| 1 | // Adapted from |
| 2 | // https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu |
| 3 | |
| 4 | // clang-format off |
| 5 | |
| 6 | #include "moe.h" |
| 7 | |
| 8 | #include "common.h" |
| 9 | #include "gemm.h" |
| 10 | |
| 11 | namespace { |
| 12 | |
| 13 | // [NOTE]: Fused MoE kernel with AMX |
| 14 | // |
| 15 | // This file contains implementations for |
| 16 | // * `moe_align_block_size` |
| 17 | // * `fused_moe` |
| 18 | // |
| 19 | // The functionality is identical to triton kernel, excepts: |
| 20 | // * fuse silu_and_mul with gemm1, therefore this kernel |
| 21 | // allocates 2 intermediate_caches instead of 3 |
| 22 | // * add `offsets` in `moe_align_block_size` which keeps track |
| 23 | // of starting offset for each M block. this is for keeping |
| 24 | // output of silu_and_mul in sorted order, thus load_A for |
| 25 | // the 2nd gemm would be contiguous, therefore we can directly |
| 26 | // load A from intermediate_cache1. |
| 27 | // |
| 28 | // TODO: |
| 29 | // 1. tune BLOCK_M and BLOCK_N (BLOCK_N * K fit L2) |
| 30 | // 2. add prefetch for load A which is indexed access |
| 31 | // 3. abstract at::native::cpublas::brgemm with WoQ gemm (M = 1 & M != 1) |
| 32 | // |
| 33 | |
| 34 | template <int BLOCK_M> |
| 35 | int moe_align_block_size( |
| 36 | int32_t* __restrict__ sorted_ids, |
| 37 | int32_t* __restrict__ expert_ids, |
| 38 | int32_t* __restrict__ topk_ids, |
| 39 | int32_t* __restrict__ total_cnts, |
| 40 | int32_t* __restrict__ cumsums, |
| 41 | int32_t* __restrict__ offsets, |
| 42 | int num_experts, |
| 43 | int numel, |
| 44 | int num_threads) { |
| 45 | #define T_INDEX(tt) total_cnts + (tt) * num_experts |
| 46 | |
| 47 | // accumulate count of expert ids locally |
| 48 | at::parallel_for(0, numel, 0, [&](int begin, int end) { |
| 49 | int tid = at::get_thread_num(); |
| 50 | int32_t* __restrict__ local_cnts = T_INDEX(tid + 1); |
| 51 | |
| 52 | for (int i = begin; i < end; ++i) { |
| 53 | local_cnts[topk_ids[i]]++; |
| 54 | } |
| 55 | }); |
| 56 | |
| 57 | using iVec = at::vec::Vectorized<int32_t>; |
| 58 | for (int t = 0; t < num_threads; ++t) { |
| 59 | at::vec::map2<int32_t>( |
| 60 | [](iVec x, iVec y) { return x + y; }, T_INDEX(t + 1), T_INDEX(t + 1), T_INDEX(t), num_experts); |
| 61 | } |
| 62 | |
| 63 | // the last row holds sums of each experts |
| 64 | int32_t* total_cnts_t_1 = T_INDEX(num_threads); |
| 65 | |
| 66 | cumsums[0] = 0; |
| 67 | for (int e = 0; e < num_experts; ++e) { |
| 68 | // accumulate `num_tokens_post_pad`, also as the expert offset |
| 69 | cumsums[e + 1] = cumsums[e] + div_up(total_cnts_t_1[e], BLOCK_M) * BLOCK_M; |
| 70 | |
| 71 | for (int k = cumsums[e]; k < cumsums[e + 1]; k += BLOCK_M) { |
| 72 | expert_ids[k / BLOCK_M] = e; |
| 73 | } |
| 74 | } |
| 75 | int num_tokens_post_pad = cumsums[num_experts]; |
| 76 | |
| 77 | at::parallel_for(0, numel, 0, [&](int begin, int end) { |
| 78 | int tid = at::get_thread_num(); |
| 79 | // thread tid offsets in `total_cnts` |
| 80 | int32_t* __restrict__ offsets = T_INDEX(tid); |
| 81 | |
| 82 | for (int i = begin; i < end; ++i) { |
| 83 | int32_t expert_id = topk_ids[i]; |
| 84 | int32_t b_offset = cumsums[expert_id]; |
| 85 | int32_t t_offset = offsets[expert_id]; |
| 86 | sorted_ids[b_offset + t_offset] = i; |
| 87 | offsets[expert_id]++; |
| 88 | } |
| 89 | }); |
| 90 | |
| 91 | // debug: the offset for thread t_1 should be identical to t_2 |
| 92 | int32_t* total_cnts_t_2 = T_INDEX(num_threads - 1); |
| 93 | for (int e = 0; e < num_experts; ++e) { |
| 94 | TORCH_CHECK(total_cnts_t_1[e] == total_cnts_t_2[e]); |
| 95 | } |
| 96 | |
| 97 | // padding value for sorted_ids: numel |
| 98 | auto sorted_id_size = [=](const int32_t* sorted_ids_ptr) { |
| 99 | for (int d = 0; d < BLOCK_M; ++d) { |
| 100 | if (sorted_ids_ptr[d] == numel) { |
| 101 | return d; |
| 102 | } |
| 103 | } |
| 104 | return BLOCK_M; |
| 105 | }; |
| 106 | |
| 107 | // offsets holds starting offset for each valida M blocks |
| 108 | // shape : [num_token_blocks + 1] |
| 109 | offsets[0] = 0; |
| 110 | const int num_token_blocks = num_tokens_post_pad / BLOCK_M; |
| 111 | at::parallel_for(0, num_token_blocks, GRAIN_SIZE / BLOCK_M, [&](int begin, int end) { |
| 112 | for (int mb = begin; mb < end; ++mb) { |
| 113 | offsets[mb + 1] = sorted_id_size(sorted_ids + mb * BLOCK_M); |
| 114 | } |
| 115 | }); |
| 116 | // TODO: do we need to vecterize this ? |
| 117 | for (int mb = 0; mb < num_token_blocks; ++mb) { |
| 118 | offsets[mb + 1] += offsets[mb]; |
| 119 | } |
| 120 | // debug: the last value of offsets should be `numel` |
| 121 | TORCH_CHECK(offsets[num_token_blocks] == numel); |
| 122 | |
| 123 | return num_tokens_post_pad; |
| 124 | } |
| 125 | |
| 126 | template <typename scalar_t, int BLOCK_M, int BLOCK_N> |
| 127 | struct tinygemm_kernel_nn2 { |
| 128 | static inline void apply( |
| 129 | const scalar_t* __restrict__ A, |
| 130 | const scalar_t* __restrict__ B0, |
| 131 | const scalar_t* __restrict__ B1, |
| 132 | scalar_t* __restrict__ C, |
| 133 | int64_t K, |
| 134 | int64_t lda, |
| 135 | int64_t ldb, |
| 136 | int64_t ldc) { |
| 137 | TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!"); |
| 138 | } |
| 139 | }; |
| 140 | |
| 141 | #if defined(CPU_CAPABILITY_AVX512) |
| 142 | template <int BLOCK_M, int BLOCK_N> |
| 143 | struct tinygemm_kernel_nn2<at::BFloat16, BLOCK_M, BLOCK_N> { |
| 144 | static inline void apply( |
| 145 | const at::BFloat16* __restrict__ A, |
| 146 | const at::BFloat16* __restrict__ B0, |
| 147 | const at::BFloat16* __restrict__ B1, |
| 148 | at::BFloat16* __restrict__ C, |
| 149 | int64_t K, |
| 150 | int64_t lda, |
| 151 | int64_t ldb, |
| 152 | int64_t ldc) { |
| 153 | constexpr int ROWS = BLOCK_M; |
| 154 | constexpr int COLS = BLOCK_N / 16; |
| 155 | |
| 156 | static_assert(COLS % 2 == 0); |
| 157 | |
| 158 | // prefetch distance |
| 159 | constexpr int PREFETCH_SIZE_K = 0; |
| 160 | |
| 161 | __m512bh va; |
| 162 | __m512bh vb0[COLS]; |
| 163 | __m512bh vb1[COLS]; |
| 164 | __m512 vc0[ROWS * COLS]; |
| 165 | __m512 vc1[ROWS * COLS]; |
| 166 | |
| 167 | auto loadc = [&](auto i) { |
| 168 | vc0[i] = _mm512_set1_ps(0.f); |
| 169 | vc1[i] = _mm512_set1_ps(0.f); |
| 170 | }; |
| 171 | Unroll<ROWS * COLS>{}(loadc); |
| 172 | |
| 173 | const int64_t K2 = K >> 1; |
| 174 | const int64_t lda2 = lda >> 1; |
| 175 | const int64_t ldb2 = ldb; // ldb * 2 >> 1; |
| 176 | const float* a_ptr = reinterpret_cast<const float*>(A); |
| 177 | const float* b0_ptr = reinterpret_cast<const float*>(B0); |
| 178 | const float* b1_ptr = reinterpret_cast<const float*>(B1); |
| 179 | |
| 180 | auto compute = [&](auto i, int64_t k) { |
| 181 | constexpr int row = i / COLS; |
| 182 | constexpr int col = i % COLS; |
| 183 | |
| 184 | if constexpr (col == 0) { |
| 185 | va = (__m512bh)(_mm512_set1_ps(a_ptr[row * lda2 + k])); |
| 186 | } |
| 187 | if constexpr (row == 0) { |
| 188 | vb0[col] = (__m512bh)(_mm512_loadu_si512(b0_ptr + k * ldb2 + col * 16)); |
| 189 | vb1[col] = (__m512bh)(_mm512_loadu_si512(b1_ptr + k * ldb2 + col * 16)); |
| 190 | if constexpr (PREFETCH_SIZE_K > 0) { |
| 191 | _mm_prefetch(b0_ptr + (k + PREFETCH_SIZE_K) * ldb2 + col * 16, _MM_HINT_T0); |
| 192 | _mm_prefetch(b1_ptr + (k + PREFETCH_SIZE_K) * ldb2 + col * 16, _MM_HINT_T0); |
| 193 | } |
| 194 | } |
| 195 | vc0[i] = _mm512_dpbf16_ps(vc0[i], va, vb0[col]); |
| 196 | vc1[i] = _mm512_dpbf16_ps(vc1[i], va, vb1[col]); |
| 197 | }; |
| 198 | for (int64_t k = 0; k < K2; ++k) { |
| 199 | Unroll<ROWS * COLS>{}(compute, k); |
| 200 | } |
| 201 | |
| 202 | auto storec = [&](auto i) { |
| 203 | constexpr int row = i / COLS; |
| 204 | constexpr int col = i % COLS; |
| 205 | // for COLS = 2, 4 use 512bit store |
| 206 | if constexpr (col % 2 == 0) { |
| 207 | __m512 x0 = vc0[row * COLS + col + 0]; |
| 208 | __m512 x1 = vc0[row * COLS + col + 1]; |
| 209 | __m512 y0 = vc1[row * COLS + col + 0]; |
| 210 | __m512 y1 = vc1[row * COLS + col + 1]; |
| 211 | x0 = _mm512_mul_ps(_mm512_rcp14_silu_ps(x0), y0); |
| 212 | x1 = _mm512_mul_ps(_mm512_rcp14_silu_ps(x1), y1); |
| 213 | |
| 214 | _mm512_storeu_si512( |
| 215 | reinterpret_cast<__m512i*>((C + row * ldc + col * 16)), |
| 216 | (__m512i)(_mm512_cvtne2ps_pbh(__m512(x1), __m512(x0)))); |
| 217 | } |
| 218 | }; |
| 219 | Unroll<ROWS * COLS>{}(storec); |
| 220 | } |
| 221 | }; |
| 222 | #endif |
| 223 | |
| 224 | #define LAUNCH_TINYGEMM_KERNEL_NN(MB_SIZE, NB_SIZE) \ |
| 225 | tinygemm_kernel_nn2<scalar_t, MB_SIZE, NB_SIZE>::apply( \ |
| 226 | A + mb_start * lda, B0 + nb_start * 2, B1 + nb_start * 2, C + mb_start * ldc + nb_start, K, lda, ldb, ldc); |
| 227 | |
| 228 | template <typename scalar_t> |
| 229 | void tinygemm_kernel( |
| 230 | const scalar_t* __restrict__ A, |
| 231 | const scalar_t* __restrict__ B0, |
| 232 | const scalar_t* __restrict__ B1, |
| 233 | scalar_t* __restrict__ C, |
| 234 | int64_t M, |
| 235 | int64_t N, |
| 236 | int64_t K, |
| 237 | int64_t lda, |
| 238 | int64_t ldb, |
| 239 | int64_t ldc) { |
| 240 | // pattern: 1-(2+2)-(8+8) |
| 241 | constexpr int64_t BLOCK_M = 4; |
| 242 | constexpr int64_t BLOCK_N = 32; |
| 243 | const int64_t MB = div_up(M, BLOCK_M); |
| 244 | const int64_t NB = div_up(N, BLOCK_N); |
| 245 | for (int mb = 0; mb < MB; ++mb) { |
| 246 | int64_t mb_start = mb * BLOCK_M; |
| 247 | int64_t mb_size = std::min(BLOCK_M, M - mb_start); |
| 248 | for (int64_t nb = 0; nb < NB; ++nb) { |
| 249 | int64_t nb_start = nb * BLOCK_N; |
| 250 | int64_t nb_size = std::min(BLOCK_N, N - nb_start); |
| 251 | |
| 252 | switch (mb_size << 4 | nb_size >> 4) { |
| 253 | // mb_size = 1 |
| 254 | case 0x12: |
| 255 | LAUNCH_TINYGEMM_KERNEL_NN(1, 32); |
| 256 | break; |
| 257 | // mb_size = 2 |
| 258 | case 0x22: |
| 259 | LAUNCH_TINYGEMM_KERNEL_NN(2, 32); |
| 260 | break; |
| 261 | // mb_size = 3 |
| 262 | case 0x32: |
| 263 | LAUNCH_TINYGEMM_KERNEL_NN(3, 32); |
| 264 | break; |
| 265 | // mb_size = 4 |
| 266 | case 0x42: |
| 267 | LAUNCH_TINYGEMM_KERNEL_NN(4, 32); |
| 268 | break; |
| 269 | default: |
| 270 | TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size"); |
| 271 | } |
| 272 | } |
| 273 | } |
| 274 | } |
| 275 | |
| 276 | template <typename scalar_t, int BLOCK_M, int BLOCK_N> |
| 277 | struct tinygemm_kernel_nn { |
| 278 | static inline void apply( |
| 279 | const scalar_t* __restrict__ A, |
| 280 | const scalar_t* __restrict__ B, |
| 281 | float* __restrict__ C, |
| 282 | int64_t K, |
| 283 | int64_t lda, |
| 284 | int64_t ldb, |
| 285 | int64_t ldc) { |
| 286 | TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!"); |
| 287 | } |
| 288 | }; |
| 289 | |
| 290 | #if defined(CPU_CAPABILITY_AVX512) |
| 291 | template <int BLOCK_M, int BLOCK_N> |
| 292 | struct tinygemm_kernel_nn<at::BFloat16, BLOCK_M, BLOCK_N> { |
| 293 | static inline void apply( |
| 294 | const at::BFloat16* __restrict__ A, |
| 295 | const at::BFloat16* __restrict__ B, |
| 296 | float* __restrict__ C, |
| 297 | int64_t K, |
| 298 | int64_t lda, |
| 299 | int64_t ldb, |
| 300 | int64_t ldc) { |
| 301 | constexpr int ROWS = BLOCK_M; |
| 302 | constexpr int COLS = BLOCK_N / 16; |
| 303 | |
| 304 | static_assert(COLS % 2 == 0); |
| 305 | |
| 306 | // prefetch distance |
| 307 | constexpr int PREFETCH_SIZE_K = 0; |
| 308 | |
| 309 | __m512bh va; |
| 310 | __m512bh vb[COLS]; |
| 311 | __m512 vc[ROWS * COLS]; |
| 312 | |
| 313 | auto loadc = [&](auto i) { vc[i] = _mm512_set1_ps(0.f); }; |
| 314 | Unroll<ROWS * COLS>{}(loadc); |
| 315 | |
| 316 | const int64_t K2 = K >> 1; |
| 317 | const int64_t lda2 = lda >> 1; |
| 318 | const int64_t ldb2 = ldb; // ldb * 2 >> 1; |
| 319 | const float* a_ptr = reinterpret_cast<const float*>(A); |
| 320 | const float* b_ptr = reinterpret_cast<const float*>(B); |
| 321 | |
| 322 | auto compute = [&](auto i, int64_t k) { |
| 323 | constexpr int row = i / COLS; |
| 324 | constexpr int col = i % COLS; |
| 325 | |
| 326 | if constexpr (col == 0) { |
| 327 | va = (__m512bh)(_mm512_set1_ps(a_ptr[row * lda2 + k])); |
| 328 | } |
| 329 | if constexpr (row == 0) { |
| 330 | vb[col] = (__m512bh)(_mm512_loadu_si512(b_ptr + k * ldb2 + col * 16)); |
| 331 | if constexpr (PREFETCH_SIZE_K > 0) { |
| 332 | _mm_prefetch(b_ptr + (k + PREFETCH_SIZE_K) * ldb2 + col * 16, _MM_HINT_T0); |
| 333 | } |
| 334 | } |
| 335 | vc[i] = _mm512_dpbf16_ps(vc[i], va, vb[col]); |
| 336 | }; |
| 337 | for (int64_t k = 0; k < K2; ++k) { |
| 338 | Unroll<ROWS * COLS>{}(compute, k); |
| 339 | } |
| 340 | |
| 341 | auto storec = [&](auto i) { |
| 342 | constexpr int row = i / COLS; |
| 343 | constexpr int col = i % COLS; |
| 344 | _mm512_storeu_ps(reinterpret_cast<__m512*>(C + row * ldc + col * 16), vc[i]); |
| 345 | }; |
| 346 | Unroll<ROWS * COLS>{}(storec); |
| 347 | } |
| 348 | }; |
| 349 | #endif |
| 350 | |
| 351 | #define LAUNCH_TINYGEMM_KERNEL_NN2(MB_SIZE, NB_SIZE) \ |
| 352 | tinygemm_kernel_nn<scalar_t, MB_SIZE, NB_SIZE>::apply( \ |
| 353 | A + mb_start * lda, B + nb_start * 2, C + mb_start * ldc + nb_start, K, lda, ldb, ldc); |
| 354 | |
| 355 | template <typename scalar_t> |
| 356 | void tinygemm_kernel( |
| 357 | const scalar_t* __restrict__ A, |
| 358 | const scalar_t* __restrict__ B, |
| 359 | float* __restrict__ C, |
| 360 | int64_t M, |
| 361 | int64_t N, |
| 362 | int64_t K, |
| 363 | int64_t lda, |
| 364 | int64_t ldb, |
| 365 | int64_t ldc) { |
| 366 | // pattern: 1-2-8 |
| 367 | constexpr int64_t BLOCK_M = 4; |
| 368 | constexpr int64_t BLOCK_N = 32; |
| 369 | const int64_t MB = div_up(M, BLOCK_M); |
| 370 | const int64_t NB = div_up(N, BLOCK_N); |
| 371 | for (int mb = 0; mb < MB; ++mb) { |
| 372 | int64_t mb_start = mb * BLOCK_M; |
| 373 | int64_t mb_size = std::min(BLOCK_M, M - mb_start); |
| 374 | for (int64_t nb = 0; nb < NB; ++nb) { |
| 375 | int64_t nb_start = nb * BLOCK_N; |
| 376 | int64_t nb_size = std::min(BLOCK_N, N - nb_start); |
| 377 | |
| 378 | switch (mb_size << 4 | nb_size >> 4) { |
| 379 | // mb_size = 1 |
| 380 | case 0x12: |
| 381 | LAUNCH_TINYGEMM_KERNEL_NN2(1, 32); |
| 382 | break; |
| 383 | // mb_size = 2 |
| 384 | case 0x22: |
| 385 | LAUNCH_TINYGEMM_KERNEL_NN2(2, 32); |
| 386 | break; |
| 387 | // mb_size = 3 |
| 388 | case 0x32: |
| 389 | LAUNCH_TINYGEMM_KERNEL_NN2(3, 32); |
| 390 | break; |
| 391 | // mb_size = 4 |
| 392 | case 0x42: |
| 393 | LAUNCH_TINYGEMM_KERNEL_NN2(4, 32); |
| 394 | break; |
| 395 | default: |
| 396 | TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size"); |
| 397 | } |
| 398 | } |
| 399 | } |
| 400 | } |
| 401 | |
| 402 | template <typename scalar_t> |
| 403 | void fused_experts_kernel_impl( |
| 404 | scalar_t* __restrict__ output, |
| 405 | scalar_t* __restrict__ ic1, |
| 406 | scalar_t* __restrict__ ic2, |
| 407 | scalar_t* __restrict__ A_tmp, |
| 408 | float* __restrict__ C_tmp, |
| 409 | const scalar_t* __restrict__ input, |
| 410 | const scalar_t* __restrict__ packed_w1, |
| 411 | const scalar_t* __restrict__ packed_w2, |
| 412 | const float* __restrict__ w1_bias, |
| 413 | const float* __restrict__ w2_bias, |
| 414 | const float* __restrict__ topk_weights, |
| 415 | const int32_t* __restrict__ sorted_ids, |
| 416 | const int32_t* __restrict__ expert_ids, |
| 417 | const int32_t* __restrict__ offsets, |
| 418 | int64_t M, |
| 419 | int64_t N, |
| 420 | int64_t K, |
| 421 | int64_t E, |
| 422 | int64_t topk, |
| 423 | int64_t num_tokens_post_pad, |
| 424 | float alpha, |
| 425 | float limit, |
| 426 | CPUActMethod act_func, |
| 427 | bool with_bias) { |
| 428 | // handle 2 tiles per block |
| 429 | constexpr int64_t BLOCK_M = block_size_m(); |
| 430 | constexpr int64_t BLOCK_N = block_size_n(); |
| 431 | |
| 432 | // stage 1: intermediate_cache1 = silu(hidden_states @ w1) |
| 433 | const int64_t MB = div_up(num_tokens_post_pad, BLOCK_M); |
| 434 | const int64_t NB = div_up(N, BLOCK_N); |
| 435 | |
| 436 | // strides for w1: [E, 2N, K] |
| 437 | TORCH_CHECK(N % BLOCK_N == 0, "Fixme when N is not multiples of ", BLOCK_N); |
| 438 | |
| 439 | const int64_t stride_e = 2 * N * K; |
| 440 | const int64_t stride_n = K; |
| 441 | |
| 442 | int64_t avg_M = std::max(int64_t(1), M * topk / E); |
| 443 | const bool use_brgemm = can_use_brgemm<scalar_t>(avg_M); |
| 444 | |
| 445 | // here we only parallel on half of 2N to fuse silu_and_mul with gemm |
| 446 | parallel_2d(MB, NB, [&](int64_t mb0, int64_t mb1, int64_t nb0, int64_t nb1) { |
| 447 | // get local pointers |
| 448 | int tid = get_thread_num(); |
| 449 | scalar_t* __restrict__ A = A_tmp + tid * BLOCK_M * K; |
| 450 | float* __restrict__ C0 = C_tmp + tid * 2 * BLOCK_M * BLOCK_N; |
| 451 | float* __restrict__ C1 = C0 + BLOCK_M * BLOCK_N; |
| 452 | |
| 453 | loop_2d<scalar_t>(mb0, mb1, nb0, nb1, BLOCK_N * K * 2, [&](int64_t mb, int64_t nb, int64_t nb_offset) { |
| 454 | // nb_upper from top half and nb_lower from bottom half |
| 455 | int64_t nb_upper = nb, nb_lower = nb + NB; |
| 456 | int64_t n_size = std::min(N - nb * BLOCK_N, BLOCK_N); |
| 457 | |
| 458 | // B shape [K, n_size] in vnni format |
| 459 | int32_t expert_id = expert_ids[mb]; |
| 460 | const scalar_t* __restrict__ B0 = packed_w1 + expert_id * stride_e + nb_upper * BLOCK_N * stride_n; |
| 461 | const scalar_t* __restrict__ B1 = packed_w1 + expert_id * stride_e + nb_lower * BLOCK_N * stride_n; |
| 462 | const float* __restrict__ B0_bias = w1_bias + expert_id * 2 * N + nb_upper * BLOCK_N; |
| 463 | const float* __restrict__ B1_bias = w1_bias + expert_id * 2 * N + nb_lower * BLOCK_N; |
| 464 | |
| 465 | int64_t m_size = offsets[mb + 1] - offsets[mb]; |
| 466 | |
| 467 | if (nb_offset == 0) { |
| 468 | // 1.a load A |
| 469 | const int32_t* A_ids = sorted_ids + mb * BLOCK_M; |
| 470 | for (int64_t m = 0; m < m_size; ++m) { |
| 471 | int32_t index = A_ids[m] / topk; |
| 472 | copy_stub(A + m * K, input + index * K, K); |
| 473 | } |
| 474 | } |
| 475 | |
| 476 | if (use_brgemm) { |
| 477 | // 1.b gemm: C0 = A @ B0 |
| 478 | at::native::cpublas::brgemm( |
| 479 | /* M */ m_size, |
| 480 | /* N */ n_size, |
| 481 | /* K */ K, |
| 482 | /* lda */ K, |
| 483 | /* ldb */ n_size, |
| 484 | /* ldc */ BLOCK_N, |
| 485 | /* add_C */ false, |
| 486 | /* A */ A, |
| 487 | /* B */ B0, |
| 488 | /* C */ C0); |
| 489 | |
| 490 | // 1.c gemm: C1 = A @ B1 |
| 491 | at::native::cpublas::brgemm( |
| 492 | /* M */ m_size, |
| 493 | /* N */ n_size, |
| 494 | /* K */ K, |
| 495 | /* lda */ K, |
| 496 | /* ldb */ n_size, |
| 497 | /* ldc */ BLOCK_N, |
| 498 | /* add_C */ false, |
| 499 | /* A */ A, |
| 500 | /* B */ B1, |
| 501 | /* C */ C1); |
| 502 | |
| 503 | } else { |
| 504 | const int64_t offset = offsets[mb]; |
| 505 | if (act_func == CPUActMethod::swiglu) { |
| 506 | tinygemm_kernel( |
| 507 | /* A */ A, |
| 508 | /* B */ B0, |
| 509 | /* C */ C0, |
| 510 | /* M */ m_size, |
| 511 | /* N */ n_size, |
| 512 | /* K */ K, |
| 513 | /* lda */ K, |
| 514 | /* ldb */ n_size, |
| 515 | /* ldc */ BLOCK_N); |
| 516 | tinygemm_kernel( |
| 517 | /* A */ A, |
| 518 | /* B */ B1, |
| 519 | /* C */ C1, |
| 520 | /* M */ m_size, |
| 521 | /* N */ n_size, |
| 522 | /* K */ K, |
| 523 | /* lda */ K, |
| 524 | /* ldb */ n_size, |
| 525 | /* ldc */ BLOCK_N); |
| 526 | } else { |
| 527 | // fused 1.bcd: silu_and_mul(A @ B0, A @ B1) |
| 528 | tinygemm_kernel( |
| 529 | /* A */ A, |
| 530 | /* B0 */ B0, |
| 531 | /* B1 */ B1, |
| 532 | /* C */ ic1 + offset * N + nb * BLOCK_N, |
| 533 | /* M */ m_size, |
| 534 | /* N */ n_size, |
| 535 | /* K */ K, |
| 536 | /* lda */ K, |
| 537 | /* ldb */ n_size, |
| 538 | /* ldc */ N); |
| 539 | } |
| 540 | } |
| 541 | if (with_bias) { |
| 542 | for (int64_t m = 0; m < m_size; ++m) { |
| 543 | add_bias_stub(C0 + m * BLOCK_N, B0_bias, n_size); |
| 544 | add_bias_stub(C1 + m * BLOCK_N, B1_bias, n_size); |
| 545 | } |
| 546 | } |
| 547 | // 1.d silu and mul |
| 548 | const int64_t offset = offsets[mb]; |
| 549 | if (act_func == CPUActMethod::silu_and_mul && use_brgemm) { |
| 550 | for (int64_t m = 0; m < m_size; ++m) { |
| 551 | silu_and_mul_stub(ic1 + (offset + m) * N + nb * BLOCK_N, C0 + m * BLOCK_N, C1 + m * BLOCK_N, BLOCK_N); |
| 552 | } |
| 553 | } else if (act_func == CPUActMethod::swiglu) { |
| 554 | for (int64_t m = 0; m < m_size; ++m) { |
| 555 | scalar_t* __restrict__ ic1_row = ic1 + (offset + m) * N; |
| 556 | clamp_sigmoid_and_mul_stub(ic1_row + nb * BLOCK_N / 2, C0 + m * BLOCK_N, BLOCK_N / 2, alpha, limit); |
| 557 | clamp_sigmoid_and_mul_stub(ic1_row + N / 2 + nb * BLOCK_N / 2, C1 + m * BLOCK_N, BLOCK_N / 2, alpha, limit); |
| 558 | } |
| 559 | } |
| 560 | }); |
| 561 | |
| 562 | if (use_brgemm) { |
| 563 | at::native::cpublas::brgemm_release(); |
| 564 | } |
| 565 | }); |
| 566 | |
| 567 | // stage 2: intermediate_cache2 = intermediate_cache1 @ w2 |
| 568 | // w2 : [E, K, N] as [E, OC, IC] |
| 569 | const int64_t OC = K; // rename K as OC |
| 570 | const int64_t IC = N; // rename N as IC |
| 571 | const int64_t MB2 = MB; |
| 572 | const int64_t NB2 = div_up(OC, BLOCK_N); |
| 573 | const int64_t stride_e2 = OC * IC; |
| 574 | const int64_t stride_oc = IC; |
| 575 | |
| 576 | // parallel on [MB2, NB2] |
| 577 | parallel_2d(MB2, NB2, [&](int64_t mb0, int64_t mb1, int64_t nb0, int64_t nb1) { |
| 578 | // get local pointers |
| 579 | int tid = get_thread_num(); |
| 580 | // we won't be using C1 for gemm2 |
| 581 | float* __restrict__ C = C_tmp + tid * 2 * BLOCK_M * BLOCK_N; |
| 582 | |
| 583 | loop_2d<scalar_t>(mb0, mb1, nb0, nb1, BLOCK_N * IC, [&](int64_t mb, int64_t nb, int64_t nb_offset) { |
| 584 | int64_t m_size = offsets[mb + 1] - offsets[mb]; |
| 585 | int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N); |
| 586 | |
| 587 | // A ptr from ic1 of [M * topk, N] in sorted order |
| 588 | // so as to avoid copy A to tmp buffer again |
| 589 | const scalar_t* __restrict__ A = ic1 + offsets[mb] * N; |
| 590 | const int32_t* A_ids = sorted_ids + mb * BLOCK_M; |
| 591 | |
| 592 | // B shape [IC, n_size] in vnni format |
| 593 | int32_t expert_id = expert_ids[mb]; |
| 594 | const scalar_t* __restrict__ B = packed_w2 + expert_id * stride_e2 + nb * BLOCK_N * stride_oc; |
| 595 | const float* __restrict__ B_bias = w2_bias + expert_id * OC + nb * BLOCK_N; |
| 596 | |
| 597 | // 2.a gemm: C = A @ B |
| 598 | if (use_brgemm) { |
| 599 | at::native::cpublas::brgemm( |
| 600 | /* M */ m_size, |
| 601 | /* N */ n_size, |
| 602 | /* K */ IC, |
| 603 | /* lda */ IC, |
| 604 | /* ldb */ n_size, |
| 605 | /* ldc */ BLOCK_N, |
| 606 | /* add_C */ false, |
| 607 | /* A */ A, |
| 608 | /* B */ B, |
| 609 | /* C */ C); |
| 610 | } else { |
| 611 | tinygemm_kernel( |
| 612 | /* A */ A, |
| 613 | /* B */ B, |
| 614 | /* C */ C, |
| 615 | /* M */ m_size, |
| 616 | /* N */ n_size, |
| 617 | /* K */ IC, |
| 618 | /* lda */ IC, |
| 619 | /* ldb */ n_size, |
| 620 | /* ldc */ BLOCK_N); |
| 621 | } |
| 622 | |
| 623 | if (with_bias) { |
| 624 | for (int64_t m = 0; m < m_size; ++m) { |
| 625 | add_bias_stub(C + m * BLOCK_N, B_bias, n_size); |
| 626 | } |
| 627 | } |
| 628 | // 2.b copy from C to ic2 in original order |
| 629 | // and also mul topk_weights in float32 |
| 630 | for (int64_t m = 0; m < m_size; ++m) { |
| 631 | int32_t index = A_ids[m]; |
| 632 | float weight = topk_weights[index]; |
| 633 | copy_mul_stub(ic2 + index * K + nb * BLOCK_N, C + m * BLOCK_N, weight, n_size); |
| 634 | } |
| 635 | }); |
| 636 | |
| 637 | if (use_brgemm) { |
| 638 | at::native::cpublas::brgemm_release(); |
| 639 | } |
| 640 | }); |
| 641 | |
| 642 | // stage 3: out = intermediate_cache2.sum(dim=1) |
| 643 | // from [M, topk, K] to [M, K] |
| 644 | at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) { |
| 645 | for (int64_t m = begin; m < end; ++m) { |
| 646 | sum_stub(output + m * K, ic2 + m * topk * K, topk, K); |
| 647 | } |
| 648 | }); |
| 649 | } |
| 650 | |
| 651 | template <typename scalar_t> |
| 652 | void shared_expert_kernel_impl( |
| 653 | scalar_t* __restrict__ output, |
| 654 | scalar_t* __restrict__ ic1, |
| 655 | float* __restrict__ C_tmp, |
| 656 | scalar_t* __restrict__ input, |
| 657 | const scalar_t* __restrict__ packed_w1, |
| 658 | const scalar_t* __restrict__ packed_w2, |
| 659 | const scalar_t* __restrict__ fused_experts_out, |
| 660 | float routed_scaling_factor, |
| 661 | int64_t M, |
| 662 | int64_t N, |
| 663 | int64_t K) { |
| 664 | // handle 2 tiles per block |
| 665 | constexpr int64_t BLOCK_M = block_size_m(); |
| 666 | constexpr int64_t BLOCK_N = block_size_n(); |
| 667 | |
| 668 | // stage 1: intermediate_cache1 = silu(hidden_states @ w1) |
| 669 | const int64_t MB = div_up(M, BLOCK_M); |
| 670 | const int64_t NB = div_up(N, BLOCK_N); |
| 671 | |
| 672 | TORCH_CHECK(N % BLOCK_N == 0, "Fixme when N is not multiples of ", BLOCK_N); |
| 673 | const int64_t stride_n = K; |
| 674 | |
| 675 | const bool use_brgemm = can_use_brgemm<scalar_t>(M); |
| 676 | |
| 677 | const bool apply_scaling_factor = fused_experts_out != nullptr; |
| 678 | |
| 679 | // here we only parallel on half of 2N to fuse silu_and_mul with gemm |
| 680 | parallel_2d(MB, NB, [&](int64_t mb0, int64_t mb1, int64_t nb0, int64_t nb1) { |
| 681 | // get local pointers |
| 682 | int tid = get_thread_num(); |
| 683 | float* __restrict__ C0 = C_tmp + tid * 2 * BLOCK_M * BLOCK_N; |
| 684 | float* __restrict__ C1 = C0 + BLOCK_M * BLOCK_N; |
| 685 | |
| 686 | loop_2d<scalar_t>(mb0, mb1, nb0, nb1, BLOCK_N * K * 2, [&](int64_t mb, int64_t nb, int64_t nb_offset) { |
| 687 | // nb_upper from top half and nb_lower from bottom half |
| 688 | int64_t nb_upper = nb, nb_lower = nb + NB; |
| 689 | int64_t n_size = std::min(N - nb * BLOCK_N, BLOCK_N); |
| 690 | int64_t m_size = std::min(M - mb * BLOCK_M, BLOCK_M); |
| 691 | |
| 692 | // A shape [m_size, K] |
| 693 | const scalar_t* A = input + mb * BLOCK_M * K; |
| 694 | |
| 695 | // B shape [K, n_size] in vnni format |
| 696 | const scalar_t* __restrict__ B0 = packed_w1 + nb_upper * BLOCK_N * stride_n; |
| 697 | const scalar_t* __restrict__ B1 = packed_w1 + nb_lower * BLOCK_N * stride_n; |
| 698 | |
| 699 | if (use_brgemm) { |
| 700 | // 1.b gemm: C0 = A @ B0 |
| 701 | at::native::cpublas::brgemm( |
| 702 | /* M */ m_size, |
| 703 | /* N */ n_size, |
| 704 | /* K */ K, |
| 705 | /* lda */ K, |
| 706 | /* ldb */ n_size, |
| 707 | /* ldc */ BLOCK_N, |
| 708 | /* add_C */ false, |
| 709 | /* A */ A, |
| 710 | /* B */ B0, |
| 711 | /* C */ C0); |
| 712 | |
| 713 | // 1.c gemm: C1 = A @ B1 |
| 714 | at::native::cpublas::brgemm( |
| 715 | /* M */ m_size, |
| 716 | /* N */ n_size, |
| 717 | /* K */ K, |
| 718 | /* lda */ K, |
| 719 | /* ldb */ n_size, |
| 720 | /* ldc */ BLOCK_N, |
| 721 | /* add_C */ false, |
| 722 | /* A */ A, |
| 723 | /* B */ B1, |
| 724 | /* C */ C1); |
| 725 | |
| 726 | // 1.d silu and mul |
| 727 | for (int64_t m = 0; m < m_size; ++m) { |
| 728 | silu_and_mul_stub(ic1 + (mb * BLOCK_M + m) * N + nb * BLOCK_N, C0 + m * BLOCK_N, C1 + m * BLOCK_N, BLOCK_N); |
| 729 | } |
| 730 | } else { |
| 731 | // fused 1.bcd: silu_and_mul(A @ B0, A @ B1) |
| 732 | tinygemm_kernel( |
| 733 | /* A */ A, |
| 734 | /* B0 */ B0, |
| 735 | /* B1 */ B1, |
| 736 | /* C */ ic1 + mb * BLOCK_M * N + nb * BLOCK_N, |
| 737 | /* M */ m_size, |
| 738 | /* N */ n_size, |
| 739 | /* K */ K, |
| 740 | /* lda */ K, |
| 741 | /* ldb */ n_size, |
| 742 | /* ldc */ N); |
| 743 | } |
| 744 | }); |
| 745 | |
| 746 | if (use_brgemm) { |
| 747 | at::native::cpublas::brgemm_release(); |
| 748 | } |
| 749 | }); |
| 750 | |
| 751 | // stage 2: output = intermediate_cache1 @ w2 |
| 752 | // w2 : [K, N] as [OC, IC] |
| 753 | const int64_t OC = K; // rename K as OC |
| 754 | const int64_t IC = N; // rename N as IC |
| 755 | const int64_t MB2 = MB; |
| 756 | const int64_t NB2 = div_up(OC, BLOCK_N); |
| 757 | const int64_t stride_oc = IC; |
| 758 | |
| 759 | // parallel on [MB2, NB2] |
| 760 | parallel_2d(MB2, NB2, [&](int64_t mb0, int64_t mb1, int64_t nb0, int64_t nb1) { |
| 761 | // get local pointers |
| 762 | int tid = get_thread_num(); |
| 763 | // we won't be using C1 for gemm2 |
| 764 | float* __restrict__ C = C_tmp + tid * 2 * BLOCK_M * BLOCK_N; |
| 765 | |
| 766 | loop_2d<scalar_t>(mb0, mb1, nb0, nb1, BLOCK_N * IC, [&](int64_t mb, int64_t nb, int64_t nb_offset) { |
| 767 | int64_t m_size = std::min(M - mb * BLOCK_M, BLOCK_M); |
| 768 | int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N); |
| 769 | |
| 770 | // A shape [m_size, IC] |
| 771 | const scalar_t* __restrict__ A = ic1 + mb * BLOCK_M * N; |
| 772 | |
| 773 | // B shape [IC, n_size] in vnni format |
| 774 | const scalar_t* __restrict__ B = packed_w2 + nb * BLOCK_N * stride_oc; |
| 775 | |
| 776 | // 2.a gemm: C = A @ B |
| 777 | if (use_brgemm) { |
| 778 | at::native::cpublas::brgemm( |
| 779 | /* M */ m_size, |
| 780 | /* N */ n_size, |
| 781 | /* K */ IC, |
| 782 | /* lda */ IC, |
| 783 | /* ldb */ n_size, |
| 784 | /* ldc */ BLOCK_N, |
| 785 | /* add_C */ false, |
| 786 | /* A */ A, |
| 787 | /* B */ B, |
| 788 | /* C */ C); |
| 789 | } else { |
| 790 | tinygemm_kernel( |
| 791 | /* A */ A, |
| 792 | /* B */ B, |
| 793 | /* C */ C, |
| 794 | /* M */ m_size, |
| 795 | /* N */ n_size, |
| 796 | /* K */ IC, |
| 797 | /* lda */ IC, |
| 798 | /* ldb */ n_size, |
| 799 | /* ldc */ BLOCK_N); |
| 800 | } |
| 801 | |
| 802 | // 2.b copy from C to output and add fused_experts_out |
| 803 | scalar_t* __restrict__ out = output + mb * BLOCK_M * K + nb * BLOCK_N; |
| 804 | const scalar_t* __restrict__ fused_out = |
| 805 | apply_scaling_factor ? fused_experts_out + mb * BLOCK_M * K + nb * BLOCK_N : nullptr; |
| 806 | for (int64_t m = 0; m < m_size; ++m) { |
| 807 | const scalar_t* __restrict__ fused_out_row = apply_scaling_factor ? (fused_out + m * K) : nullptr; |
| 808 | add_mul_stub(out + m * K, C + m * BLOCK_N, fused_out_row, routed_scaling_factor, n_size); |
| 809 | } |
| 810 | }); |
| 811 | |
| 812 | if (use_brgemm) { |
| 813 | at::native::cpublas::brgemm_release(); |
| 814 | } |
| 815 | }); |
| 816 | } |
| 817 | |
| 818 | } // anonymous namespace |
| 819 | |
| 820 | // common checks |
| 821 | template <CPUQuantMethod quant> |
| 822 | static inline void check_moe_scales( |
| 823 | const std::optional<at::Tensor>& w1_scale, |
| 824 | const std::optional<at::Tensor>& w2_scale, |
| 825 | const std::optional<std::vector<int64_t>> block_size) { |
| 826 | if constexpr (quant == CPUQuantMethod::INT8_W8A8) { |
| 827 | TORCH_CHECK(w1_scale.has_value(), "missing w1_scale for int8 w8a8."); |
| 828 | TORCH_CHECK(w2_scale.has_value(), "missing w2_scale for int8 w8a8."); |
| 829 | } else if constexpr (quant == CPUQuantMethod::FP8_W8A16) { |
| 830 | TORCH_CHECK(w1_scale.has_value(), "missing w1_scale for fp8 w8a16."); |
| 831 | TORCH_CHECK(w2_scale.has_value(), "missing w2_scale for fp8 w8a16."); |
| 832 | TORCH_CHECK(block_size.has_value(), "missing block_size for fp8 w8a16."); |
| 833 | TORCH_CHECK(block_size.value().size() == 2, "expect block_size.size() to be 2."); |
| 834 | } else if constexpr (quant == CPUQuantMethod::MXFP4) { |
| 835 | TORCH_CHECK(w1_scale.has_value(), "missing w1_scale for mxfp4."); |
| 836 | TORCH_CHECK(w2_scale.has_value(), "missing w2_scale for mxfp4."); |
| 837 | TORCH_CHECK(w1_scale.value().scalar_type() == at::kByte, "expect w1_scale to be uint8."); |
| 838 | TORCH_CHECK(w2_scale.value().scalar_type() == at::kByte, "expect w2_scale to be uint8."); |
| 839 | } else if constexpr (quant == CPUQuantMethod::FP8_W8A8) { |
| 840 | TORCH_CHECK(w1_scale.has_value(), "missing w1_scale for fp8 w8a8."); |
| 841 | TORCH_CHECK(w2_scale.has_value(), "missing w2_scale for fp8 w8a8."); |
| 842 | TORCH_CHECK(block_size.has_value(), "missing block_size for fp8 w8a8."); |
| 843 | TORCH_CHECK(block_size.value().size() == 2, "expect block_size.size() to be 2."); |
| 844 | } |
| 845 | } |
| 846 | |
| 847 | static inline void check_moe_scales( |
| 848 | int64_t moe_comp_method, |
| 849 | const std::optional<at::Tensor>& w1_scale, |
| 850 | const std::optional<at::Tensor>& w2_scale, |
| 851 | const std::optional<std::vector<int64_t>> block_size) { |
| 852 | if (moe_comp_method == CPUQuantMethod::INT8_W8A8) { |
| 853 | check_moe_scales<CPUQuantMethod::INT8_W8A8>(w1_scale, w2_scale, block_size); |
| 854 | } else if (moe_comp_method == CPUQuantMethod::FP8_W8A16) { |
| 855 | check_moe_scales<CPUQuantMethod::FP8_W8A16>(w1_scale, w2_scale, block_size); |
| 856 | } else if (moe_comp_method == CPUQuantMethod::MXFP4) { |
| 857 | check_moe_scales<CPUQuantMethod::MXFP4>(w1_scale, w2_scale, block_size); |
| 858 | } else if (moe_comp_method == CPUQuantMethod::FP8_W8A8) { |
| 859 | check_moe_scales<CPUQuantMethod::FP8_W8A8>(w1_scale, w2_scale, block_size); |
| 860 | } |
| 861 | } |
| 862 | |
| 863 | // CHECK_MOE_SCALES_FP8: validate scale shapes for FP8_W8A16. |
| 864 | // w1s shape [E, 2N/block_N, K/block_K, block_N]: DIM0=1, DIM1=2 |
| 865 | // w2s shape [E, K/block_N, N/block_K]: DIM0=1, DIM1=2 |
| 866 | #define CHECK_MOE_SCALES_FP8(DIM0, DIM1) \ |
| 867 | auto w1s = w1_scale.value(); \ |
| 868 | auto w2s = w2_scale.value(); \ |
| 869 | auto block_size_val = block_size.value(); \ |
| 870 | int64_t block_size_N = block_size_val[0]; \ |
| 871 | int64_t block_size_K = block_size_val[1]; \ |
| 872 | TORCH_CHECK(w1s.size(DIM0) == div_up(2 * N, block_size_N)); \ |
| 873 | TORCH_CHECK(w1s.size(DIM1) == div_up(K, block_size_K)); \ |
| 874 | TORCH_CHECK(w2s.size(DIM0) == div_up(K, block_size_N)); \ |
| 875 | TORCH_CHECK(w2s.size(DIM1) == div_up(N, block_size_K)) |
| 876 | |
| 877 | // CHECK_MOE_SCALES_FP8_W8A8: extract scale info for FP8_W8A8 without shape |
| 878 | // validation, because w1s is pre-packed by float8_linear_prepack_cpu with |
| 879 | // shape [E, 2N/BLOCK_N, G, BLOCK_N] where BLOCK_N=32 != block_size_N(128). |
| 880 | // This matches the sglang upstream behavior (checks are commented out there). |
| 881 | #define CHECK_MOE_SCALES_FP8_W8A8(DIM0, DIM1) \ |
| 882 | auto w1s = w1_scale.value(); \ |
| 883 | auto w2s = w2_scale.value(); \ |
| 884 | auto block_size_val = block_size.value(); \ |
| 885 | int64_t block_size_N = block_size_val[0]; \ |
| 886 | int64_t block_size_K = block_size_val[1] |
| 887 | |
| 888 | // hidden_states: [M, K] |
| 889 | // w1: [E, 2N, K] or [E, 2N, K / 2] for uint8 |
| 890 | // w2: [E, K, N] or [E, K, N / 2] for uint8 |
| 891 | // topk_weights: [M, topk] |
| 892 | // topk_ids: [M, topk] (int32_t) |
| 893 | // |
| 894 | |
| 895 | void fused_experts_cpu( |
| 896 | at::Tensor& out, |
| 897 | at::Tensor& hidden_states, |
| 898 | at::Tensor& w1, |
| 899 | at::Tensor& w2, |
| 900 | at::Tensor& topk_weights, |
| 901 | at::Tensor& topk_ids, |
| 902 | int64_t moe_comp_method, |
| 903 | const std::optional<at::Tensor>& w1_scale, |
| 904 | const std::optional<at::Tensor>& w2_scale, |
| 905 | const std::optional<at::Tensor>& w1_zero, |
| 906 | const std::optional<at::Tensor>& w2_zero, |
| 907 | const std::optional<at::Tensor>& a1_scale, |
| 908 | const std::optional<std::vector<int64_t>> block_size, |
| 909 | const std::optional<at::Tensor>& w1_bias, |
| 910 | const std::optional<at::Tensor>& w2_bias, |
| 911 | const std::optional<double>& alpha, |
| 912 | const std::optional<double>& limit, |
| 913 | bool is_vnni) { |
| 914 | auto packed_w1 = is_vnni ? w1 : convert_weight_packed(w1); |
| 915 | auto packed_w2 = is_vnni ? w2 : convert_weight_packed(w2); |
| 916 | |
| 917 | constexpr int64_t BLOCK_M = block_size_m(); |
| 918 | constexpr int64_t BLOCK_N = block_size_n(); |
| 919 | |
| 920 | // FP8_W8A8 hidden_states input is FP8; scalar_t should be BF16 for output |
| 921 | const auto st = (moe_comp_method == CPUQuantMethod::FP8_W8A8) |
| 922 | ? at::kBFloat16 |
| 923 | : hidden_states.scalar_type(); |
| 924 | // TODO: fused_topk_torch_native (CPU fallback for models like MiniMax) |
| 925 | // returns int64 topk_ids; fused_experts_cpu requires int32. Remove the typecast after topk kernel is provided |
| 926 | auto topk_ids_ = topk_ids.scalar_type() == at::kInt ? topk_ids : topk_ids.to(at::kInt); |
| 927 | |
| 928 | CHECK_INPUT(hidden_states); |
| 929 | CHECK_INPUT(w1); |
| 930 | CHECK_INPUT(w2); |
| 931 | CHECK_EQ(topk_weights.sizes(), topk_ids_.sizes()); |
| 932 | CHECK_DIM(2, hidden_states); |
| 933 | if (moe_comp_method == CPUQuantMethod::INT4_W4A8 && is_vnni) { |
| 934 | CHECK_DIM(4, w1); |
| 935 | CHECK_DIM(4, w2); |
| 936 | } else if (moe_comp_method == CPUQuantMethod::FP8_W8A8) { |
| 937 | // w1 is prepackaged by float8_linear_prepack_cpu (5D), skip w1 dim check |
| 938 | CHECK_DIM(3, w2); |
| 939 | } else { |
| 940 | CHECK_DIM(3, w1); |
| 941 | CHECK_DIM(3, w2); |
| 942 | } |
| 943 | CHECK_DIM(2, topk_weights); |
| 944 | CHECK_DIM(2, topk_ids_); |
| 945 | CHECK_EQ(topk_ids_.scalar_type(), at::kInt); |
| 946 | |
| 947 | // TODO: support topk_weights to be bf16 or fp16 in the kernel. |
| 948 | // The topk_weights of llama4 is computed via Llama4MoE:custom_routing_function and is bf16/fp16 |
| 949 | // while the kernel currently only supports it to be float32 |
| 950 | auto topk_weights_ = topk_weights.to(at::kFloat); |
| 951 | CHECK_EQ(topk_weights_.scalar_type(), at::kFloat); |
| 952 | |
| 953 | int64_t M = hidden_states.size(0); |
| 954 | int64_t K = hidden_states.size(1); |
| 955 | int64_t N = moe_comp_method == CPUQuantMethod::INT4_W4A8 |
| 956 | ? w1_scale.value().size(1) * w1_scale.value().size(3) / 2 |
| 957 | : (moe_comp_method == CPUQuantMethod::FP8_W8A8 ? w2.size(2) : w1.size(1) / 2); |
| 958 | int64_t E = w1.size(0); |
| 959 | int64_t topk = topk_weights_.size(1); |
| 960 | |
| 961 | // we use int32_t compensation for int8 w8a8 |
| 962 | int64_t packed_K = get_row_size(static_cast<CPUQuantMethod>(moe_comp_method), K); |
| 963 | int64_t packed_N = get_row_size(static_cast<CPUQuantMethod>(moe_comp_method), N); |
| 964 | |
| 965 | // check weight shapes |
| 966 | CHECK_EQ(w2.size(0), E); |
| 967 | if (moe_comp_method == CPUQuantMethod::FP8_W8A8) { |
| 968 | CHECK_EQ(w2.size(1), K); |
| 969 | CHECK_EQ(packed_w2.size(2), packed_N); |
| 970 | } else if (!(moe_comp_method == CPUQuantMethod::INT4_W4A8)) { |
| 971 | CHECK_EQ(w2.size(1), K); |
| 972 | CHECK_EQ(packed_w1.size(2), packed_K / (moe_comp_method == CPUQuantMethod::INT4_W4A8 ? 2 : 1)); |
| 973 | CHECK_EQ(packed_w2.size(2), packed_N / (moe_comp_method == CPUQuantMethod::INT4_W4A8 ? 2 : 1)); |
| 974 | } |
| 975 | // check scales |
| 976 | check_moe_scales(moe_comp_method, w1_scale, w2_scale, block_size); |
| 977 | |
| 978 | at::Tensor& out_hidden_states = out; |
| 979 | |
| 980 | // NB: worst case is each expert holds a block with remainder of 1 |
| 981 | // 1. sorted_ids : [M * topk + E * (BLOCK_M - 1)] |
| 982 | // 2. expert_ids : [max_num_blocks] |
| 983 | // 3. total_cnts : [T + 1, E] |
| 984 | // 4. cumsums : [E + 1] |
| 985 | // 5. offsets : [max_num_blocks + 1] |
| 986 | // |
| 987 | int num_threads = at::get_num_threads(); |
| 988 | int64_t max_num_tokens_padded = M * topk + E * (BLOCK_M - 1); |
| 989 | int64_t max_num_blocks = div_up(max_num_tokens_padded, BLOCK_M); |
| 990 | auto buffer = at::empty( |
| 991 | {max_num_tokens_padded + max_num_blocks + (num_threads + 1) * E + (E + 1) + (max_num_blocks + 1)}, |
| 992 | topk_ids_.options()); |
| 993 | |
| 994 | int32_t* __restrict__ sorted_ids = buffer.data_ptr<int32_t>(); |
| 995 | int32_t* __restrict__ expert_ids = sorted_ids + max_num_tokens_padded; |
| 996 | int32_t* __restrict__ total_cnts = expert_ids + max_num_blocks; |
| 997 | int32_t* __restrict__ cumsums = total_cnts + (num_threads + 1) * E; |
| 998 | int32_t* __restrict__ offsets = cumsums + (E + 1); |
| 999 | |
| 1000 | // init sorted_ids with `numel` as the padding number |
| 1001 | // init expert_ids with `num_experts` |
| 1002 | int64_t numel = M * topk; |
| 1003 | at::parallel_for(0, max_num_blocks, GRAIN_SIZE / BLOCK_M, [&](int64_t begin, int64_t end) { |
| 1004 | int64_t m_start = begin * BLOCK_M; |
| 1005 | int64_t m_size = std::min((end - begin) * BLOCK_M, max_num_tokens_padded - m_start); |
| 1006 | fill_stub(sorted_ids + m_start, (int32_t)numel, m_size); |
| 1007 | fill_stub(expert_ids + begin, (int32_t)E, end - begin); |
| 1008 | }); |
| 1009 | // zero total_cnts and cumsums |
| 1010 | at::parallel_for(0, (num_threads + 1) * E + (E + 1), GRAIN_SIZE, [&](int64_t begin, int64_t end) { |
| 1011 | fill_stub(total_cnts + begin, 0, end - begin); |
| 1012 | }); |
| 1013 | |
| 1014 | // align experts index |
| 1015 | int64_t num_tokens_post_pad = moe_align_block_size<BLOCK_M>( |
| 1016 | sorted_ids, expert_ids, topk_ids_.data_ptr<int32_t>(), total_cnts, cumsums, offsets, E, numel, num_threads); |
| 1017 | |
| 1018 | // unlike triton kernel, we fuse silu with gemm1 so only need 2 intermediate_caches: |
| 1019 | // 1. intermediate_cache1 : [M * topk, N] |
| 1020 | // 2. intermediate_cache2 : [M * topk, K] |
| 1021 | // 3. A_tmp : [T, BLOCK_M * K] |
| 1022 | // 4. C_tmp : [T, 2 * BLOCK_M * BLOCK_N] |
| 1023 | // |
| 1024 | // for int8 w8a8: |
| 1025 | // 5. Aq_tmp : [M, K] or [M * topk, N] |
| 1026 | // 6. As_tmp : [M * topk] |
| 1027 | // |
| 1028 | // for fp8 w8a16, mxfp4, and fp8 w8a8: |
| 1029 | // 7. intermediate_cache0 : [M * topk, 2N] |
| 1030 | // 8. B_tmp : [T, MAX_CACHE_BLOCK_SIZE, BLOCK_N, std::max(K, N)] |
| 1031 | // |
| 1032 | // for fp8 w8a8 additionally: |
| 1033 | // 9. Ukernel_tmp : [T, 2 * BLOCK_M * BLOCK_N] (float32 for FP8×FP8 brgemm) |
| 1034 | // |
| 1035 | int64_t buffer_size_nbytes = |
| 1036 | M * topk * N * 2 + M * topk * K * 2 + |
| 1037 | num_threads * BLOCK_M * K * |
| 1038 | (moe_comp_method == CPUQuantMethod::INT8_W8A8 | moe_comp_method == CPUQuantMethod::INT4_W4A8 |
| 1039 | | moe_comp_method == CPUQuantMethod::FP8_W8A8 ? 1 : 2) + |
| 1040 | num_threads * 2 * BLOCK_M * BLOCK_N * sizeof(float); |
| 1041 | |
| 1042 | if (moe_comp_method == CPUQuantMethod::INT8_W8A8) { |
| 1043 | buffer_size_nbytes += std::max(M * K, M * topk * N) + M * topk * sizeof(float); |
| 1044 | } |
| 1045 | if (moe_comp_method == CPUQuantMethod::FP8_W8A16 || moe_comp_method == CPUQuantMethod::MXFP4 |
| 1046 | || moe_comp_method == CPUQuantMethod::FP8_W8A8) { |
| 1047 | buffer_size_nbytes += M * topk * 2 * N * 2 + num_threads * MAX_CACHE_BLOCK_SIZE * BLOCK_N * std::max(K, N) * 2; |
| 1048 | } |
| 1049 | if (moe_comp_method == CPUQuantMethod::FP8_W8A8) { |
| 1050 | // Ukernel buffer: float32 workspace per thread for FP8×FP8 brgemm |
| 1051 | buffer_size_nbytes += num_threads * 2 * BLOCK_M * BLOCK_N * sizeof(float); |
| 1052 | #ifndef CPUBLAS_BRGEMM_F8F8F32 |
| 1053 | // BF16 fallback: also need dqA [T, BLOCK_M, BLOCK_K] + dqB [T, BLOCK_K, BLOCK_N] |
| 1054 | buffer_size_nbytes += num_threads * (BLOCK_M * BLOCK_K + BLOCK_K * BLOCK_N) * 2; |
| 1055 | #endif |
| 1056 | } |
| 1057 | if (moe_comp_method == CPUQuantMethod::INT4_W4A8) { |
| 1058 | buffer_size_nbytes += M * topk * 2 * N * 2 + std::max(M * K, M * topk * N) + M * topk * sizeof(float) + |
| 1059 | num_threads * 2 * get_4bit_block_k_size(K / w1_scale.value().size(2)) * BLOCK_N; |
| 1060 | } |
| 1061 | auto buffer2 = at::empty({buffer_size_nbytes}, hidden_states.options().dtype(at::kChar)); |
| 1062 | |
| 1063 | AT_DISPATCH_REDUCED_FLOATING_TYPES(st, "fused_experts_kernel_impl", [&] { |
| 1064 | scalar_t* __restrict__ intermediate_cache1 = (scalar_t*)((void*)(buffer2.data_ptr<int8_t>())); |
| 1065 | scalar_t* __restrict__ intermediate_cache2 = intermediate_cache1 + M * topk * N; |
| 1066 | |
| 1067 | if (moe_comp_method == CPUQuantMethod::INT8_W8A8) { |
| 1068 | uint8_t* __restrict__ A_tmp = (uint8_t*)((void*)(intermediate_cache2 + M * topk * K)); |
| 1069 | float* __restrict__ C_tmp = (float*)((void*)(A_tmp + num_threads * BLOCK_M * K)); |
| 1070 | uint8_t* __restrict__ Aq_tmp = (uint8_t*)((void*)(C_tmp + num_threads * 2 * BLOCK_M * BLOCK_N)); |
| 1071 | float* __restrict__ As_tmp = (float*)((void*)(Aq_tmp + std::max(M * K, M * topk * N))); |
| 1072 | |
| 1073 | auto w1s = w1_scale.value(); |
| 1074 | auto w2s = w2_scale.value(); |
| 1075 | TORCH_CHECK(w1s.numel() == E * 2 * N); |
| 1076 | TORCH_CHECK(w2s.numel() == E * K); |
| 1077 | |
| 1078 | fused_experts_int8_kernel_impl<scalar_t>( |
| 1079 | out_hidden_states.data_ptr<scalar_t>(), |
| 1080 | intermediate_cache1, |
| 1081 | intermediate_cache2, |
| 1082 | A_tmp, |
| 1083 | C_tmp, |
| 1084 | Aq_tmp, |
| 1085 | As_tmp, |
| 1086 | hidden_states.data_ptr<scalar_t>(), |
| 1087 | packed_w1.data_ptr<int8_t>(), |
| 1088 | packed_w2.data_ptr<int8_t>(), |
| 1089 | w1s.data_ptr<float>(), |
| 1090 | w2s.data_ptr<float>(), |
| 1091 | topk_weights_.data_ptr<float>(), |
| 1092 | sorted_ids, |
| 1093 | expert_ids, |
| 1094 | offsets, |
| 1095 | M, |
| 1096 | N, |
| 1097 | K, |
| 1098 | E, |
| 1099 | topk, |
| 1100 | num_tokens_post_pad); |
| 1101 | } else if (moe_comp_method == CPUQuantMethod::FP8_W8A16) { |
| 1102 | // here we just ignore C_tmp as it is not used |
| 1103 | scalar_t* __restrict__ A_tmp = (scalar_t*)((void*)(intermediate_cache2 + M * topk * K)); |
| 1104 | float* __restrict__ C_tmp = (float*)((void*)(A_tmp + num_threads * BLOCK_M * K)); |
| 1105 | scalar_t* __restrict__ intermediate_cache0 = (scalar_t*)((void*)(C_tmp + num_threads * 2 * BLOCK_M * BLOCK_N)); |
| 1106 | scalar_t* __restrict__ B_tmp = (scalar_t*)((void*)(intermediate_cache0 + M * topk * 2 * N)); |
| 1107 | bool with_bias = w1_bias.has_value(); |
| 1108 | auto act_func = alpha.has_value() && limit.has_value() ? CPUActMethod::swiglu : CPUActMethod::silu_and_mul; |
| 1109 | |
| 1110 | CHECK_MOE_SCALES_FP8(1, 2); |
| 1111 | fused_experts_fp_kernel_impl<scalar_t, at::Float8_e4m3fn, float, false>( |
| 1112 | out_hidden_states.data_ptr<scalar_t>(), |
| 1113 | intermediate_cache0, |
| 1114 | intermediate_cache1, |
| 1115 | intermediate_cache2, |
| 1116 | A_tmp, |
| 1117 | B_tmp, |
| 1118 | C_tmp, |
| 1119 | hidden_states.data_ptr<scalar_t>(), |
| 1120 | packed_w1.data_ptr<at::Float8_e4m3fn>(), |
| 1121 | packed_w2.data_ptr<at::Float8_e4m3fn>(), |
| 1122 | with_bias ? w1_bias.value().data_ptr<float>() : nullptr, |
| 1123 | with_bias ? w2_bias.value().data_ptr<float>() : nullptr, |
| 1124 | w1s.data_ptr<float>(), |
| 1125 | w2s.data_ptr<float>(), |
| 1126 | block_size_N, |
| 1127 | block_size_K, |
| 1128 | topk_weights_.data_ptr<float>(), |
| 1129 | sorted_ids, |
| 1130 | expert_ids, |
| 1131 | offsets, |
| 1132 | M, |
| 1133 | N, |
| 1134 | K, |
| 1135 | E, |
| 1136 | topk, |
| 1137 | num_tokens_post_pad, |
| 1138 | alpha.has_value() ? float(alpha.value()) : 0, |
| 1139 | limit.has_value() ? float(limit.value()) : 0, |
| 1140 | act_func, |
| 1141 | with_bias); |
| 1142 | } else if (moe_comp_method == CPUQuantMethod::MXFP4) { |
| 1143 | scalar_t* __restrict__ A_tmp = (scalar_t*)((void*)(intermediate_cache2 + M * topk * K)); |
| 1144 | float* __restrict__ C_tmp = (float*)((void*)(A_tmp + num_threads * BLOCK_M * K)); |
| 1145 | scalar_t* __restrict__ intermediate_cache0 = (scalar_t*)((void*)(C_tmp + num_threads * 2 * BLOCK_M * BLOCK_N)); |
| 1146 | scalar_t* __restrict__ B_tmp = (scalar_t*)((void*)(intermediate_cache0 + M * topk * 2 * N)); |
| 1147 | bool with_bias = w1_bias.has_value(); |
| 1148 | auto act_func = alpha.has_value() && limit.has_value() ? CPUActMethod::swiglu : CPUActMethod::silu_and_mul; |
| 1149 | |
| 1150 | // mxfp4 supports only group size of 32 (2^5) |
| 1151 | constexpr int64_t group_size = 32; |
| 1152 | auto w1s = w1_scale.value(); |
| 1153 | auto w2s = w2_scale.value(); |
| 1154 | TORCH_CHECK(w1s.numel() == E * 2 * N * K / group_size, "w1_scale size mismatch"); |
| 1155 | TORCH_CHECK(w2s.numel() == E * K * N / group_size, "w2_scale size mismatch"); |
| 1156 | fused_experts_fp_kernel_impl<scalar_t, uint8_t, uint8_t, true>( |
| 1157 | out_hidden_states.data_ptr<scalar_t>(), |
| 1158 | intermediate_cache0, |
| 1159 | intermediate_cache1, |
| 1160 | intermediate_cache2, |
| 1161 | A_tmp, |
| 1162 | B_tmp, |
| 1163 | C_tmp, |
| 1164 | hidden_states.data_ptr<scalar_t>(), |
| 1165 | packed_w1.data_ptr<uint8_t>(), |
| 1166 | packed_w2.data_ptr<uint8_t>(), |
| 1167 | with_bias ? w1_bias.value().data_ptr<float>() : nullptr, |
| 1168 | with_bias ? w2_bias.value().data_ptr<float>() : nullptr, |
| 1169 | w1s.data_ptr<uint8_t>(), |
| 1170 | w2s.data_ptr<uint8_t>(), |
| 1171 | /*block_size_N*/ 1, |
| 1172 | /*block_size_K*/ group_size, |
| 1173 | topk_weights_.data_ptr<float>(), |
| 1174 | sorted_ids, |
| 1175 | expert_ids, |
| 1176 | offsets, |
| 1177 | M, |
| 1178 | N, |
| 1179 | K, |
| 1180 | E, |
| 1181 | topk, |
| 1182 | num_tokens_post_pad, |
| 1183 | alpha.has_value() ? float(alpha.value()) : 0, |
| 1184 | limit.has_value() ? float(limit.value()) : 0, |
| 1185 | act_func, |
| 1186 | with_bias); |
| 1187 | } else if (moe_comp_method == CPUQuantMethod::INT4_W4A8) { |
| 1188 | uint8_t* __restrict__ A_tmp = (uint8_t*)((void*)(intermediate_cache2 + M * topk * K)); |
| 1189 | float* __restrict__ C_tmp = (float*)((void*)(A_tmp + num_threads * BLOCK_M * K)); |
| 1190 | scalar_t* __restrict__ intermediate_cache0 = (scalar_t*)((void*)(C_tmp + num_threads * 2 * BLOCK_M * BLOCK_N)); |
| 1191 | uint8_t* __restrict__ Aq_tmp = (uint8_t*)((void*)(intermediate_cache0 + M * topk * 2 * N)); |
| 1192 | float* __restrict__ As_tmp = (float*)((void*)(Aq_tmp + std::max(M * K, M * topk * N))); |
| 1193 | int8_t* __restrict__ dqB_tmp = (int8_t*)((void*)(As_tmp + M * topk)); |
| 1194 | |
| 1195 | // weight + compensation shape = [Nc, Kc, block_n * block_k / 2 + block_n*sizeof(int32_t)] |
| 1196 | // scales/qzeros shape = [E, Nc, G, block_n] |
| 1197 | int64_t num_groups = w1_scale.value().size(2); |
| 1198 | const int group_size = K / num_groups; |
| 1199 | // TODO: check scales and zeros |
| 1200 | fused_experts_int4_w4a8_kernel_impl<scalar_t>( |
| 1201 | out_hidden_states.data_ptr<scalar_t>(), |
| 1202 | intermediate_cache0, |
| 1203 | intermediate_cache1, |
| 1204 | intermediate_cache2, |
| 1205 | A_tmp, |
| 1206 | Aq_tmp, |
| 1207 | As_tmp, |
| 1208 | nullptr, |
| 1209 | C_tmp, |
| 1210 | dqB_tmp, |
| 1211 | hidden_states.data_ptr<scalar_t>(), |
| 1212 | packed_w1.data_ptr<uint8_t>(), |
| 1213 | packed_w2.data_ptr<uint8_t>(), |
| 1214 | w1_zero.value().data_ptr<int8_t>(), |
| 1215 | w2_zero.value().data_ptr<int8_t>(), |
| 1216 | w1_scale.value().data_ptr<float>(), |
| 1217 | w2_scale.value().data_ptr<float>(), |
| 1218 | group_size, |
| 1219 | topk_weights.data_ptr<float>(), |
| 1220 | sorted_ids, |
| 1221 | expert_ids, |
| 1222 | offsets, |
| 1223 | M, |
| 1224 | N, |
| 1225 | K, |
| 1226 | E, |
| 1227 | topk, |
| 1228 | num_tokens_post_pad); |
| 1229 | } else if (moe_comp_method == CPUQuantMethod::FP8_W8A8) { |
| 1230 | // Stage 1: FP8×FP8 GEMM (hidden_states is FP8 input) |
| 1231 | // Stage 2: BF16×FP8 GEMM (W8A16 path for down-projection) |
| 1232 | at::Float8_e4m3fn* __restrict__ A_tmp = |
| 1233 | (at::Float8_e4m3fn*)((void*)(intermediate_cache2 + M * topk * K)); |
| 1234 | float* __restrict__ C_tmp = (float*)((void*)(A_tmp + num_threads * BLOCK_M * K)); |
| 1235 | scalar_t* __restrict__ intermediate_cache0 = |
| 1236 | (scalar_t*)((void*)(C_tmp + num_threads * 2 * BLOCK_M * BLOCK_N)); |
| 1237 | scalar_t* __restrict__ B_tmp = |
| 1238 | (scalar_t*)((void*)(intermediate_cache0 + M * topk * 2 * N)); |
| 1239 | float* __restrict__ Ukernel_tmp = |
| 1240 | (float*)((void*)(B_tmp + num_threads * MAX_CACHE_BLOCK_SIZE * BLOCK_N * std::max(K, N))); |
| 1241 | |
| 1242 | CHECK_MOE_SCALES_FP8_W8A8(1, 2); |
| 1243 | fused_experts_fp8_a8_kernel_impl( |
| 1244 | out_hidden_states.data_ptr<scalar_t>(), |
| 1245 | intermediate_cache0, |
| 1246 | intermediate_cache1, |
| 1247 | intermediate_cache2, |
| 1248 | A_tmp, |
| 1249 | B_tmp, |
| 1250 | C_tmp, |
| 1251 | Ukernel_tmp, |
| 1252 | hidden_states.data_ptr<at::Float8_e4m3fn>(), |
| 1253 | packed_w1.data_ptr<at::Float8_e4m3fn>(), |
| 1254 | packed_w2.data_ptr<at::Float8_e4m3fn>(), |
| 1255 | a1_scale.has_value() ? a1_scale.value().data_ptr<float>() : nullptr, |
| 1256 | w1s.data_ptr<float>(), |
| 1257 | w2s.data_ptr<float>(), |
| 1258 | block_size_N, |
| 1259 | block_size_K, |
| 1260 | topk_weights_.data_ptr<float>(), |
| 1261 | sorted_ids, |
| 1262 | expert_ids, |
| 1263 | offsets, |
| 1264 | M, |
| 1265 | N, |
| 1266 | K, |
| 1267 | E, |
| 1268 | topk, |
| 1269 | num_tokens_post_pad); |
| 1270 | } else { |
| 1271 | scalar_t* __restrict__ A_tmp = intermediate_cache2 + M * topk * K; |
| 1272 | float* __restrict__ C_tmp = (float*)((void*)(A_tmp + num_threads * BLOCK_M * K)); |
| 1273 | bool with_bias = w1_bias.has_value(); |
| 1274 | auto act_func = alpha.has_value() && limit.has_value() ? CPUActMethod::swiglu : CPUActMethod::silu_and_mul; |
| 1275 | |
| 1276 | fused_experts_kernel_impl<scalar_t>( |
| 1277 | out_hidden_states.data_ptr<scalar_t>(), |
| 1278 | intermediate_cache1, |
| 1279 | intermediate_cache2, |
| 1280 | A_tmp, |
| 1281 | C_tmp, |
| 1282 | hidden_states.data_ptr<scalar_t>(), |
| 1283 | packed_w1.data_ptr<scalar_t>(), |
| 1284 | packed_w2.data_ptr<scalar_t>(), |
| 1285 | with_bias ? w1_bias.value().data_ptr<float>() : nullptr, |
| 1286 | with_bias ? w2_bias.value().data_ptr<float>() : nullptr, |
| 1287 | topk_weights_.data_ptr<float>(), |
| 1288 | sorted_ids, |
| 1289 | expert_ids, |
| 1290 | offsets, |
| 1291 | M, |
| 1292 | N, |
| 1293 | K, |
| 1294 | E, |
| 1295 | topk, |
| 1296 | num_tokens_post_pad, |
| 1297 | alpha.has_value() ? float(alpha.value()) : 0, |
| 1298 | limit.has_value() ? float(limit.value()) : 0, |
| 1299 | act_func, |
| 1300 | with_bias); |
| 1301 | } |
| 1302 | }); |
| 1303 | } |
| 1304 | |
| 1305 | // shared expert kernel |
| 1306 | // |
| 1307 | // hidden_states: [M, K] |
| 1308 | // w1: [2N, K] |
| 1309 | // w2: [K, N] |
| 1310 | // fused_experts_out |
| 1311 | at::Tensor shared_expert_cpu( |
| 1312 | at::Tensor& hidden_states, |
| 1313 | at::Tensor& w1, |
| 1314 | at::Tensor& w2, |
| 1315 | const std::optional<at::Tensor>& fused_experts_out, |
| 1316 | const std::optional<double> routed_scaling_factor, |
| 1317 | bool inplace, |
| 1318 | bool use_int8_w8a8, |
| 1319 | bool use_fp8_w8a16, |
| 1320 | const std::optional<at::Tensor>& w1_scale, |
| 1321 | const std::optional<at::Tensor>& w2_scale, |
| 1322 | const std::optional<std::vector<int64_t>> block_size, |
| 1323 | bool is_vnni) { |
| 1324 | auto packed_w1 = is_vnni ? w1 : convert_weight_packed(w1); |
| 1325 | auto packed_w2 = is_vnni ? w2 : convert_weight_packed(w2); |
| 1326 | |
| 1327 | constexpr int64_t BLOCK_M = block_size_m(); |
| 1328 | constexpr int64_t BLOCK_N = block_size_n(); |
| 1329 | |
| 1330 | double routed_scaling_factor_value = 0; |
| 1331 | if (routed_scaling_factor.has_value()) { |
| 1332 | TORCH_CHECK(fused_experts_out.has_value(), "shared_expert_cpu: expect fused_experts_out."); |
| 1333 | const auto fused_experts_out_tensor = fused_experts_out.value(); |
| 1334 | routed_scaling_factor_value = routed_scaling_factor.value(); |
| 1335 | CHECK_INPUT(fused_experts_out_tensor); |
| 1336 | CHECK_EQ(hidden_states.sizes(), fused_experts_out_tensor.sizes()); |
| 1337 | } |
| 1338 | |
| 1339 | const auto st = hidden_states.scalar_type(); |
| 1340 | CHECK_INPUT(hidden_states); |
| 1341 | CHECK_INPUT(w1); |
| 1342 | CHECK_INPUT(w2); |
| 1343 | CHECK_DIM(2, hidden_states); |
| 1344 | CHECK_DIM(2, w1); |
| 1345 | CHECK_DIM(2, w2); |
| 1346 | CHECK_EQ(hidden_states.scalar_type(), st); |
| 1347 | |
| 1348 | int64_t M = hidden_states.size(0); |
| 1349 | int64_t K = hidden_states.size(1); |
| 1350 | int64_t N = w1.size(0) / 2; |
| 1351 | |
| 1352 | // we use int32_t compensation for int8 w8a8 |
| 1353 | int64_t packed_K = get_row_size(K, use_int8_w8a8); |
| 1354 | int64_t packed_N = get_row_size(N, use_int8_w8a8); |
| 1355 | |
| 1356 | // check weight shapes |
| 1357 | CHECK_EQ(w2.size(0), K); |
| 1358 | CHECK_EQ(packed_w1.size(1), packed_K); |
| 1359 | CHECK_EQ(packed_w2.size(1), packed_N); |
| 1360 | |
| 1361 | // check scales |
| 1362 | if (use_int8_w8a8) { |
| 1363 | check_moe_scales<CPUQuantMethod::INT8_W8A8>(w1_scale, w2_scale, block_size); |
| 1364 | } else if (use_fp8_w8a16) { |
| 1365 | check_moe_scales<CPUQuantMethod::FP8_W8A16>(w1_scale, w2_scale, block_size); |
| 1366 | } |
| 1367 | |
| 1368 | at::Tensor out_hidden_states = inplace ? hidden_states : at::empty_like(hidden_states); |
| 1369 | |
| 1370 | // unlike triton kernel, we fuse silu with gemm1 so only need 2 intermediate_caches: |
| 1371 | // 1. intermediate_cache1 : [M, N] |
| 1372 | // 2. C_tmp : [T, 2 * BLOCK_M * BLOCK_N] |
| 1373 | // |
| 1374 | // for int8 w8a8: |
| 1375 | // 3. Aq_tmp : [M, K] or [M, N] |
| 1376 | // 4. As_tmp : [M] |
| 1377 | // |
| 1378 | // for fp8 w8a16: |
| 1379 | // 5. intermediate_cache0 : [M, 2N] |
| 1380 | // 6. B_tmp: [T, MAX_CACHE_BLOCK_SIZE, BLOCK_M, max(K, N)] |
| 1381 | // |
| 1382 | int num_threads = at::get_num_threads(); |
| 1383 | int64_t buffer_size_nbytes = M * N * 2 + num_threads * 2 * BLOCK_M * BLOCK_N * sizeof(float); |
| 1384 | |
| 1385 | if (use_int8_w8a8) { |
| 1386 | buffer_size_nbytes += std::max(M * K, M * N) + M * sizeof(float); |
| 1387 | } |
| 1388 | if (use_fp8_w8a16) { |
| 1389 | buffer_size_nbytes += M * 2 * N * 2 + num_threads * MAX_CACHE_BLOCK_SIZE * BLOCK_M * std::max(K, N) * 2; |
| 1390 | } |
| 1391 | |
| 1392 | auto buffer = at::empty({buffer_size_nbytes}, hidden_states.options().dtype(at::kChar)); |
| 1393 | AT_DISPATCH_REDUCED_FLOATING_TYPES(st, "share_experts_kernel_impl", [&] { |
| 1394 | scalar_t* __restrict__ intermediate_cache1 = (scalar_t*)((void*)(buffer.data_ptr<int8_t>())); |
| 1395 | float* __restrict__ C_tmp = (float*)((void*)(intermediate_cache1 + M * N)); |
| 1396 | |
| 1397 | if (use_int8_w8a8) { |
| 1398 | uint8_t* __restrict__ Aq_tmp = (uint8_t*)((void*)(C_tmp + num_threads * 2 * BLOCK_M * BLOCK_N)); |
| 1399 | float* __restrict__ As_tmp = (float*)((void*)(Aq_tmp + std::max(M * K, M * N))); |
| 1400 | |
| 1401 | auto w1s = w1_scale.value(); |
| 1402 | auto w2s = w2_scale.value(); |
| 1403 | TORCH_CHECK(w1s.numel() == 2 * N); |
| 1404 | TORCH_CHECK(w2s.numel() == K); |
| 1405 | |
| 1406 | shared_expert_int8_kernel_impl<scalar_t>( |
| 1407 | out_hidden_states.data_ptr<scalar_t>(), |
| 1408 | intermediate_cache1, |
| 1409 | C_tmp, |
| 1410 | Aq_tmp, |
| 1411 | As_tmp, |
| 1412 | hidden_states.data_ptr<scalar_t>(), |
| 1413 | packed_w1.data_ptr<int8_t>(), |
| 1414 | packed_w2.data_ptr<int8_t>(), |
| 1415 | w1s.data_ptr<float>(), |
| 1416 | w2s.data_ptr<float>(), |
| 1417 | conditional_data_ptr<scalar_t>(fused_experts_out), |
| 1418 | routed_scaling_factor_value, |
| 1419 | M, |
| 1420 | N, |
| 1421 | K); |
| 1422 | } else if (use_fp8_w8a16) { |
| 1423 | scalar_t* __restrict__ intermediate_cache0 = (scalar_t*)((void*)(C_tmp + num_threads * 2 * BLOCK_M * BLOCK_N)); |
| 1424 | scalar_t* __restrict__ B_tmp = (scalar_t*)((void*)(intermediate_cache0 + M * 2 * N)); |
| 1425 | |
| 1426 | CHECK_MOE_SCALES_FP8(0, 1); |
| 1427 | shared_expert_fp8_kernel_impl<scalar_t>( |
| 1428 | out_hidden_states.data_ptr<scalar_t>(), |
| 1429 | intermediate_cache0, |
| 1430 | intermediate_cache1, |
| 1431 | B_tmp, |
| 1432 | C_tmp, |
| 1433 | hidden_states.data_ptr<scalar_t>(), |
| 1434 | packed_w1.data_ptr<at::Float8_e4m3fn>(), |
| 1435 | packed_w2.data_ptr<at::Float8_e4m3fn>(), |
| 1436 | w1s.data_ptr<float>(), |
| 1437 | w2s.data_ptr<float>(), |
| 1438 | block_size_N, |
| 1439 | block_size_K, |
| 1440 | conditional_data_ptr<scalar_t>(fused_experts_out), |
| 1441 | routed_scaling_factor_value, |
| 1442 | M, |
| 1443 | N, |
| 1444 | K); |
| 1445 | } else { |
| 1446 | shared_expert_kernel_impl<scalar_t>( |
| 1447 | out_hidden_states.data_ptr<scalar_t>(), |
| 1448 | intermediate_cache1, |
| 1449 | C_tmp, |
| 1450 | hidden_states.data_ptr<scalar_t>(), |
| 1451 | packed_w1.data_ptr<scalar_t>(), |
| 1452 | packed_w2.data_ptr<scalar_t>(), |
| 1453 | conditional_data_ptr<scalar_t>(fused_experts_out), |
| 1454 | routed_scaling_factor_value, |
| 1455 | M, |
| 1456 | N, |
| 1457 | K); |
| 1458 | } |
| 1459 | }); |
| 1460 | return out_hidden_states; |
| 1461 | } |