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
// quantized_matmul_id_mm.metal — MoE-routed GGML-quantized matrix-matrix kernels.
//
// Ports llama.cpp's `kernel_mul_mm_id_<qtype>_f32`
// (ggml/src/ggml-metal/ggml-metal.metal:9650) and its preprocessing helper
// `kernel_mul_mm_id_map0_ne20_<N>` (same file:9584) to mlx-native.
//
// Why mm_id?  The MoE variant of quantized_matmul needs to run the same
// expert weight tile against many tokens that routed to that expert.  The
// existing `kernel_mul_mv_id_<qtype>_f32` re-reads each expert's weight
// blocks once per routed (token, slot) pair.  The mm_id variant stages a
// 64x32 tile of the expert's weights into threadgroup shared memory once
// and reuses it across a 32-row block of the expert's routed tokens —
// the same win as the dense mm kernel, but per-expert.
//
// Two-stage dispatch (matches llama.cpp):
//
//   1. `kernel_mul_mm_id_map0_ne20_8` — scans the (token, slot) -> expert
//       id table, for each expert builds a contiguous list of the packed
//       `(token_idx * ne20 + slot_idx)` values that routed to it, plus a
//       per-expert count.  Scratch buffers: `hids` `[n_experts, n_tokens]`
//       (int32, row-major), `htpe` `[n_experts]` (uint32).
//
//   2. `kernel_mul_mm_id_<qtype>_f32` — dispatched with one Z-per-expert.
//       Each threadgroup owns one (N-tile, M-tile, expert_id) and short-
//       circuits when the expert's routed-token count is below its M-tile
//       start.  Inside a tile, ALL 32 M-rows belong to the same expert,
//       so the 64x32 weight tile loaded into shmem is valid for every row.
//
// The preprocessing step is the key to making mm profitable at MoE:
// without it, 32 consecutive output rows in a tile could route to 32
// different experts, defeating weight reuse.
//
// Port rules identical to the dense mm port (see quantized_matmul_mm.metal
// for detail).  Both kernels are bit-compatible with the llama.cpp source;
// output tolerance-level matches with the existing mv_id kernel.
//
// Portions derived from llama.cpp (MIT).  Copyright the llama.cpp Authors.

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

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

#define QK_NL 16

// ---- Params for map0 ----
//
// Maps llama.cpp's `ggml_metal_kargs_mul_mm_id_map0`
// (ggml-metal-impl.h:483).  We omit `ne02` (== n_experts, used for a
// threadgroup-size assert upstream — we pass it host-side and dispatch
// exactly `n_experts` threads per threadgroup to mirror the upstream
// `ntg = n_experts` launch).

struct GgmlMatmulIdMm_Map0Params {
    int32_t  ne10;    // unused, kept for struct symmetry
    int32_t  ne11;    // n_expert_used (bcast), always == ne20 in our case
    uint64_t nb11;    // unused
    uint64_t nb12;    // unused
    int32_t  ne21;    // n_tokens
    int32_t  ne20;    // n_expert_used (== top_k)
    uint64_t nb21;    // bytes per token in the source ids table
                      //  (= ne20 * sizeof(int32_t) for our layout)
};

// ---- Params for mul_mm_id ----
//
// Maps llama.cpp's `ggml_metal_kargs_mul_mm_id`
// (ggml-metal-impl.h:494).

struct GgmlMatmulIdMm_MmParams {
    int32_t  ne00;   // K
    int32_t  ne02;   // n_experts
    uint64_t nb01;   // bytes per weight row (within one expert)
    uint64_t nb02;   // bytes per expert weight slab (== nb01 * ne01)
    uint64_t nb03;   // unused
    int32_t  ne11;   // n_expert_used (bcast, == ne20)
    uint64_t nb10;   // = sizeof(float) = 4
    uint64_t nb11;   // bytes per input row = K * sizeof(float)
    uint64_t nb12;   // bytes per input batch (n_tokens * nb11)
    uint64_t nb13;   // unused
    int32_t  ne20;   // n_expert_used (== top_k)
    int32_t  ne21;   // n_tokens
    int32_t  ne0;    // N (per-expert output rows)
    int32_t  ne1;    // batch (always == ne21 * ne20 in our layout)
    int16_t  r2;     // 1
    int16_t  r3;     // 1
    int16_t  _pad0;
    int16_t  _pad1;
};

// ---- block_q struct layouts (byte-for-byte 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;

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;

// ---- Dequantize helpers (identical to quantized_matmul_mm.metal) ----

template <typename type4x4>
void dequantize_q4_0(device const block_q4_0 * xb, short il, thread type4x4 & reg) {
    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
    const float d1 = il ? (xb->d / 16.h) : xb->d;
    const float d2 = d1 / 256.f;
    const float md = -8.h * xb->d;
    const ushort mask0 = il ? 0x00F0 : 0x000F;
    const ushort mask1 = mask0 << 8;

    float4x4 reg_f;
    for (int i = 0; i < 8; i++) {
        reg_f[i/2][2*(i%2) + 0] = d1 * (qs[i] & mask0) + md;
        reg_f[i/2][2*(i%2) + 1] = d2 * (qs[i] & mask1) + md;
    }
    reg = (type4x4) reg_f;
}

template <typename type4x4>
void dequantize_q8_0(device const block_q8_0 * xb, short il, thread type4x4 & reg) {
    device const int8_t * qs = ((device const int8_t *)xb->qs);
    const float d = xb->d;

    float4x4 reg_f;
    for (int i = 0; i < 16; i++) {
        reg_f[i/4][i%4] = (qs[i + 16*il] * d);
    }
    reg = (type4x4) reg_f;
}

template <typename type4x4>
void dequantize_q6_K(device const block_q6_K * xb, short il, thread type4x4 & reg) {
    const half d_all = xb->d;
    device const uint16_t * ql = (device const uint16_t *)xb->ql;
    device const uint16_t * qh = (device const uint16_t *)xb->qh;
    device const int8_t * scales = (device const int8_t *)xb->scales;

    ql = ql + 32*(il/8) + 16*((il/2)&1) + 8*(il&1);
    qh = qh + 16*(il/8) + 8*(il&1);
    float sc = scales[(il%2) + 2 * ((il/2))];
    il = (il/2) & 3;

    const uint32_t kmask1 = il>1 ? (il>2 ? 0xC0C0C0C0 : 0x30303030) : (il>0 ? 0x0C0C0C0C : 0x03030303);
    const uint32_t kmask2 = il>1 ? 0xF0F0F0F0                       : 0x0F0F0F0F;
    const float ml = d_all * sc * 32.f;
    const float dl0 = d_all * sc;
    const float dl1 = dl0 / 256.f;
    const float dl2 = dl0 / (256.f * 256.f);
    const float dl3 = dl0 / (256.f * 256.f * 256.f);
    const uint8_t shr_h = il>2 ? 2 : 0;
    const uint8_t shl_h = il>1 ? 0 : (il>0 ? 2 : 4);
    const uint8_t shr_l = il>1 ? 4 : 0;

    float4x4 reg_f;
    for (int i = 0; i < 4; ++i) {
        const uint32_t  low = (ql[2*i] | (uint32_t)(ql[2*i+1] << 16)) & kmask2;
        const uint32_t high = (qh[2*i] | (uint32_t)(qh[2*i+1] << 16)) & kmask1;
        const uint32_t q = ((high << shl_h) >> shr_h) | (low >> shr_l);
        reg_f[i][0] = dl0 *  ((half)(q & 0xFF))      - ml;
        reg_f[i][1] = dl1 * ((float)(q & 0xFF00))    - ml;
        reg_f[i][2] = dl2 * ((float)(q & 0xFF0000))  - ml;
        reg_f[i][3] = dl3 * ((float)(q & 0xFF000000))- ml;
    }
    reg = (type4x4) reg_f;
}

// ====================================================================
// map0 — per-expert routed-token list builder
// ====================================================================
//
// Input:  src2 — per-token expert ids `[n_tokens, n_expert_used]` int32.
// Output: htpe — per-expert routed count  `[n_experts]` uint32.
// Output: hids — per-expert routed-token list `[n_experts, n_tokens]` int32.
//         Each slot holds `(token_idx * ne20 + slot_idx)` packed.
//
// Dispatch geometry (matches llama.cpp:9584):
//   threadgroups = (1, 1, 1)
//   threads_per_threadgroup = (n_experts, 1, 1)     (ntg == ne02)
//
// One thread per expert id (ide = tpitg).  Threads scan the token axis
// in blocks of `ntg` tokens, stage each token's `ne20` expert-ids into
// threadgroup shmem, then each thread scans that block looking for its
// own expert-id and appends entries into its slot in `hids`.

template<short ne20>
kernel void hf2q_mul_mm_id_map0_impl(
        constant GgmlMatmulIdMm_Map0Params & args [[buffer(0)]],
        device const char * src2 [[buffer(1)]],
        device       char * htpe [[buffer(2)]],
        device       char * hids [[buffer(3)]],
        threadgroup  char * shmem [[threadgroup(0)]],
        ushort tpitg[[thread_position_in_threadgroup]],
        ushort   ntg[[threads_per_threadgroup]]) {
    const short ide = tpitg;    // expert id owned by this thread

    uint32_t n_all = 0;
    device int32_t * ids_i32 = (device int32_t *) hids + ide * args.ne21;

    for (int i21 = 0; i21 < args.ne21; i21 += ntg) {
        if (i21 + tpitg < args.ne21) {
            device const int32_t * src2_i32 =
                (device const int32_t *) (src2 + (i21 + tpitg) * args.nb21);
            threadgroup uint16_t * sids = (threadgroup uint16_t *) shmem + tpitg * ne20;
            for (short i20 = 0; i20 < ne20; i20++) {
                sids[i20] = src2_i32[i20];
            }
        }
        threadgroup_barrier(mem_flags::mem_threadgroup);

        for (short t = 0; t < ntg; t++) {
            if (i21 + t >= args.ne21) break;
            threadgroup const uint16_t * sids = (threadgroup const uint16_t *) shmem + t * ne20;

            short sel = 0;
            for (short i20 = 0; i20 < ne20; i20++) {
                sel += (sids[i20] == ide) * (i20 + 1);
            }
            ids_i32[n_all] = (i21 + t) * ne20 + sel - 1;
            n_all += sel > 0;
        }
        threadgroup_barrier(mem_flags::mem_threadgroup);
    }

    device uint32_t * tpe_u32 = (device uint32_t *) (htpe);
    tpe_u32[ide] = n_all;
}

// Gemma 4 uses top_k = 8 for the MoE gate_up call and top_k = 1 for the
// MoE down call (where the gate_up output is flattened to
// [seq_len*top_k] rows, each one routed to a single expert via the
// expert-id table).  Both must hit the mm_id path — without ne20_1, the
// down call falls back to mv_id and re-reads each expert's weights once
// per (seq_len * top_k) row, defeating the whole point of mm.
template [[host_name("kernel_mul_mm_id_map0_ne20_1")]]
kernel void hf2q_mul_mm_id_map0_impl<1>(
    constant GgmlMatmulIdMm_Map0Params &,
    device const char *, device char *, device char *,
    threadgroup char *, ushort, ushort);

template [[host_name("kernel_mul_mm_id_map0_ne20_8")]]
kernel void hf2q_mul_mm_id_map0_impl<8>(
    constant GgmlMatmulIdMm_Map0Params &,
    device const char *, device char *, device char *,
    threadgroup char *, ushort, ushort);

// ====================================================================
// mul_mm_id — MoE matmul using the map0-produced expert token lists
// ====================================================================
//
// Dispatch geometry (matches llama.cpp:2389):
//   threadgroups = (ceil(neh1/32), ceil(N/64), n_experts)
//   threads_per_threadgroup = (128, 1, 1)
//
// where `neh1 = htpe[expert]` (count of tokens routed to this expert).
//
// Each threadgroup owns:
//   * im       = tgpig.z     — expert index
//   * r0-block = tgpig.y     — output-N tile index
//   * r1-block = tgpig.x     — input-M (per-expert) tile index
//
// The kernel short-circuits inactive (expert, r1) combinations via the
// `r1 >= neh1` check — some threadgroups may exit early but the grid
// launch is uniform.
//
// Per-iteration of the K loop: stage a 64x32 A-tile (from the expert's
// weight slab) + 32x32 B-tile (from input rows selected via hids) into
// threadgroup memory, then run 8 simdgroup MMA ops per simdgroup.  Four
// simdgroups cooperate on one 64x32 output tile.
//
// Write-back staging ALWAYS goes through shmem because consecutive M-rows
// in this tile belong to the same expert but may correspond to different
// (token, slot) output rows (determined by hids[im * ne21 + r1 + j]).

template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
kernel void hf2q_mul_mm_id_impl(
        constant GgmlMatmulIdMm_MmParams & args [[buffer(0)]],
        device const char * src0 [[buffer(1)]],
        device const char * src1 [[buffer(2)]],
        device const char * htpe [[buffer(3)]],
        device const char * hids [[buffer(4)]],
        device       char * dst  [[buffer(5)]],
        threadgroup  char * shmem [[threadgroup(0)]],
        uint3  tgpig[[threadgroup_position_in_grid]],
        ushort tiitg[[thread_index_in_threadgroup]],
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {

    threadgroup half  * sa = (threadgroup half  *)(shmem);
    threadgroup float * sb = (threadgroup float *)(shmem + 4096);

    constexpr int NR0 = 64;
    constexpr int NR1 = 32;
    constexpr int NK  = 32;
    constexpr int NL0 = NK/16;  // 2
    constexpr int NL1 = NK/8;   // 4

    const int im = tgpig.z;
    const int r0 = tgpig.y * NR0;
    const int r1 = tgpig.x * NR1;

    device const uint32_t * tpe_u32 = (device const uint32_t *) (htpe);
    device const int32_t  * ids_i32 = (device const int32_t  *) (hids);

    const int32_t neh1 = tpe_u32[im];

    // Early exit: this expert has fewer routed tokens than our tile's
    // M-base.  Whole threadgroup returns.
    if (r1 >= neh1) return;

    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
    const short nr1 = (    neh1 - r1 < NR1) ? (    neh1 - r1) : NR1;

    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1;
    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1;

    const short il0 = (tiitg % NL0);
    short il = il0;

    // (token_idx, slot_idx) of the row this thread owns.
    const int id = ids_i32[im * args.ne21 + r1 + lr1];
    const short i11 = (id % args.ne20) % args.ne11;   // slot index
    const short i12 = (id / args.ne20);               // token index
    const short i13 = 0;

    // Base of expert `im`'s weight slab.
    const uint64_t offset0 = im*args.nb02 + i13*args.nb03;
    const short    offset1 = il0/nl;

    device const block_q * x =
        (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;

    const short iy = 8*(tiitg % NL1);

    device const float * y = (device const float *)(src1
        + args.nb13*i13
        + args.nb12*i12
        + args.nb11*i11
        + args.nb10*iy);

    simdgroup_half8x8  ma[4];
    simdgroup_float8x8 mb[2];
    simdgroup_float8x8 mc[8];

    for (short i = 0; i < 8; i++){
        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
    }

    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
        // ---- A tile dequantize + stage ----
        {
            half4x4 temp_a;
            dequantize_func(x, il, temp_a);

            threadgroup_barrier(mem_flags::mem_threadgroup);

            for (short i = 0; i < 16; i++) {
                const short sx = 2*il0 + i/8;
                const short sy = (tiitg/NL0)/8;
                const short lx = (tiitg/NL0)%8;
                const short ly = i%8;
                const short ib = 8*sx + sy;
                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
            }
        }

        // ---- B tile stage (per-element, bounds-checked for K tail) ----
        for (short i = 0; i < 8; ++i) {
            const short sx = (tiitg%NL1);
            const short sy = (tiitg/NL1)/8;
            const short lx = i;
            const short ly = (tiitg/NL1)%8;
            const short ib = 4*sx + sy;
            *(sb + 64*ib + 8*ly + lx) =
                (loop_k + iy + i < args.ne00) ? *((device float *) y + i) : 0.f;
        }

        il = (il + 2 < nl) ? il + 2 : il % 2;
        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;

        y += NK;

        threadgroup_barrier(mem_flags::mem_threadgroup);

        // ---- MMA accumulate ----
        threadgroup const half  * lsma = (sa + 4*64*(sgitg%2));
        threadgroup const float * lsmb = (sb + 2*64*(sgitg/2));

        for (short ik = 0; ik < NK/8; ik++) {
            simdgroup_barrier(mem_flags::mem_none);
            for (short i = 0; i < 4; i++) {
                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
            }
            simdgroup_barrier(mem_flags::mem_none);
            for (short i = 0; i < 2; i++) {
                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
            }
            simdgroup_barrier(mem_flags::mem_none);
            for (short i = 0; i < 8; i++) {
                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
            }
            lsma += 8*64;
            lsmb += 4*64;
        }
    }

    // ---- Write-back (always through shmem staging: see header note) ----
    threadgroup_barrier(mem_flags::mem_threadgroup);

    threadgroup float * temp_str = ((threadgroup float *) shmem)
        + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;

    for (short i = 0; i < 8; i++) {
        simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
    }

    threadgroup_barrier(mem_flags::mem_threadgroup);

    // Each simdgroup `sgitg` strides through the tile in steps of 4 rows.
    // For each owned row j (within the tile), look up its (token, slot)
    // via hids, compute the destination address in dst, and copy.
    for (short j = sgitg; j < nr1; j += 4) {
        const int id = ids_i32[im*args.ne21 + r1 + j];

        const short ide = id % args.ne20;  // slot index (output row-major)
        const short idt = id / args.ne20;  // token index (output batch)

        device float  * D  = (device float  *) dst + r0 + ide*args.ne0 + idt*args.ne1*args.ne0;
        device float4 * D4 = (device float4 *) D;

        threadgroup float  * C  = (threadgroup float *) shmem + j*NR0;
        threadgroup float4 * C4 = (threadgroup float4 *) C;

        int i = tiisg;
        for (; i < nr0/4; i += 32) {
            *(D4 + i) = *(C4 + i);
        }

        i = (4*(nr0/4)) + tiisg;
        for (; i < nr0; i += 32) {
            *(D + i) = *(C + i);
        }
    }
}

// Template instantiations for the three quant types our GGUF uses.

template [[host_name("kernel_mul_mm_id_q4_0_f32")]]
kernel void hf2q_mul_mm_id_impl<block_q4_0, 2, dequantize_q4_0>(
    constant GgmlMatmulIdMm_MmParams &,
    device const char *, device const char *, device const char *, device const char *,
    device char *, threadgroup char *, uint3, ushort, ushort, ushort);

template [[host_name("kernel_mul_mm_id_q8_0_f32")]]
kernel void hf2q_mul_mm_id_impl<block_q8_0, 2, dequantize_q8_0>(
    constant GgmlMatmulIdMm_MmParams &,
    device const char *, device const char *, device const char *, device const char *,
    device char *, threadgroup char *, uint3, ushort, ushort, ushort);

template [[host_name("kernel_mul_mm_id_q6_K_f32")]]
kernel void hf2q_mul_mm_id_impl<block_q6_K, QK_NL, dequantize_q6_K>(
    constant GgmlMatmulIdMm_MmParams &,
    device const char *, device const char *, device const char *, device const char *,
    device char *, threadgroup char *, uint3, ushort, ushort, ushort);