llama_cpp 0.2.0 → 0.2.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (19) hide show
  1. checksums.yaml +4 -4
  2. data/CHANGELOG.md +7 -0
  3. data/examples/README.md +60 -0
  4. data/examples/chat.rb +195 -0
  5. data/ext/llama_cpp/llama_cpp.cpp +52 -0
  6. data/ext/llama_cpp/src/ggml-cuda.cu +697 -130
  7. data/ext/llama_cpp/src/ggml-cuda.h +4 -1
  8. data/ext/llama_cpp/src/ggml-metal.h +1 -0
  9. data/ext/llama_cpp/src/ggml-metal.m +548 -497
  10. data/ext/llama_cpp/src/ggml-metal.metal +425 -122
  11. data/ext/llama_cpp/src/ggml-opencl.cpp +3 -32
  12. data/ext/llama_cpp/src/ggml-opencl.h +1 -2
  13. data/ext/llama_cpp/src/ggml.c +1904 -303
  14. data/ext/llama_cpp/src/ggml.h +126 -2
  15. data/ext/llama_cpp/src/llama.cpp +212 -108
  16. data/ext/llama_cpp/src/llama.h +12 -3
  17. data/lib/llama_cpp/version.rb +2 -2
  18. data/sig/llama_cpp.rbs +3 -0
  19. metadata +4 -2
data/ext/llama_cpp/src/ggml-metal.m CHANGED
@@ -52,14 +52,18 @@ struct ggml_metal_context {
52
52
  GGML_METAL_DECL_KERNEL(get_rows_q4_0);
53
53
  GGML_METAL_DECL_KERNEL(get_rows_q4_1);
54
54
  GGML_METAL_DECL_KERNEL(get_rows_q2_k);
55
+ GGML_METAL_DECL_KERNEL(get_rows_q3_k);
55
56
  GGML_METAL_DECL_KERNEL(get_rows_q4_k);
57
+ GGML_METAL_DECL_KERNEL(get_rows_q5_k);
56
58
  GGML_METAL_DECL_KERNEL(get_rows_q6_k);
57
59
  GGML_METAL_DECL_KERNEL(rms_norm);
58
60
  GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
59
61
  GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
60
62
  GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32);
61
63
  GGML_METAL_DECL_KERNEL(mul_mat_q2_k_f32);
64
+ GGML_METAL_DECL_KERNEL(mul_mat_q3_k_f32);
62
65
  GGML_METAL_DECL_KERNEL(mul_mat_q4_k_f32);
66
+ GGML_METAL_DECL_KERNEL(mul_mat_q5_k_f32);
63
67
  GGML_METAL_DECL_KERNEL(mul_mat_q6_k_f32);
64
68
  GGML_METAL_DECL_KERNEL(rope);
65
69
  GGML_METAL_DECL_KERNEL(cpy_f32_f16);
@@ -86,6 +90,7 @@ struct ggml_metal_context * ggml_metal_init(void) {
86
90
 
87
91
  ctx->device = MTLCreateSystemDefaultDevice();
88
92
  ctx->queue = [ctx->device newCommandQueue];
93
+ ctx->n_buffers = 0;
89
94
 
90
95
  // determine if we can use MPS
91
96
  if (MPSSupportsMTLDevice(ctx->device)) {
@@ -152,14 +157,18 @@ struct ggml_metal_context * ggml_metal_init(void) {
152
157
  GGML_METAL_ADD_KERNEL(get_rows_q4_0);
153
158
  GGML_METAL_ADD_KERNEL(get_rows_q4_1);
154
159
  GGML_METAL_ADD_KERNEL(get_rows_q2_k);
160
+ GGML_METAL_ADD_KERNEL(get_rows_q3_k);
155
161
  GGML_METAL_ADD_KERNEL(get_rows_q4_k);
162
+ GGML_METAL_ADD_KERNEL(get_rows_q5_k);
156
163
  GGML_METAL_ADD_KERNEL(get_rows_q6_k);
157
164
  GGML_METAL_ADD_KERNEL(rms_norm);
158
165
  GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
159
166
  GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
160
167
  GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32);
161
168
  GGML_METAL_ADD_KERNEL(mul_mat_q2_k_f32);
169
+ GGML_METAL_ADD_KERNEL(mul_mat_q3_k_f32);
162
170
  GGML_METAL_ADD_KERNEL(mul_mat_q4_k_f32);
171
+ GGML_METAL_ADD_KERNEL(mul_mat_q5_k_f32);
163
172
  GGML_METAL_ADD_KERNEL(mul_mat_q6_k_f32);
164
173
  GGML_METAL_ADD_KERNEL(rope);
165
174
  GGML_METAL_ADD_KERNEL(cpy_f32_f16);
@@ -275,509 +284,551 @@ void ggml_metal_get_tensor(
275
284
 
276
285
  void ggml_metal_graph_compute(
277
286
  struct ggml_metal_context * ctx,
278
- struct ggml_cgraph * gf) {
287
+ struct ggml_cgraph * gf) {
279
288
  metal_printf("%s: evaluating graph\n", __func__);
280
289
 
281
- size_t offs_src0 = 0;
282
- size_t offs_src1 = 0;
283
- size_t offs_dst = 0;
284
-
285
- id<MTLCommandBuffer> command_buffer = [ctx->queue commandBuffer];
286
- id<MTLComputeCommandEncoder> encoder = nil;
287
-
288
- for (int i = 0; i < gf->n_nodes; ++i) {
289
- //metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
290
-
291
- struct ggml_tensor * src0 = gf->nodes[i]->src0;
292
- struct ggml_tensor * src1 = gf->nodes[i]->src1;
293
- struct ggml_tensor * dst = gf->nodes[i];
294
-
295
- const int64_t ne00 = src0 ? src0->ne[0] : 0;
296
- const int64_t ne01 = src0 ? src0->ne[1] : 0;
297
- const int64_t ne02 = src0 ? src0->ne[2] : 0;
298
- const int64_t ne03 = src0 ? src0->ne[3] : 0;
299
-
300
- const uint64_t nb00 = src0 ? src0->nb[0] : 0;
301
- const uint64_t nb01 = src0 ? src0->nb[1] : 0;
302
- const uint64_t nb02 = src0 ? src0->nb[2] : 0;
303
- const uint64_t nb03 = src0 ? src0->nb[3] : 0;
304
-
305
- const int64_t ne10 = src1 ? src1->ne[0] : 0;
306
- const int64_t ne11 = src1 ? src1->ne[1] : 0;
307
- const int64_t ne12 = src1 ? src1->ne[2] : 0;
308
- const int64_t ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
309
-
310
- const uint64_t nb10 = src1 ? src1->nb[0] : 0;
311
- const uint64_t nb11 = src1 ? src1->nb[1] : 0;
312
- const uint64_t nb12 = src1 ? src1->nb[2] : 0;
313
- const uint64_t nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13);
314
-
315
- const int64_t ne0 = dst ? dst->ne[0] : 0;
316
- const int64_t ne1 = dst ? dst->ne[1] : 0;
317
- const int64_t ne2 = dst ? dst->ne[2] : 0;
318
- const int64_t ne3 = dst ? dst->ne[3] : 0;
319
-
320
- const uint64_t nb0 = dst ? dst->nb[0] : 0;
321
- const uint64_t nb1 = dst ? dst->nb[1] : 0;
322
- const uint64_t nb2 = dst ? dst->nb[2] : 0;
323
- const uint64_t nb3 = dst ? dst->nb[3] : 0;
324
-
325
- const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
326
- const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
327
- const enum ggml_type dstt = dst ? dst->type : GGML_TYPE_COUNT;
328
-
329
- id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(ctx, src0, &offs_src0) : nil;
330
- id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
331
- id<MTLBuffer> id_dst = dst ? ggml_metal_get_buffer(ctx, dst, &offs_dst) : nil;
332
-
333
- //metal_printf("%s: op - %s\n", __func__, ggml_op_name(dst->op));
334
- //if (src0) {
335
- // metal_printf("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
336
- // ggml_is_contiguous(src0), src0->name);
337
- //}
338
- //if (src1) {
339
- // metal_printf("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
340
- // ggml_is_contiguous(src1), src1->name);
341
- //}
342
- //if (dst) {
343
- // metal_printf("%s: dst - %4s [%5lld, %5lld, %5lld], 1, %s\n", __func__, ggml_type_name(dstt), ne0, ne1, ne2,
344
- // dst->name);
345
- //}
346
-
347
- switch (dst->op) {
348
- case GGML_OP_RESHAPE:
349
- case GGML_OP_VIEW:
350
- case GGML_OP_TRANSPOSE:
351
- case GGML_OP_PERMUTE:
352
- {
353
- // noop
354
- } break;
355
- case GGML_OP_ADD:
356
- {
357
- if (encoder == nil) {
358
- encoder = [command_buffer computeCommandEncoder];
359
- }
360
-
361
- [encoder setComputePipelineState:ctx->pipeline_add];
362
- [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
363
- [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
364
- [encoder setBuffer:id_dst offset:offs_dst atIndex:2];
365
-
366
- const int64_t n = ggml_nelements(dst);
367
-
368
- [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
369
- } break;
370
- case GGML_OP_MUL:
371
- {
372
- if (encoder == nil) {
373
- encoder = [command_buffer computeCommandEncoder];
374
- }
375
-
376
- if (ggml_nelements(src1) == ne10) {
377
- // src1 is a row
378
- [encoder setComputePipelineState:ctx->pipeline_mul_row];
379
- } else {
380
- [encoder setComputePipelineState:ctx->pipeline_mul];
381
- }
382
- [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
383
- [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
384
- [encoder setBuffer:id_dst offset:offs_dst atIndex:2];
385
- [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
386
-
387
- const int64_t n = ggml_nelements(dst);
388
-
389
- [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
390
- } break;
391
- case GGML_OP_SCALE:
392
- {
393
- if (encoder == nil) {
394
- encoder = [command_buffer computeCommandEncoder];
395
- }
396
-
397
- const float scale = *(const float *) src1->data;
398
-
399
- [encoder setComputePipelineState:ctx->pipeline_scale];
400
- [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
401
- [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
402
- [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
403
-
404
- const int64_t n = ggml_nelements(dst);
405
-
406
- [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
407
- } break;
408
- case GGML_OP_SILU:
409
- {
410
- if (encoder == nil) {
411
- encoder = [command_buffer computeCommandEncoder];
412
- }
413
-
414
- [encoder setComputePipelineState:ctx->pipeline_silu];
415
- [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
416
- [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
417
-
418
- const int64_t n = ggml_nelements(dst);
419
-
420
- [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
421
- } break;
422
- case GGML_OP_RELU:
423
- {
424
- if (encoder == nil) {
425
- encoder = [command_buffer computeCommandEncoder];
426
- }
427
-
428
- [encoder setComputePipelineState:ctx->pipeline_relu];
429
- [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
430
- [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
431
-
432
- const int64_t n = ggml_nelements(dst);
433
-
434
- [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
435
- } break;
436
- case GGML_OP_GELU:
437
- {
438
- if (encoder == nil) {
439
- encoder = [command_buffer computeCommandEncoder];
440
- }
441
-
442
- [encoder setComputePipelineState:ctx->pipeline_gelu];
443
- [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
444
- [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
445
-
446
- const int64_t n = ggml_nelements(dst);
447
-
448
- [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
449
- } break;
450
- case GGML_OP_SOFT_MAX:
451
- {
452
- if (encoder == nil) {
453
- encoder = [command_buffer computeCommandEncoder];
454
- }
455
-
456
- const int nth = 32;
457
-
458
- [encoder setComputePipelineState:ctx->pipeline_soft_max];
459
- [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
460
- [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
461
- [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
462
- [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
463
- [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
464
- [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
465
-
466
- [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
467
- } break;
468
- case GGML_OP_DIAG_MASK_INF:
469
- {
470
- if (encoder == nil) {
471
- encoder = [command_buffer computeCommandEncoder];
472
- }
473
-
474
- const int n_past = ((int32_t *)(src1->data))[0];
475
-
476
- [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
477
- [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
478
- [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
479
- [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
480
- [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
481
- [encoder setBytes:&n_past length:sizeof(int) atIndex:4];
482
-
483
- [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
484
- } break;
485
- case GGML_OP_MUL_MAT:
486
- {
487
- // TODO: needs to be updated after PR: https://github.com/ggerganov/ggml/pull/224
488
-
489
- GGML_ASSERT(ne00 == ne10);
490
- GGML_ASSERT(ne02 == ne12);
491
-
492
- if (ggml_is_contiguous(src0) &&
493
- ggml_is_contiguous(src1) &&
494
- (src0t == GGML_TYPE_F32 || src0t == GGML_TYPE_F16) && ne11 > 1) {
495
-
496
- if (encoder != nil) {
497
- [encoder endEncoding];
498
- encoder = nil;
499
- }
500
-
501
- MPSDataType src0dt = src0t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
502
- MPSDataType src1dt = src1t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
503
-
504
- // for F32 x F32 we use MPS
505
- MPSMatrixDescriptor * desc0 = [MPSMatrixDescriptor
506
- matrixDescriptorWithRows:ne01 columns:ne00 rowBytes:src0->nb[1] dataType:src0dt];
507
-
508
- MPSMatrixDescriptor * desc1 = [MPSMatrixDescriptor
509
- matrixDescriptorWithRows:ne11 columns:ne10 rowBytes:src1->nb[1] dataType:src1dt];
510
-
511
- MPSMatrixDescriptor * desc = [MPSMatrixDescriptor
512
- matrixDescriptorWithRows:ne1 columns:ne0 rowBytes:dst->nb[1] dataType:MPSDataTypeFloat32];
513
-
514
- MPSMatrixMultiplication * mul = [[MPSMatrixMultiplication alloc]
515
- initWithDevice:ctx->device transposeLeft:false transposeRight:true
516
- resultRows:ne11 resultColumns:ne01 interiorColumns:ne00 alpha:1.0 beta:0.0];
517
-
518
- // we need to do ne02 multiplications
519
- // TODO: is there a way to do this in parallel - currently very slow ..
520
- // TODO: might be possible to offload part of the computation to ANE using Accelerate's CBLAS
521
- for (int64_t i02 = 0; i02 < ne02; ++i02) {
522
- size_t offs_src0_cur = offs_src0 + i02*nb02;
523
- size_t offs_src1_cur = offs_src1 + i02*nb12;
524
- size_t offs_dst_cur = offs_dst + i02*nb2;
525
-
526
- MPSMatrix * mat_src0 = [[MPSMatrix alloc] initWithBuffer:id_src0 offset:offs_src0_cur descriptor:desc0];
527
- MPSMatrix * mat_src1 = [[MPSMatrix alloc] initWithBuffer:id_src1 offset:offs_src1_cur descriptor:desc1];
528
- MPSMatrix * mat_dst = [[MPSMatrix alloc] initWithBuffer:id_dst offset:offs_dst_cur descriptor:desc ];
529
-
530
- [mul encodeToCommandBuffer:command_buffer leftMatrix:mat_src1 rightMatrix:mat_src0 resultMatrix:mat_dst];
531
- }
532
- } else {
533
- if (encoder == nil) {
534
- encoder = [command_buffer computeCommandEncoder];
535
- }
536
-
537
- int nth0 = 32;
538
- int nth1 = 1;
539
-
540
- // use custom matrix x vector kernel
541
- switch (src0t) {
542
- case GGML_TYPE_F16:
543
- {
544
- GGML_ASSERT(ne02 == ne12);
545
-
546
- nth0 = 64;
547
- nth1 = 1;
548
- [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
549
- } break;
550
- case GGML_TYPE_Q4_0:
551
- {
552
- GGML_ASSERT(ne02 == 1);
553
- GGML_ASSERT(ne12 == 1);
554
-
555
- nth0 = 8;
556
- nth1 = 8;
557
- [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_0_f32];
558
- } break;
559
- case GGML_TYPE_Q4_1:
560
- {
561
- GGML_ASSERT(ne02 == 1);
562
- GGML_ASSERT(ne12 == 1);
563
-
564
- nth0 = 8;
565
- nth1 = 8;
566
- [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_1_f32];
567
- } break;
568
- case GGML_TYPE_Q2_K:
569
- {
570
- GGML_ASSERT(ne02 == 1);
571
- GGML_ASSERT(ne12 == 1);
572
-
573
- nth0 = 4;
574
- nth1 = 16;
575
- [encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_k_f32];
576
- } break;
577
- case GGML_TYPE_Q4_K:
578
- {
579
- GGML_ASSERT(ne02 == 1);
580
- GGML_ASSERT(ne12 == 1);
581
-
582
- nth0 = 4;
583
- nth1 = 16;
584
- [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_k_f32];
585
- } break;
586
- case GGML_TYPE_Q6_K:
587
- {
588
- GGML_ASSERT(ne02 == 1);
589
- GGML_ASSERT(ne12 == 1);
590
-
591
- nth0 = 4;
592
- nth1 = 16;
593
- [encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_k_f32];
594
- } break;
595
- default:
596
- {
597
- fprintf(stderr, "Asserting on type %d\n",(int)src0t);
598
- GGML_ASSERT(false && "not implemented");
599
- }
600
- };
601
-
602
-
603
- [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
604
- [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
605
- [encoder setBuffer:id_dst offset:offs_dst atIndex:2];
606
- [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
607
- [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
608
- [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:5];
609
- [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:6];
610
- [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:7];
611
- [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:8];
612
- [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:9];
613
- [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:10];
614
- [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:11];
615
- [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:12];
616
- [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:13];
617
- [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:14];
618
-
619
- if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1) {
620
- [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
621
- [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
622
- } else if (src0t == GGML_TYPE_Q2_K) {
623
- [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
624
- [encoder dispatchThreadgroups:MTLSizeMake(ne01, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
625
- } else if (src0t == GGML_TYPE_Q4_K) {
626
- [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
627
- [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
628
- } else if (src0t == GGML_TYPE_Q6_K) {
629
- [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
630
- [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
631
- } else {
632
- [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
633
- [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
634
- }
635
- }
636
- } break;
637
- case GGML_OP_GET_ROWS:
638
- {
639
- if (encoder == nil) {
640
- encoder = [command_buffer computeCommandEncoder];
641
- }
642
-
643
- switch (src0->type) {
644
- case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_get_rows_f16]; break;
645
- case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
646
- case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
647
- case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_k]; break;
648
- case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_k]; break;
649
- case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_k]; break;
650
- default: GGML_ASSERT(false && "not implemented");
651
- }
652
-
653
- [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
654
- [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
655
- [encoder setBuffer:id_dst offset:offs_dst atIndex:2];
656
- [encoder setBytes:&(src0->ne[0]) length:sizeof( int64_t) atIndex:3];
657
- [encoder setBytes:&(src0->nb[1]) length:sizeof(uint64_t) atIndex:4];
658
- [encoder setBytes:&(dst->nb[1]) length:sizeof(uint64_t) atIndex:5];
659
-
660
- const int64_t n = ggml_nelements(src1);
661
-
662
- [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
663
- } break;
664
- case GGML_OP_RMS_NORM:
665
- {
666
- if (encoder == nil) {
667
- encoder = [command_buffer computeCommandEncoder];
668
- }
669
-
670
- const float eps = 1e-6f;
671
-
672
- const int nth = 256;
673
-
674
- [encoder setComputePipelineState:ctx->pipeline_rms_norm];
675
- [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
676
- [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
677
- [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
678
- [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
679
- [encoder setBytes:&eps length:sizeof( float) atIndex:4];
680
- [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
681
-
682
- const int64_t nrows = ggml_nrows(src0);
683
-
684
- [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
685
- } break;
686
- case GGML_OP_ROPE:
687
- {
688
- if (encoder == nil) {
689
- encoder = [command_buffer computeCommandEncoder];
690
- }
691
-
692
- const int n_dims = ((int32_t *) src1->data)[1];
693
- const int mode = ((int32_t *) src1->data)[2];
694
-
695
- const int n_past = ((int32_t *)(src1->data))[0];
696
-
697
- [encoder setComputePipelineState:ctx->pipeline_rope];
698
- [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
699
- [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
700
- [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
701
- [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
702
- [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
703
- [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
704
- [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
705
- [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
706
- [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
707
- [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
708
- [encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10];
709
- [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11];
710
- [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12];
711
- [encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13];
712
- [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14];
713
- [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15];
714
- [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16];
715
- [encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17];
716
- [encoder setBytes:&n_past length:sizeof( int) atIndex:18];
717
- [encoder setBytes:&n_dims length:sizeof( int) atIndex:19];
718
- [encoder setBytes:&mode length:sizeof( int) atIndex:20];
719
-
720
- [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
721
- } break;
722
- case GGML_OP_CPY:
723
- {
724
- if (encoder == nil) {
725
- encoder = [command_buffer computeCommandEncoder];
726
- }
727
-
728
- const int nth = 32;
729
-
730
- switch (src0t) {
731
- case GGML_TYPE_F32:
732
- {
733
- switch (dstt) {
734
- case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f16]; break;
735
- case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f32]; break;
736
- default: GGML_ASSERT(false && "not implemented");
737
- };
738
- } break;
739
- default: GGML_ASSERT(false && "not implemented");
740
- }
741
-
742
- [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
743
- [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
744
- [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
745
- [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
746
- [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
747
- [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
748
- [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
749
- [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
750
- [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
751
- [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
752
- [encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10];
753
- [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11];
754
- [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12];
755
- [encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13];
756
- [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14];
757
- [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15];
758
- [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16];
759
- [encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17];
760
-
761
- [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
762
- } break;
763
- default:
764
- fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
765
- GGML_ASSERT(false);
766
- }
767
- }
290
+ // create multiple command buffers and enqueue them
291
+ // then, we encode the graph into the command buffers in parallel
292
+
293
+ const int n_cb = gf->n_threads;
294
+
295
+ NSMutableArray * command_buffers = [NSMutableArray arrayWithCapacity:n_cb];
296
+
297
+ for (int i = 0; i < n_cb; ++i) {
298
+ command_buffers[i] = [ctx->queue commandBuffer];
768
299
 
769
- if (encoder != nil) {
770
- [encoder endEncoding];
771
- encoder = nil;
300
+ // enqueue the command buffers in order to specify their execution order
301
+ [command_buffers[i] enqueue];
772
302
  }
773
303
 
774
- [command_buffer commit];
775
- [command_buffer waitUntilCompleted];
304
+ // TODO: is this the best way to start threads?
305
+ dispatch_queue_t queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT);
306
+
307
+ for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
308
+ const int n_nodes_per_cb = (gf->n_nodes + n_cb - 1) / n_cb;
309
+
310
+ dispatch_async(queue, ^{
311
+ size_t offs_src0 = 0;
312
+ size_t offs_src1 = 0;
313
+ size_t offs_dst = 0;
314
+
315
+ id<MTLCommandBuffer> command_buffer = command_buffers[cb_idx];
316
+
317
+ id<MTLComputeCommandEncoder> encoder = nil;
318
+
319
+ const int node_start = (cb_idx + 0) * n_nodes_per_cb;
320
+ const int node_end = (cb_idx == n_cb - 1) ? gf->n_nodes : (cb_idx + 1) * n_nodes_per_cb;
321
+
322
+ for (int i = node_start; i < node_end; ++i) {
323
+ metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
324
+
325
+ struct ggml_tensor * src0 = gf->nodes[i]->src0;
326
+ struct ggml_tensor * src1 = gf->nodes[i]->src1;
327
+ struct ggml_tensor * dst = gf->nodes[i];
328
+
329
+ const int64_t ne00 = src0 ? src0->ne[0] : 0;
330
+ const int64_t ne01 = src0 ? src0->ne[1] : 0;
331
+ const int64_t ne02 = src0 ? src0->ne[2] : 0;
332
+ const int64_t ne03 = src0 ? src0->ne[3] : 0;
333
+
334
+ const uint64_t nb00 = src0 ? src0->nb[0] : 0;
335
+ const uint64_t nb01 = src0 ? src0->nb[1] : 0;
336
+ const uint64_t nb02 = src0 ? src0->nb[2] : 0;
337
+ const uint64_t nb03 = src0 ? src0->nb[3] : 0;
338
+
339
+ const int64_t ne10 = src1 ? src1->ne[0] : 0;
340
+ const int64_t ne11 = src1 ? src1->ne[1] : 0;
341
+ const int64_t ne12 = src1 ? src1->ne[2] : 0;
342
+ const int64_t ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
343
+
344
+ const uint64_t nb10 = src1 ? src1->nb[0] : 0;
345
+ const uint64_t nb11 = src1 ? src1->nb[1] : 0;
346
+ const uint64_t nb12 = src1 ? src1->nb[2] : 0;
347
+ const uint64_t nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13);
348
+
349
+ const int64_t ne0 = dst ? dst->ne[0] : 0;
350
+ const int64_t ne1 = dst ? dst->ne[1] : 0;
351
+ const int64_t ne2 = dst ? dst->ne[2] : 0;
352
+ const int64_t ne3 = dst ? dst->ne[3] : 0;
353
+
354
+ const uint64_t nb0 = dst ? dst->nb[0] : 0;
355
+ const uint64_t nb1 = dst ? dst->nb[1] : 0;
356
+ const uint64_t nb2 = dst ? dst->nb[2] : 0;
357
+ const uint64_t nb3 = dst ? dst->nb[3] : 0;
358
+
359
+ const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
360
+ const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
361
+ const enum ggml_type dstt = dst ? dst->type : GGML_TYPE_COUNT;
362
+
363
+ id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(ctx, src0, &offs_src0) : nil;
364
+ id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
365
+ id<MTLBuffer> id_dst = dst ? ggml_metal_get_buffer(ctx, dst, &offs_dst) : nil;
366
+
367
+ //metal_printf("%s: op - %s\n", __func__, ggml_op_name(dst->op));
368
+ //if (src0) {
369
+ // metal_printf("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
370
+ // ggml_is_contiguous(src0), src0->name);
371
+ //}
372
+ //if (src1) {
373
+ // metal_printf("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
374
+ // ggml_is_contiguous(src1), src1->name);
375
+ //}
376
+ //if (dst) {
377
+ // metal_printf("%s: dst - %4s [%5lld, %5lld, %5lld], 1, %s\n", __func__, ggml_type_name(dstt), ne0, ne1, ne2,
378
+ // dst->name);
379
+ //}
380
+
381
+ switch (dst->op) {
382
+ case GGML_OP_RESHAPE:
383
+ case GGML_OP_VIEW:
384
+ case GGML_OP_TRANSPOSE:
385
+ case GGML_OP_PERMUTE:
386
+ {
387
+ // noop
388
+ } break;
389
+ case GGML_OP_ADD:
390
+ {
391
+ if (encoder == nil) {
392
+ encoder = [command_buffer computeCommandEncoder];
393
+ }
394
+
395
+ [encoder setComputePipelineState:ctx->pipeline_add];
396
+ [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
397
+ [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
398
+ [encoder setBuffer:id_dst offset:offs_dst atIndex:2];
399
+
400
+ const int64_t n = ggml_nelements(dst);
401
+
402
+ [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
403
+ } break;
404
+ case GGML_OP_MUL:
405
+ {
406
+ if (encoder == nil) {
407
+ encoder = [command_buffer computeCommandEncoder];
408
+ }
409
+
410
+ if (ggml_nelements(src1) == ne10) {
411
+ // src1 is a row
412
+ [encoder setComputePipelineState:ctx->pipeline_mul_row];
413
+ } else {
414
+ [encoder setComputePipelineState:ctx->pipeline_mul];
415
+ }
416
+ [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
417
+ [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
418
+ [encoder setBuffer:id_dst offset:offs_dst atIndex:2];
419
+ [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
420
+
421
+ const int64_t n = ggml_nelements(dst);
422
+
423
+ [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
424
+ } break;
425
+ case GGML_OP_SCALE:
426
+ {
427
+ if (encoder == nil) {
428
+ encoder = [command_buffer computeCommandEncoder];
429
+ }
430
+
431
+ const float scale = *(const float *) src1->data;
432
+
433
+ [encoder setComputePipelineState:ctx->pipeline_scale];
434
+ [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
435
+ [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
436
+ [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
437
+
438
+ const int64_t n = ggml_nelements(dst);
439
+
440
+ [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
441
+ } break;
442
+ case GGML_OP_SILU:
443
+ {
444
+ if (encoder == nil) {
445
+ encoder = [command_buffer computeCommandEncoder];
446
+ }
447
+
448
+ [encoder setComputePipelineState:ctx->pipeline_silu];
449
+ [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
450
+ [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
451
+
452
+ const int64_t n = ggml_nelements(dst);
453
+
454
+ [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
455
+ } break;
456
+ case GGML_OP_RELU:
457
+ {
458
+ if (encoder == nil) {
459
+ encoder = [command_buffer computeCommandEncoder];
460
+ }
461
+
462
+ [encoder setComputePipelineState:ctx->pipeline_relu];
463
+ [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
464
+ [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
465
+
466
+ const int64_t n = ggml_nelements(dst);
467
+
468
+ [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
469
+ } break;
470
+ case GGML_OP_GELU:
471
+ {
472
+ if (encoder == nil) {
473
+ encoder = [command_buffer computeCommandEncoder];
474
+ }
475
+
476
+ [encoder setComputePipelineState:ctx->pipeline_gelu];
477
+ [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
478
+ [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
479
+
480
+ const int64_t n = ggml_nelements(dst);
481
+
482
+ [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
483
+ } break;
484
+ case GGML_OP_SOFT_MAX:
485
+ {
486
+ if (encoder == nil) {
487
+ encoder = [command_buffer computeCommandEncoder];
488
+ }
489
+
490
+ const int nth = 32;
491
+
492
+ [encoder setComputePipelineState:ctx->pipeline_soft_max];
493
+ [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
494
+ [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
495
+ [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
496
+ [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
497
+ [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
498
+ [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
499
+
500
+ [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
501
+ } break;
502
+ case GGML_OP_DIAG_MASK_INF:
503
+ {
504
+ if (encoder == nil) {
505
+ encoder = [command_buffer computeCommandEncoder];
506
+ }
507
+
508
+ const int n_past = ((int32_t *)(src1->data))[0];
509
+
510
+ [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
511
+ [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
512
+ [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
513
+ [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
514
+ [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
515
+ [encoder setBytes:&n_past length:sizeof(int) atIndex:4];
516
+
517
+ [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
518
+ } break;
519
+ case GGML_OP_MUL_MAT:
520
+ {
521
+ // TODO: needs to be updated after PR: https://github.com/ggerganov/ggml/pull/224
522
+
523
+ GGML_ASSERT(ne00 == ne10);
524
+ GGML_ASSERT(ne02 == ne12);
525
+
526
+ if (ggml_is_contiguous(src0) &&
527
+ ggml_is_contiguous(src1) &&
528
+ (src0t == GGML_TYPE_F32 || src0t == GGML_TYPE_F16) && ne11 > 1) {
529
+
530
+ if (encoder != nil) {
531
+ [encoder endEncoding];
532
+ encoder = nil;
533
+ }
776
534
 
777
- {
778
- const double time_elapsed = [command_buffer GPUEndTime] - [command_buffer GPUStartTime];
779
- UNUSED(time_elapsed);
535
+ MPSDataType src0dt = src0t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
536
+ MPSDataType src1dt = src1t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
537
+
538
+ // for F32 x F32 we use MPS
539
+ MPSMatrixDescriptor * desc0 = [MPSMatrixDescriptor
540
+ matrixDescriptorWithRows:ne01 columns:ne00 rowBytes:src0->nb[1] dataType:src0dt];
541
+
542
+ MPSMatrixDescriptor * desc1 = [MPSMatrixDescriptor
543
+ matrixDescriptorWithRows:ne11 columns:ne10 rowBytes:src1->nb[1] dataType:src1dt];
544
+
545
+ MPSMatrixDescriptor * desc = [MPSMatrixDescriptor
546
+ matrixDescriptorWithRows:ne1 columns:ne0 rowBytes:dst->nb[1] dataType:MPSDataTypeFloat32];
547
+
548
+ MPSMatrixMultiplication * mul = [[MPSMatrixMultiplication alloc]
549
+ initWithDevice:ctx->device transposeLeft:false transposeRight:true
550
+ resultRows:ne11 resultColumns:ne01 interiorColumns:ne00 alpha:1.0 beta:0.0];
551
+
552
+ // we need to do ne02 multiplications
553
+ // TODO: is there a way to do this in parallel - currently very slow ..
554
+ // TODO: might be possible to offload part of the computation to ANE using Accelerate's CBLAS
555
+ for (int64_t i02 = 0; i02 < ne02; ++i02) {
556
+ size_t offs_src0_cur = offs_src0 + i02*nb02;
557
+ size_t offs_src1_cur = offs_src1 + i02*nb12;
558
+ size_t offs_dst_cur = offs_dst + i02*nb2;
780
559
 
781
- metal_printf("%s: time elapsed = %f ms\n", __func__, time_elapsed * 1000.0);
560
+ MPSMatrix * mat_src0 = [[MPSMatrix alloc] initWithBuffer:id_src0 offset:offs_src0_cur descriptor:desc0];
561
+ MPSMatrix * mat_src1 = [[MPSMatrix alloc] initWithBuffer:id_src1 offset:offs_src1_cur descriptor:desc1];
562
+ MPSMatrix * mat_dst = [[MPSMatrix alloc] initWithBuffer:id_dst offset:offs_dst_cur descriptor:desc ];
563
+
564
+ [mul encodeToCommandBuffer:command_buffer leftMatrix:mat_src1 rightMatrix:mat_src0 resultMatrix:mat_dst];
565
+ }
566
+ } else {
567
+ if (encoder == nil) {
568
+ encoder = [command_buffer computeCommandEncoder];
569
+ }
570
+
571
+ int nth0 = 32;
572
+ int nth1 = 1;
573
+
574
+ // use custom matrix x vector kernel
575
+ switch (src0t) {
576
+ case GGML_TYPE_F16:
577
+ {
578
+ GGML_ASSERT(ne02 == ne12);
579
+
580
+ nth0 = 64;
581
+ nth1 = 1;
582
+ [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
583
+ } break;
584
+ case GGML_TYPE_Q4_0:
585
+ {
586
+ GGML_ASSERT(ne02 == 1);
587
+ GGML_ASSERT(ne12 == 1);
588
+
589
+ nth0 = 8;
590
+ nth1 = 8;
591
+ [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_0_f32];
592
+ } break;
593
+ case GGML_TYPE_Q4_1:
594
+ {
595
+ GGML_ASSERT(ne02 == 1);
596
+ GGML_ASSERT(ne12 == 1);
597
+
598
+ nth0 = 8;
599
+ nth1 = 8;
600
+ [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_1_f32];
601
+ } break;
602
+ case GGML_TYPE_Q2_K:
603
+ {
604
+ GGML_ASSERT(ne02 == 1);
605
+ GGML_ASSERT(ne12 == 1);
606
+
607
+ nth0 = 4;
608
+ nth1 = 16;
609
+ [encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_k_f32];
610
+ } break;
611
+ case GGML_TYPE_Q3_K:
612
+ {
613
+ GGML_ASSERT(ne02 == 1);
614
+ GGML_ASSERT(ne12 == 1);
615
+
616
+ nth0 = 4;
617
+ nth1 = 16;
618
+ [encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_k_f32];
619
+ } break;
620
+ case GGML_TYPE_Q4_K:
621
+ {
622
+ GGML_ASSERT(ne02 == 1);
623
+ GGML_ASSERT(ne12 == 1);
624
+
625
+ nth0 = 4;
626
+ nth1 = 16;
627
+ [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_k_f32];
628
+ } break;
629
+ case GGML_TYPE_Q5_K:
630
+ {
631
+ GGML_ASSERT(ne02 == 1);
632
+ GGML_ASSERT(ne12 == 1);
633
+
634
+ nth0 = 4;
635
+ nth1 = 16;
636
+ [encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_k_f32];
637
+ } break;
638
+ case GGML_TYPE_Q6_K:
639
+ {
640
+ GGML_ASSERT(ne02 == 1);
641
+ GGML_ASSERT(ne12 == 1);
642
+
643
+ nth0 = 4;
644
+ nth1 = 16;
645
+ [encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_k_f32];
646
+ } break;
647
+ default:
648
+ {
649
+ fprintf(stderr, "Asserting on type %d\n",(int)src0t);
650
+ GGML_ASSERT(false && "not implemented");
651
+ }
652
+ };
653
+
654
+ [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
655
+ [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
656
+ [encoder setBuffer:id_dst offset:offs_dst atIndex:2];
657
+ [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
658
+ [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
659
+ [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:5];
660
+ [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:6];
661
+ [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:7];
662
+ [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:8];
663
+ [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:9];
664
+ [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:10];
665
+ [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:11];
666
+ [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:12];
667
+ [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:13];
668
+ [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:14];
669
+
670
+ if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1) {
671
+ [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
672
+ [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
673
+ }
674
+ else if (src0t == GGML_TYPE_Q2_K ||
675
+ src0t == GGML_TYPE_Q3_K ||
676
+ src0t == GGML_TYPE_Q4_K ||
677
+ src0t == GGML_TYPE_Q5_K ||
678
+ src0t == GGML_TYPE_Q6_K) {
679
+ [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
680
+ [encoder dispatchThreadgroups:MTLSizeMake(ne01, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
681
+ } else {
682
+ [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
683
+ [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
684
+ }
685
+ }
686
+ } break;
687
+ case GGML_OP_GET_ROWS:
688
+ {
689
+ if (encoder == nil) {
690
+ encoder = [command_buffer computeCommandEncoder];
691
+ }
692
+
693
+ switch (src0->type) {
694
+ case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_get_rows_f16]; break;
695
+ case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
696
+ case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
697
+ case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_k]; break;
698
+ case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q3_k]; break;
699
+ case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_k]; break;
700
+ case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_k]; break;
701
+ case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_k]; break;
702
+ default: GGML_ASSERT(false && "not implemented");
703
+ }
704
+
705
+ [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
706
+ [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
707
+ [encoder setBuffer:id_dst offset:offs_dst atIndex:2];
708
+ [encoder setBytes:&(src0->ne[0]) length:sizeof( int64_t) atIndex:3];
709
+ [encoder setBytes:&(src0->nb[1]) length:sizeof(uint64_t) atIndex:4];
710
+ [encoder setBytes:&(dst->nb[1]) length:sizeof(uint64_t) atIndex:5];
711
+
712
+ const int64_t n = ggml_nelements(src1);
713
+
714
+ [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
715
+ } break;
716
+ case GGML_OP_RMS_NORM:
717
+ {
718
+ if (encoder == nil) {
719
+ encoder = [command_buffer computeCommandEncoder];
720
+ }
721
+
722
+ const float eps = 1e-6f;
723
+
724
+ const int nth = 256;
725
+
726
+ [encoder setComputePipelineState:ctx->pipeline_rms_norm];
727
+ [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
728
+ [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
729
+ [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
730
+ [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
731
+ [encoder setBytes:&eps length:sizeof( float) atIndex:4];
732
+ [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
733
+
734
+ const int64_t nrows = ggml_nrows(src0);
735
+
736
+ [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
737
+ } break;
738
+ case GGML_OP_ROPE:
739
+ {
740
+ if (encoder == nil) {
741
+ encoder = [command_buffer computeCommandEncoder];
742
+ }
743
+
744
+ const int n_dims = ((int32_t *) src1->data)[1];
745
+ const int mode = ((int32_t *) src1->data)[2];
746
+
747
+ const int n_past = ((int32_t *)(src1->data))[0];
748
+
749
+ [encoder setComputePipelineState:ctx->pipeline_rope];
750
+ [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
751
+ [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
752
+ [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
753
+ [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
754
+ [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
755
+ [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
756
+ [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
757
+ [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
758
+ [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
759
+ [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
760
+ [encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10];
761
+ [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11];
762
+ [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12];
763
+ [encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13];
764
+ [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14];
765
+ [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15];
766
+ [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16];
767
+ [encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17];
768
+ [encoder setBytes:&n_past length:sizeof( int) atIndex:18];
769
+ [encoder setBytes:&n_dims length:sizeof( int) atIndex:19];
770
+ [encoder setBytes:&mode length:sizeof( int) atIndex:20];
771
+
772
+ [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
773
+ } break;
774
+ case GGML_OP_CPY:
775
+ {
776
+ if (encoder == nil) {
777
+ encoder = [command_buffer computeCommandEncoder];
778
+ }
779
+
780
+ const int nth = 32;
781
+
782
+ switch (src0t) {
783
+ case GGML_TYPE_F32:
784
+ {
785
+ switch (dstt) {
786
+ case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f16]; break;
787
+ case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f32]; break;
788
+ default: GGML_ASSERT(false && "not implemented");
789
+ };
790
+ } break;
791
+ default: GGML_ASSERT(false && "not implemented");
792
+ }
793
+
794
+ [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
795
+ [encoder setBuffer:id_dst offset:offs_dst atIndex:1];
796
+ [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
797
+ [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
798
+ [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
799
+ [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
800
+ [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
801
+ [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
802
+ [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
803
+ [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
804
+ [encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10];
805
+ [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11];
806
+ [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12];
807
+ [encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13];
808
+ [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14];
809
+ [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15];
810
+ [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16];
811
+ [encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17];
812
+
813
+ [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
814
+ } break;
815
+ default:
816
+ fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
817
+ GGML_ASSERT(false);
818
+ }
819
+ }
820
+
821
+ if (encoder != nil) {
822
+ [encoder endEncoding];
823
+ encoder = nil;
824
+ }
825
+
826
+ [command_buffer commit];
827
+ });
782
828
  }
829
+
830
+ // wait for all threads to finish
831
+ dispatch_barrier_sync(queue, ^{});
832
+
833
+ [command_buffers[n_cb - 1] waitUntilCompleted];
783
834
  }