdgroup `s` owns same BT row-tile [s*8, s*8+8)
//       × all 4 BV col-tiles → 4 accumulator fragments per simdgroup
//       (bo_acc reused from 2A — already gated, scale applied at store).
//     * BT-reduction: 8 8×8 BT-tiles. Per BT-tile: load 1 bA_bf16 frag
//       (bf16, bA_stage layout [BT, BT] row-major, BT cols innermost,
//       row stride = BT) + 4 v-frags (bf16, v_new is [BT, BV] within
//       a chunk-V-tile, row stride = v_t_stride); 4 MMAs.
//
//   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.
//
// # Numerical precision
//
//   Inputs / stored intermediates: bf16. Accumulators: f32. 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-137 oracle still passes at < 1e-6).
//
// # CRITICAL ORDERING DETAILS (per iter-1.5 lesson)
//
// 1. The bf16 cast of bA_acc happens AFTER the mask, BEFORE the dot with
//    v (FLA line 137). Moving or skipping it changes numerics. The MMA
//    reformulation preserves this: bA_acc lives in f32 frags during 2B,
//    is gated/masked in f32 (per-thread element extract), then bf16-cast
//    and published to threadgroup `bA_stage` BEFORE 2C reads it.
// 2. The mask is `>=` (causal+diag) — NOT strict `>` like kkt. The
//    diagonal element corresponds to attention to the same token and is
//    included.
// 3. Two `* scale` multiplications on FLA line 137: BOTH `bo_acc * scale`
//    AND `dot_result * scale`. Mathematically equivalent to one global
//    scale on the sum; we apply scale once at the final bf16 store.
// 4. h dimension order is [V, K] per (chunk, head) — `b_o += dot(b_q,
//    trans(b_h))` per FLA :111.
//
// # Threadgroup memory budget (post-iter-2 MMA, BT=64)
//
//   bA_stage : BT * BT * 2 bytes (bf16) = 64 * 64 * 2 = 8 KB
//   Total: 8 KB (down from 28 KB iter 3 — bo_acc and ba_acc now live in
//   simdgroup-matrix accumulator registers, not threadgroup memory).
//
// # K-tile count is hard-coded — DO NOT generalize
//
//   The K-reduction loop runs `num_k_tiles = 16` (= K/8 at K=128). 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/8u` collapsed inter_state's section 2f from
//   1.08 ms to 3.40 ms (3.15× regression, measured 2026-04-27, iter 1.5).
//   So MAX_K is narrowed to 128 in src/ops/gated_delta_net_chunk_o.rs;
//   K=192/256 dispatch is rejected by validate(). To support K=192/256
//   later, port FLA's bank-split which keeps each kernel's K-tile count
//   compile-time-known per bank.
//
// # Memory layouts (innermost-first)
//
//   q:     [B, T, Hg, K]    bf16  — K innermost
//   k:     [B, T, Hg, K]    bf16  — K innermost
//   v_new: [B, T, H,  V]    bf16  — V innermost
//   h:     [B, NT, H, V, K] bf16  — K innermost (per chunk-head)
//   g:     [B, T, H]        f32   — H innermost (chunk-cumsumed)
//   o:     [B, T, H, V]     bf16  — V innermost (output)
//
// # Threading
//
//   Grid: (NV, NT, B*H) where NV = V/BV
//        program_id(0) -> i_v (V-tile index)
//        program_id(1) -> i_t (chunk index)
//        program_id(2) -> i_bh (flattened batch * head)
//          -> i_b = i_bh / H,  i_h = i_bh % H
//
//   Threadgroup: TG_THREADS = 256 threads (8 simdgroups × 32 lanes).
//
// # Buffer bindings
//
//   buffer(0): q        bf16
//   buffer(1): k        bf16
//   buffer(2): v        bf16  (= v_new from iter 1)
//   buffer(3): h        bf16
//   buffer(4): g        f32
//   buffer(5): o        bf16  (output)
//   buffer(6): params   uint[11] = [B, T, Hg, H, K, V, BT, NT, BK, BV]
//                                   plus float[1] scale at byte offset 40.

constant uint TG_THREADS = 256u;
constant uint NSG = 8u;        // simdgroups per threadgroup (= TG_THREADS / 32)

kernel void gated_delta_net_chunk_o_bf16(
    device const bfloat *q          [[buffer(0)]],
    device const bfloat *k          [[buffer(1)]],
    device const bfloat *v          [[buffer(2)]],
    device const bfloat *h          [[buffer(3)]],
    device const float  *g          [[buffer(4)]],
    device bfloat       *o          [[buffer(5)]],
    device const uint   *params     [[buffer(6)]],
    threadgroup bfloat  *bA_stage   [[threadgroup(0)]],   // [BT, BT] bf16, 8 KB
    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];
    // params[8] = BK, params[9] = BV — retained for host/test compatibility,
    // not used by this MMA-rewritten kernel (BV = 32 is hard-coded via the
    // 4 BV col-tile count; BK is irrelevant since the K-loop is per 8-tile).

    // scale is packed as float in the same buffer at u32 index 10
    // (host writes scale.to_bits() at u32[10]).
    const float scale = as_type<float>(params[10]);

    const uint i_v_tile = tgid.x;
    const uint i_t      = tgid.y;
    const uint i_bh     = tgid.z;
    const uint tid      = tid3.x;

    const uint i_b = i_bh / H;
    const uint i_h = i_bh - i_b * H;

    if (i_b >= B || i_h >= H || i_t >= NT) return;

    const uint nv  = V / 32u;
    if (i_v_tile >= nv) return;

    const uint group_ratio = H / Hg;
    const uint kh          = i_h / group_ratio;
    const uint t_start     = i_t * BT;
    const uint v_off       = i_v_tile * 32u;

    // ---- Strides (in elements) ----
    // q, k: [B, T, Hg, K]
    const uint qk_t_stride   = Hg * K;
    const uint qk_seq_stride = T * qk_t_stride;
    // v, o: [B, T, H, V]
    const uint v_t_stride    = H * V;
    const uint v_seq_stride  = T * v_t_stride;
    // h: [B, NT, H, V, K]
    const uint h_v_stride    = K;
    const uint h_h_stride    = V * h_v_stride;
    const uint h_nt_stride   = H * h_h_stride;
    const uint h_seq_stride  = NT * h_nt_stride;
    // g: [B, T, H]
    const uint g_t_stride    = H;
    const uint g_seq_stride  = T * g_t_stride;

    // Per-(b, h, i_t, i_v) bases.
    const uint q_chunk_base = i_b * qk_seq_stride
                            + t_start * qk_t_stride
                            + kh * K;
    const uint k_chunk_base = i_b * qk_seq_stride
                            + t_start * qk_t_stride
                            + kh * K;
    const uint v_chunk_base = i_b * v_seq_stride
                            + t_start * v_t_stride
                            + i_h * V;
    const uint h_chunk_base = i_b * h_seq_stride
                            + i_t * h_nt_stride
                            + i_h * h_h_stride;
    const uint o_chunk_base = i_b * v_seq_stride
                            + t_start * v_t_stride
                            + i_h * V;
    const uint g_base       = i_b * g_seq_stride + i_h;

    // 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. Initialise MMA accumulator frags.
    //    bo_acc: 4 frags per sg (1 BT row-tile × 4 BV col-tiles).
    //    bA_acc: 8 frags per sg (1 BT row-tile × 8 BT col-tiles).
    // ===================================================================
    simdgroup_matrix<float, 8, 8> bo_acc[4];
    simdgroup_matrix<float, 8, 8> bA_acc[8];
    for (uint c = 0; c < 4; ++c) {
        bo_acc[c] = simdgroup_matrix<float, 8, 8>(0);
    }
    for (uint c = 0; c < 8; ++c) {
        bA_acc[c] = simdgroup_matrix<float, 8, 8>(0);
    }

    // ===================================================================
    // 2. K-reduction loop (16 K-tiles of 8). Sections 2A and 2B fused —
    //    they share the same q-frag (sgid's BT row-tile, current K-tile).
    //    Per FLA :93-113.
    //
    //    HARD-CODED at 16 (= K=128 / 8). MAX_K=128 in validate(); see
    //    iter 1.5 lesson on compile-time loop bounds.
    // ===================================================================
    {
        const uint sg_bt_origin = sgid * 8u;     // BT row this sg owns
        const device bfloat *q_row_base =
            q + q_chunk_base + sg_bt_origin * qk_t_stride;
        const device bfloat *h_base = h + h_chunk_base + v_off * h_v_stride;
        const device bfloat *k_row_base = k + k_chunk_base;

        const uint num_k_tiles = 16u;
        for (uint kt = 0; kt < num_k_tiles; ++kt) {
            const uint k_origin = kt * 8u;

            simdgroup_barrier(mem_flags::mem_none);

            // Load 1 q-frag : BT rows [sg_bt_origin, +8), K cols [k_origin, +8).
            simdgroup_matrix<bfloat, 8, 8> q_frag;
            simdgroup_load(q_frag, q_row_base + k_origin,
                           qk_t_stride, 0, false);

            // 2A: load 4 h^T-frags — all 4 BV col-tiles.
            //   h layout is [BV-rows, K] row-major (K-innermost, row stride = K).
            //   h^T frag at (k_origin..+8, bv_origin..+8) reads
            //   h[bv_origin+j][k_origin+i] — base = h_base + bv_origin*K + k_origin
            //   with row stride K and transpose=true.
            simdgroup_matrix<bfloat, 8, 8> h_frag_T[4];
            for (uint j = 0; j < 4; ++j) {
                const uint bv_origin = j * 8u;
                const device bfloat *h_ptr =
                    h_base + bv_origin * h_v_stride + k_origin;
                simdgroup_load(h_frag_T[j], h_ptr, h_v_stride, 0, true);
            }

            // 2B: load 8 k^T-frags — all 8 BT col-tiles.
            //   k layout is [BT, K] row-major (K-innermost, row stride = qk_t_stride).
            //   k^T frag at (k_origin..+8, bt_col..+8) reads
            //   k[bt_col+j][k_origin+i] — base = k_row_base + bt_col*qk_t_stride
            //   + k_origin with row stride qk_t_stride and transpose=true.
            simdgroup_matrix<bfloat, 8, 8> k_frag_T[8];
            for (uint j = 0; j < 8; ++j) {
                const uint bt_col = j * 8u;
                const device bfloat *k_ptr =
                    k_row_base + bt_col * qk_t_stride + k_origin;
                simdgroup_load(k_frag_T[j], k_ptr, qk_t_stride, 0, true);
            }

            simdgroup_barrier(mem_flags::mem_none);

            // 2A: 4 MMAs — bo_acc[j] += q_frag @ h_frag_T[j].
            for (uint j = 0; j < 4; ++j) {
                simdgroup_multiply_accumulate(
                    bo_acc[j], q_frag, h_frag_T[j], bo_acc[j]);
            }
            // 2B: 8 MMAs — bA_acc[j] += q_frag @ k_frag_T[j].
            for (uint j = 0; j < 8; ++j) {
                simdgroup_multiply_accumulate(
                    bA_acc[j], q_frag, k_frag_T[j], bA_acc[j]);
            }
        }
    }

    // ===================================================================
    // 3. Apply gate to bo_acc + gate/mask/bf16-stage bA_acc.
    //    Per-thread element extract from MMA frags.
    //
    //    bo_acc[i, j] *= exp(g[t_start + i])
    //    bA_acc[i, j] *= exp(g[t_start + i] - g[t_start + j])
    //    bA_acc[i, j] = (i >= j) ? bA_acc[i, j] : 0    (FLA :122-125 `>=`)
    //    bA_stage[i, j] = bfloat(bA_acc[i, j])         (FLA :137)
    // ===================================================================
    const uint sg_bt_origin = sgid * 8u;
    const uint bt_local = sg_bt_origin + fm;       // BT row index this lane owns
    const float g_i = g[g_base + (t_start + bt_local) * g_t_stride];
    const float exp_g_i = metal::exp(g_i);

    // 3A: gate bo_acc in registers.
    for (uint j = 0; j < 4; ++j) {
        thread auto &elems = bo_acc[j].thread_elements();
        elems[0] = float(elems[0]) * exp_g_i;
        elems[1] = float(elems[1]) * exp_g_i;
    }

    // 3B: gate + mask + bf16-stage bA_acc.
    //   Each lane owns 2 elements per frag at (fm, fn, fn+1).
    //   Across 8 col-tiles: bt_col = j*8 + fn, j*8 + fn + 1 (16 total cells/lane).
    for (uint j = 0; j < 8; ++j) {
        thread auto &elems = bA_acc[j].thread_elements();
        const uint bt_col0 = j * 8u + fn;
        const uint bt_col1 = bt_col0 + 1u;
        const float g_j0 = g[g_base + (t_start + bt_col0) * g_t_stride];
        const float g_j1 = g[g_base + (t_start + bt_col1) * g_t_stride];
        float v0 = float(elems[0]) * metal::exp(g_i - g_j0);
        float v1 = float(elems[1]) * metal::exp(g_i - g_j1);
        // FLA :124 — INCLUSIVE causal mask (`>=`, NOT strict).
        v0 = (bt_local >= bt_col0) ? v0 : 0.0f;
        v1 = (bt_local >= bt_col1) ? v1 : 0.0f;
        // bf16 round-trip + publish to threadgroup mem.
        // bA_stage layout: [BT, BT] row-major (BT cols innermost, row stride = BT).
        bA_stage[bt_local * BT + bt_col0] = bfloat(v0);
        bA_stage[bt_local * BT + bt_col1] = bfloat(v1);
    }
    threadgroup_barrier(mem_flags::mem_threadgroup);

    // ===================================================================
    // 4. Section 2C: bo_acc += bA_stage @ v_new  (8 BT-tiles of 8).
    //    Each sg owns same BT row-tile [sg_bt_origin, +8) × 4 BV col-tiles.
    //    Per BT-tile: 1 bA_stage frag (BT-row sgid, 8 BT cols) + 4 v-frags
    //    (8 BT rows × 8 BV cols each, 4 col-tiles total) → 4 MMAs.
    //
    //    HARD-CODED at 8 (= BT=64 / 8). BT=64 enforced by validate().
    // ===================================================================
    {
        // bA_stage frag origin row = sg_bt_origin (same as q's BT row-tile).
        const threadgroup bfloat *bA_row_base = bA_stage + sg_bt_origin * BT;
        const device bfloat *v_base = v + v_chunk_base + v_off;

        const uint num_bt_tiles = 8u;
        for (uint bj = 0; bj < num_bt_tiles; ++bj) {
            const uint bt_origin = bj * 8u;

            simdgroup_barrier(mem_flags::mem_none);

            // Load 1 bA-frag : BT rows [sg_bt_origin, +8), BT cols
            // [bt_origin, +8). Row stride = BT.
            simdgroup_matrix<bfloat, 8, 8> bA_frag;
            simdgroup_load(bA_frag, bA_row_base + bt_origin, BT, 0, false);

            // Load 4 v-frags — 8 BT rows [bt_origin, +8) × 8 BV cols each.
            //   v_new[B, T, H, V] row-major V-innermost; row stride = v_t_stride.
            simdgroup_matrix<bfloat, 8, 8> v_frag[4];
            const device bfloat *v_row_base =
                v_base + bt_origin * v_t_stride;
            for (uint j = 0; j < 4; ++j) {
                const uint bv_origin = j * 8u;
                simdgroup_load(v_frag[j], v_row_base + bv_origin,
                               v_t_stride, 0, false);
            }

            simdgroup_barrier(mem_flags::mem_none);

            // 4 MMAs: bo_acc[j] += bA_frag @ v_frag[j].
            for (uint j = 0; j < 4; ++j) {
                simdgroup_multiply_accumulate(
                    bo_acc[j], bA_frag, v_frag[j], bo_acc[j]);
            }
        }
    }

    // ===================================================================
    // 5. Final scale + bf16 store of bo_acc to o.
    //    Per FLA :137: o = (bo_acc_pre_2C + dot_2C) * scale, applied as
    //    a single multiply on the post-2C accumulator (mathematically
    //    equivalent to the spec's two `* scale` terms — distributive law
    //    over the sum).
    //
    //    Each lane owns 2 elements per frag at (fm, fn, fn+1) across 4
    //    BV col-tiles = 8 cells.
    // ===================================================================
    {
        const uint o_row_base = o_chunk_base + bt_local * v_t_stride + v_off;
        for (uint j = 0; j < 4; ++j) {
            thread auto &elems = bo_acc[j].thread_elements();
            const uint bv0 = j * 8u + fn;
            const uint bv1 = bv0 + 1u;
            const float val0 = float(elems[0]) * scale;
            const float val1 = float(elems[1]) * scale;
            o[o_row_base + bv0] = bfloat(val0);
            o[o_row_base + bv1] = bfloat(val1);
        }
    }
}