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
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
// Derived from candle-metal-kernels (Apache-2.0) kernel_mul_mv_id template
// and mlx-native's quantized_matmul_ggml kernels.
// Combines GGML block format dequantization with expert-indexed dispatch.
//
// Original sources:
//   candle-metal-kernels/src/metal_src/quantized.metal:7544-7618 (kernel_mul_mv_id)
//   candle-metal-kernels/src/metal_src/quantized.metal:90-293    (Q4_0, Q8_0, Q6_K kernels)
//   mlx-native/src/shaders/quantized_matmul_ggml.metal           (GGML block dequant)
//
// This kernel performs expert-indexed (MoE) quantized matrix-vector multiply:
//   For each (token, slot) pair:
//     expert_id = ids[token * top_k + slot]
//     output[token*top_k + slot, :] = matmul(input[token, :], weight[expert_id])
//
// The key insight: instead of dispatching one kernel per expert, we dispatch once
// for ALL (token, slot) pairs. The kernel uses the ids buffer to route each output
// row to the correct expert's weight slice.
//
// Copyright the candle Authors and llama.cpp Authors.
// See LICENSE-APACHE-candle in this directory.

#include <metal_stdlib>
using namespace metal;

// ---- Constants (must match quantized_matmul_ggml.metal) ----

#define QK4_0 32
#define QK8_0 32
#define QK_K  256

#define N_DST       4
#define N_SIMDGROUP 2
#define N_SIMDWIDTH 32

// ---- Parameters for expert-indexed GGML matmul ----

struct GgmlMatvecIdParams {
    int64_t ne00;           // K: input dimension
    int64_t ne01;           // N: output dimension per expert
    int64_t ne02;           // 1 (unused, kept for struct compat)
    int64_t ne10;           // K: input dimension (redundant, == ne00)
    int64_t ne12;           // 1 (unused)
    int64_t ne0;            // N: output stride
    int64_t ne1;            // total output rows = n_tokens * top_k
    uint    r2;             // 1
    uint    r3;             // 1
    uint    top_k;          // experts per token
    uint    n_tokens;       // number of input tokens
    int64_t expert_stride;  // bytes between expert weight slices
};

// K_SCALE_SIZE: bytes used for scales+mins in Q4_K and Q5_K super-blocks.
#define K_SCALE_SIZE 12

// ---- GGML block struct definitions (byte-for-byte with GGUF) ----

typedef struct {
    half    d;
    uint8_t qs[QK4_0 / 2];
} block_q4_0;

typedef struct {
    half   d;
    int8_t qs[QK8_0];
} block_q8_0;

// Q5_K: 256 values per block, 176 bytes per block.
// Layout: [half d][half dmin][uint8_t scales[12]][uint8_t qh[32]][uint8_t qs[128]]
typedef struct {
    half    d;                    // super-block scale for quantized scales
    half    dmin;                 // super-block scale for quantized mins
    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
    uint8_t qh[QK_K/8];           // quants, high bit   (32 bytes)
    uint8_t qs[QK_K/2];           // quants, low 4 bits (128 bytes)
} block_q5_K;

typedef struct {
    uint8_t ql[QK_K/2];
    uint8_t qh[QK_K/4];
    int8_t  scales[QK_K/16];
    half    d;
} block_q6_K;

// ---- Q4_0 dot product helper (identical to quantized_matmul_ggml.metal) ----

inline float block_q4_0_dot_y(
    device const block_q4_0 * qb,
    float sumy,
    thread float * yl,
    int il
) {
    float d = qb->d;
    float2 acc = 0.f;
    device const uint16_t * qs = ((device const uint16_t *)qb + 1 + il/2);
    for (int i = 0; i < 8; i += 2) {
        acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
                + yl[i + 1] * (qs[i / 2] & 0x0F00);
        acc[1] += yl[i + 8] * (qs[i / 2] & 0x00F0)
                + yl[i + 9] * (qs[i / 2] & 0xF000);
    }
    return d * (sumy * -8.f + acc[0] + acc[1]);
}

// ====================================================================
// Q4_0 expert-indexed mat-vec kernel
// ====================================================================
//
// For each output row r (where r = token*top_k + slot):
//   expert_id = ids[r]   (ids is [n_tokens * top_k] flat, pre-expanded)
//   src0_cur  = src0 + expert_id * expert_stride
//   output[r] = matmul(src1[token], src0_cur)
//
// Dispatch geometry: threadgroups=(ceil(N/8), n_tokens*top_k, 1), tg=(8,8,1)
//
// Routing index is in dim Y, NOT dim Z (despite llama.cpp's mul_mv_id grid
// using ne123 in z at ggml-metal-ops.cpp:2452).  Tested 2026-04-26 on M5 Max
// dwq46 64-token decode: switching kernel to read tgpig.z + dispatcher
// MTLSize::new(N/8, 1, m) regressed throughput from 114 t/s to 90.9 t/s
// (-19%).  Apple GPU's threadgroup scheduler distributes this dispatch
// shape better via Y than Z — 7th confirmed static-evidence kernel
// hypothesis falsified per `project_metal_compiler_auto_optimizes_static_levers.md`.

kernel void kernel_mul_mv_id_q4_0_f32(
    device const  char  * src0   [[buffer(0)]],
    device const float  * src1   [[buffer(1)]],
    device       float  * dst    [[buffer(2)]],
    device const  uint  * ids    [[buffer(3)]],
    constant GgmlMatvecIdParams & p [[buffer(4)]],
    uint3 tgpig [[threadgroup_position_in_grid]],
    uint  tiisg [[thread_index_in_simdgroup]],
    uint  sgitg [[simdgroup_index_in_threadgroup]]
) {
    const int nr  = N_DST;
    const int nsg = N_SIMDGROUP;
    const int nw  = N_SIMDWIDTH;

    const int nb = p.ne00 / QK4_0;
    const int r0 = tgpig.x;
    const int output_row = tgpig.y;  // flat index into output: token*top_k + slot

    // Bounds check
    if (output_row >= (int)p.ne1) return;

    // Determine which token this output row belongs to and which expert
    const uint token_idx = output_row / p.top_k;
    const uint expert_id = ids[output_row];

    const int first_row = (r0 * nsg + sgitg) * nr;

    // Point to the expert's weight slice
    device const block_q4_0 * x = (device const block_q4_0 *)((device const char *)src0 + expert_id * p.expert_stride) + first_row * nb;

    // Point to the input row for this token
    device const float * y = src1 + token_idx * p.ne10;

    float yl[16];
    float sumf[nr] = {0.f};

    const int ix = tiisg / 2;
    const int il = (tiisg % 2) * 8;

    device const float * yb = y + ix * QK4_0 + il;

    for (int ib = ix; ib < nb; ib += nw/2) {
        float sumy = 0;
        for (int i = 0; i < 8; i += 2) {
            sumy += yb[i] + yb[i+1];
            yl[i+0] = yb[i+0];
            yl[i+1] = yb[i+1] / 256.f;
            sumy += yb[i+16] + yb[i+17];
            yl[i+8] = yb[i+16] / 16.f;
            yl[i+9] = yb[i+17] / 4096.f;
        }

        for (int row = 0; row < nr; row++) {
            sumf[row] += block_q4_0_dot_y(x + ib + row*nb, sumy, yl, il);
        }

        yb += QK4_0 * 16;
    }

    for (int row = 0; row < nr; ++row) {
        const float tot = simd_sum(sumf[row]);
        if (tiisg == 0 && first_row + row < p.ne01) {
            dst[output_row * p.ne0 + first_row + row] = tot;
        }
    }
}

// ====================================================================
// Q4_0 fused-SwiGLU expert-indexed mat-vec kernel
// ====================================================================
//
// Computes: dst[r][n] = sum_k(dequant(W_q4_0[expert_id][n][k])
//                              * (silu(gate[r][k]) * up[r][k]))
//
// where r = token*top_k + slot, expert_id = ids[r].
//
// Replaces the dispatch sequence:
//   silu_mul_f32(gate, up → h_all)        # 1 dispatch + memory_barrier
//   kernel_mul_mv_id_q4_0_f32(W, h_all → dst)  # 1 dispatch
//
// with a single dispatch that reads gate + up directly and computes
// swiglu inline before the dot product.  Closes ~5-10µs/layer × 40
// layers ≈ 0.3-0.4ms/token of CPU dispatch overhead in the dwq46
// decode hot path (ADR-012 §Optimize / Task #15).
//
// Buffer layout:
//   buffer(0): src0   - Q4_0 packed weight, [n_experts, N, K/QK4_0] blocks
//   buffer(1): gate   - f32 [n_tokens*top_k, K]
//   buffer(2): up     - f32 [n_tokens*top_k, K]
//   buffer(3): dst    - f32 [n_tokens*top_k, N]
//   buffer(4): ids    - u32 [n_tokens*top_k]
//   buffer(5): params - GgmlMatvecIdParams
//
// Dispatch geometry: identical to kernel_mul_mv_id_q4_0_f32 —
// threadgroups=(ceil(N/8), n_tokens*top_k, 1), tg=(8, 8, 1).
kernel void kernel_mul_mv_id_q4_0_f32_swiglu(
    device const  char  * src0   [[buffer(0)]],
    device const float  * gate   [[buffer(1)]],
    device const float  * up     [[buffer(2)]],
    device       float  * dst    [[buffer(3)]],
    device const  uint  * ids    [[buffer(4)]],
    constant GgmlMatvecIdParams & p [[buffer(5)]],
    uint3 tgpig [[threadgroup_position_in_grid]],
    uint  tiisg [[thread_index_in_simdgroup]],
    uint  sgitg [[simdgroup_index_in_threadgroup]]
) {
    const int nr  = N_DST;
    const int nsg = N_SIMDGROUP;
    const int nw  = N_SIMDWIDTH;

    const int nb = p.ne00 / QK4_0;
    const int r0 = tgpig.x;
    const int output_row = tgpig.y;  // flat index into output: token*top_k + slot

    if (output_row >= (int)p.ne1) return;

    // For expert_down in the decode-time MoE pipeline, the input row IS
    // the output row index (one h_all row per (token, expert_slot) pair),
    // not token_idx.  Each (token, expert_slot) has its own gate/up vectors.
    const uint expert_id = ids[output_row];
    const uint input_row = output_row;  // gate/up are pre-routed per (token, slot).

    const int first_row = (r0 * nsg + sgitg) * nr;

    // Expert's weight slice.
    device const block_q4_0 * x = (device const block_q4_0 *)((device const char *)src0 + expert_id * p.expert_stride) + first_row * nb;

    // Per-row gate and up vectors.
    device const float * gate_y = gate + input_row * p.ne10;
    device const float * up_y   = up   + input_row * p.ne10;

    float yl[16];
    float sumf[nr] = {0.f};

    const int ix = tiisg / 2;
    const int il = (tiisg % 2) * 8;

    device const float * gb = gate_y + ix * QK4_0 + il;
    device const float * ub = up_y   + ix * QK4_0 + il;

    for (int ib = ix; ib < nb; ib += nw/2) {
        float sumy = 0;
        // Compute swiglu = silu(gate) * up = gate * sigmoid(gate) * up
        // for each of the 16 active elements per simdthread, reusing the
        // same yl[] / sumy aggregation as the unfused kernel.
        for (int i = 0; i < 8; i += 2) {
            // Lane block 0 (i=0, i+1).
            float g0 = gb[i+0];
            float g1 = gb[i+1];
            float u0 = ub[i+0];
            float u1 = ub[i+1];
            float s0 = (g0 / (1.0f + metal::exp(-g0))) * u0;
            float s1 = (g1 / (1.0f + metal::exp(-g1))) * u1;
            sumy += s0 + s1;
            yl[i+0] = s0;
            yl[i+1] = s1 / 256.f;

            // Lane block 1 (i+16, i+17).
            float g2 = gb[i+16];
            float g3 = gb[i+17];
            float u2 = ub[i+16];
            float u3 = ub[i+17];
            float s2 = (g2 / (1.0f + metal::exp(-g2))) * u2;
            float s3 = (g3 / (1.0f + metal::exp(-g3))) * u3;
            sumy += s2 + s3;
            yl[i+8] = s2 / 16.f;
            yl[i+9] = s3 / 4096.f;
        }

        for (int row = 0; row < nr; row++) {
            sumf[row] += block_q4_0_dot_y(x + ib + row*nb, sumy, yl, il);
        }

        gb += QK4_0 * 16;
        ub += QK4_0 * 16;
    }

    for (int row = 0; row < nr; ++row) {
        const float tot = simd_sum(sumf[row]);
        if (tiisg == 0 && first_row + row < p.ne01) {
            dst[output_row * p.ne0 + first_row + row] = tot;
        }
    }
}

// ====================================================================
// Q8_0 expert-indexed mat-vec kernel
// ====================================================================

#define NB_Q8_0 8

kernel void kernel_mul_mv_id_q8_0_f32(
    device const  char  * src0   [[buffer(0)]],
    device const float  * src1   [[buffer(1)]],
    device       float  * dst    [[buffer(2)]],
    device const  uint  * ids    [[buffer(3)]],
    constant GgmlMatvecIdParams & p [[buffer(4)]],
    uint3 tgpig [[threadgroup_position_in_grid]],
    uint  tiisg [[thread_index_in_simdgroup]],
    uint  sgitg [[simdgroup_index_in_threadgroup]]
) {
    const int nr  = N_DST;
    const int nsg = N_SIMDGROUP;
    const int nw  = N_SIMDWIDTH;

    const int nb = p.ne00 / QK8_0;
    const int r0 = tgpig.x;
    const int output_row = tgpig.y;

    if (output_row >= (int)p.ne1) return;

    const uint token_idx = output_row / p.top_k;
    const uint expert_id = ids[output_row];

    const int first_row = (r0 * nsg + sgitg) * nr;

    device const block_q8_0 * x = (device const block_q8_0 *)((device const char *)src0 + expert_id * p.expert_stride) + first_row * nb;
    device const float      * y = src1 + token_idx * p.ne10;

    float yl[NB_Q8_0];
    float sumf[nr] = {0.f};

    const int ix = tiisg / 4;
    const int il = tiisg % 4;

    device const float * yb = y + ix * QK8_0 + NB_Q8_0 * il;

    for (int ib = ix; ib < nb; ib += nw/4) {
        for (int i = 0; i < NB_Q8_0; ++i) {
            yl[i] = yb[i];
        }

        for (int row = 0; row < nr; row++) {
            device const int8_t * qs = x[ib + row*nb].qs + NB_Q8_0 * il;
            float sumq = 0.f;
            for (int iq = 0; iq < NB_Q8_0; ++iq) {
                sumq += qs[iq] * yl[iq];
            }
            sumf[row] += sumq * x[ib + row*nb].d;
        }

        yb += NB_Q8_0 * nw;
    }

    for (int row = 0; row < nr; ++row) {
        const float tot = simd_sum(sumf[row]);
        if (tiisg == 0 && first_row + row < p.ne01) {
            dst[output_row * p.ne0 + first_row + row] = tot;
        }
    }
}

// ====================================================================
// Q5_K expert-indexed mat-vec kernel
// ====================================================================
//
// Dispatch geometry (same as Q6_K): threadgroups = (ceil(N/2), n_tokens*top_k, 1)
//   tgpig.x = weight-row-pair index        (two rows: 2*r0 + sgitg)
//   tgpig.y = flat output row              (token*top_k + slot)
//   sgitg   = selects which of the two rows this simdgroup processes
//
// Ported from candle-metal-kernels kernel_mul_mv_q5_K_f32_impl with
// the expert-routing indirection from the Q6_K _id kernel above.
// Copyright the candle Authors (Apache-2.0) and llama.cpp Authors (MIT).

kernel void kernel_mul_mv_id_q5_K_f32(
    device const  char  * src0   [[buffer(0)]],
    device const float  * src1   [[buffer(1)]],
    device       float  * dst    [[buffer(2)]],
    device const  uint  * ids    [[buffer(3)]],
    constant GgmlMatvecIdParams & p [[buffer(4)]],
    uint3 tgpig [[threadgroup_position_in_grid]],
    uint  tiisg [[thread_index_in_simdgroup]],
    uint  sgitg [[simdgroup_index_in_threadgroup]]
) {
    const int nb = p.ne00 / QK_K;

    const int64_t r0          = tgpig.x;
    const int     output_row  = tgpig.y;  // flat: token*top_k + slot

    if (output_row >= (int)p.ne1) return;

    const uint token_idx = output_row / p.top_k;
    const uint expert_id = ids[output_row];

    // Each threadgroup covers weight-row pair (2*r0, 2*r0+1);
    // sgitg selects which row this simdgroup computes.
    const int row = 2 * (int)r0 + (int)sgitg;

    // Point to the expert's weight slice and the token's input row.
    device const block_q5_K * x  = (device const block_q5_K *)((device const char *)src0 + expert_id * p.expert_stride) + row * nb;
    device const float      * yy = src1 + token_idx * p.ne10;

    float sumf = 0.f;

    const uint16_t kmask1 = 0x3f3f;
    const uint16_t kmask2 = 0x0f0f;
    const uint16_t kmask3 = 0xc0c0;

    const int tid = tiisg / 4;
    const int ix  = tiisg % 4;
    const int iq  = tid / 4;
    const int ir  = tid % 4;
    const int n   = 8;

    const int l0       = n * ir;
    const int q_offset = 32 * iq + l0;
    const int y_offset = 64 * iq + l0;

    const uint8_t hm1 = 1u << (2 * iq);
    const uint8_t hm2 = hm1 << 1;
    const uint8_t hm3 = hm1 << 4;
    const uint8_t hm4 = hm2 << 4;

    uint16_t sc16[4];
    thread const uint8_t * sc8 = (thread const uint8_t *)sc16;

    device const float * y1 = yy + ix * QK_K + y_offset;

    for (int i = ix; i < nb; i += 4) {
        device const uint8_t  * q1 = x[i].qs + q_offset;
        device const uint8_t  * q2 = q1 + 64;
        device const uint8_t  * qh = x[i].qh + l0;
        device const half     * dh = &x[i].d;
        // scales array is uint8_t[12]; cast to uint16_t[6] for the
        // sc16 decoding identical to the reference candle kernel.
        device const uint16_t * a  = (device const uint16_t *)x[i].scales + iq;

        device const float * y2 = y1 + 128;
        float yl[16], yh[16];
        float4 sumy = {0.f, 0.f, 0.f, 0.f};
        for (int l = 0; l < n; ++l) {
            yl[l+0] = y1[l +  0]; sumy[0] += yl[l+0];
            yl[l+8] = y1[l + 32]; sumy[1] += yl[l+8];
            yh[l+0] = y2[l +  0]; sumy[2] += yh[l+0];
            yh[l+8] = y2[l + 32]; sumy[3] += yh[l+8];
        }

        sc16[0] = a[0] & kmask1;
        sc16[1] = a[2] & kmask1;
        sc16[2] = ((a[4] >> 0) & kmask2) | ((a[0] & kmask3) >> 2);
        sc16[3] = ((a[4] >> 4) & kmask2) | ((a[2] & kmask3) >> 2);

        float4 acc1 = {0.f, 0.f, 0.f, 0.f};
        float4 acc2 = {0.f, 0.f, 0.f, 0.f};
        for (int l = 0; l < n; ++l) {
            uint8_t h = qh[l];
            acc1[0] += yl[l+0] * (float)(q1[l] & 0x0F);
            acc1[1] += yl[l+8] * (float)(q1[l] & 0xF0);
            acc1[2] += yh[l+0] * (float)(q2[l] & 0x0F);
            acc1[3] += yh[l+8] * (float)(q2[l] & 0xF0);
            acc2[0] += (h & hm1) ? yl[l+0] : 0.f;
            acc2[1] += (h & hm2) ? yl[l+8] : 0.f;
            acc2[2] += (h & hm3) ? yh[l+0] : 0.f;
            acc2[3] += (h & hm4) ? yh[l+8] : 0.f;
        }

        const float dall = (float)dh[0];
        const float dmin = (float)dh[1];
        sumf += dall * ((float)sc8[0] * (acc1[0]        + 16.f * acc2[0]) +
                        (float)sc8[1] * (acc1[1] / 16.f + 16.f * acc2[1]) +
                        (float)sc8[4] * (acc1[2]        + 16.f * acc2[2]) +
                        (float)sc8[5] * (acc1[3] / 16.f + 16.f * acc2[3])) -
               dmin * (sumy[0] * (float)sc8[2] + sumy[1] * (float)sc8[3] +
                       sumy[2] * (float)sc8[6] + sumy[3] * (float)sc8[7]);

        y1 += 4 * QK_K;
    }

    const float tot = simd_sum(sumf);
    if (tiisg == 0 && row < (int)p.ne01) {
        dst[output_row * p.ne0 + row] = tot;
    }
}

// ====================================================================
// Q6_K expert-indexed mat-vec kernel
// ====================================================================

kernel void kernel_mul_mv_id_q6_K_f32(
    device const  char  * src0   [[buffer(0)]],
    device const float  * src1   [[buffer(1)]],
    device       float  * dst    [[buffer(2)]],
    device const  uint  * ids    [[buffer(3)]],
    constant GgmlMatvecIdParams & p [[buffer(4)]],
    uint3 tgpig [[threadgroup_position_in_grid]],
    uint  tiisg [[thread_index_in_simdgroup]],
    uint  sgitg [[simdgroup_index_in_threadgroup]]
) {
    const uint8_t kmask1 = 0x03;
    const uint8_t kmask2 = 0x0C;
    const uint8_t kmask3 = 0x30;
    const uint8_t kmask4 = 0xC0;

    const int nb = p.ne00 / QK_K;

    const int64_t r0 = tgpig.x;
    const int output_row_base = tgpig.y;

    if (output_row_base >= (int)p.ne1) return;

    // For Q6_K, each threadgroup handles 2 rows (one per SIMD group).
    // But we need to handle the _id dimension: output_row_base is the
    // flat output-row index. We process row = 2*r0 + sgitg within the
    // weight matrix, and the output goes to output_row_base.
    //
    // Actually, the Q6_K dispatch geometry is different from Q4_0/Q8_0.
    // In the non-id version: threadgroups = (ceil(N/2), M, 1)
    // Each threadgroup handles 2 adjacent weight rows (r0*2 + sgitg).
    //
    // For the _id version: threadgroups = (ceil(N/2), n_tokens*top_k, 1)
    // tgpig.y = output_row (flat: token*top_k + slot)
    // tgpig.x = weight row pair index
    // sgitg selects within the pair: row = 2*r0 + sgitg

    const uint token_idx = output_row_base / p.top_k;
    const uint expert_id = ids[output_row_base];

    const int row = 2 * r0 + sgitg;

    device const block_q6_K * x  = (device const block_q6_K *)((device const char *)src0 + expert_id * p.expert_stride) + row * nb;
    device const float      * yy = src1 + token_idx * p.ne10;

    float sumf = 0;

    const int tid  = tiisg / 2;
    const int ix   = tiisg % 2;
    const int ip   = tid / 8;
    const int il   = tid % 8;
    const int n    = 4;
    const int l0   = n * il;
    const int is   = 8*ip + l0/16;

    const int y_offset   = 128*ip + l0;
    const int q_offset_l = 64*ip + l0;
    const int q_offset_h = 32*ip + l0;

    for (int i = ix; i < nb; i += 2) {
        device const uint8_t * q1 = x[i].ql + q_offset_l;
        device const uint8_t * q2 = q1 + 32;
        device const uint8_t * qh = x[i].qh + q_offset_h;
        device const int8_t  * sc = x[i].scales + is;

        device const float * y = yy + i * QK_K + y_offset;

        const float dall = x[i].d;

        float4 sums = {0.f, 0.f, 0.f, 0.f};
        for (int l = 0; l < n; ++l) {
            sums[0] += y[l+ 0] * ((int8_t)((q1[l] & 0xF) | ((qh[l] & kmask1) << 4)) - 32);
            sums[1] += y[l+32] * ((int8_t)((q2[l] & 0xF) | ((qh[l] & kmask2) << 2)) - 32);
            sums[2] += y[l+64] * ((int8_t)((q1[l]  >> 4) | ((qh[l] & kmask3) << 0)) - 32);
            sums[3] += y[l+96] * ((int8_t)((q2[l]  >> 4) | ((qh[l] & kmask4) >> 2)) - 32);
        }

        sumf += dall * (sums[0] * sc[0] + sums[1] * sc[2] + sums[2] * sc[4] + sums[3] * sc[6]);
    }

    const float tot = simd_sum(sumf);
    if (tiisg == 0 && row < (int)p.ne01) {
        dst[output_row_base * p.ne0 + row] = tot;
    }
}