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
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
|
extern "C" {
#include "embed_cuda.h"
#include "embed_internal.h"
}
#include <cuda_runtime.h>
#include <cuda_bf16.h>
#include <cublas_v2.h>
#include <limits.h>
#include <math.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
typedef struct {
float *d; // device buffer; holds __nv_bfloat16 elements when bf16
int rows;
int cols;
int bf16; // 1 if d holds __nv_bfloat16, 0 if F32
} cuda_matrix_t;
typedef struct {
cuda_matrix_t qkv, wo;
float *q_norm;
float *k_norm;
float *input_norm;
float *post_attn_norm;
cuda_matrix_t gate_up_proj, down_proj;
} cuda_layer_t;
struct embed_cuda_ctx {
embed_model_t *cpu;
embed_config_t config;
cudaStream_t stream;
cublasHandle_t blas;
cuda_matrix_t embed_tokens;
cuda_layer_t *layers;
float *norm;
int seq_cap;
int batch_cap;
int max_seq_cap;
float *x;
float *x_norm;
float *qkv;
float *attn_out;
float *ffn_gate_up;
void *act_bf16; // BF16 cast of a GEMM activation operand (bf16 weights)
float *pooled_out;
float *rope_cos;
float *rope_sin;
int *token_ids;
int *offsets;
int *offsets_host;
int *positions;
float *kexp;
float *vexp;
float *attn_scores;
void *attn_probs; // BF16 softmax output (bf16 weights)
long long attn_scores_elems;
long long attn_probs_elems;
// Equal-length micro-batch attention: pointer arrays for batched GEMMs,
// built once per forward pass (layer-invariant buffer addresses).
// attn_G is the number of sequences whose score tensors fit the scratch
// budget at once; 0 disables the batched path for this forward.
const void **attn_ptrs; // device: [K|Q|S|V|P|O] x batch x n_heads
const void **attn_ptrs_host;
int attn_ptrs_cap; // capacity in pointer entries
int attn_G;
int attn_L;
int *span_starts;
int *span_lens;
int pooled_rows_cap;
int span_cap;
};
// Scratch budget for the batched-attention score tensors (F32 elements).
enum { CUDA_ATTN_SCORES_BUDGET = 64 * 1024 * 1024 };
// Above this sequence length the per-sequence strided-batched loop beats the
// pointer-array batched GEMMs (L4: batched wins <=80tok, ties ~140, loses
// >=256; both paths are identical precision).
enum { CUDA_ATTN_BATCHED_MAX_LEN = 192 };
#define CUDA_CHECK(expr) do { \
cudaError_t _e = (expr); \
if (_e != cudaSuccess) { \
fprintf(stderr, "cuda: %s failed: %s\n", #expr, \
cudaGetErrorString(_e)); \
return -1; \
} \
} while (0)
#define CUBLAS_CHECK(expr) do { \
cublasStatus_t _s = (expr); \
if (_s != CUBLAS_STATUS_SUCCESS) { \
fprintf(stderr, "cuda: %s failed: cublas status %d\n", \
#expr, (int)_s); \
return -1; \
} \
} while (0)
static int parse_gemm_compute(const char *mode, cublasComputeType_t *out)
{
if (!mode || !mode[0] || !strcmp(mode, "f32")) {
*out = CUBLAS_COMPUTE_32F;
return 0;
}
if (!strcmp(mode, "tf32")) {
*out = CUBLAS_COMPUTE_32F_FAST_TF32;
return 0;
}
if (!strcmp(mode, "bf16")) {
*out = CUBLAS_COMPUTE_32F_FAST_16BF;
return 0;
}
if (!strcmp(mode, "16f")) {
*out = CUBLAS_COMPUTE_32F_FAST_16F;
return 0;
}
return -1;
}
static cublasComputeType_t g_gemm_compute = CUBLAS_COMPUTE_32F;
static int g_gemm_compute_set = 0;
int embed_cuda_set_fast_gemm(const char *mode)
{
cublasComputeType_t ct;
if (parse_gemm_compute(mode, &ct) != 0)
return -1;
g_gemm_compute = ct;
g_gemm_compute_set = 1;
return 0;
}
static void cuda_matrix_free(cuda_matrix_t *m)
{
if (!m) return;
cudaFree(m->d);
memset(m, 0, sizeof(*m));
}
static float bf16_to_f32(uint16_t v)
{
uint32_t u = ((uint32_t)v) << 16;
float f;
memcpy(&f, &u, sizeof(f));
return f;
}
// Round F32 to BF16 (truncate the low 16 bits, round to nearest even). NaNs are
// preserved. BF16 shares F32's 8-bit exponent, so this never overflows.
static uint16_t f32_to_bf16(float f)
{
uint32_t x;
memcpy(&x, &f, sizeof(x));
if ((x & 0x7fffffffu) > 0x7f800000u) // NaN
return (uint16_t)((x >> 16) | 0x0040u);
x += 0x7fffu + ((x >> 16) & 1u); // round to nearest even
return (uint16_t)(x >> 16);
}
// Store weights as native BF16 on the device when set (halves weight memory and
// feeds BF16 tensor cores); default keeps exact F32 for F32 model files. When
// the flag was never set explicitly, the loader follows the model file dtype,
// matching the CPU and MLX backends.
static int g_weights_bf16 = 0;
static int g_weights_bf16_set = 0;
int embed_cuda_set_weights_bf16(int on)
{
g_weights_bf16 = on ? 1 : 0;
g_weights_bf16_set = 1;
return 0;
}
static int copy_weight_host_f32(float *dst, const embed_weight_ref_t *w,
size_t count)
{
if (!dst || !w || !w->data) return -1;
if (w->dtype == DTYPE_F32) {
memcpy(dst, w->data, count * sizeof(float));
return 0;
}
if (w->dtype == DTYPE_BF16) {
const uint16_t *src = (const uint16_t *)w->data;
for (size_t i = 0; i < count; i++)
dst[i] = bf16_to_f32(src[i]);
return 0;
}
return -1;
}
// Upload a host F32 weight buffer to the device, as BF16 when g_weights_bf16 is
// set (rounding each value) or as F32 otherwise. Sets m->d, rows, cols, bf16.
static int upload_weight(cuda_matrix_t *m, const float *src, size_t count,
int rows, int cols)
{
cudaError_t e;
if (g_weights_bf16) {
uint16_t *tmp16 = (uint16_t *)malloc(count * sizeof(uint16_t));
if (!tmp16) return -1;
for (size_t i = 0; i < count; i++)
tmp16[i] = f32_to_bf16(src[i]);
e = cudaMalloc((void **)&m->d, count * sizeof(__nv_bfloat16));
if (e != cudaSuccess) {
fprintf(stderr, "cudaMalloc weight failed: %s\n", cudaGetErrorString(e));
free(tmp16);
return -1;
}
e = cudaMemcpy(m->d, tmp16, count * sizeof(__nv_bfloat16),
cudaMemcpyHostToDevice);
free(tmp16);
m->bf16 = 1;
} else {
e = cudaMalloc((void **)&m->d, count * sizeof(float));
if (e != cudaSuccess) {
fprintf(stderr, "cudaMalloc weight failed: %s\n", cudaGetErrorString(e));
return -1;
}
e = cudaMemcpy(m->d, src, count * sizeof(float), cudaMemcpyHostToDevice);
m->bf16 = 0;
}
if (e != cudaSuccess) {
fprintf(stderr, "cudaMemcpy weight failed: %s\n", cudaGetErrorString(e));
cuda_matrix_free(m);
return -1;
}
m->rows = rows;
m->cols = cols;
return 0;
}
static int load_matrix(cuda_matrix_t *m, const embed_weight_ref_t *w,
int rows, int cols)
{
if (!m || !w || rows <= 0 || cols <= 0) return -1;
size_t count = (size_t)rows * (size_t)cols;
float *tmp = (float *)malloc(count * sizeof(float));
if (!tmp) return -1;
if (copy_weight_host_f32(tmp, w, count) != 0) {
free(tmp);
return -1;
}
int r = upload_weight(m, tmp, count, rows, cols);
free(tmp);
return r;
}
static int load_qkv_matrix(cuda_matrix_t *m, const embed_layer_t *src,
const embed_config_t *c)
{
if (!m || !src || !c) return -1;
int rows = c->q_dim + 2 * c->kv_dim;
int cols = c->hidden_size;
size_t q_count = (size_t)c->q_dim * cols;
size_t kv_count = (size_t)c->kv_dim * cols;
size_t total = q_count + 2 * kv_count;
float *tmp = (float *)malloc(total * sizeof(float));
if (!tmp) return -1;
if (copy_weight_host_f32(tmp, &src->wq, q_count) != 0 ||
copy_weight_host_f32(tmp + q_count, &src->wk, kv_count) != 0 ||
copy_weight_host_f32(tmp + q_count + kv_count, &src->wv, kv_count) != 0) {
free(tmp);
return -1;
}
int r = upload_weight(m, tmp, total, rows, cols);
free(tmp);
return r;
}
static int load_gate_up_matrix(cuda_matrix_t *m, const embed_layer_t *src,
const embed_config_t *c)
{
if (!m || !src || !c) return -1;
int rows = 2 * c->intermediate_size;
int cols = c->hidden_size;
size_t proj_count = (size_t)c->intermediate_size * cols;
size_t total = 2 * proj_count;
float *tmp = (float *)malloc(total * sizeof(float));
if (!tmp) return -1;
if (copy_weight_host_f32(tmp, &src->gate_proj, proj_count) != 0 ||
copy_weight_host_f32(tmp + proj_count, &src->up_proj, proj_count) != 0) {
free(tmp);
return -1;
}
int r = upload_weight(m, tmp, total, rows, cols);
free(tmp);
return r;
}
static int load_vector(float **out, const float *src, int n)
{
if (!out || !src || n <= 0) return -1;
cudaError_t e = cudaMalloc((void **)out, (size_t)n * sizeof(float));
if (e != cudaSuccess) {
fprintf(stderr, "cudaMalloc vector failed: %s\n", cudaGetErrorString(e));
return -1;
}
e = cudaMemcpy(*out, src, (size_t)n * sizeof(float), cudaMemcpyHostToDevice);
if (e != cudaSuccess) {
fprintf(stderr, "cudaMemcpy vector failed: %s\n", cudaGetErrorString(e));
cudaFree(*out);
*out = NULL;
return -1;
}
return 0;
}
__global__ static void embed_lookup_kernel(float *x, const int *ids,
const void *emb, int emb_bf16,
int total, int hidden, int vocab)
{
int idx = blockIdx.x * blockDim.x + threadIdx.x;
int count = total * hidden;
if (idx >= count) return;
int tok = idx / hidden;
int d = idx - tok * hidden;
int id = ids[tok];
if (id < 0 || id >= vocab) {
x[idx] = 0.0f;
return;
}
size_t off = (size_t)id * hidden + d;
x[idx] = emb_bf16 ? __bfloat162float(((const __nv_bfloat16 *)emb)[off])
: ((const float *)emb)[off];
}
// Cast n F32 values to BF16 (used to feed BF16-weight projection GEMMs).
__global__ static void cast_f32_to_bf16_kernel(__nv_bfloat16 *out,
const float *in, size_t n)
{
size_t i = (size_t)blockIdx.x * blockDim.x + threadIdx.x;
if (i < n) out[i] = __float2bfloat16(in[i]);
}
__device__ static float block_sum(float v)
{
__shared__ float partial[8];
__shared__ float total;
int lane = threadIdx.x & 31;
int warp = threadIdx.x >> 5;
int warps = blockDim.x >> 5;
for (int offset = 16; offset > 0; offset >>= 1)
v += __shfl_down_sync(0xffffffff, v, offset);
if (lane == 0) partial[warp] = v;
__syncthreads();
if (warp == 0) {
v = lane < warps ? partial[lane] : 0.0f;
for (int offset = 16; offset > 0; offset >>= 1)
v += __shfl_down_sync(0xffffffff, v, offset);
if (lane == 0) total = v;
}
__syncthreads();
return total;
}
__device__ static inline void store_act(float *out, size_t i, float v)
{
out[i] = v;
}
__device__ static inline void store_act(__nv_bfloat16 *out, size_t i, float v)
{
out[i] = __float2bfloat16(v);
}
// OUT_T is float, or __nv_bfloat16 when the result feeds a BF16-weight
// projection GEMM directly (same rounding as a separate cast kernel).
template <typename OUT_T>
__global__ static void rms_norm_kernel(OUT_T *out, const float *x,
const float *weight, int rows,
int dim, float eps)
{
int row = blockIdx.x;
int tid = threadIdx.x;
float sum = 0.0f;
for (int d = tid; d < dim; d += blockDim.x) {
float v = x[(size_t)row * dim + d];
sum += v * v;
}
float total = block_sum(sum);
float inv = rsqrtf(total / (float)dim + eps);
for (int d = tid; d < dim; d += blockDim.x)
store_act(out, (size_t)row * dim + d,
x[(size_t)row * dim + d] * inv * weight[d]);
}
// One grid covers Q and K: blockIdx.y < n_heads selects the Q head with
// q_norm/q_offset, the remaining blocks the K head with k_norm/k_offset.
__global__ static void rms_norm_rope_qk_kernel(float *x, const float *q_norm,
const float *k_norm,
const int *positions,
const float *cosv,
const float *sinv,
int rows, int n_heads,
int n_kv_heads,
int head_dim,
int row_stride,
int q_offset, int k_offset,
float eps)
{
__shared__ float partial[4];
__shared__ float total;
int row = blockIdx.x;
int head = blockIdx.y;
int tid = threadIdx.x;
int lane = tid & 31;
int warp = tid >> 5;
int half = head_dim / 2;
int pair_d = tid < half ? tid + half : tid - half;
float sign = tid < half ? -1.0f : 1.0f;
const float *weight = head < n_heads ? q_norm : k_norm;
int base_offset = head < n_heads ? q_offset : k_offset;
if (head >= n_heads) head -= n_heads;
size_t base = (size_t)row * row_stride + base_offset + head * head_dim;
float v = tid < head_dim ? x[base + tid] : 0.0f;
float pair_v = tid < head_dim ? x[base + pair_d] : 0.0f;
float sum = v * v;
for (int offset = 16; offset > 0; offset >>= 1)
sum += __shfl_down_sync(0xffffffff, sum, offset);
if (lane == 0) partial[warp] = sum;
__syncthreads();
if (warp == 0) {
sum = lane < 4 ? partial[lane] : 0.0f;
for (int offset = 16; offset > 0; offset >>= 1)
sum += __shfl_down_sync(0xffffffff, sum, offset);
if (lane == 0) total = sum;
}
__syncthreads();
if (tid < head_dim) {
float inv = rsqrtf(total / (float)head_dim + eps);
int pos = positions[row];
float c = cosv[(size_t)pos * head_dim + tid];
float s = sinv[(size_t)pos * head_dim + tid];
float a = v * inv * weight[tid];
float b = pair_v * inv * weight[pair_d];
x[base + tid] = a * c + sign * b * s;
}
}
// Writes SiLU(gate)*up to out with out_stride. The F32 path writes in place
// over the gate half (out = gate_up, out_stride = 2*intermediate); the BF16
// path writes a packed [rows x intermediate] BF16 matrix for the down GEMM.
template <typename OUT_T>
__global__ static void silu_mul_packed_kernel(OUT_T *out, int out_stride,
const float *gate_up, int rows,
int intermediate)
{
int i = blockIdx.x * blockDim.x + threadIdx.x;
int n = rows * intermediate;
if (i < n) {
int row = i / intermediate;
int d = i - row * intermediate;
size_t base = (size_t)row * (2 * intermediate);
float g = gate_up[base + d];
float u = gate_up[base + intermediate + d];
store_act(out, (size_t)row * out_stride + d,
(g / (1.0f + expf(-g))) * u);
}
}
// Expand K and V from the n_kv_heads layout to a contiguous [total x q_dim]
// per-query-head layout (GQA: each kv head repeated n_heads/n_kv_heads times)
// so the attention GEMMs can stride uniformly over query heads.
// V_T is float, or __nv_bfloat16 when V feeds a BF16 @V GEMM directly.
template <typename V_T>
__global__ static void attn_expand_kv_kernel(
float *kexp, V_T *vexp, const float *qkv, int total, int n_heads,
int n_kv_heads, int head_dim, int qkv_dim, int k_offset, int v_offset)
{
int q_dim = n_heads * head_dim;
size_t idx = (size_t)blockIdx.x * blockDim.x + threadIdx.x;
size_t count = (size_t)total * q_dim;
if (idx >= count) return;
int t = (int)(idx / q_dim);
int rem = (int)(idx - (size_t)t * q_dim);
int h = rem / head_dim;
int e = rem - h * head_dim;
int kv = h / (n_heads / n_kv_heads);
const float *base = qkv + (size_t)t * qkv_dim;
kexp[idx] = base[k_offset + kv * head_dim + e];
store_act(vexp, idx, base[v_offset + kv * head_dim + e]);
}
// Row softmax over the key dimension for one sequence's score tensor laid out
// as scores[head * L * L + i * L + j]; normalizes across j for each (head, i).
// One block per (head, query) row; scores are pre-scaled by the QK^T GEMM.
// P_T is float with probs == scores (in place), or __nv_bfloat16 when the
// probabilities feed a BF16 @V GEMM.
template <typename P_T>
__global__ static void attn_softmax_kernel(float *scores, P_T *probs, int L)
{
float *s = scores + (size_t)blockIdx.x * L;
P_T *p = probs + (size_t)blockIdx.x * L;
int tid = threadIdx.x;
__shared__ float red[256];
float m = -3.402823466e+38F;
for (int j = tid; j < L; j += blockDim.x) m = fmaxf(m, s[j]);
red[tid] = m;
__syncthreads();
for (int o = blockDim.x / 2; o > 0; o >>= 1) {
if (tid < o) red[tid] = fmaxf(red[tid], red[tid + o]);
__syncthreads();
}
m = red[0];
__syncthreads();
float sum = 0.0f;
for (int j = tid; j < L; j += blockDim.x) {
float e = __expf(s[j] - m);
s[j] = e;
sum += e;
}
red[tid] = sum;
__syncthreads();
for (int o = blockDim.x / 2; o > 0; o >>= 1) {
if (tid < o) red[tid] += red[tid + o];
__syncthreads();
}
float inv = 1.0f / red[0];
for (int j = tid; j < L; j += blockDim.x)
store_act(p, (size_t)j, s[j] * inv);
}
template <typename P_T>
__global__ static void attn_causal_softmax_kernel(float *scores, P_T *probs,
int L)
{
float *s = scores + (size_t)blockIdx.x * L;
P_T *p = probs + (size_t)blockIdx.x * L;
int query = blockIdx.x % L;
int keys = query + 1;
int tid = threadIdx.x;
__shared__ float red[256];
float m = -3.402823466e+38F;
for (int j = tid; j < keys; j += blockDim.x) m = fmaxf(m, s[j]);
red[tid] = m;
__syncthreads();
for (int o = blockDim.x / 2; o > 0; o >>= 1) {
if (tid < o) red[tid] = fmaxf(red[tid], red[tid + o]);
__syncthreads();
}
m = red[0];
__syncthreads();
float sum = 0.0f;
for (int j = tid; j < keys; j += blockDim.x) {
float e = __expf(s[j] - m);
s[j] = e;
sum += e;
}
red[tid] = sum;
__syncthreads();
for (int o = blockDim.x / 2; o > 0; o >>= 1) {
if (tid < o) red[tid] += red[tid + o];
__syncthreads();
}
float inv = 1.0f / red[0];
for (int j = tid; j < L; j += blockDim.x)
store_act(p, (size_t)j, j < keys ? s[j] * inv : 0.0f);
}
__global__ static void mean_pool_kernel(float *out, const float *x,
const int *offsets, int batch,
int hidden)
{
int b = blockIdx.x;
int tid = threadIdx.x;
int start = offsets[b];
int end = offsets[b + 1];
int len = end - start;
for (int d = tid; d < hidden; d += blockDim.x) {
float sum = 0.0f;
for (int t = start; t < end; t++)
sum += x[(size_t)t * hidden + d];
out[(size_t)b * hidden + d] = sum / (float)len;
}
}
__global__ static void last_pool_kernel(float *out, const float *x,
const int *offsets, int batch,
int hidden)
{
int b = blockIdx.x;
int tid = threadIdx.x;
int last = offsets[b + 1] - 1;
for (int d = tid; d < hidden; d += blockDim.x)
out[(size_t)b * hidden + d] = x[(size_t)last * hidden + d];
}
__global__ static void span_pool_kernel(float *out, const float *x,
const int *starts, const int *lens,
int n_spans, int hidden)
{
int s = blockIdx.x;
int tid = threadIdx.x;
int start = starts[s];
int len = lens[s];
for (int d = tid; d < hidden; d += blockDim.x) {
float sum = 0.0f;
for (int t = 0; t < len; t++)
sum += x[(size_t)(start + t) * hidden + d];
out[(size_t)s * hidden + d] = sum / (float)len;
}
}
static int launch_check(void)
{
cudaError_t e = cudaGetLastError();
if (e != cudaSuccess) {
fprintf(stderr, "cuda kernel failed: %s\n", cudaGetErrorString(e));
return -1;
}
return 0;
}
// Compute type for CUDA GEMMs. Default is exact F32 (== cublasSgemm).
// EMBED_CUDA_GEMM_MODE={bf16,tf32,16f} selects a reduced-precision tensor-core
// matmul that keeps F32 inputs/outputs and F32 accumulation but rounds operands
// internally.
static cublasComputeType_t gemm_compute(void)
{
static int env_init = 0;
if (!g_gemm_compute_set && !env_init) {
const char *e = getenv("EMBED_CUDA_GEMM_MODE");
if (e && parse_gemm_compute(e, &g_gemm_compute) == 0)
g_gemm_compute_set = 1;
env_init = 1;
}
return g_gemm_compute;
}
// y = w @ x (+ beta*y). With BF16 weights both GEMM inputs are BF16 and
// accumulation/output stay F32: when x_is_bf16 the activation operand was
// already produced in BF16 by the upstream kernel, otherwise the F32 operand
// is cast into ctx->act_bf16 first. With F32 weights this is the exact path,
// honoring gemm_compute().
static int linear_ex(embed_cuda_ctx_t *ctx, const cuda_matrix_t *w,
const void *x, int x_is_bf16, float *y, int rows,
int in_dim, int out_dim, int x_stride, float beta)
{
const float alpha = 1.0f;
if (w->bf16) {
const void *xb = x;
if (!x_is_bf16) {
size_t n = (size_t)rows * (size_t)x_stride;
int threads = 256;
cast_f32_to_bf16_kernel<<<(n + threads - 1) / threads, threads, 0,
ctx->stream>>>(
(__nv_bfloat16 *)ctx->act_bf16, (const float *)x, n);
if (launch_check() != 0) return -1;
xb = ctx->act_bf16;
}
CUBLAS_CHECK(cublasGemmEx(ctx->blas, CUBLAS_OP_T, CUBLAS_OP_N,
out_dim, rows, in_dim,
&alpha, w->d, CUDA_R_16BF, in_dim,
xb, CUDA_R_16BF, x_stride,
&beta, y, CUDA_R_32F, out_dim,
CUBLAS_COMPUTE_32F, CUBLAS_GEMM_DEFAULT));
return 0;
}
CUBLAS_CHECK(cublasGemmEx(ctx->blas, CUBLAS_OP_T, CUBLAS_OP_N,
out_dim, rows, in_dim,
&alpha, w->d, CUDA_R_32F, in_dim,
x, CUDA_R_32F, x_stride,
&beta, y, CUDA_R_32F, out_dim,
gemm_compute(), CUBLAS_GEMM_DEFAULT));
return 0;
}
static int linear(embed_cuda_ctx_t *ctx, const cuda_matrix_t *w,
const float *x, float *y, int rows, int in_dim, int out_dim)
{
return linear_ex(ctx, w, x, 0, y, rows, in_dim, out_dim, in_dim, 0.0f);
}
// x points at activations already stored as BF16 (e.g. ctx->act_bf16).
static int linear_bf16x(embed_cuda_ctx_t *ctx, const cuda_matrix_t *w,
const void *x, float *y, int rows, int in_dim,
int out_dim, int x_stride, float beta)
{
return linear_ex(ctx, w, x, 1, y, rows, in_dim, out_dim, x_stride, beta);
}
static int linear_accum(embed_cuda_ctx_t *ctx, const cuda_matrix_t *w,
const float *x, float *y,
int rows, int in_dim, int out_dim)
{
return linear_ex(ctx, w, x, 0, y, rows, in_dim, out_dim, in_dim, 1.0f);
}
static int gemm_strided_batched(cublasHandle_t blas, cublasOperation_t transa,
cublasOperation_t transb, int m, int n, int k,
const float *alpha, const float *A, int lda,
long long strideA, const float *B, int ldb,
long long strideB, const float *beta,
float *C, int ldc, long long strideC,
int batch_count)
{
cublasComputeType_t ct = gemm_compute();
if (ct == CUBLAS_COMPUTE_32F) {
CUBLAS_CHECK(cublasSgemmStridedBatched(
blas, transa, transb, m, n, k, alpha, A, lda, strideA, B, ldb,
strideB, beta, C, ldc, strideC, batch_count));
} else {
CUBLAS_CHECK(cublasGemmStridedBatchedEx(
blas, transa, transb, m, n, k,
alpha, A, CUDA_R_32F, lda, strideA,
B, CUDA_R_32F, ldb, strideB,
beta, C, CUDA_R_32F, ldc, strideC,
batch_count, ct, CUBLAS_GEMM_DEFAULT));
}
return 0;
}
// Attention for one micro-batch. Expands K/V to the per-query-head layout,
// then for each sequence runs batched-over-heads Q@K^T (scaled) -> row softmax
// -> @V via cuBLAS. Requires ctx->kexp, ctx->vexp ([total x q_dim]) and
// ctx->attn_scores ([n_heads x max_seq x max_seq]) to be allocated. h_offsets
// is the host copy of the packed sequence boundaries.
// Prepare the equal-length batched-attention path for one forward pass:
// grow the score/probability scratch to hold attn_G sequences at once and
// upload the pointer arrays for the two batched GEMMs (buffer addresses are
// layer-invariant). attn_G stays 0 - per-sequence fallback - for ragged
// batches, single sequences, or when one sequence's scores exceed the
// budget.
static int attn_batched_setup(embed_cuda_ctx_t *ctx, int batch, int q_offset,
int qkv_dim)
{
const embed_config_t *c = &ctx->config;
ctx->attn_G = 0;
if (batch < 2) return 0;
int L = ctx->offsets_host[1] - ctx->offsets_host[0];
if (L > CUDA_ATTN_BATCHED_MAX_LEN) return 0;
for (int b = 1; b < batch; b++)
if (ctx->offsets_host[b + 1] - ctx->offsets_host[b] != L) return 0;
int H = c->n_heads, hd = c->head_dim, q_dim = c->q_dim;
long long per = (long long)H * L * L;
if (per <= 0 || per > CUDA_ATTN_SCORES_BUDGET / 2) return 0;
int G = (int)(CUDA_ATTN_SCORES_BUDGET / per);
if (G > batch) G = batch;
if (G < 2) return 0;
int bf16 = g_weights_bf16;
long long want = (long long)G * per;
if (want > ctx->attn_scores_elems) {
cudaFree(ctx->attn_scores);
ctx->attn_scores = NULL;
ctx->attn_scores_elems = 0;
CUDA_CHECK(cudaMalloc((void **)&ctx->attn_scores,
(size_t)want * sizeof(float)));
ctx->attn_scores_elems = want;
}
if (bf16 && want > ctx->attn_probs_elems) {
cudaFree(ctx->attn_probs);
ctx->attn_probs = NULL;
ctx->attn_probs_elems = 0;
CUDA_CHECK(cudaMalloc(&ctx->attn_probs,
(size_t)want * sizeof(__nv_bfloat16)));
ctx->attn_probs_elems = want;
}
int entries = 6 * batch * H;
if (entries > ctx->attn_ptrs_cap) {
cudaFree((void *)ctx->attn_ptrs);
free((void *)ctx->attn_ptrs_host);
ctx->attn_ptrs = NULL;
ctx->attn_ptrs_cap = 0;
ctx->attn_ptrs_host =
(const void **)malloc((size_t)entries * sizeof(void *));
if (!ctx->attn_ptrs_host) return -1;
CUDA_CHECK(cudaMalloc((void **)&ctx->attn_ptrs,
(size_t)entries * sizeof(void *)));
ctx->attn_ptrs_cap = entries;
}
const void **K = ctx->attn_ptrs_host;
const void **Q = K + (size_t)batch * H;
const void **S = Q + (size_t)batch * H;
const void **V = S + (size_t)batch * H;
const void **P = V + (size_t)batch * H;
const void **O = P + (size_t)batch * H;
const __nv_bfloat16 *vb = (const __nv_bfloat16 *)ctx->vexp;
const __nv_bfloat16 *pb = (const __nv_bfloat16 *)ctx->attn_probs;
const __nv_bfloat16 *ab = (const __nv_bfloat16 *)ctx->act_bf16;
for (int b = 0; b < batch; b++) {
// Score/probability rows cycle every G sequences (chunked reuse).
long long srow = (long long)(b % G) * per;
size_t tok = (size_t)b * L;
for (int h = 0; h < H; h++) {
size_t i = (size_t)b * H + h;
K[i] = ctx->kexp + tok * q_dim + (size_t)h * hd;
Q[i] = ctx->qkv + tok * qkv_dim + q_offset + (size_t)h * hd;
S[i] = ctx->attn_scores + srow + (long long)h * L * L;
if (bf16) {
V[i] = vb + tok * q_dim + (size_t)h * hd;
P[i] = pb + srow + (long long)h * L * L;
O[i] = ab + tok * q_dim + (size_t)h * hd;
} else {
V[i] = ctx->vexp + tok * q_dim + (size_t)h * hd;
P[i] = S[i];
O[i] = ctx->attn_out + tok * q_dim + (size_t)h * hd;
}
}
}
CUDA_CHECK(cudaMemcpy((void *)ctx->attn_ptrs, ctx->attn_ptrs_host,
(size_t)entries * sizeof(void *),
cudaMemcpyHostToDevice));
ctx->attn_G = G;
ctx->attn_L = L;
return 0;
}
// out_bf16 != NULL selects the BF16-weight layout: V and the softmax
// probabilities are stored BF16 and the @V GEMM emits BF16 attention output
// at out_bf16 (stride q_dim), ready for the wo projection without a cast.
static int cuda_attention_gemm(embed_cuda_ctx_t *ctx, const int *h_offsets,
int batch, int q_offset, int k_offset,
int v_offset, int qkv_dim, float scale,
__nv_bfloat16 *out_bf16)
{
const embed_config_t *c = &ctx->config;
int q_dim = c->q_dim;
int hd = c->head_dim;
int H = c->n_heads;
int total = h_offsets[batch];
int causal = c->attention_mode == EMBED_ATTENTION_CAUSAL;
const float alpha = scale, beta = 0.0f, one = 1.0f;
int threads = 256;
size_t exp_count = (size_t)total * q_dim;
if (out_bf16) {
attn_expand_kv_kernel<<<(exp_count + threads - 1) / threads, threads,
0, ctx->stream>>>(
ctx->kexp, (__nv_bfloat16 *)ctx->vexp, ctx->qkv, total, H,
c->n_kv_heads, hd, qkv_dim, k_offset, v_offset);
} else {
attn_expand_kv_kernel<<<(exp_count + threads - 1) / threads, threads,
0, ctx->stream>>>(
ctx->kexp, ctx->vexp, ctx->qkv, total, H, c->n_kv_heads, hd,
qkv_dim, k_offset, v_offset);
}
if (launch_check() != 0) return -1;
if (ctx->attn_G >= 2) {
int G = ctx->attn_G, L = ctx->attn_L;
const void *const *K = (const void *const *)ctx->attn_ptrs;
const void *const *Q = K + (size_t)batch * H;
void *const *S = (void *const *)(Q + (size_t)batch * H);
const void *const *V = (const void *const *)S + (size_t)batch * H;
const void *const *P = V + (size_t)batch * H;
void *const *O = (void *const *)(P + (size_t)batch * H);
cudaDataType pv = out_bf16 ? CUDA_R_16BF : CUDA_R_32F;
cublasComputeType_t av_ct =
out_bf16 ? CUBLAS_COMPUTE_32F : gemm_compute();
for (int s = 0; s < batch; s += G) {
int g = batch - s < G ? batch - s : G;
int n = g * H;
size_t off = (size_t)s * H;
/* scores = scale * Q @ K^T, batched over (seq, head) */
CUBLAS_CHECK(cublasGemmBatchedEx(
ctx->blas, CUBLAS_OP_T, CUBLAS_OP_N, L, L, hd,
&alpha, K + off, CUDA_R_32F, q_dim,
Q + off, CUDA_R_32F, qkv_dim,
&beta, S + off, CUDA_R_32F, L,
n, gemm_compute(), CUBLAS_GEMM_DEFAULT));
if (out_bf16) {
if (causal)
attn_causal_softmax_kernel<<<n * L, 256, 0, ctx->stream>>>(
ctx->attn_scores,
(__nv_bfloat16 *)ctx->attn_probs, L);
else
attn_softmax_kernel<<<n * L, 256, 0, ctx->stream>>>(
ctx->attn_scores,
(__nv_bfloat16 *)ctx->attn_probs, L);
} else {
if (causal)
attn_causal_softmax_kernel<<<n * L, 256, 0, ctx->stream>>>(
ctx->attn_scores, ctx->attn_scores, L);
else
attn_softmax_kernel<<<n * L, 256, 0, ctx->stream>>>(
ctx->attn_scores, ctx->attn_scores, L);
}
if (launch_check() != 0) return -1;
/* O = P @ V, batched over (seq, head) */
CUBLAS_CHECK(cublasGemmBatchedEx(
ctx->blas, CUBLAS_OP_N, CUBLAS_OP_N, hd, L, L,
&one, V + off, pv, q_dim,
P + off, pv, L,
&beta, O + off, pv, q_dim,
n, av_ct, CUBLAS_GEMM_DEFAULT));
}
return 0;
}
for (int b = 0; b < batch; b++) {
int start = h_offsets[b];
int L = h_offsets[b + 1] - start;
const float *Q = ctx->qkv + (size_t)start * qkv_dim + q_offset;
const float *K = ctx->kexp + (size_t)start * q_dim;
/* scores[h] = scale * Q[h] @ K[h]^T (row-major C = A @ B^T) */
if (gemm_strided_batched(
ctx->blas, CUBLAS_OP_T, CUBLAS_OP_N, L, L, hd,
&alpha, K, q_dim, hd, Q, qkv_dim, hd,
&beta, ctx->attn_scores, L, (long long)L * L, H) != 0)
return -1;
if (out_bf16) {
__nv_bfloat16 *P = (__nv_bfloat16 *)ctx->attn_probs;
const __nv_bfloat16 *V =
(const __nv_bfloat16 *)ctx->vexp + (size_t)start * q_dim;
__nv_bfloat16 *O = out_bf16 + (size_t)start * q_dim;
if (causal)
attn_causal_softmax_kernel<<<H * L, 256, 0, ctx->stream>>>(
ctx->attn_scores, P, L);
else
attn_softmax_kernel<<<H * L, 256, 0, ctx->stream>>>(
ctx->attn_scores, P, L);
if (launch_check() != 0) return -1;
/* O[h] = P[h] @ V[h] (row-major C = A @ B) */
CUBLAS_CHECK(cublasGemmStridedBatchedEx(
ctx->blas, CUBLAS_OP_N, CUBLAS_OP_N, hd, L, L,
&one, V, CUDA_R_16BF, q_dim, hd,
P, CUDA_R_16BF, L, (long long)L * L,
&beta, O, CUDA_R_16BF, q_dim, hd,
H, CUBLAS_COMPUTE_32F, CUBLAS_GEMM_DEFAULT));
} else {
const float *V = ctx->vexp + (size_t)start * q_dim;
float *O = ctx->attn_out + (size_t)start * q_dim;
if (causal)
attn_causal_softmax_kernel<<<H * L, 256, 0, ctx->stream>>>(
ctx->attn_scores, ctx->attn_scores, L);
else
attn_softmax_kernel<<<H * L, 256, 0, ctx->stream>>>(
ctx->attn_scores, ctx->attn_scores, L);
if (launch_check() != 0) return -1;
/* O[h] = P[h] @ V[h] (row-major C = A @ B) */
if (gemm_strided_batched(
ctx->blas, CUBLAS_OP_N, CUBLAS_OP_N, hd, L, L,
&one, V, q_dim, hd, ctx->attn_scores, L,
(long long)L * L, &beta, O, q_dim, hd, H) != 0)
return -1;
}
}
return 0;
}
static int ensure_buffers(embed_cuda_ctx_t *ctx, int total, int batch, int max_seq)
{
const embed_config_t *c = &ctx->config;
if (total > ctx->seq_cap) {
cudaFree(ctx->x);
cudaFree(ctx->x_norm);
cudaFree(ctx->qkv);
cudaFree(ctx->attn_out);
cudaFree(ctx->ffn_gate_up);
cudaFree(ctx->token_ids);
cudaFree(ctx->positions);
cudaFree(ctx->kexp);
cudaFree(ctx->vexp);
cudaFree(ctx->act_bf16);
ctx->x = ctx->x_norm = ctx->qkv = NULL;
ctx->attn_out = ctx->ffn_gate_up = NULL;
ctx->token_ids = ctx->positions = NULL;
ctx->kexp = ctx->vexp = NULL;
ctx->act_bf16 = NULL;
CUDA_CHECK(cudaMalloc((void **)&ctx->x, (size_t)total * c->hidden_size * sizeof(float)));
CUDA_CHECK(cudaMalloc((void **)&ctx->x_norm, (size_t)total * c->hidden_size * sizeof(float)));
CUDA_CHECK(cudaMalloc((void **)&ctx->qkv, (size_t)total * (c->q_dim + 2 * c->kv_dim) * sizeof(float)));
CUDA_CHECK(cudaMalloc((void **)&ctx->attn_out, (size_t)total * c->q_dim * sizeof(float)));
CUDA_CHECK(cudaMalloc((void **)&ctx->ffn_gate_up,
(size_t)total * 2 * c->intermediate_size * sizeof(float)));
CUDA_CHECK(cudaMalloc((void **)&ctx->token_ids, (size_t)total * sizeof(int)));
CUDA_CHECK(cudaMalloc((void **)&ctx->positions, (size_t)total * sizeof(int)));
CUDA_CHECK(cudaMalloc((void **)&ctx->kexp, (size_t)total * c->q_dim * sizeof(float)));
CUDA_CHECK(cudaMalloc((void **)&ctx->vexp, (size_t)total * c->q_dim * sizeof(float)));
if (g_weights_bf16)
CUDA_CHECK(cudaMalloc(&ctx->act_bf16,
(size_t)total * 2 * c->intermediate_size * sizeof(__nv_bfloat16)));
ctx->seq_cap = total;
}
if (batch + 1 > ctx->batch_cap) {
cudaFree(ctx->offsets);
free(ctx->offsets_host);
CUDA_CHECK(cudaMalloc((void **)&ctx->offsets, (size_t)(batch + 1) * sizeof(int)));
ctx->offsets_host = (int *)malloc((size_t)(batch + 1) * sizeof(int));
if (!ctx->offsets_host) return -1;
ctx->batch_cap = batch + 1;
}
if (max_seq > ctx->max_seq_cap) {
cudaFree(ctx->rope_cos);
cudaFree(ctx->rope_sin);
long long score_elems = (long long)c->n_heads * max_seq * max_seq;
if (score_elems > ctx->attn_scores_elems) {
cudaFree(ctx->attn_scores);
ctx->attn_scores = NULL;
ctx->attn_scores_elems = 0;
CUDA_CHECK(cudaMalloc((void **)&ctx->attn_scores,
(size_t)score_elems * sizeof(float)));
ctx->attn_scores_elems = score_elems;
}
if (g_weights_bf16 && score_elems > ctx->attn_probs_elems) {
cudaFree(ctx->attn_probs);
ctx->attn_probs = NULL;
ctx->attn_probs_elems = 0;
CUDA_CHECK(cudaMalloc(&ctx->attn_probs,
(size_t)score_elems *
sizeof(__nv_bfloat16)));
ctx->attn_probs_elems = score_elems;
}
float *hcos = (float *)malloc((size_t)max_seq * c->head_dim * sizeof(float));
float *hsin = (float *)malloc((size_t)max_seq * c->head_dim * sizeof(float));
if (!hcos || !hsin) {
free(hcos);
free(hsin);
return -1;
}
int half = c->head_dim / 2;
for (int pos = 0; pos < max_seq; pos++) {
for (int d = 0; d < half; d++) {
float inv = 1.0f / powf(c->rope_theta,
(float)(2 * d) / (float)c->head_dim);
float angle = (float)pos * inv;
float cv = cosf(angle);
float sv = sinf(angle);
hcos[(size_t)pos * c->head_dim + d] = cv;
hcos[(size_t)pos * c->head_dim + half + d] = cv;
hsin[(size_t)pos * c->head_dim + d] = sv;
hsin[(size_t)pos * c->head_dim + half + d] = sv;
}
}
CUDA_CHECK(cudaMalloc((void **)&ctx->rope_cos,
(size_t)max_seq * c->head_dim * sizeof(float)));
CUDA_CHECK(cudaMalloc((void **)&ctx->rope_sin,
(size_t)max_seq * c->head_dim * sizeof(float)));
CUDA_CHECK(cudaMemcpyAsync(ctx->rope_cos, hcos,
(size_t)max_seq * c->head_dim * sizeof(float),
cudaMemcpyHostToDevice, ctx->stream));
CUDA_CHECK(cudaMemcpyAsync(ctx->rope_sin, hsin,
(size_t)max_seq * c->head_dim * sizeof(float),
cudaMemcpyHostToDevice, ctx->stream));
free(hcos);
free(hsin);
ctx->max_seq_cap = max_seq;
}
return 0;
}
static int ensure_pooled_rows(embed_cuda_ctx_t *ctx, int rows)
{
if (rows <= ctx->pooled_rows_cap) return 0;
const embed_config_t *c = &ctx->config;
cudaFree(ctx->pooled_out);
ctx->pooled_out = NULL;
CUDA_CHECK(cudaMalloc((void **)&ctx->pooled_out,
(size_t)rows * c->hidden_size * sizeof(float)));
ctx->pooled_rows_cap = rows;
return 0;
}
static int ensure_span_buffers(embed_cuda_ctx_t *ctx, int n_spans)
{
if (n_spans <= ctx->span_cap) return 0;
cudaFree(ctx->span_starts);
cudaFree(ctx->span_lens);
ctx->span_starts = NULL;
ctx->span_lens = NULL;
CUDA_CHECK(cudaMalloc((void **)&ctx->span_starts,
(size_t)n_spans * sizeof(int)));
CUDA_CHECK(cudaMalloc((void **)&ctx->span_lens,
(size_t)n_spans * sizeof(int)));
ctx->span_cap = n_spans;
return 0;
}
static int load_layer(cuda_layer_t *dst, const embed_layer_t *src,
const embed_config_t *c)
{
return
load_qkv_matrix(&dst->qkv, src, c) ||
load_matrix(&dst->wo, &src->wo, c->hidden_size, c->q_dim) ||
load_vector(&dst->q_norm, src->q_norm, c->head_dim) ||
load_vector(&dst->k_norm, src->k_norm, c->head_dim) ||
load_vector(&dst->input_norm, src->input_norm, c->hidden_size) ||
load_vector(&dst->post_attn_norm, src->post_attn_norm, c->hidden_size) ||
load_gate_up_matrix(&dst->gate_up_proj, src, c) ||
load_matrix(&dst->down_proj, &src->down_proj,
c->hidden_size, c->intermediate_size);
}
static void free_layer(cuda_layer_t *l)
{
cuda_matrix_free(&l->qkv);
cuda_matrix_free(&l->wo);
cudaFree(l->q_norm);
cudaFree(l->k_norm);
cudaFree(l->input_norm);
cudaFree(l->post_attn_norm);
cuda_matrix_free(&l->gate_up_proj);
cuda_matrix_free(&l->down_proj);
}
embed_cuda_ctx_t *embed_cuda_load(const char *model_dir)
{
embed_model_t *cpu = embed_model_load(model_dir);
if (!cpu) return NULL;
embed_cuda_ctx_t *ctx = (embed_cuda_ctx_t *)calloc(1, sizeof(*ctx));
if (!ctx) {
embed_model_free(cpu);
return NULL;
}
ctx->cpu = cpu;
ctx->config = cpu->config;
/* Without an explicit --cuda-weight-dtype, store weights in the model
* file's dtype: BF16 snapshots load as BF16 (bit-exact pass-through),
* F32 snapshots keep the exact F32 default. */
if (!g_weights_bf16_set) {
g_weights_bf16 = cpu->weights.embed_tokens.dtype == DTYPE_BF16;
if (g_weights_bf16)
fprintf(stderr, "cuda: BF16 model file; storing weights as BF16 "
"(use --cuda-weight-dtype f32 to override)\n");
}
if (cudaStreamCreate(&ctx->stream) != cudaSuccess ||
cublasCreate(&ctx->blas) != CUBLAS_STATUS_SUCCESS ||
cublasSetStream(ctx->blas, ctx->stream) != CUBLAS_STATUS_SUCCESS) {
embed_cuda_free(ctx);
return NULL;
}
const embed_config_t *c = &ctx->config;
ctx->layers = (cuda_layer_t *)calloc((size_t)c->n_layers, sizeof(*ctx->layers));
if (!ctx->layers) {
embed_cuda_free(ctx);
return NULL;
}
if (load_matrix(&ctx->embed_tokens, &cpu->weights.embed_tokens,
c->vocab_size, c->hidden_size) != 0 ||
load_vector(&ctx->norm, cpu->weights.norm, c->hidden_size) != 0) {
embed_cuda_free(ctx);
return NULL;
}
for (int i = 0; i < c->n_layers; i++) {
if (load_layer(&ctx->layers[i], &cpu->weights.layers[i], c) != 0) {
fprintf(stderr, "cuda: failed to load layer %d\n", i);
embed_cuda_free(ctx);
return NULL;
}
}
cudaStreamSynchronize(ctx->stream);
return ctx;
}
void embed_cuda_free(embed_cuda_ctx_t *ctx)
{
if (!ctx) return;
cuda_matrix_free(&ctx->embed_tokens);
if (ctx->layers) {
for (int i = 0; i < ctx->config.n_layers; i++)
free_layer(&ctx->layers[i]);
free(ctx->layers);
}
cudaFree(ctx->norm);
cudaFree(ctx->x);
cudaFree(ctx->x_norm);
cudaFree(ctx->qkv);
cudaFree(ctx->attn_out);
cudaFree(ctx->ffn_gate_up);
cudaFree(ctx->pooled_out);
cudaFree(ctx->rope_cos);
cudaFree(ctx->rope_sin);
cudaFree(ctx->token_ids);
cudaFree(ctx->offsets);
free(ctx->offsets_host);
cudaFree(ctx->positions);
cudaFree(ctx->kexp);
cudaFree(ctx->vexp);
cudaFree(ctx->act_bf16);
cudaFree(ctx->attn_scores);
cudaFree(ctx->attn_probs);
cudaFree((void *)ctx->attn_ptrs);
free((void *)ctx->attn_ptrs_host);
cudaFree(ctx->span_starts);
cudaFree(ctx->span_lens);
if (ctx->blas) cublasDestroy(ctx->blas);
if (ctx->stream) cudaStreamDestroy(ctx->stream);
embed_model_free(ctx->cpu);
free(ctx);
}
const embed_config_t *embed_cuda_config(const embed_cuda_ctx_t *ctx)
{
return ctx ? &ctx->config : NULL;
}
static int cuda_forward_batch(embed_cuda_ctx_t *ctx, const embed_input_t *inputs,
int batch)
{
if (!ctx || !inputs || batch <= 0) return -1;
const embed_config_t *c = &ctx->config;
int *h_offsets = (int *)malloc((size_t)(batch + 1) * sizeof(int));
if (!h_offsets) return -1;
int total = 0, max_seq = 0;
for (int b = 0; b < batch; b++) {
if (!inputs[b].ids || inputs[b].n_tokens <= 0) {
free(h_offsets);
return -1;
}
h_offsets[b] = total;
total += inputs[b].n_tokens;
if (inputs[b].n_tokens > max_seq) max_seq = inputs[b].n_tokens;
}
h_offsets[batch] = total;
int *h_ids = (int *)malloc((size_t)total * sizeof(int));
int *h_pos = (int *)malloc((size_t)total * sizeof(int));
if (!h_offsets || !h_ids || !h_pos) {
free(h_offsets); free(h_ids); free(h_pos);
return -1;
}
for (int b = 0; b < batch; b++) {
int off = h_offsets[b];
memcpy(h_ids + off, inputs[b].ids,
(size_t)inputs[b].n_tokens * sizeof(int));
for (int i = 0; i < inputs[b].n_tokens; i++)
h_pos[off + i] = i;
}
if (ensure_buffers(ctx, total, batch, max_seq) != 0) {
free(h_offsets); free(h_ids); free(h_pos);
return -1;
}
memcpy(ctx->offsets_host, h_offsets, (size_t)(batch + 1) * sizeof(int));
cudaError_t ce = cudaMemcpyAsync(ctx->token_ids, h_ids,
(size_t)total * sizeof(int),
cudaMemcpyHostToDevice, ctx->stream);
if (ce == cudaSuccess)
ce = cudaMemcpyAsync(ctx->positions, h_pos,
(size_t)total * sizeof(int),
cudaMemcpyHostToDevice, ctx->stream);
if (ce == cudaSuccess)
ce = cudaMemcpyAsync(ctx->offsets, h_offsets,
(size_t)(batch + 1) * sizeof(int),
cudaMemcpyHostToDevice, ctx->stream);
free(h_offsets); free(h_ids); free(h_pos);
if (ce != cudaSuccess) {
fprintf(stderr, "cuda input copy failed: %s\n", cudaGetErrorString(ce));
return -1;
}
int threads = 256;
int blocks_hidden = (total * c->hidden_size + threads - 1) / threads;
embed_lookup_kernel<<<blocks_hidden, threads, 0, ctx->stream>>>(
ctx->x, ctx->token_ids, ctx->embed_tokens.d, ctx->embed_tokens.bf16,
total, c->hidden_size, c->vocab_size);
if (launch_check() != 0) return -1;
float scale = 1.0f / sqrtf((float)c->head_dim);
int q_offset = 0;
int k_offset = c->q_dim;
int v_offset = c->q_dim + c->kv_dim;
int qkv_dim = c->q_dim + 2 * c->kv_dim;
// With BF16 weights every producer (norm, SiLU, attention @V GEMM)
// stores its result as BF16 directly into ctx->act_bf16, which the next
// projection GEMM consumes before any later producer reuses the buffer
// (single stream). No cast launches remain in the layer loop.
__nv_bfloat16 *act = (__nv_bfloat16 *)ctx->act_bf16;
if (attn_batched_setup(ctx, batch, q_offset, qkv_dim) != 0)
return -1;
for (int layer = 0; layer < c->n_layers; layer++) {
cuda_layer_t *l = &ctx->layers[layer];
int wbf16 = l->qkv.bf16;
if (wbf16) {
rms_norm_kernel<<<total, 256, 0, ctx->stream>>>(
act, ctx->x, l->input_norm, total,
c->hidden_size, c->rms_norm_eps);
if (launch_check() != 0) return -1;
if (linear_bf16x(ctx, &l->qkv, act, ctx->qkv,
total, c->hidden_size, qkv_dim,
c->hidden_size, 0.0f) != 0)
return -1;
} else {
rms_norm_kernel<<<total, 256, 0, ctx->stream>>>(
ctx->x_norm, ctx->x, l->input_norm, total,
c->hidden_size, c->rms_norm_eps);
if (launch_check() != 0) return -1;
if (linear(ctx, &l->qkv, ctx->x_norm, ctx->qkv,
total, c->hidden_size, qkv_dim) != 0)
return -1;
}
rms_norm_rope_qk_kernel<<<dim3(total, c->n_heads + c->n_kv_heads),
128, 0, ctx->stream>>>(
ctx->qkv, l->q_norm, l->k_norm, ctx->positions, ctx->rope_cos,
ctx->rope_sin, total, c->n_heads, c->n_kv_heads, c->head_dim,
qkv_dim, q_offset, k_offset, c->rms_norm_eps);
if (launch_check() != 0) return -1;
if (cuda_attention_gemm(ctx, ctx->offsets_host, batch, q_offset,
k_offset, v_offset, qkv_dim, scale,
wbf16 ? act : NULL) != 0)
return -1;
if (wbf16) {
if (linear_bf16x(ctx, &l->wo, act, ctx->x,
total, c->q_dim, c->hidden_size,
c->q_dim, 1.0f) != 0)
return -1;
} else {
if (linear_accum(ctx, &l->wo, ctx->attn_out, ctx->x,
total, c->q_dim, c->hidden_size) != 0)
return -1;
}
int inter_count = total * c->intermediate_size;
if (wbf16) {
rms_norm_kernel<<<total, 256, 0, ctx->stream>>>(
act, ctx->x, l->post_attn_norm, total,
c->hidden_size, c->rms_norm_eps);
if (launch_check() != 0) return -1;
if (linear_bf16x(ctx, &l->gate_up_proj, act, ctx->ffn_gate_up,
total, c->hidden_size, 2 * c->intermediate_size,
c->hidden_size, 0.0f) != 0)
return -1;
silu_mul_packed_kernel<<<(inter_count + threads - 1) / threads,
threads, 0, ctx->stream>>>(
act, c->intermediate_size, ctx->ffn_gate_up, total,
c->intermediate_size);
if (launch_check() != 0) return -1;
if (linear_bf16x(ctx, &l->down_proj, act, ctx->x,
total, c->intermediate_size, c->hidden_size,
c->intermediate_size, 1.0f) != 0)
return -1;
} else {
rms_norm_kernel<<<total, 256, 0, ctx->stream>>>(
ctx->x_norm, ctx->x, l->post_attn_norm, total,
c->hidden_size, c->rms_norm_eps);
if (launch_check() != 0) return -1;
if (linear(ctx, &l->gate_up_proj, ctx->x_norm,
ctx->ffn_gate_up, total, c->hidden_size,
2 * c->intermediate_size) != 0)
return -1;
silu_mul_packed_kernel<<<(inter_count + threads - 1) / threads,
threads, 0, ctx->stream>>>(
ctx->ffn_gate_up, 2 * c->intermediate_size, ctx->ffn_gate_up,
total, c->intermediate_size);
if (launch_check() != 0) return -1;
if (linear_ex(ctx, &l->down_proj, ctx->ffn_gate_up, 0, ctx->x,
total, c->intermediate_size, c->hidden_size,
2 * c->intermediate_size, 1.0f) != 0)
return -1;
}
}
rms_norm_kernel<<<total, 256, 0, ctx->stream>>>(
ctx->x, ctx->x, ctx->norm, total, c->hidden_size, c->rms_norm_eps);
if (launch_check() != 0) return -1;
return 0;
}
int embed_cuda_encode_batch(embed_cuda_ctx_t *ctx, const embed_input_t *inputs,
int batch, float *out_embeddings)
{
if (!ctx || !inputs || batch <= 0 || !out_embeddings) return -1;
const embed_config_t *c = &ctx->config;
if (cuda_forward_batch(ctx, inputs, batch) != 0)
return -1;
if (ensure_pooled_rows(ctx, batch) != 0)
return -1;
if (c->pooling_mode == EMBED_POOL_LAST_TOKEN)
last_pool_kernel<<<batch, 256, 0, ctx->stream>>>(
ctx->pooled_out, ctx->x, ctx->offsets, batch, c->hidden_size);
else
mean_pool_kernel<<<batch, 256, 0, ctx->stream>>>(
ctx->pooled_out, ctx->x, ctx->offsets, batch, c->hidden_size);
if (launch_check() != 0) return -1;
cudaError_t ce = cudaMemcpyAsync(out_embeddings, ctx->pooled_out,
(size_t)batch * c->hidden_size * sizeof(float),
cudaMemcpyDeviceToHost, ctx->stream);
if (ce == cudaSuccess)
ce = cudaStreamSynchronize(ctx->stream);
if (ce != cudaSuccess) {
fprintf(stderr, "cuda output copy failed: %s\n", cudaGetErrorString(ce));
return -1;
}
if (c->normalize_embeddings) {
for (int b = 0; b < batch; b++) {
if (embed_l2_normalize(
out_embeddings + (size_t)b * c->hidden_size,
c->hidden_size) != 0)
return -1;
}
}
return 0;
}
int embed_cuda_encode_spans_batch(embed_cuda_ctx_t *ctx,
const embed_context_input_t *inputs,
int batch, float *out_embeddings)
{
if (!ctx || !inputs || batch <= 0 || !out_embeddings) return -1;
const embed_config_t *c = &ctx->config;
int total_spans = 0;
int total_tokens = 0;
for (int b = 0; b < batch; b++) {
const embed_context_input_t *input = &inputs[b];
int n_tokens = input->input.n_tokens;
if (!input->input.ids || n_tokens <= 0 || !input->spans ||
input->n_spans <= 0 || total_spans > INT_MAX - input->n_spans ||
total_tokens > INT_MAX - n_tokens)
return -1;
for (int s = 0; s < input->n_spans; s++) {
int start = input->spans[s].start;
int len = input->spans[s].n_tokens;
if (start < 0 || len <= 0 || start > n_tokens ||
len > n_tokens - start)
return -1;
}
total_spans += input->n_spans;
total_tokens += n_tokens;
}
embed_input_t *packed = (embed_input_t *)malloc((size_t)batch * sizeof(*packed));
int *h_starts = (int *)malloc((size_t)total_spans * sizeof(int));
int *h_lens = (int *)malloc((size_t)total_spans * sizeof(int));
if (!packed || !h_starts || !h_lens) {
free(packed);
free(h_starts);
free(h_lens);
return -1;
}
int token_offset = 0;
int span_offset = 0;
for (int b = 0; b < batch; b++) {
const embed_context_input_t *input = &inputs[b];
packed[b] = input->input;
for (int s = 0; s < input->n_spans; s++) {
h_starts[span_offset] = token_offset + input->spans[s].start;
h_lens[span_offset] = input->spans[s].n_tokens;
span_offset++;
}
token_offset += input->input.n_tokens;
}
int rc = -1;
cudaError_t ce;
if (cuda_forward_batch(ctx, packed, batch) != 0)
goto cleanup;
if (ensure_pooled_rows(ctx, total_spans) != 0 ||
ensure_span_buffers(ctx, total_spans) != 0)
goto cleanup;
ce = cudaMemcpyAsync(ctx->span_starts, h_starts,
(size_t)total_spans * sizeof(int),
cudaMemcpyHostToDevice, ctx->stream);
if (ce == cudaSuccess)
ce = cudaMemcpyAsync(ctx->span_lens, h_lens,
(size_t)total_spans * sizeof(int),
cudaMemcpyHostToDevice, ctx->stream);
if (ce != cudaSuccess) {
fprintf(stderr, "cuda span copy failed: %s\n", cudaGetErrorString(ce));
goto cleanup;
}
span_pool_kernel<<<total_spans, 256, 0, ctx->stream>>>(
ctx->pooled_out, ctx->x, ctx->span_starts, ctx->span_lens,
total_spans, c->hidden_size);
if (launch_check() != 0)
goto cleanup;
ce = cudaMemcpyAsync(out_embeddings, ctx->pooled_out,
(size_t)total_spans * c->hidden_size * sizeof(float),
cudaMemcpyDeviceToHost, ctx->stream);
if (ce == cudaSuccess)
ce = cudaStreamSynchronize(ctx->stream);
if (ce != cudaSuccess) {
fprintf(stderr, "cuda span output copy failed: %s\n",
cudaGetErrorString(ce));
goto cleanup;
}
rc = 0;
cleanup:
free(h_lens);
free(h_starts);
free(packed);
return rc;
}
int embed_cuda_encode_into(embed_cuda_ctx_t *ctx, const int *token_ids,
int n_tokens, float *out_embedding)
{
embed_input_t input = { token_ids, n_tokens };
return embed_cuda_encode_batch(ctx, &input, 1, out_embedding);
}
float *embed_cuda_encode(embed_cuda_ctx_t *ctx, const int *token_ids,
int n_tokens)
{
if (!ctx) return NULL;
int hidden = ctx->config.hidden_size;
float *out = (float *)malloc((size_t)hidden * sizeof(float));
if (!out) return NULL;
if (embed_cuda_encode_into(ctx, token_ids, n_tokens, out) != 0) {
free(out);
return NULL;
}
return out;
}
|