Instructions to use VladHong/llama-cpp-K2-FAST2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use VladHong/llama-cpp-K2-FAST2 with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf VladHong/llama-cpp-K2-FAST2 # Run inference directly in the terminal: llama cli -hf VladHong/llama-cpp-K2-FAST2
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf VladHong/llama-cpp-K2-FAST2 # Run inference directly in the terminal: llama cli -hf VladHong/llama-cpp-K2-FAST2
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-app # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama serve -hf VladHong/llama-cpp-K2-FAST2 # Run inference directly in the terminal: ./build/bin/llama cli -hf VladHong/llama-cpp-K2-FAST2
Use Docker
docker model run hf.co/VladHong/llama-cpp-K2-FAST2
- Llama new
- LM Studio
- Jan
- Ollama
How to use VladHong/llama-cpp-K2-FAST2 with Ollama:
ollama run hf.co/VladHong/llama-cpp-K2-FAST2
- Unsloth Desktop
- Docker Model Runner
How to use VladHong/llama-cpp-K2-FAST2 with Docker Model Runner:
docker model run hf.co/VladHong/llama-cpp-K2-FAST2
- Lemonade
How to use VladHong/llama-cpp-K2-FAST2 with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull VladHong/llama-cpp-K2-FAST2
Run and chat with the model
lemonade run user.llama-cpp-K2-FAST2-{{QUANT_TAG}}List all available models
lemonade list
- Atomic Chat
File size: 53,741 Bytes
c59fd93 | 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 | #include "common.hpp"
#include "ggml-sycl/presets.hpp"
#include "ggml.h"
#include "element_wise.hpp"
#define SYCL_GLOBAL_ID_LOOP(K, ITEM) \
for (auto i = ITEM.get_global_id(0); i < (size_t)K; i += ITEM.get_global_range(0))
#define SYCL_LOCAL_ID_CALC(ITEM, IDX) \
(ITEM.get_local_range(IDX) * ITEM.get_group(IDX) + ITEM.get_local_id(IDX))
static void acc_f32(const char * x, const char * y, float * dst, const int64_t ne,
const int64_t ne0, const int64_t ne1, const int64_t ne2,
const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,
const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
const int64_t s11, const int64_t s12, const int64_t s13, const int64_t offset) {
auto item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
const int64_t i = SYCL_LOCAL_ID_CALC(item_ct1, 2);
if (i >= ne) {
return;
}
int64_t src1_idx = i - offset;
int64_t tmp = src1_idx;
const int64_t i13 = tmp / s13;
tmp -= i13 * s13;
const int64_t i12 = tmp / s12;
tmp -= i12 * s12;
const int64_t i11 = tmp / s11;
tmp -= i11 * s11;
const int64_t i10 = tmp;
int64_t tmp_dst = i;
const int64_t i3 = tmp_dst / (ne2*ne1*ne0);
tmp_dst -= i3 * (ne2*ne1*ne0);
const int64_t i2 = tmp_dst / (ne1*ne0);
tmp_dst -= i2 * (ne1*ne0);
const int64_t i1 = tmp_dst / ne0;
tmp_dst -= i1 * ne0;
const int64_t i0 = tmp_dst;
float val = *(const float *) (x + i0*nb00 + i1*nb01 + i2*nb02 + i3*nb03);
if (src1_idx >= 0 && i10 < ne10 && i11 < ne11 && i12 < ne12 && i13 < ne13) {
val += *(const float *) (y + i10*nb10 + i11*nb11 + i12*nb12 + i13*nb13);
}
dst[i] = val;
}
/* Unary OP funcs */
template<typename T>
static __dpct_inline__ T op_sgn(T x) {
return x > static_cast<T>(0.f) ? static_cast<T>(1.f) : ((x < static_cast<T>(0.f) ? static_cast<T>(-1.f) : static_cast<T>(0.f)));
}
template<typename T>
static __dpct_inline__ T op_abs(T x) {
if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
return sycl::ext::oneapi::experimental::fabs(x); // or experimental namespace if needed
} else {
return sycl::fabs(x);
}
}
template<typename T>
static __dpct_inline__ T op_expm1(T x) {
if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
return static_cast<sycl::ext::oneapi::bfloat16>(
sycl::expm1(static_cast<float>(x))
);
} else {
return sycl::expm1(x);
}
}
template<typename T>
static __dpct_inline__ T op_elu(T x) {
return (x > static_cast<T>(0.f)) ? x : op_expm1(x);
}
template<typename T>
static __dpct_inline__ T op_erf(T x) {
if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
return static_cast<sycl::ext::oneapi::bfloat16>(
sycl::erf(static_cast<float>(x))
);
} else {
return sycl::erf(x);
}
}
template<typename T>
static __dpct_inline__ T op_gelu_erf(T x) {
const T SQRT_2_INV = static_cast<T>(0.70710678118654752440084436210484f);
return static_cast<T>(0.5f) * x * (static_cast<T>(1.0f) + op_erf(x * SQRT_2_INV));
}
template<typename T>
static __dpct_inline__ T op_gelu_quick(T x) {
const T GELU_QUICK_COEF_LOCAL = static_cast<T>(-1.702f);
return x * (static_cast<T>(1.0f) / (static_cast<T>(1.0f) + op_exp(GELU_QUICK_COEF_LOCAL * x)));
}
template<typename T>
static __dpct_inline__ T op_relu(T x) {
if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
return sycl::ext::oneapi::experimental::fmax(x, static_cast<T>(0));
} else {
return sycl::fmax(x, static_cast<T>(0));
}
}
template<typename T>
static __dpct_inline__ T op_sigmoid(T x) {
return static_cast<T>(1.0f) / (static_cast<T>(1.0f) + op_exp(-x));
}
template<typename T>
static __dpct_inline__ T op_sqrt(T x) {
if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
return sycl::ext::oneapi::experimental::sqrt(x);
} else {
return sycl::sqrt(x);
}
}
template<typename T>
static __dpct_inline__ T op_sin(T x) {
if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
return sycl::ext::oneapi::experimental::sin(x);
} else {
return sycl::sin(x);
}
}
template<typename T>
static __dpct_inline__ T op_cos(T x) {
if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
return sycl::ext::oneapi::experimental::cos(x);
} else {
return sycl::cos(x);
}
}
template<typename T>
static __dpct_inline__ T op_hardsigmoid(T x) {
if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
return sycl::ext::oneapi::experimental::fmin(
static_cast<T>(1.0f), sycl::ext::oneapi::experimental::fmax(
static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
} else {
return sycl::fmin(static_cast<T>(1.0f),
sycl::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
}
}
template<typename T>
static __dpct_inline__ T op_hardswish(T x) {
if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
return x * sycl::ext::oneapi::experimental::fmin(static_cast<T>(1.0f), sycl::ext::oneapi::experimental::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
} else {
return x * sycl::fmin(static_cast<T>(1.0f), sycl::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
}
}
template<typename T>
static __dpct_inline__ T op_log(T x) {
if (x <= static_cast<T>(0)) {
return neg_infinity<T>();
}
if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
return sycl::ext::oneapi::experimental::log(x);
} else {
return sycl::log(x);
}
}
template<typename T>
static __dpct_inline__ T op_softplus(T x) {
const float xf = (float) x;
const float ax = op_abs(xf);
const float m = sycl::fmax(xf, 0.0f);
const float y = m + sycl::log1p(sycl::exp(-ax));
return (T) y;
}
template<typename T>
static __dpct_inline__ T op_neg(T x) {
return -x;
}
template<typename T>
static __dpct_inline__ T op_step(T x) {
return (x > static_cast<T>(0.0f)) ? static_cast<T>(1.0f) : static_cast<T>(0.0f);
}
template<typename T>
static __dpct_inline__ T op_leaky_relu(T x, float negative_slope) {
T neg_slope_T = static_cast<T>(negative_slope);
if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
return sycl::ext::oneapi::experimental::fmax(x, static_cast<T>(0)) +
sycl::ext::oneapi::experimental::fmin(x, static_cast<T>(0.0f)) * neg_slope_T;
} else {
return sycl::fmax(x, static_cast<T>(0)) +
sycl::fmin(x, static_cast<T>(0.0f)) * neg_slope_T;
}
}
template<typename T>
static __dpct_inline__ T op_xielu(T x, float alpha_n, float alpha_p, float beta, float eps) {
const float xi = static_cast<float>(x);
const float gate_pos = (xi > 0.0f);
const float y_pos = alpha_p * xi * xi + beta * xi;
const float min_v_eps = sycl::fmin(xi, eps);
const float y_neg = (sycl::expm1(min_v_eps) - xi) * alpha_n + beta * xi;
const float out = gate_pos * y_pos + (1.0f - gate_pos) * y_neg;
return static_cast<T>(out);
}
template<typename T>
static __dpct_inline__ T op_sqr(T x) {
return x * x;
}
template<typename T>
static __dpct_inline__ T op_clamp(T x, float min_val, float max_val) {
return x < static_cast<T>(min_val) ? static_cast<T>(min_val) : (x > static_cast<T>(max_val) ? static_cast<T>(max_val) : x);
}
template<typename T>
static __dpct_inline__ T op_floor(T x) {
if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
return sycl::ext::oneapi::experimental::floor(x);
} else {
return sycl::floor(x);
}
}
template<typename T>
static __dpct_inline__ T op_ceil(T x) {
if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
return sycl::ext::oneapi::experimental::ceil(x);
} else {
return sycl::ceil(x);
}
}
template<typename T>
static __dpct_inline__ T op_round(T x) {
if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
return static_cast<sycl::ext::oneapi::bfloat16>(
sycl::round(static_cast<float>(x))
);
} else {
return sycl::round(x);
}
}
template<typename T>
static __dpct_inline__ T op_trunc(T x) {
if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
return sycl::ext::oneapi::experimental::trunc(x);
} else {
return sycl::trunc(x);
}
}
template<typename T, typename F>
static void unary_op_flat_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> & item_ct1, F func) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
dst[i] = func(x[i]);
}
}
template<typename T, typename F>
static void unary_op_generic_kernel(
const T * x,
T * dst,
const int k,
const sycl::uint3 ne0_fd, const sycl::uint3 ne1_fd, const sycl::uint3 ne2_fd,
const size_t nb0, const size_t nb1, const size_t nb2, const size_t nb3,
const size_t nbd0, const size_t nbd1, const size_t nbd2, const size_t nbd3,
const sycl::nd_item<1> & item_ct1,
F func) {
// 32-bit index math: k is int, so every logical index fits u32. 64-bit integer div/mod is
// emulated on Xe and dominates this kernel otherwise, and even the 32-bit divide is worth
// avoiding -- the divisors are launch-invariant, so the magic numbers are precomputed
// host-side and each division becomes a multiply-high plus a shift.
// Byte offsets are widened back to size_t only for the final address math.
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
sycl::uint2 dm = fast_div_modulo((uint32_t) i, ne0_fd);
const uint32_t i0 = dm.y();
dm = fast_div_modulo(dm.x(), ne1_fd);
const uint32_t i1 = dm.y();
dm = fast_div_modulo(dm.x(), ne2_fd);
const uint32_t i2 = dm.y();
const uint32_t i3 = dm.x();
const char * src_base = (const char *) x;
char * dst_base = (char *) dst;
const T * srcp = (const T *)(src_base + (size_t) i0*nb0 + (size_t) i1*nb1 + (size_t) i2*nb2 + (size_t) i3*nb3 );
T * dstp = (T *)(dst_base + (size_t) i0*nbd0 + (size_t) i1*nbd1 + (size_t) i2*nbd2 + (size_t) i3*nbd3);
*dstp = func(*srcp);
}
}
template<typename T>
static void unary_op_sqrt_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
dst[i] = op_sqrt(x[i]);
}
}
template<typename T>
static void unary_op_sin_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
dst[i] = op_sin(x[i]);
}
}
template<typename T>
static void unary_op_cos_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
dst[i] = op_cos(x[i]);
}
}
template<typename T>
static void unary_op_log_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
dst[i] = op_log(x[i]);
}
}
template<typename T>
static void unary_op_leaky_relu_kernel(const T * x, T * dst, const int k, float negative_slope, const sycl::nd_item<1> &item_ct1) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
dst[i] = op_leaky_relu(x[i], negative_slope);
}
}
template<typename T>
static void unary_op_xielu_kernel(const T * x, T * dst, const int k, float alpha_n, float alpha_p, float beta, float eps, const sycl::nd_item<1> &item_ct1) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
dst[i] = op_xielu(x[i], alpha_n, alpha_p, beta, eps);
}
}
template<typename T>
static void unary_op_sqr_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
dst[i] = op_sqr(x[i]);
}
}
template<typename T>
static void unary_op_clamp_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1, float min_val, float max_val) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
dst[i] = op_clamp(x[i], min_val, max_val);
}
}
template<typename T>
static void unary_op_ceil_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
dst[i] = op_ceil(x[i]);
}
}
template<typename T>
static void clamp(const T * x, T * dst, const float min, const float max, const int k,
const sycl::nd_item<1> &item_ct1) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
dst[i] = x[i] < static_cast<T>(min) ? static_cast<T>(min) : (x[i] > static_cast<T>(max) ? static_cast<T>(max) : x[i]);
}
}
template<typename T, typename F>
static void unary_gated_op_flat_kernel(const T * x, const T * g, T * dst, const uint64_t k, const sycl::nd_item<1> & item_ct1, F func) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
dst[i] = func(x[i]) * g[i];
}
}
template<typename T, typename F>
static void unary_gated_op_generic_kernel(
const T * x,
const T * g,
T * dst,
const uint64_t k,
const sycl::uint3 n_fd,
const uint64_t o0,
const uint64_t o1,
const sycl::nd_item<1> & item_ct1,
F func) {
// rows of n columns at strides o0 and o1: two halves of one fused tensor, or two tensors
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
const sycl::uint2 rc = fast_div_modulo((uint32_t) i, n_fd);
const int64_t j0 = rc.x() * o0 + rc.y();
const int64_t j1 = o0 == o1 ? j0 : rc.x() * o1 + rc.y();
dst[i] = func(x[j0]) * g[j1];
}
}
// Fused UNARY + MUL. Unlike the gated ops above, `x` and `g` are separate tensors of the
// same shape; `o0`/`o1` are their row strides in elements, so a half-view needs no repack.
// `dst` is contiguous and indexed flat. Math is done in f32, as the CPU and CUDA references do.
template<typename T, typename F>
static void unary_mul_flat_kernel(const T * x, const T * g, T * dst, const int64_t k, const sycl::nd_item<1> &item_ct1, F op) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
dst[i] = (T) (op((float) x[i]) * (float) g[i]);
}
}
template<typename T, typename F>
static void unary_mul_strided_kernel(const T * x, const T * g, T * dst, const int64_t k, const sycl::uint3 n_fd, const int64_t o0, const int64_t o1, const sycl::nd_item<1> &item_ct1, F op) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
const sycl::uint2 rc = fast_div_modulo((uint32_t) i, n_fd);
const int64_t j0 = rc.x() * o0 + rc.y();
const int64_t j1 = o0 == o1 ? j0 : rc.x() * o1 + rc.y();
dst[i] = (T) (op((float) x[j0]) * (float) g[j1]);
}
}
template<typename T, typename F>
static void unary_mul_sycl(const T * x, const T * g, T * dst, const int64_t k, const int64_t n, const int64_t o0, const int64_t o1, queue_ptr main_stream, F op) {
const size_t num_blocks = ceil_div((size_t) k, (size_t) SYCL_GLU_BLOCK_SIZE);
const sycl::nd_range<1> range(num_blocks * sycl::range<1>(SYCL_GLU_BLOCK_SIZE), sycl::range<1>(SYCL_GLU_BLOCK_SIZE));
// o0 == o1 == n makes (i/n)*o0 + (i%n) == i, so the strided kernel degenerates to the flat one
if (o0 == n && o1 == n) {
main_stream->parallel_for(range, [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
unary_mul_flat_kernel(x, g, dst, k, item_ct1, op);
});
return;
}
// 32-bit fastdiv, exact only below 2^31; ggml_sycl_can_fuse() already declined past that
GGML_ASSERT(k < ((int64_t) 1 << 31));
const sycl::uint3 n_fd = init_fastdiv_values((uint32_t) n);
main_stream->parallel_for(range, [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
unary_mul_strided_kernel(x, g, dst, k, n_fd, o0, o1, item_ct1, op);
});
}
namespace ggml_sycl_detail {
static void acc_f32_sycl(const char *x, const char *y, float *dst,
const int64_t n_elements,
const int64_t ne0, const int64_t ne1, const int64_t ne2,
const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,
const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
const int64_t s1, const int64_t s2, const int64_t s3,
const int64_t offset, queue_ptr stream) {
const int num_blocks = (n_elements + SYCL_ACC_BLOCK_SIZE - 1) / SYCL_ACC_BLOCK_SIZE;
stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE),
sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)),
[=](sycl::nd_item<3> /*item_ct1*/) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
acc_f32(x, y, dst, n_elements,
ne0, ne1, ne2,
nb00, nb01, nb02, nb03,
ne10, ne11, ne12, ne13,
nb10, nb11, nb12, nb13,
s1, s2, s3, offset);
});
}
template<typename T>
static void arange_kernel(T * dst, const int k, T start, T step,
const sycl::nd_item<1> &item_ct1) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
dst[i] = start + static_cast<T>(i) * step;
}
}
template<typename KernelInvoker, typename... Args>
static inline void dispatch_ggml_sycl_op_unary(ggml_backend_sycl_context & ctx, ggml_tensor * dst, KernelInvoker kernel_invoker, Args&&... args) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16 || dst->src[0]->type == GGML_TYPE_BF16);
GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_BF16);
GGML_ASSERT(dst->src[0]->type == dst->type);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
switch (dst->type) {
case GGML_TYPE_F16:
{
auto data_pts = cast_data<sycl::half>(dst);
kernel_invoker(data_pts.src, data_pts.dst, (int)ggml_nelements(dst->src[0]), main_stream, std::forward<Args>(args)...);
break;
}
#ifdef GGML_SYCL_HAS_BF16
case GGML_TYPE_BF16:
{
auto data_pts = cast_data<sycl::ext::oneapi::bfloat16>(dst);
kernel_invoker(data_pts.src, data_pts.dst, (int)ggml_nelements(dst->src[0]), main_stream, std::forward<Args>(args)...);
break;
}
#endif
case GGML_TYPE_F32:
{
auto data_pts = cast_data<float>(dst);
kernel_invoker(data_pts.src, data_pts.dst, (int)ggml_nelements(dst->src[0]), main_stream, std::forward<Args>(args)...);
break;
}
default:
GGML_ABORT("GGML tensor type not supported!\n");
}
}
template<typename KernelInvoker, typename... Args>
static inline void dispatch_ggml_sycl_op_fused_glu(ggml_backend_sycl_context & ctx, ggml_tensor * dst, KernelInvoker kernel_invoker, Args&&... args) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
GGML_ASSERT(dst->src[0]->type == dst->type);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
const int64_t nc = src1 ? src0->ne[0] : src0->ne[0] / 2;;
GGML_ASSERT(dst->ne[0] == nc);
GGML_ASSERT(ggml_is_contiguous_1(dst->src[0]));
GGML_ASSERT(ggml_is_contiguous(dst));
// The fused GLU kernels index with 32-bit fastdiv, which is exact only for indices below
// 2^31. A dst that large is ~8 GB at f32, and the grid sizing already narrows to 32 bits,
// so assert the bound rather than carry a second code path for it.
GGML_ASSERT(ggml_nelements(dst) < ((int64_t) 1 << 31));
const int32_t swapped = ((const int32_t *) dst->op_params)[1];
void * src0_d = src0->data;
void * src1_d = src1 ? src1->data : src0->data;
const int64_t src0_o = src0->nb[1];
const int64_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
void * dst_d = dst->data;
if (src1) {
GGML_ASSERT(ggml_is_contiguous_1(src1));
GGML_ASSERT(src1->nb[0] == ggml_element_size(src1));
GGML_ASSERT(src1->ne[0] == nc);
GGML_ASSERT(src0->type == src1->type);
}
switch (dst->type) {
case GGML_TYPE_F16:
{
sycl::half * src0_p = (sycl::half *) src0_d;
sycl::half * src1_p = (sycl::half *) src1_d;
if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}
kernel_invoker(src0_p,
src1_p,
(sycl::half *) dst_d,
ggml_nelements(dst),
nc,
src0_o / sizeof(sycl::half),
src1_o / sizeof(sycl::half),
main_stream,
std::forward<Args>(args)...);
break;
}
case GGML_TYPE_F32:
{
float * src0_p = (float *) src0_d;
float * src1_p = (float *) src1_d;
if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}
kernel_invoker(src0_p,
src1_p,
(float *) dst_d,
ggml_nelements(dst),
nc,
src0_o / sizeof(float),
src1_o / sizeof(float),
main_stream,
std::forward<Args>(args)...);
break;
}
default:
GGML_ABORT("GGML tensor type not supported!\n");
}
}
template<typename F>
static inline void ggml_sycl_op_unary(
ggml_backend_sycl_context & ctx, ggml_tensor * dst, F func) {
ggml_tensor * src0 = dst->src[0];
const int64_t ne0 = dst->ne[0];
const int64_t ne1 = dst->ne[1];
const int64_t ne2 = dst->ne[2];
const size_t nb0 = src0->nb[0];
const size_t nb1 = src0->nb[1];
const size_t nb2 = src0->nb[2];
const size_t nb3 = src0->nb[3];
const size_t nbd0 = dst->nb[0];
const size_t nbd1 = dst->nb[1];
const size_t nbd2 = dst->nb[2];
const size_t nbd3 = dst->nb[3];
// Hot unary ops (FFN/GDN silu, sigmoid, ...) run on contiguous tensors;
// skip the strided index math entirely for them.
const bool contiguous = ggml_is_contiguous(src0) && ggml_is_contiguous(dst);
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[=](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, 256);
if (contiguous) {
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
sycl::range<1>(256)),
[=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
unary_op_flat_kernel(src, dst_ptr, k_elements, item_ct1, func);
});
} else {
// Launch-invariant divisors: compute the magic numbers once on the host so the
// kernel never issues an integer divide. Only the strided path needs them.
const sycl::uint3 ne0_fd = init_fastdiv_values((uint32_t) ne0);
const sycl::uint3 ne1_fd = init_fastdiv_values((uint32_t) ne1);
const sycl::uint3 ne2_fd = init_fastdiv_values((uint32_t) ne2);
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
sycl::range<1>(256)),
[=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
unary_op_generic_kernel(
src, dst_ptr, k_elements,
ne0_fd, ne1_fd, ne2_fd,
nb0, nb1, nb2, nb3,
nbd0, nbd1, nbd2, nbd3,
item_ct1,
func
);
});
}
});
}
template<typename F>
static inline void ggml_sycl_op_unary_gated(
ggml_backend_sycl_context & ctx, ggml_tensor * dst, F func) {
dispatch_ggml_sycl_op_fused_glu(ctx, dst,
[func](const auto * x_ptr, const auto * g_ptr, auto * dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
const uint32_t num_blocks = (uint32_t) ceil_div(k, SYCL_GLU_BLOCK_SIZE);
const sycl::nd_range<1> launch_range(num_blocks * sycl::range<1>(SYCL_GLU_BLOCK_SIZE),
sycl::range<1>(SYCL_GLU_BLOCK_SIZE));
// o0 == n and o1 == n make the index math the identity, so index flat
// note: not ggml_is_contiguous - a fused [gate|up] src0 is contiguous with o0 == 2n
if (o0 == n && o1 == n) {
main_stream->parallel_for(launch_range,
[=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
unary_gated_op_flat_kernel(x_ptr, g_ptr, dst_ptr, k, item_ct1, func);
});
} else {
// launch-invariant divisor, and only this path needs it
const sycl::uint3 n_fd = init_fastdiv_values((uint32_t) n);
main_stream->parallel_for(launch_range,
[=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
unary_gated_op_generic_kernel(x_ptr, g_ptr, dst_ptr, k, n_fd, o0, o1, item_ct1, func);
});
}
});
}
static inline void ggml_sycl_op_arange(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(dst->type == GGML_TYPE_F32);
float start, stop, step;
memcpy(&start, dst->op_params, sizeof(float));
memcpy(&stop, (float *) dst->op_params + 1, sizeof(float));
memcpy(&step, (float *) dst->op_params + 2, sizeof(float));
dpct::queue_ptr stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
float * dst_ptr = (float *)dst->data;
const int k = (int)ggml_nelements(dst);
const int num_blocks = ceil_div(k, SYCL_ARANGE_BLOCK_SIZE);
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_ARANGE_BLOCK_SIZE),
sycl::range<1>(SYCL_ARANGE_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
arange_kernel(dst_ptr, k, start, step, item_ct1);
});
}
} // namespace ggml_sycl_detail
static inline void ggml_sycl_op_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_sgn(x);
});
}
static inline void ggml_sycl_op_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_abs(x);
});
}
static inline void ggml_sycl_op_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_elu(x);
});
}
static inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_silu(x);
});
}
static inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_gelu(x);
});
}
static inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_gelu_quick(x);
});
}
static inline void ggml_sycl_op_gelu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_gelu_erf(x);
});
}
static inline void ggml_sycl_op_tanh(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_tanh(x);
});
}
static inline void ggml_sycl_op_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_relu(x);
});
}
static inline void ggml_sycl_op_hardsigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_hardsigmoid(x);
});
}
static inline void ggml_sycl_op_hardswish(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_hardswish(x);
});
}
static inline void ggml_sycl_op_exp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_exp(x);
});
}
static inline void ggml_sycl_op_expm1(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_expm1(x);
});
}
static inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_EXP_BLOCK_SIZE); // Using EXP block size
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_EXP_BLOCK_SIZE),
sycl::range<1>(SYCL_EXP_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
unary_op_log_kernel(src, dst_ptr, k_elements, item_ct1);
});
});
}
static inline void ggml_sycl_op_softplus(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_softplus(x);
});
}
static inline void ggml_sycl_op_neg(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_neg(x);
});
}
static inline void ggml_sycl_op_step(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_step(x);
});
}
static inline void ggml_sycl_op_sigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_sigmoid(x);
});
}
static inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_SQRT_BLOCK_SIZE);
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SQRT_BLOCK_SIZE),
sycl::range<1>(SYCL_SQRT_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
unary_op_sqrt_kernel(src, dst_ptr, k_elements, item_ct1);
});
});
}
static inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_SIN_BLOCK_SIZE);
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SIN_BLOCK_SIZE),
sycl::range<1>(SYCL_SIN_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
unary_op_sin_kernel(src, dst_ptr, k_elements, item_ct1);
});
});
}
static inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_SIN_BLOCK_SIZE); // Using SIN block size
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SIN_BLOCK_SIZE),
sycl::range<1>(SYCL_SIN_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
unary_op_cos_kernel(src, dst_ptr, k_elements, item_ct1);
});
});
}
static inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
float negative_slope;
memcpy(&negative_slope, dst->op_params, sizeof(float));
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream, float slope) {
const int num_blocks = ceil_div(k_elements, SYCL_RELU_BLOCK_SIZE);
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_RELU_BLOCK_SIZE),
sycl::range<1>(SYCL_RELU_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
unary_op_leaky_relu_kernel(src, dst_ptr, k_elements, slope, item_ct1);
});
}, negative_slope);
}
static inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_SQR_BLOCK_SIZE);
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SQR_BLOCK_SIZE),
sycl::range<1>(SYCL_SQR_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
unary_op_sqr_kernel(src, dst_ptr, k_elements, item_ct1);
});
});
}
static inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
float min_val;
float max_val;
memcpy(&min_val, dst->op_params, sizeof(float));
memcpy(&max_val, (float *) dst->op_params + 1, sizeof(float));
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream, float min_arg, float max_arg) {
const int num_blocks = ceil_div(k_elements, SYCL_CLAMP_BLOCK_SIZE);
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_CLAMP_BLOCK_SIZE),
sycl::range<1>(SYCL_CLAMP_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
clamp(src, dst_ptr, min_arg, max_arg, k_elements, item_ct1);
});
}, min_val, max_val);
}
static inline void ggml_sycl_op_xielu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
const float alpha_n = ggml_get_op_params_f32(dst, 1);
const float alpha_p = ggml_get_op_params_f32(dst, 2);
const float beta = ggml_get_op_params_f32(dst, 3);
const float eps = ggml_get_op_params_f32(dst, 4);
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream, float alpha_n_arg, float alpha_p_arg, float beta_arg, float eps_arg) {
const int num_blocks = ceil_div(k_elements, SYCL_RELU_BLOCK_SIZE);
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_RELU_BLOCK_SIZE),
sycl::range<1>(SYCL_RELU_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
unary_op_xielu_kernel(src, dst_ptr, k_elements, alpha_n_arg, alpha_p_arg, beta_arg, eps_arg, item_ct1);
});
}, alpha_n, alpha_p, beta, eps);
}
static inline void ggml_sycl_op_floor(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_floor(x);
});
}
static inline void ggml_sycl_op_ceil(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_ceil(x);
});
}
static inline void ggml_sycl_op_round(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_round(x);
});
}
static inline void ggml_sycl_op_trunc(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
return op_trunc(x);
});
}
static inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
const char * src0_d = (const char *) src0->data;
const char * src1_d = (const char *) src1->data;
float * dst_d = (float *) dst->data;
dpct::queue_ptr stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
GGML_ASSERT(dst->nb[0] == ggml_element_size(dst));
GGML_ASSERT(ggml_is_contiguously_allocated(dst));
GGML_ASSERT(ggml_are_same_shape(src0, dst));
const int64_t s1 = (int64_t) ((const int32_t *) dst->op_params)[0] / (int64_t) sizeof(float);
const int64_t s2 = (int64_t) ((const int32_t *) dst->op_params)[1] / (int64_t) sizeof(float);
const int64_t s3 = (int64_t) ((const int32_t *) dst->op_params)[2] / (int64_t) sizeof(float);
const int64_t offset = (int64_t) ((const int32_t *) dst->op_params)[3] / (int64_t) sizeof(float);
ggml_sycl_detail::acc_f32_sycl(src0_d, src1_d, dst_d, ggml_nelements(dst),
dst->ne[0], dst->ne[1], dst->ne[2],
src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],
s1, s2, s3, offset, stream);
}
static inline void ggml_sycl_op_geglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary_gated(ctx, dst, [](auto x) {
return op_gelu(x);
});
}
static inline void ggml_sycl_op_reglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary_gated(ctx, dst, [](auto x) {
return op_relu(x);
});
}
static inline void ggml_sycl_op_swiglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary_gated(ctx, dst, [](auto x) {
return op_silu(x);
});
}
// dst = op(unary_node->src[0]) * other, written straight to the MUL output, saving the
// standalone unary launch. Preconditions come from ggml_sycl_can_fuse(); re-asserted here.
void ggml_sycl_op_unary_mul_fused(ggml_backend_sycl_context & ctx, ggml_tensor * unary_node, ggml_tensor * mul_node) {
scope_op_debug_print scope_dbg_print(__func__, mul_node, /*num_src=*/2);
const ggml_tensor * x = unary_node->src[0];
const ggml_tensor * g = (mul_node->src[0] == unary_node) ? mul_node->src[1] : mul_node->src[0];
// g is picked by elimination; ggml_can_fuse()'s single-use rule rules out MUL(unary, unary)
GGML_ASSERT(g != unary_node);
GGML_ASSERT(x->type == g->type && x->type == mul_node->type);
GGML_ASSERT(ggml_are_same_shape(x, g) && ggml_are_same_shape(x, mul_node));
GGML_ASSERT(ggml_is_contiguous_1(x) && ggml_is_contiguous_1(g));
// dst is indexed flat
GGML_ASSERT(ggml_is_contiguous(mul_node));
queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const int64_t k = ggml_nelements(mul_node);
const int64_t n = mul_node->ne[0];
const auto dispatch_type = [&](auto op) {
switch (mul_node->type) {
case GGML_TYPE_F32:
unary_mul_sycl((const float *) x->data, (const float *) g->data, (float *) mul_node->data,
k, n, x->nb[1] / sizeof(float), g->nb[1] / sizeof(float), main_stream, op);
break;
case GGML_TYPE_F16:
unary_mul_sycl((const sycl::half *) x->data, (const sycl::half *) g->data, (sycl::half *) mul_node->data,
k, n, x->nb[1] / sizeof(sycl::half), g->nb[1] / sizeof(sycl::half), main_stream, op);
break;
default:
GGML_ABORT("fused unary+mul: unsupported type %s", ggml_type_name(mul_node->type));
}
};
switch (ggml_get_unary_op(unary_node)) {
case GGML_UNARY_OP_SILU: dispatch_type([](float v) { return op_silu(v); }); break;
case GGML_UNARY_OP_SIGMOID: dispatch_type([](float v) { return op_sigmoid(v); }); break;
case GGML_UNARY_OP_SOFTPLUS: dispatch_type([](float v) { return op_softplus(v); }); break;
default:
GGML_ABORT("fused unary+mul: unsupported unary op %s", ggml_unary_op_name(ggml_get_unary_op(unary_node)));
}
}
__dpct_inline__ float ggml_sycl_op_swiglu_oai_single(float x, float g, float alpha = 1.702f, float limit = 7.0f) {
x = sycl::fmin(x, limit);
g = sycl::fmax(sycl::fmin(g, limit), -limit);
float out_glu = x / (1.0f + sycl::native::exp(-x * alpha));
out_glu = out_glu * (1.0f + g);
return out_glu;
}
template <typename T>
static void swiglu_oai_kernel(const T * x, const T * g, T * dst, const int64_t k,
const int64_t n, const int64_t o0, const int64_t o1,
float alpha, float limit, sycl::nd_item<3> item_ct1) {
const int64_t i = int64_t(item_ct1.get_local_range(2)) * item_ct1.get_group(2) + item_ct1.get_local_id(2);
if (i >= k) {
return;
}
const int64_t j0 = (i / n) * o0 + (i % n);
const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
float xi = x[j0];
float gi = g[j1];
dst[i] = ggml_sycl_op_swiglu_oai_single(xi, gi, alpha, limit);
}
template <typename T>
static void swiglu_oai_sycl(const T * x,
const T * g,
T * dst,
const int64_t k,
const int64_t n,
const int64_t o0,
const int64_t o1,
const float alpha,
const float limit,
dpct::queue_ptr stream) {
const int64_t num_blocks = (k + SYCL_GLU_BLOCK_SIZE - 1) / SYCL_GLU_BLOCK_SIZE;
stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GLU_BLOCK_SIZE),
sycl::range<3>(1, 1, SYCL_GLU_BLOCK_SIZE)),
[=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
swiglu_oai_kernel(x, g, dst, k, n, o0, o1, alpha, limit, item_ct1);
});
}
void ggml_sycl_op_swiglu_oai(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
void * src0_d = src0->data;
void * src1_d = src1 ? src1->data : src0->data;
const int64_t src0_o = src0->nb[1];
const int64_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
void * dst_d = dst->data;
const int64_t nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
dpct::queue_ptr stream = ctx.stream();
GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(src0->nb[0] == ggml_element_size(src0));
GGML_ASSERT(ggml_is_contiguous(dst));
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
GGML_ASSERT(src0->type == dst->type);
GGML_ASSERT(dst->ne[0] == nc);
GGML_ASSERT(ggml_nrows(dst) == ggml_nrows(src0));
if (src1) {
GGML_ASSERT(ggml_is_contiguous_1(src1));
GGML_ASSERT(src1->nb[0] == ggml_element_size(src1));
GGML_ASSERT(src1->ne[0] == nc);
GGML_ASSERT(src0->type == src1->type);
}
//const int32_t swapped = ((const int32_t *) dst->op_params)[1];
const int32_t swapped = ggml_get_op_params_i32(dst, 1);
const float alpha = ggml_get_op_params_f32(dst, 2);
const float limit = ggml_get_op_params_f32(dst, 3);
float * src0_p = (float *) src0_d;
float * src1_p = (float *) src1_d;
if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}
swiglu_oai_sycl(src0_p, src1_p, (float *)dst_d, ggml_nelements(dst), nc, src0_o / sizeof(float), src1_o / sizeof(float), alpha, limit, stream);
}
static inline void ggml_sycl_op_geglu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary_gated(ctx, dst, [](auto x) {
return op_gelu_erf(x);
});
}
static inline void ggml_sycl_op_geglu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_detail::ggml_sycl_op_unary_gated(ctx, dst, [](auto x) {
return op_gelu_quick(x);
});
}
void ggml_sycl_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_sqrt(ctx, dst);
}
void ggml_sycl_sin(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_sin(ctx, dst);
}
void ggml_sycl_cos(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_cos(ctx, dst);
}
void ggml_sycl_acc(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
ggml_sycl_op_acc(ctx, dst);
}
void ggml_sycl_gelu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_gelu(ctx, dst);
}
void ggml_sycl_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_silu(ctx, dst);
}
void ggml_sycl_gelu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_gelu_quick(ctx, dst);
}
void ggml_sycl_gelu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_gelu_erf(ctx, dst);
}
void ggml_sycl_tanh(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_tanh(ctx, dst);
}
void ggml_sycl_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_relu(ctx, dst);
}
void ggml_sycl_sigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_sigmoid(ctx, dst);
}
void ggml_sycl_hardsigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_hardsigmoid(ctx, dst);
}
void ggml_sycl_hardswish(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_hardswish(ctx, dst);
}
void ggml_sycl_exp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_exp(ctx, dst);
}
void ggml_sycl_expm1(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_expm1(ctx, dst);
}
void ggml_sycl_log(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_log(ctx, dst);
}
void ggml_sycl_softplus(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_softplus(ctx, dst);
}
void ggml_sycl_neg(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_neg(ctx, dst);
}
void ggml_sycl_step(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_step(ctx, dst);
}
void ggml_sycl_leaky_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_leaky_relu(ctx, dst);
}
void ggml_sycl_sqr(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_sqr(ctx, dst);
}
void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_clamp(ctx, dst);
}
void ggml_sycl_xielu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_xielu(ctx, dst);
}
void ggml_sycl_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_sgn(ctx, dst);
}
void ggml_sycl_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_abs(ctx, dst);
}
void ggml_sycl_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_elu(ctx, dst);
}
void ggml_sycl_geglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_geglu(ctx, dst);
}
void ggml_sycl_reglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_reglu(ctx, dst);
}
void ggml_sycl_swiglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_swiglu(ctx, dst);
}
void ggml_sycl_swiglu_oai(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_swiglu_oai(ctx, dst);
}
void ggml_sycl_geglu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_geglu_erf(ctx, dst);
}
void ggml_sycl_geglu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_geglu_quick(ctx, dst);
}
void ggml_sycl_arange(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/0);
ggml_sycl_detail::ggml_sycl_op_arange(ctx, dst);
}
void ggml_sycl_floor(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_floor(ctx, dst);
}
void ggml_sycl_ceil(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_ceil(ctx, dst);
}
void ggml_sycl_round(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_round(ctx, dst);
}
void ggml_sycl_trunc(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
ggml_sycl_op_trunc(ctx, dst);
}
|