| 1 | // Adapted from |
| 2 | // https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu |
| 3 | |
| 4 | // clang-format off |
| 5 | |
| 6 | #include "common.h" |
| 7 | #include "gemm.h" |
| 8 | #include "vec.h" |
| 9 | |
| 10 | namespace { |
| 11 | |
| 12 | // [NOTE] TODO list for this kernel: |
| 13 | // 1. tune the value for BLOCK_N |
| 14 | // 2. planning for {batches, num_heads, num_kv_splits} |
| 15 | // and use actual num_kv_splits for small seq length |
| 16 | // 3. try fast impl of `.tanh()` |
| 17 | // 4. provide amx kernel for index_gemm_kernel_nn when M = 16 |
| 18 | // |
| 19 | |
| 20 | #if defined(CPU_CAPABILITY_AVX512) |
| 21 | // key: from [N, 32] to [32/2, N, 2] |
| 22 | // val: from [N, 32] to [N/2, 32, 2] |
| 23 | template <typename scalar_t, typename index_t> |
| 24 | inline void pack_vnni_Nx32( |
| 25 | scalar_t* __restrict__ dst0, |
| 26 | scalar_t* __restrict__ dst1, |
| 27 | const scalar_t* __restrict__ src, |
| 28 | const index_t* __restrict__ ind, |
| 29 | int N, |
| 30 | int ld_src, |
| 31 | int ld_dst0, |
| 32 | int ld_dst1, |
| 33 | bool convert_v) { |
| 34 | __m512i vinputs[16]; |
| 35 | int n = 0; |
| 36 | for (; n < N; ++n) { |
| 37 | vinputs[n] = _mm512_loadu_si512(src + ind[n] * ld_src); |
| 38 | } |
| 39 | // padding with zero to avoid uninitialized vectors |
| 40 | for (; n < 16; ++n) { |
| 41 | vinputs[n] = _mm512_set1_epi32(0); |
| 42 | } |
| 43 | |
| 44 | // pack value, skip 64 elems for deepseek |
| 45 | // handle 2 vectors at a time from [2, 32] to [32, 2] |
| 46 | if (convert_v) { |
| 47 | for (int n = 0; n < 16; n += 2) { |
| 48 | __m512i d0, d1; |
| 49 | std::tie(d0, d1) = transpose_2x32_16bit(vinputs[n], vinputs[n + 1]); |
| 50 | _mm512_storeu_si512(dst1 + (n >> 1) * ld_dst1 * 2, d0); |
| 51 | _mm512_storeu_si512(dst1 + (n >> 1) * ld_dst1 * 2 + 32, d1); |
| 52 | } |
| 53 | } |
| 54 | |
| 55 | // pack key |
| 56 | transpose_16x16_32bit(vinputs); |
| 57 | |
| 58 | const __mmask16 vmask = (1 << N) - 1; |
| 59 | for (int k = 0; k < 16; ++k) { |
| 60 | _mm512_mask_storeu_epi32(dst0 + k * ld_dst0 * 2, vmask, vinputs[k]); |
| 61 | } |
| 62 | } |
| 63 | #endif |
| 64 | |
| 65 | // [NOTE]: MLA vnni format conversion |
| 66 | // |
| 67 | // here we apply same strategy as `FlashMLA`: |
| 68 | // each kv_cache is loaded once and packed twice (L2 cache hit) |
| 69 | // |
| 70 | // * for key: from [N, K/2, 2] to [K/2, N, 2] |
| 71 | // * for value: from [N/2, 2, Kv] to [N/2, Kv, 2] |
| 72 | // |
| 73 | template <typename scalar_t, typename index_t> |
| 74 | void pack_vnni( |
| 75 | scalar_t* __restrict__ dst0, |
| 76 | scalar_t* __restrict__ dst1, |
| 77 | const scalar_t* __restrict__ src, |
| 78 | const index_t* __restrict__ ind, |
| 79 | int N, |
| 80 | int K, |
| 81 | int Kv, |
| 82 | int ld_src, |
| 83 | int ld_dst0, |
| 84 | int ld_dst1) { |
| 85 | #if defined(CPU_CAPABILITY_AVX512) |
| 86 | const int NB = div_up(N, 16); |
| 87 | const int KB = K / 32; // no remainder |
| 88 | const int KBv = Kv / 32; // no remainder |
| 89 | |
| 90 | for (int nb = 0; nb < NB; ++nb) { |
| 91 | for (int kb = 0; kb < KB; ++kb) { |
| 92 | // handle 16x512bits each block |
| 93 | int nb_size = std::min(N - nb * 16, 16); |
| 94 | pack_vnni_Nx32<scalar_t, index_t>( |
| 95 | /* dst0 */ dst0 + ((kb * 32) >> 1) * ld_dst0 * 2 + nb * 16 * 2, |
| 96 | /* dst1 */ dst1 + ((nb * 16) >> 1) * ld_dst1 * 2 + kb * 32 * 2, |
| 97 | /* src */ src + kb * 32, |
| 98 | /* ind */ ind + nb * 16, |
| 99 | /* N */ nb_size, |
| 100 | /* ld_src */ ld_src, |
| 101 | /* ld_dst0 */ ld_dst0, |
| 102 | /* ld_dst1 */ ld_dst1, |
| 103 | /* cvt_v */ kb < KBv); |
| 104 | } |
| 105 | } |
| 106 | #else |
| 107 | for (int n = 0; n < N; ++n) { |
| 108 | index_t index = ind[n]; |
| 109 | for (int k = 0; k < K / 2; ++k) { |
| 110 | for (int d = 0; d < 2; ++d) { |
| 111 | dst0[k * ld_dst0 * 2 + n * 2 + d] = src[index * ld_src + k * 2 + d]; |
| 112 | } |
| 113 | } |
| 114 | } |
| 115 | // from [N/2, 2, K] to [N/2, K, 2] |
| 116 | for (int n = 0; n < (N >> 1) * 2; n += 2) { |
| 117 | index_t index0 = ind[n + 0]; |
| 118 | index_t index1 = ind[n + 1]; |
| 119 | for (int k = 0; k < Kv; ++k) { |
| 120 | dst1[(n >> 1) * ld_dst1 * 2 + k * 2 + 0] = src[index0 * ld_src + k]; |
| 121 | dst1[(n >> 1) * ld_dst1 * 2 + k * 2 + 1] = src[index1 * ld_src + k]; |
| 122 | } |
| 123 | } |
| 124 | if (N % 2 != 0) { |
| 125 | index_t index = ind[N - 1]; |
| 126 | for (int k = 0; k < Kv; ++k) { |
| 127 | dst1[(N >> 1) * ld_dst1 * 2 + k * 2 + 0] = src[index * ld_src + k]; |
| 128 | dst1[(N >> 1) * ld_dst1 * 2 + k * 2 + 1] = 0; |
| 129 | } |
| 130 | } |
| 131 | #endif |
| 132 | } |
| 133 | |
| 134 | template <typename scalar_t> |
| 135 | inline void fill_stub(scalar_t* __restrict__ out, float val, int64_t size) { |
| 136 | using Vec = at::vec::Vectorized<scalar_t>; |
| 137 | constexpr int kVecSize = Vec::size(); |
| 138 | const Vec data_vec = Vec(static_cast<scalar_t>(val)); |
| 139 | int64_t d = 0; |
| 140 | #pragma GCC unroll 4 |
| 141 | for (; d <= size - kVecSize; d += kVecSize) { |
| 142 | data_vec.store(out + d); |
| 143 | } |
| 144 | if (size - d > 0) { |
| 145 | data_vec.store(out + d, size - d); |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | template <typename scalar_t> |
| 150 | inline void copy_stub(scalar_t* __restrict__ out, const float* __restrict__ acc, float s, int64_t size) { |
| 151 | using bVec = at::vec::Vectorized<scalar_t>; |
| 152 | using fVec = at::vec::Vectorized<float>; |
| 153 | constexpr int kVecSize = bVec::size(); |
| 154 | const fVec s_fvec = fVec(s); |
| 155 | int64_t d = 0; |
| 156 | #pragma GCC unroll 4 |
| 157 | for (; d <= size - kVecSize; d += kVecSize) { |
| 158 | auto [a_fvec0, a_fvec1] = load_float_vec2(acc + d); |
| 159 | a_fvec0 = a_fvec0 * s_fvec; |
| 160 | a_fvec1 = a_fvec1 * s_fvec; |
| 161 | bVec out_bvec = convert_from_float_ext<scalar_t>(a_fvec0, a_fvec1); |
| 162 | out_bvec.store(out + d); |
| 163 | } |
| 164 | for (; d < size; ++d) { |
| 165 | out[d] = static_cast<scalar_t>(acc[d] * s); |
| 166 | } |
| 167 | } |
| 168 | |
| 169 | template <typename scalar_t> |
| 170 | inline void copy_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ src, int64_t size) { |
| 171 | using bVec = at::vec::Vectorized<scalar_t>; |
| 172 | constexpr int kVecSize = bVec::size(); |
| 173 | int64_t d = 0; |
| 174 | #pragma GCC unroll 4 |
| 175 | for (; d <= size - kVecSize; d += kVecSize) { |
| 176 | bVec out_bvec = bVec::loadu(src + d); |
| 177 | out_bvec.store(out + d); |
| 178 | } |
| 179 | for (; d < size; ++d) { |
| 180 | out[d] = src[d]; |
| 181 | } |
| 182 | } |
| 183 | |
| 184 | template <typename scalar_t, int BLOCK_N> |
| 185 | inline void copy_stub(scalar_t* __restrict__ out, const float* __restrict__ input) { |
| 186 | static_assert(BLOCK_N % 32 == 0); |
| 187 | using bVec = at::vec::Vectorized<scalar_t>; |
| 188 | using fVec = at::vec::Vectorized<float>; |
| 189 | |
| 190 | constexpr int COLS = BLOCK_N / 16; |
| 191 | auto store = [&](auto i) { |
| 192 | constexpr int col = i % COLS; |
| 193 | // for COLS = 2, 4 use 512bit store |
| 194 | if constexpr (col % 2 == 0) { |
| 195 | auto [a_fvec0, a_fvec1] = load_float_vec2(input + col * 16); |
| 196 | bVec out_bvec = convert_from_float_ext<scalar_t>(a_fvec0, a_fvec1); |
| 197 | out_bvec.store(out + col * 16); |
| 198 | } |
| 199 | }; |
| 200 | Unroll<COLS>{}(store); |
| 201 | } |
| 202 | |
| 203 | // GEMM handles query @ key (indexed) x scale |
| 204 | // A : [M, K] |
| 205 | // B : [N, K] indexed |
| 206 | // C : [M, N] |
| 207 | // |
| 208 | template <typename scalar_t, typename index_t, int BLOCK_M, int BLOCK_N> |
| 209 | struct tinygemm_kernel_nt { |
| 210 | static inline void apply( |
| 211 | const scalar_t* __restrict__ A, |
| 212 | const scalar_t* __restrict__ B, |
| 213 | float* __restrict__ C, |
| 214 | const index_t* __restrict__ indices, |
| 215 | float scale, |
| 216 | int64_t lda, |
| 217 | int64_t ldb, |
| 218 | int64_t ldc, |
| 219 | int64_t K, |
| 220 | int64_t max_tokens) { |
| 221 | for (int64_t m = 0; m < BLOCK_M; ++m) { |
| 222 | for (int64_t n = 0; n < BLOCK_N; ++n) { |
| 223 | float sum = 0.f; |
| 224 | int64_t b_idx = indices[n]; |
| 225 | TORCH_CHECK(b_idx < max_tokens, "token index out of scope!"); |
| 226 | for (int64_t k = 0; k < K; ++k) { |
| 227 | sum += scale * static_cast<float>(A[m * lda + k]) * static_cast<float>(B[b_idx * ldb + k]); |
| 228 | } |
| 229 | C[m * ldc + n] = sum; |
| 230 | } |
| 231 | } |
| 232 | } |
| 233 | }; |
| 234 | |
| 235 | #if defined(CPU_CAPABILITY_AVX512) |
| 236 | template <typename index_t, int BLOCK_M, int BLOCK_N> |
| 237 | struct tinygemm_kernel_nt<at::BFloat16, index_t, BLOCK_M, BLOCK_N> { |
| 238 | static inline void apply( |
| 239 | const at::BFloat16* __restrict__ A, |
| 240 | const at::BFloat16* __restrict__ B, |
| 241 | float* __restrict__ C, |
| 242 | const index_t* __restrict__ indices, |
| 243 | float scale, |
| 244 | int64_t lda, |
| 245 | int64_t ldb, |
| 246 | int64_t ldc, |
| 247 | int64_t K, |
| 248 | int64_t max_tokens) { |
| 249 | constexpr int ROWS = BLOCK_M; |
| 250 | constexpr int COLS = BLOCK_N; |
| 251 | |
| 252 | __m512bh va; |
| 253 | __m512bh vb[COLS]; |
| 254 | __m512 vc[ROWS * COLS]; |
| 255 | __m512 vscale = _mm512_set1_ps(scale); |
| 256 | |
| 257 | auto loadc = [&](auto i) { vc[i] = _mm512_setzero_ps(); }; |
| 258 | Unroll<ROWS * COLS>{}(loadc); |
| 259 | |
| 260 | // for main loop |
| 261 | auto compute = [&](auto i, int64_t k) { |
| 262 | constexpr int row = i / COLS; |
| 263 | constexpr int col = i % COLS; |
| 264 | |
| 265 | if constexpr (col == 0) { |
| 266 | va = (__m512bh)(_mm512_loadu_si512(A + row * lda + k)); |
| 267 | } |
| 268 | if constexpr (row == 0) { |
| 269 | if constexpr (col + 1 < COLS) { |
| 270 | int64_t b_idx_prefetch = indices[col + 1]; |
| 271 | _mm_prefetch(B + b_idx_prefetch * ldb + k, _MM_HINT_T0); |
| 272 | } |
| 273 | int64_t b_idx = indices[col]; |
| 274 | TORCH_CHECK(b_idx < max_tokens, "token index out of scope!"); |
| 275 | vb[col] = (__m512bh)(_mm512_loadu_si512(B + b_idx * ldb + k)); |
| 276 | } |
| 277 | vc[i] = _mm512_dpbf16_ps(vc[i], va, vb[col]); |
| 278 | }; |
| 279 | |
| 280 | // for remainder |
| 281 | auto compute2 = [&](auto i, int64_t k, __mmask32 mask) { |
| 282 | constexpr int row = i / COLS; |
| 283 | constexpr int col = i % COLS; |
| 284 | |
| 285 | if constexpr (col == 0) { |
| 286 | va = (__m512bh)(_mm512_maskz_loadu_epi16(mask, A + row * lda + k)); |
| 287 | } |
| 288 | if constexpr (row == 0) { |
| 289 | int64_t b_idx = indices[col]; |
| 290 | TORCH_CHECK(b_idx < max_tokens, "token index out of scope!"); |
| 291 | vb[col] = (__m512bh)(_mm512_maskz_loadu_epi16(mask, B + b_idx * ldb + k)); |
| 292 | } |
| 293 | vc[i] = _mm512_dpbf16_ps(vc[i], va, vb[col]); |
| 294 | }; |
| 295 | |
| 296 | int64_t k = 0; |
| 297 | for (; k <= K - 32; k += 32) { |
| 298 | Unroll<ROWS * COLS>{}(compute, k); |
| 299 | } |
| 300 | int64_t count = K - k; |
| 301 | if (count > 0) { |
| 302 | __mmask32 mask = (1ULL << count) - 1; |
| 303 | Unroll<ROWS * COLS>{}(compute2, k, mask); |
| 304 | } |
| 305 | |
| 306 | auto storec = [&](auto i) { |
| 307 | constexpr int row = i / COLS; |
| 308 | constexpr int col = i % COLS; |
| 309 | C[row * ldc + col] = _mm512_reduce_add_ps(_mm512_mul_ps(vc[i], vscale)); |
| 310 | }; |
| 311 | Unroll<ROWS * COLS>{}(storec); |
| 312 | } |
| 313 | }; |
| 314 | #endif |
| 315 | |
| 316 | #if defined(CPU_CAPABILITY_AVX512) |
| 317 | template <typename index_t, int BLOCK_M, int BLOCK_N> |
| 318 | struct tinygemm_kernel_nt<at::Half, index_t, BLOCK_M, BLOCK_N> { |
| 319 | static inline void apply( |
| 320 | const at::Half* __restrict__ A, |
| 321 | const at::Half* __restrict__ B, |
| 322 | float* __restrict__ C, |
| 323 | const index_t* __restrict__ indices, |
| 324 | float scale, |
| 325 | int64_t lda, |
| 326 | int64_t ldb, |
| 327 | int64_t ldc, |
| 328 | int64_t K, |
| 329 | int64_t max_tokens) { |
| 330 | constexpr int ROWS = BLOCK_M; |
| 331 | constexpr int COLS = BLOCK_N; |
| 332 | |
| 333 | __m512 va0, va1; |
| 334 | __m512 vb0[COLS], vb1[COLS]; |
| 335 | __m512 vc[ROWS * COLS]; |
| 336 | __m512 vscale = _mm512_set1_ps(scale); |
| 337 | |
| 338 | auto loadc = [&](auto i) { vc[i] = _mm512_setzero_ps(); }; |
| 339 | Unroll<ROWS * COLS>{}(loadc); |
| 340 | |
| 341 | auto compute = [&](auto i, int64_t k) { |
| 342 | constexpr int row = i / COLS; |
| 343 | constexpr int col = i % COLS; |
| 344 | |
| 345 | if constexpr (col == 0) { |
| 346 | __m512i a16 = _mm512_loadu_si512((__m512i const*)(A + row * lda + k)); |
| 347 | va0 = CVT_FP16_TO_FP32(_mm512_extracti32x8_epi32(a16, 0)); |
| 348 | va1 = CVT_FP16_TO_FP32(_mm512_extracti32x8_epi32(a16, 1)); |
| 349 | } |
| 350 | |
| 351 | if constexpr (row == 0) { |
| 352 | int64_t b_idx = indices[col]; |
| 353 | TORCH_CHECK(b_idx < max_tokens, "token index out of scope!"); |
| 354 | __m512i b16 = _mm512_loadu_si512((__m512i const*)(B + b_idx * ldb + k)); |
| 355 | vb0[col] = CVT_FP16_TO_FP32(_mm512_extracti32x8_epi32(b16, 0)); |
| 356 | vb1[col] = CVT_FP16_TO_FP32(_mm512_extracti32x8_epi32(b16, 1)); |
| 357 | } |
| 358 | |
| 359 | vc[i] = _mm512_fmadd_ps(va0, vb0[col], _mm512_fmadd_ps(va1, vb1[col], vc[i])); |
| 360 | }; |
| 361 | |
| 362 | auto compute2 = [&](auto i, int64_t k, __mmask32 mask) { |
| 363 | constexpr int row = i / COLS; |
| 364 | constexpr int col = i % COLS; |
| 365 | |
| 366 | if constexpr (col == 0) { |
| 367 | __m512i a16 = _mm512_maskz_loadu_epi16(mask, (const void*)(A + row * lda + k)); |
| 368 | va0 = CVT_FP16_TO_FP32(_mm512_extracti32x8_epi32(a16, 0)); |
| 369 | va1 = CVT_FP16_TO_FP32(_mm512_extracti32x8_epi32(a16, 1)); |
| 370 | } |
| 371 | |
| 372 | if constexpr (row == 0) { |
| 373 | int64_t b_idx = indices[col]; |
| 374 | TORCH_CHECK(b_idx < max_tokens, "token index out of scope!"); |
| 375 | __m512i b16 = _mm512_maskz_loadu_epi16(mask, (const void*)(B + b_idx * ldb + k)); |
| 376 | vb0[col] = CVT_FP16_TO_FP32(_mm512_extracti32x8_epi32(b16, 0)); |
| 377 | vb1[col] = CVT_FP16_TO_FP32(_mm512_extracti32x8_epi32(b16, 1)); |
| 378 | } |
| 379 | |
| 380 | vc[i] = _mm512_fmadd_ps(va0, vb0[col], _mm512_fmadd_ps(va1, vb1[col], vc[i])); |
| 381 | }; |
| 382 | |
| 383 | int64_t k = 0; |
| 384 | for (; k <= K - 32; k += 32) { |
| 385 | Unroll<ROWS * COLS>{}(compute, k); |
| 386 | } |
| 387 | int64_t count = K - k; |
| 388 | if (count > 0) { |
| 389 | __mmask32 mask = (1ULL << count) - 1; |
| 390 | Unroll<ROWS * COLS>{}(compute2, k, mask); |
| 391 | } |
| 392 | |
| 393 | auto storec = [&](auto i) { |
| 394 | constexpr int row = i / COLS; |
| 395 | constexpr int col = i % COLS; |
| 396 | C[row * ldc + col] = _mm512_reduce_add_ps(_mm512_mul_ps(vc[i], vscale)); |
| 397 | }; |
| 398 | Unroll<ROWS * COLS>{}(storec); |
| 399 | } |
| 400 | }; |
| 401 | #endif |
| 402 | |
| 403 | #define LAUNCH_TINYGEMM_KERNEL_NT(MB_SIZE, NB_SIZE) \ |
| 404 | tinygemm_kernel_nt<scalar_t, index_t, MB_SIZE, NB_SIZE>::apply( \ |
| 405 | A + mb_start * lda, B, C + mb_start * ldc + nb_start, indices + nb_start, scale, lda, ldb, ldc, K, max_tokens); |
| 406 | |
| 407 | // this is used when N isn't multiple of 16, |
| 408 | // N corresponds to `head_size_v` which should be 16x |
| 409 | template <typename scalar_t, typename index_t> |
| 410 | inline void tinygemm_kernel_nn_scalar( |
| 411 | const float* __restrict__ A, |
| 412 | const scalar_t* __restrict__ B, |
| 413 | float* __restrict__ C, |
| 414 | const index_t* __restrict__ indices, |
| 415 | const float* __restrict__ scale, |
| 416 | int64_t M, |
| 417 | int64_t N, |
| 418 | int64_t K, |
| 419 | int64_t lda, |
| 420 | int64_t ldb, |
| 421 | int64_t ldc, |
| 422 | int64_t max_tokens) { |
| 423 | for (int64_t m = 0; m < M; ++m) { |
| 424 | for (int64_t n = 0; n < N; ++n) { |
| 425 | C[m * ldc + n] *= scale[m]; |
| 426 | for (int64_t k = 0; k < K; ++k) { |
| 427 | int64_t b_idx = indices[k]; |
| 428 | TORCH_CHECK(b_idx < max_tokens, "token index out of scope!"); |
| 429 | C[m * ldc + n] += A[m * lda + k] * static_cast<float>(B[b_idx * ldb + n]); |
| 430 | } |
| 431 | } |
| 432 | } |
| 433 | } |
| 434 | |
| 435 | // GEMM handles v' * scale + attn @ value (indexed) |
| 436 | // A : [M, K] |
| 437 | // B : [K, N] indexed |
| 438 | // C ：[M, N] |
| 439 | // |
| 440 | template <typename scalar_t, typename index_t, int BLOCK_M, int BLOCK_N> |
| 441 | struct tinygemm_kernel_nn { |
| 442 | static inline void apply( |
| 443 | const float* __restrict__ A, |
| 444 | const scalar_t* __restrict__ B, |
| 445 | float* __restrict__ C, |
| 446 | const index_t* __restrict__ indices, |
| 447 | const float* __restrict__ scale, |
| 448 | int64_t lda, |
| 449 | int64_t ldb, |
| 450 | int64_t ldc, |
| 451 | int64_t K, |
| 452 | int64_t max_tokens) { |
| 453 | tinygemm_kernel_nn_scalar(A, B, C, indices, scale, BLOCK_M, BLOCK_N, K, lda, ldb, ldc, max_tokens); |
| 454 | } |
| 455 | }; |
| 456 | |
| 457 | #if defined(CPU_CAPABILITY_AVX512) |
| 458 | template <typename index_t, int BLOCK_M, int BLOCK_N> |
| 459 | struct tinygemm_kernel_nn<at::BFloat16, index_t, BLOCK_M, BLOCK_N> { |
| 460 | static inline void apply( |
| 461 | const float* __restrict__ A, |
| 462 | const at::BFloat16* __restrict__ B, |
| 463 | float* __restrict__ C, |
| 464 | const index_t* __restrict__ indices, |
| 465 | const float* __restrict__ scale, |
| 466 | int64_t lda, |
| 467 | int64_t ldb, |
| 468 | int64_t ldc, |
| 469 | int64_t K, |
| 470 | int64_t max_tokens) { |
| 471 | constexpr int ROWS = BLOCK_M; |
| 472 | constexpr int COLS = BLOCK_N / 16; |
| 473 | |
| 474 | __m512 va; |
| 475 | __m512 vb[COLS]; |
| 476 | __m512 vc[ROWS * COLS]; |
| 477 | __m512 vscale; |
| 478 | |
| 479 | auto loadc = [&](auto i) { |
| 480 | constexpr int row = i / COLS; |
| 481 | constexpr int col = i % COLS; |
| 482 | #pragma GCC diagnostic push |
| 483 | #pragma GCC diagnostic ignored "-Warray-bounds" |
| 484 | if constexpr (col == 0) { |
| 485 | vscale = _mm512_set1_ps(scale[row]); |
| 486 | } |
| 487 | #pragma GCC diagnostic pop |
| 488 | vc[i] = _mm512_loadu_ps(C + row * ldc + col * 16); |
| 489 | vc[i] = _mm512_mul_ps(vc[i], vscale); |
| 490 | }; |
| 491 | Unroll<ROWS * COLS>{}(loadc); |
| 492 | |
| 493 | auto compute = [&](auto i, int64_t k) { |
| 494 | constexpr int row = i / COLS; |
| 495 | constexpr int col = i % COLS; |
| 496 | |
| 497 | if constexpr (col == 0) { |
| 498 | va = _mm512_set1_ps(A[row * lda + k]); |
| 499 | } |
| 500 | if constexpr (row == 0) { |
| 501 | if (k + 1 < K) { |
| 502 | int64_t b_idx_prefetch = indices[k + 1]; |
| 503 | _mm_prefetch(B + b_idx_prefetch * ldb + col * 16, _MM_HINT_T0); |
| 504 | } |
| 505 | int64_t b_idx = indices[k]; |
| 506 | TORCH_CHECK(b_idx < max_tokens, "token index out of scope!"); |
| 507 | |
| 508 | // for COLS = 2, 4, 6, 8 use 512 bit load |
| 509 | // for COLS = 1, 3, 5, 7 use 256 bit load |
| 510 | if constexpr (COLS % 2 == 0) { |
| 511 | if constexpr (col % 2 == 0) { |
| 512 | __m512i b16 = _mm512_loadu_si512(reinterpret_cast<const __m512i*>(B + b_idx * ldb + col * 16)); |
| 513 | vb[col + 0] = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32(b16, 0)); |
| 514 | vb[col + 1] = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32(b16, 1)); |
| 515 | } |
| 516 | } else { |
| 517 | __m256i b16 = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(B + b_idx * ldb + col * 16)); |
| 518 | vb[col] = CVT_BF16_TO_FP32(b16); |
| 519 | } |
| 520 | } |
| 521 | vc[i] = _mm512_fmadd_ps(va, vb[col], vc[i]); |
| 522 | }; |
| 523 | |
| 524 | for (int64_t k = 0; k < K; ++k) { |
| 525 | Unroll<ROWS * COLS>{}(compute, k); |
| 526 | } |
| 527 | |
| 528 | auto storec = [&](auto i) { |
| 529 | constexpr int row = i / COLS; |
| 530 | constexpr int col = i % COLS; |
| 531 | _mm512_storeu_ps(C + row * ldc + col * 16, vc[i]); |
| 532 | }; |
| 533 | Unroll<ROWS * COLS>{}(storec); |
| 534 | } |
| 535 | }; |
| 536 | #endif |
| 537 | |
| 538 | #if defined(CPU_CAPABILITY_AVX512) |
| 539 | template <typename index_t, int BLOCK_M, int BLOCK_N> |
| 540 | struct tinygemm_kernel_nn<at::Half, index_t, BLOCK_M, BLOCK_N> { |
| 541 | static inline void apply( |
| 542 | const float* __restrict__ A, |
| 543 | const at::Half* __restrict__ B, |
| 544 | float* __restrict__ C, |
| 545 | const index_t* __restrict__ indices, |
| 546 | const float* __restrict__ scale, |
| 547 | int64_t lda, |
| 548 | int64_t ldb, |
| 549 | int64_t ldc, |
| 550 | int64_t K, |
| 551 | int64_t max_tokens) { |
| 552 | constexpr int ROWS = BLOCK_M; |
| 553 | constexpr int COLS = BLOCK_N / 16; |
| 554 | |
| 555 | __m512 va; |
| 556 | __m512 vb[COLS]; |
| 557 | __m512 vc[ROWS * COLS]; |
| 558 | __m512 vscale; |
| 559 | |
| 560 | auto loadc = [&](auto i) { |
| 561 | constexpr int row = i / COLS; |
| 562 | constexpr int col = i % COLS; |
| 563 | #pragma GCC diagnostic push |
| 564 | #pragma GCC diagnostic ignored "-Warray-bounds" |
| 565 | if constexpr (col == 0) { |
| 566 | vscale = _mm512_set1_ps(scale[row]); |
| 567 | } |
| 568 | #pragma GCC diagnostic pop |
| 569 | vc[i] = _mm512_loadu_ps(C + row * ldc + col * 16); |
| 570 | vc[i] = _mm512_mul_ps(vc[i], vscale); |
| 571 | }; |
| 572 | Unroll<ROWS * COLS>{}(loadc); |
| 573 | |
| 574 | auto compute = [&](auto i, int64_t k) { |
| 575 | constexpr int row = i / COLS; |
| 576 | constexpr int col = i % COLS; |
| 577 | |
| 578 | if constexpr (col == 0) { |
| 579 | va = _mm512_set1_ps(A[row * lda + k]); |
| 580 | } |
| 581 | if constexpr (row == 0) { |
| 582 | if (k + 1 < K) { |
| 583 | int64_t b_idx_prefetch = indices[k + 1]; |
| 584 | _mm_prefetch(B + b_idx_prefetch * ldb + col * 16, _MM_HINT_T0); |
| 585 | } |
| 586 | int64_t b_idx = indices[k]; |
| 587 | TORCH_CHECK(b_idx < max_tokens, "token index out of scope!"); |
| 588 | |
| 589 | // for COLS = 2, 4, 6, 8 use 512 bit load |
| 590 | // for COLS = 1, 3, 5, 7 use 256 bit load |
| 591 | if constexpr (COLS % 2 == 0) { |
| 592 | if constexpr (col % 2 == 0) { |
| 593 | __m512i b16 = _mm512_loadu_si512(reinterpret_cast<const __m512i*>(B + b_idx * ldb + col * 16)); |
| 594 | vb[col + 0] = CVT_FP16_TO_FP32(_mm512_extracti32x8_epi32(b16, 0)); |
| 595 | vb[col + 1] = CVT_FP16_TO_FP32(_mm512_extracti32x8_epi32(b16, 1)); |
| 596 | } |
| 597 | } else { |
| 598 | __m256i b16 = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(B + b_idx * ldb + col * 16)); |
| 599 | vb[col] = CVT_FP16_TO_FP32(b16); |
| 600 | } |
| 601 | } |
| 602 | vc[i] = _mm512_fmadd_ps(va, vb[col], vc[i]); |
| 603 | }; |
| 604 | |
| 605 | for (int64_t k = 0; k < K; ++k) { |
| 606 | Unroll<ROWS * COLS>{}(compute, k); |
| 607 | } |
| 608 | |
| 609 | auto storec = [&](auto i) { |
| 610 | constexpr int row = i / COLS; |
| 611 | constexpr int col = i % COLS; |
| 612 | _mm512_storeu_ps(C + row * ldc + col * 16, vc[i]); |
| 613 | }; |
| 614 | Unroll<ROWS * COLS>{}(storec); |
| 615 | } |
| 616 | }; |
| 617 | #endif |
| 618 | |
| 619 | #define LAUNCH_TINYGEMM_KERNEL_NN(MB_SIZE, NB_SIZE) \ |
| 620 | tinygemm_kernel_nn<scalar_t, index_t, MB_SIZE, NB_SIZE>::apply( \ |
| 621 | A + mb_start * lda, \ |
| 622 | B + nb_start, \ |
| 623 | C + mb_start * ldc + nb_start, \ |
| 624 | indices, \ |
| 625 | scale + mb_start, \ |
| 626 | lda, \ |
| 627 | ldb, \ |
| 628 | ldc, \ |
| 629 | K, \ |
| 630 | max_tokens); |
| 631 | |
| 632 | template <typename scalar_t, typename index_t> |
| 633 | void index_gemm_kernel_nt( |
| 634 | const scalar_t* __restrict__ A, |
| 635 | const scalar_t* __restrict__ B, |
| 636 | float* __restrict__ C, |
| 637 | const index_t* __restrict__ indices, |
| 638 | float scale, |
| 639 | int64_t M, |
| 640 | int64_t N, |
| 641 | int64_t K, |
| 642 | int64_t lda, |
| 643 | int64_t ldb, |
| 644 | int64_t ldc, |
| 645 | int64_t max_tokens) { |
| 646 | // pattern: 1-8-8 |
| 647 | if (M == 1) { |
| 648 | constexpr int64_t BLOCK_N = 8; |
| 649 | const int64_t NB = div_up(N, BLOCK_N); |
| 650 | int64_t mb_start = 0, lda = 1, ldc = 1; |
| 651 | |
| 652 | for (int64_t nb = 0; nb < NB; ++nb) { |
| 653 | int64_t nb_start = nb * BLOCK_N; |
| 654 | int64_t nb_size = std::min(BLOCK_N, N - nb_start); |
| 655 | |
| 656 | switch (nb_size) { |
| 657 | case 1: |
| 658 | LAUNCH_TINYGEMM_KERNEL_NT(1, 1); |
| 659 | break; |
| 660 | case 2: |
| 661 | LAUNCH_TINYGEMM_KERNEL_NT(1, 2); |
| 662 | break; |
| 663 | case 3: |
| 664 | LAUNCH_TINYGEMM_KERNEL_NT(1, 3); |
| 665 | break; |
| 666 | case 4: |
| 667 | LAUNCH_TINYGEMM_KERNEL_NT(1, 4); |
| 668 | break; |
| 669 | case 5: |
| 670 | LAUNCH_TINYGEMM_KERNEL_NT(1, 5); |
| 671 | break; |
| 672 | case 6: |
| 673 | LAUNCH_TINYGEMM_KERNEL_NT(1, 6); |
| 674 | break; |
| 675 | case 7: |
| 676 | LAUNCH_TINYGEMM_KERNEL_NT(1, 7); |
| 677 | break; |
| 678 | case 8: |
| 679 | LAUNCH_TINYGEMM_KERNEL_NT(1, 8); |
| 680 | break; |
| 681 | default: |
| 682 | TORCH_CHECK(false, "Unexpected block size, 1x", "nb_size"); |
| 683 | } |
| 684 | } |
| 685 | return; |
| 686 | } |
| 687 | |
| 688 | // default pattern: 1-6-24 |
| 689 | // FP16 pattern: 2-8-16 |
| 690 | constexpr int64_t BLOCK_M = 4; |
| 691 | constexpr int64_t BLOCK_N = std::is_same_v<scalar_t, at::Half> ? 4 : 6; |
| 692 | const int64_t MB = div_up(M, BLOCK_M); |
| 693 | const int64_t NB = div_up(N, BLOCK_N); |
| 694 | |
| 695 | for (int64_t mb = 0; mb < MB; ++mb) { |
| 696 | int64_t mb_start = mb * BLOCK_M; |
| 697 | int64_t mb_size = std::min(BLOCK_M, M - mb_start); |
| 698 | for (int64_t nb = 0; nb < NB; ++nb) { |
| 699 | int64_t nb_start = nb * BLOCK_N; |
| 700 | int64_t nb_size = std::min(BLOCK_N, N - nb_start); |
| 701 | |
| 702 | switch (mb_size << 4 | nb_size) { |
| 703 | // mb_size = 1 |
| 704 | case 0x11: |
| 705 | LAUNCH_TINYGEMM_KERNEL_NT(1, 1); |
| 706 | break; |
| 707 | case 0x12: |
| 708 | LAUNCH_TINYGEMM_KERNEL_NT(1, 2); |
| 709 | break; |
| 710 | case 0x13: |
| 711 | LAUNCH_TINYGEMM_KERNEL_NT(1, 3); |
| 712 | break; |
| 713 | case 0x14: |
| 714 | LAUNCH_TINYGEMM_KERNEL_NT(1, 4); |
| 715 | break; |
| 716 | case 0x15: |
| 717 | LAUNCH_TINYGEMM_KERNEL_NT(1, 5); |
| 718 | break; |
| 719 | case 0x16: |
| 720 | LAUNCH_TINYGEMM_KERNEL_NT(1, 6); |
| 721 | break; |
| 722 | // mb_size = 2 |
| 723 | case 0x21: |
| 724 | LAUNCH_TINYGEMM_KERNEL_NT(2, 1); |
| 725 | break; |
| 726 | case 0x22: |
| 727 | LAUNCH_TINYGEMM_KERNEL_NT(2, 2); |
| 728 | break; |
| 729 | case 0x23: |
| 730 | LAUNCH_TINYGEMM_KERNEL_NT(2, 3); |
| 731 | break; |
| 732 | case 0x24: |
| 733 | LAUNCH_TINYGEMM_KERNEL_NT(2, 4); |
| 734 | break; |
| 735 | case 0x25: |
| 736 | LAUNCH_TINYGEMM_KERNEL_NT(2, 5); |
| 737 | break; |
| 738 | case 0x26: |
| 739 | LAUNCH_TINYGEMM_KERNEL_NT(2, 6); |
| 740 | break; |
| 741 | // mb_size = 3 |
| 742 | case 0x31: |
| 743 | LAUNCH_TINYGEMM_KERNEL_NT(3, 1); |
| 744 | break; |
| 745 | case 0x32: |
| 746 | LAUNCH_TINYGEMM_KERNEL_NT(3, 2); |
| 747 | break; |
| 748 | case 0x33: |
| 749 | LAUNCH_TINYGEMM_KERNEL_NT(3, 3); |
| 750 | break; |
| 751 | case 0x34: |
| 752 | LAUNCH_TINYGEMM_KERNEL_NT(3, 4); |
| 753 | break; |
| 754 | case 0x35: |
| 755 | LAUNCH_TINYGEMM_KERNEL_NT(3, 5); |
| 756 | break; |
| 757 | case 0x36: |
| 758 | LAUNCH_TINYGEMM_KERNEL_NT(3, 6); |
| 759 | break; |
| 760 | // mb_size = 4 |
| 761 | case 0x41: |
| 762 | LAUNCH_TINYGEMM_KERNEL_NT(4, 1); |
| 763 | break; |
| 764 | case 0x42: |
| 765 | LAUNCH_TINYGEMM_KERNEL_NT(4, 2); |
| 766 | break; |
| 767 | case 0x43: |
| 768 | LAUNCH_TINYGEMM_KERNEL_NT(4, 3); |
| 769 | break; |
| 770 | case 0x44: |
| 771 | LAUNCH_TINYGEMM_KERNEL_NT(4, 4); |
| 772 | break; |
| 773 | case 0x45: |
| 774 | LAUNCH_TINYGEMM_KERNEL_NT(4, 5); |
| 775 | break; |
| 776 | case 0x46: |
| 777 | LAUNCH_TINYGEMM_KERNEL_NT(4, 6); |
| 778 | break; |
| 779 | default: |
| 780 | TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size"); |
| 781 | } |
| 782 | } |
| 783 | } |
| 784 | } |
| 785 | |
| 786 | template <typename scalar_t, typename index_t> |
| 787 | void index_gemm_kernel_nn( |
| 788 | const float* __restrict__ A, |
| 789 | const scalar_t* __restrict__ B, |
| 790 | float* __restrict__ C, |
| 791 | const index_t* __restrict__ indices, |
| 792 | float* __restrict__ scale, |
| 793 | int64_t M, |
| 794 | int64_t N, |
| 795 | int64_t K, |
| 796 | int64_t lda, |
| 797 | int64_t ldb, |
| 798 | int64_t ldc, |
| 799 | int64_t max_tokens) { |
| 800 | constexpr int kVecSize = 16; |
| 801 | if ((N & (kVecSize - 1)) != 0) { |
| 802 | tinygemm_kernel_nn_scalar(A, B, C, indices, scale, M, N, K, lda, ldb, ldc, max_tokens); |
| 803 | return; |
| 804 | } |
| 805 | |
| 806 | // pattern: 1-8-8 |
| 807 | if (M == 1) { |
| 808 | constexpr int64_t BLOCK_N = 8 * kVecSize; |
| 809 | const int64_t NB = div_up(N, BLOCK_N); |
| 810 | int64_t mb_start = 0, lda = 1, ldc = 1; |
| 811 | |
| 812 | for (int64_t nb = 0; nb < NB; ++nb) { |
| 813 | int64_t nb_start = nb * BLOCK_N; |
| 814 | int64_t nb_size = std::min(BLOCK_N, N - nb_start); |
| 815 | |
| 816 | switch (nb_size >> 4) { |
| 817 | case 1: |
| 818 | LAUNCH_TINYGEMM_KERNEL_NN(1, 16); |
| 819 | break; |
| 820 | case 2: |
| 821 | LAUNCH_TINYGEMM_KERNEL_NN(1, 32); |
| 822 | break; |
| 823 | case 3: |
| 824 | LAUNCH_TINYGEMM_KERNEL_NN(1, 48); |
| 825 | break; |
| 826 | case 4: |
| 827 | LAUNCH_TINYGEMM_KERNEL_NN(1, 64); |
| 828 | break; |
| 829 | case 5: |
| 830 | LAUNCH_TINYGEMM_KERNEL_NN(1, 80); |
| 831 | break; |
| 832 | case 6: |
| 833 | LAUNCH_TINYGEMM_KERNEL_NN(1, 96); |
| 834 | break; |
| 835 | case 7: |
| 836 | LAUNCH_TINYGEMM_KERNEL_NN(1, 112); |
| 837 | break; |
| 838 | case 8: |
| 839 | LAUNCH_TINYGEMM_KERNEL_NN(1, 128); |
| 840 | break; |
| 841 | default: |
| 842 | TORCH_CHECK(false, "Unexpected block size, 1x", "nb_size"); |
| 843 | } |
| 844 | } |
| 845 | return; |
| 846 | } |
| 847 | |
| 848 | constexpr int64_t BLOCK_M = 4; |
| 849 | constexpr int64_t BLOCK_N = 6 * kVecSize; |
| 850 | const int64_t MB = div_up(M, BLOCK_M); |
| 851 | const int64_t NB = div_up(N, BLOCK_N); |
| 852 | |
| 853 | for (int64_t mb = 0; mb < MB; ++mb) { |
| 854 | int64_t mb_start = mb * BLOCK_M; |
| 855 | int64_t mb_size = std::min(BLOCK_M, M - mb_start); |
| 856 | for (int64_t nb = 0; nb < NB; ++nb) { |
| 857 | int64_t nb_start = nb * BLOCK_N; |
| 858 | int64_t nb_size = std::min(BLOCK_N, N - nb_start); |
| 859 | |
| 860 | switch (mb_size << 4 | nb_size >> 4) { |
| 861 | // mb_size = 1 |
| 862 | case 0x11: |
| 863 | LAUNCH_TINYGEMM_KERNEL_NN(1, 16); |
| 864 | break; |
| 865 | case 0x12: |
| 866 | LAUNCH_TINYGEMM_KERNEL_NN(1, 32); |
| 867 | break; |
| 868 | case 0x13: |
| 869 | LAUNCH_TINYGEMM_KERNEL_NN(1, 48); |
| 870 | break; |
| 871 | case 0x14: |
| 872 | LAUNCH_TINYGEMM_KERNEL_NN(1, 64); |
| 873 | break; |
| 874 | case 0x15: |
| 875 | LAUNCH_TINYGEMM_KERNEL_NN(1, 80); |
| 876 | break; |
| 877 | case 0x16: |
| 878 | LAUNCH_TINYGEMM_KERNEL_NN(1, 96); |
| 879 | break; |
| 880 | // mb_size = 2 |
| 881 | case 0x21: |
| 882 | LAUNCH_TINYGEMM_KERNEL_NN(2, 16); |
| 883 | break; |
| 884 | case 0x22: |
| 885 | LAUNCH_TINYGEMM_KERNEL_NN(2, 32); |
| 886 | break; |
| 887 | case 0x23: |
| 888 | LAUNCH_TINYGEMM_KERNEL_NN(2, 48); |
| 889 | break; |
| 890 | case 0x24: |
| 891 | LAUNCH_TINYGEMM_KERNEL_NN(2, 64); |
| 892 | break; |
| 893 | case 0x25: |
| 894 | LAUNCH_TINYGEMM_KERNEL_NN(2, 80); |
| 895 | break; |
| 896 | case 0x26: |
| 897 | LAUNCH_TINYGEMM_KERNEL_NN(2, 96); |
| 898 | break; |
| 899 | // mb_size = 3 |
| 900 | case 0x31: |
| 901 | LAUNCH_TINYGEMM_KERNEL_NN(3, 16); |
| 902 | break; |
| 903 | case 0x32: |
| 904 | LAUNCH_TINYGEMM_KERNEL_NN(3, 32); |
| 905 | break; |
| 906 | case 0x33: |
| 907 | LAUNCH_TINYGEMM_KERNEL_NN(3, 48); |
| 908 | break; |
| 909 | case 0x34: |
| 910 | LAUNCH_TINYGEMM_KERNEL_NN(3, 64); |
| 911 | break; |
| 912 | case 0x35: |
| 913 | LAUNCH_TINYGEMM_KERNEL_NN(3, 80); |
| 914 | break; |
| 915 | case 0x36: |
| 916 | LAUNCH_TINYGEMM_KERNEL_NN(3, 96); |
| 917 | break; |
| 918 | // mb_size = 4 |
| 919 | case 0x41: |
| 920 | LAUNCH_TINYGEMM_KERNEL_NN(4, 16); |
| 921 | break; |
| 922 | case 0x42: |
| 923 | LAUNCH_TINYGEMM_KERNEL_NN(4, 32); |
| 924 | break; |
| 925 | case 0x43: |
| 926 | LAUNCH_TINYGEMM_KERNEL_NN(4, 48); |
| 927 | break; |
| 928 | case 0x44: |
| 929 | LAUNCH_TINYGEMM_KERNEL_NN(4, 64); |
| 930 | break; |
| 931 | case 0x45: |
| 932 | LAUNCH_TINYGEMM_KERNEL_NN(4, 80); |
| 933 | break; |
| 934 | case 0x46: |
| 935 | LAUNCH_TINYGEMM_KERNEL_NN(4, 96); |
| 936 | break; |
| 937 | default: |
| 938 | TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size"); |
| 939 | } |
| 940 | } |
| 941 | } |
| 942 | } |
| 943 | |
| 944 | template <typename scalar_t> |
| 945 | void decode_set_kv_buffer( |
| 946 | scalar_t* __restrict__ k_buffer, |
| 947 | scalar_t* __restrict__ v_buffer, |
| 948 | const scalar_t* __restrict__ key, |
| 949 | const scalar_t* __restrict__ value, |
| 950 | const int64_t* __restrict__ loc, |
| 951 | int64_t batches, |
| 952 | int64_t num_heads_kv, |
| 953 | int64_t head_size, |
| 954 | int64_t head_size_v, |
| 955 | int64_t k_strideN, |
| 956 | int64_t k_strideH, |
| 957 | int64_t v_strideN, |
| 958 | int64_t v_strideH, |
| 959 | int64_t nk_strideN, |
| 960 | int64_t nk_strideH, |
| 961 | int64_t nv_strideN, |
| 962 | int64_t nv_strideH, |
| 963 | bool is_mla) { |
| 964 | at::parallel_for(0, batches * num_heads_kv, 0, [&](int64_t begin, int64_t end) { |
| 965 | int64_t bs{0}, head_kv_id{0}; |
| 966 | data_index_init(begin, bs, batches, head_kv_id, num_heads_kv); |
| 967 | |
| 968 | for (int64_t i = begin; i < end; i++) { |
| 969 | int64_t loc_val = loc[bs]; |
| 970 | scalar_t* k_buffer_ptr = k_buffer + loc_val * k_strideN + head_kv_id * k_strideH; |
| 971 | const scalar_t* new_key_ptr = key + bs * nk_strideN + head_kv_id * nk_strideH; |
| 972 | copy_stub<scalar_t>(k_buffer_ptr, new_key_ptr, head_size); |
| 973 | if (!is_mla) { |
| 974 | scalar_t* v_buffer_ptr = v_buffer + loc_val * v_strideN + head_kv_id * v_strideH; |
| 975 | const scalar_t* new_value_ptr = value + bs * nv_strideN + head_kv_id * nv_strideH; |
| 976 | copy_stub<scalar_t>(v_buffer_ptr, new_value_ptr, head_size_v); |
| 977 | } |
| 978 | |
| 979 | // move to the next index |
| 980 | data_index_step(bs, batches, head_kv_id, num_heads_kv); |
| 981 | } |
| 982 | }); |
| 983 | } |
| 984 | |
| 985 | template <typename scalar_t> |
| 986 | void decode_accumulate_kv_splits( |
| 987 | scalar_t* __restrict__ output, |
| 988 | float* __restrict__ attn_logits, |
| 989 | const scalar_t* __restrict__ sinks_ptr, |
| 990 | int64_t batches, |
| 991 | int64_t num_heads, |
| 992 | int64_t head_size_v, |
| 993 | int64_t num_kv_splits, |
| 994 | int64_t l_stride1, |
| 995 | int64_t l_stride2, |
| 996 | bool has_sink) { |
| 997 | using Vec = at::vec::Vectorized<float>; |
| 998 | |
| 999 | // parallel on [batches, num_heads] |
| 1000 | at::parallel_for(0, batches * num_heads, 0, [&](int64_t begin, int64_t end) { |
| 1001 | int64_t bi{0}, ni{0}; |
| 1002 | data_index_init(begin, bi, batches, ni, num_heads); |
| 1003 | // NB: here we use logits[b][h][0] as acc, since |
| 1004 | // for the first kv split (kv_id == 0): |
| 1005 | // m_delta = std::exp(-inf) = 0 |
| 1006 | // e_logic = std::exp(0) = 1 |
| 1007 | // acc = acc * m_delta + tv * e_logic = tv |
| 1008 | for (int64_t i = begin; i < end; ++i) { |
| 1009 | float* __restrict__ acc = attn_logits + i * l_stride1; |
| 1010 | |
| 1011 | float s_prime = 0.f; |
| 1012 | float m_prime = -std::numeric_limits<scalar_t>::infinity(); |
| 1013 | |
| 1014 | // update acc with from each kv_split |
| 1015 | for (int64_t kv_id = 0; kv_id < num_kv_splits; ++kv_id) { |
| 1016 | float* __restrict__ tv = acc + kv_id * l_stride2; |
| 1017 | const float tlogic = (acc + kv_id * l_stride2)[head_size_v]; |
| 1018 | |
| 1019 | float m_i = std::max(tlogic, m_prime); |
| 1020 | float m_delta = std::exp(m_prime - m_i); |
| 1021 | float e_logic = std::exp(tlogic - m_i); |
| 1022 | if (kv_id != 0) { |
| 1023 | at::vec::map2<float>( |
| 1024 | [m_delta, e_logic](Vec x, Vec y) { return x * Vec(m_delta) + y * Vec(e_logic); }, |
| 1025 | acc, |
| 1026 | acc, |
| 1027 | tv, |
| 1028 | head_size_v); |
| 1029 | } |
| 1030 | |
| 1031 | s_prime = s_prime * m_delta + e_logic; |
| 1032 | m_prime = m_i; |
| 1033 | } |
| 1034 | if (has_sink) { |
| 1035 | s_prime += std::exp(sinks_ptr[ni] - m_prime); |
| 1036 | } |
| 1037 | copy_stub<scalar_t>(output + i * head_size_v, acc, 1 / s_prime, head_size_v); |
| 1038 | // move to the next index |
| 1039 | data_index_step(bi, batches, ni, num_heads); |
| 1040 | } |
| 1041 | }); |
| 1042 | } |
| 1043 | |
| 1044 | template <typename scalar_t, typename index_t, int64_t BLOCK_N> |
| 1045 | void decode_attention_kernel_impl( |
| 1046 | scalar_t* __restrict__ output, |
| 1047 | float* __restrict__ attn_logits, |
| 1048 | const scalar_t* __restrict__ query, |
| 1049 | const scalar_t* __restrict__ k_buffer, |
| 1050 | const scalar_t* __restrict__ v_buffer, |
| 1051 | const index_t* __restrict__ req_to_token, |
| 1052 | const int64_t* __restrict__ req_pool_indices, |
| 1053 | const int64_t* __restrict__ seq_lens, |
| 1054 | const int64_t* __restrict__ encoder_lens, |
| 1055 | const scalar_t* __restrict__ sinks, |
| 1056 | int64_t batches, |
| 1057 | int64_t num_heads, |
| 1058 | int64_t head_size, |
| 1059 | int64_t head_size_v, |
| 1060 | int64_t num_kv_splits, |
| 1061 | int64_t q_strideM, |
| 1062 | int64_t q_strideH, |
| 1063 | int64_t k_strideN, |
| 1064 | int64_t k_strideH, |
| 1065 | int64_t v_strideN, |
| 1066 | int64_t v_strideH, |
| 1067 | float sm_scale, |
| 1068 | float logit_cap, |
| 1069 | int64_t max_num_reqs, |
| 1070 | int64_t max_context_len, |
| 1071 | int64_t max_total_num_tokens, |
| 1072 | int64_t sliding_window_size, |
| 1073 | bool is_cross_attn, |
| 1074 | bool has_encoder_lens, |
| 1075 | bool has_sink) { |
| 1076 | using Vec = at::vec::Vectorized<float>; |
| 1077 | |
| 1078 | // strides |
| 1079 | const int64_t l_stride1 = num_kv_splits * (head_size_v + 1); |
| 1080 | const int64_t l_stride2 = head_size_v + 1; |
| 1081 | |
| 1082 | const bool has_logit_cap = logit_cap > 0; |
| 1083 | float rlogit_cap = has_logit_cap ? 1 / logit_cap : 0.f; |
| 1084 | |
| 1085 | // parallel on [batches, num_heads, num_kv_splits] |
| 1086 | at::parallel_for(0, batches * num_heads * num_kv_splits, 0, [&](int64_t begin, int64_t end) { |
| 1087 | int64_t bs{0}, head_id{0}, kv_id{0}; |
| 1088 | data_index_init(begin, bs, batches, head_id, num_heads, kv_id, num_kv_splits); |
| 1089 | |
| 1090 | // s_prime and s_delta |
| 1091 | alignas(64) float s_i[BLOCK_N]; |
| 1092 | float* __restrict__ s_delta = s_i; |
| 1093 | |
| 1094 | for (int64_t i = begin; i < end; ++i) { |
| 1095 | // get query |
| 1096 | const scalar_t* __restrict__ q_ptr = query + bs * q_strideM + head_id * q_strideH; |
| 1097 | |
| 1098 | // get key/value |
| 1099 | int64_t seq_len_kv = is_cross_attn ? encoder_lens[bs] : seq_lens[bs]; |
| 1100 | int64_t req_pool_id = req_pool_indices[bs]; |
| 1101 | int64_t kv_offset = (has_encoder_lens && (!is_cross_attn)) ? encoder_lens[bs] : 0; |
| 1102 | if (sliding_window_size > 0 && seq_len_kv > sliding_window_size) { |
| 1103 | kv_offset = seq_len_kv - sliding_window_size; |
| 1104 | seq_len_kv = sliding_window_size; |
| 1105 | } |
| 1106 | TORCH_CHECK(seq_len_kv <= max_context_len, "seq_len_kv out of scope!"); |
| 1107 | TORCH_CHECK(req_pool_id < max_num_reqs, "req_pool_id out of scope!"); |
| 1108 | |
| 1109 | const int64_t SPLIT_SIZE = div_up(seq_len_kv, num_kv_splits); |
| 1110 | const int64_t kv_start = kv_id * SPLIT_SIZE; |
| 1111 | const int64_t kv_end = std::min(kv_start + SPLIT_SIZE, seq_len_kv); |
| 1112 | |
| 1113 | float m_prime = -std::numeric_limits<float>::infinity(); |
| 1114 | float s_prime = 0.f; |
| 1115 | |
| 1116 | // get v_prime, and init to zero |
| 1117 | float* __restrict__ v_prime = attn_logits + i * (head_size_v + 1); |
| 1118 | fill_stub(v_prime, 0.f, head_size_v); |
| 1119 | |
| 1120 | // loop over K and V sequence with BLOCK_N |
| 1121 | for (int64_t n = kv_start; n < kv_end; n += BLOCK_N) { |
| 1122 | int64_t n_size = std::min(BLOCK_N, kv_end - n); |
| 1123 | |
| 1124 | // calculate s_i <- scale * Q @ K |
| 1125 | index_gemm_kernel_nt<scalar_t, index_t>( |
| 1126 | /* A */ q_ptr, |
| 1127 | /* B */ k_buffer + head_id * k_strideH, |
| 1128 | /* C */ s_i, |
| 1129 | /* ind */ req_to_token + req_pool_id * max_context_len + n + kv_offset, |
| 1130 | /* scl */ sm_scale, |
| 1131 | /* M */ 1, |
| 1132 | /* N */ n_size, |
| 1133 | /* K */ head_size, |
| 1134 | /* lda */ 1, |
| 1135 | /* ldb */ k_strideN, |
| 1136 | /* ldc */ 1, |
| 1137 | /* mtt */ max_total_num_tokens); |
| 1138 | |
| 1139 | // TODO: `tanh` from torch uses sleef u10, going to be slow |
| 1140 | if (has_logit_cap) { |
| 1141 | at::vec::map<float>( |
| 1142 | [logit_cap, rlogit_cap](Vec x) { return Vec(logit_cap) * (x * Vec(rlogit_cap)).tanh(); }, |
| 1143 | s_i, |
| 1144 | s_i, |
| 1145 | n_size); |
| 1146 | } |
| 1147 | |
| 1148 | // m_i: max value per row |
| 1149 | float m_i = at::vec::reduce_all<float>([](Vec& x, Vec& y) { return at::vec::maximum(x, y); }, s_i, n_size); |
| 1150 | m_i = std::max(m_i, m_prime); |
| 1151 | |
| 1152 | // m_delta <- exp(m' - m_i) |
| 1153 | float m_delta = std::exp(m_prime - m_i); |
| 1154 | |
| 1155 | // s_delta <- exp(s_i - m_i) |
| 1156 | at::vec::map<float>([m_i](Vec x) { return (x - Vec(m_i)).exp_u20(); }, s_delta, s_i, n_size); |
| 1157 | |
| 1158 | // s' <- s' * m_delta + sum(s_delta) |
| 1159 | s_prime *= m_delta; |
| 1160 | s_prime += at::vec::reduce_all<float>([](Vec& x, Vec& y) { return x + y; }, s_delta, n_size); |
| 1161 | |
| 1162 | m_prime = m_i; |
| 1163 | |
| 1164 | // calculate V' <- s_delta @ V + V' * m_delta |
| 1165 | index_gemm_kernel_nn<scalar_t, index_t>( |
| 1166 | /* A */ s_delta, |
| 1167 | /* B */ v_buffer + head_id * v_strideH, |
| 1168 | /* C */ v_prime, |
| 1169 | /* ind */ req_to_token + req_pool_id * max_context_len + n + kv_offset, |
| 1170 | /* scl */ &m_delta, |
| 1171 | /* M */ 1, |
| 1172 | /* N */ head_size_v, |
| 1173 | /* K */ n_size, |
| 1174 | /* lda */ 1, |
| 1175 | /* ldb */ v_strideN, |
| 1176 | /* ldc */ 1, |
| 1177 | /* mtt */ max_total_num_tokens); |
| 1178 | } // loop with KV blocks |
| 1179 | |
| 1180 | // only update v' when kv_split_size > 0 |
| 1181 | if (kv_end > kv_start) { |
| 1182 | float s = 1 / s_prime; |
| 1183 | at::vec::map<float>([s](Vec out) { return out * Vec(s); }, v_prime, v_prime, head_size_v); |
| 1184 | |
| 1185 | v_prime[head_size_v] = m_prime + std::log(s_prime); |
| 1186 | } else { |
| 1187 | v_prime[head_size_v] = -std::numeric_limits<float>::infinity(); |
| 1188 | } |
| 1189 | |
| 1190 | // move to the next index |
| 1191 | data_index_step(bs, batches, head_id, num_heads, kv_id, num_kv_splits); |
| 1192 | } |
| 1193 | }); |
| 1194 | |
| 1195 | decode_accumulate_kv_splits( |
| 1196 | output, attn_logits, sinks, batches, num_heads, head_size_v, num_kv_splits, l_stride1, l_stride2, has_sink); |
| 1197 | } // MHA |
| 1198 | |
| 1199 | template <typename scalar_t, typename index_t, int64_t BLOCK_N> |
| 1200 | void decode_attention_mla_kernel_impl( |
| 1201 | scalar_t* __restrict__ output, |
| 1202 | float* __restrict__ attn_logits, |
| 1203 | const scalar_t* __restrict__ query, |
| 1204 | const scalar_t* __restrict__ k_buffer, |
| 1205 | const scalar_t* __restrict__ v_buffer, |
| 1206 | const index_t* __restrict__ req_to_token, |
| 1207 | const int64_t* __restrict__ req_pool_indices, |
| 1208 | const int64_t* __restrict__ seq_lens, |
| 1209 | scalar_t* __restrict__ buffer, |
| 1210 | const scalar_t* __restrict__ sinks, |
| 1211 | int64_t batches, |
| 1212 | int64_t num_heads, |
| 1213 | int64_t head_size, |
| 1214 | int64_t head_size_v, |
| 1215 | int64_t num_kv_splits, |
| 1216 | int64_t q_strideM, |
| 1217 | int64_t q_strideH, |
| 1218 | int64_t k_strideN, |
| 1219 | int64_t k_strideH, |
| 1220 | int64_t v_strideN, |
| 1221 | int64_t v_strideH, |
| 1222 | float sm_scale, |
| 1223 | float logit_cap, |
| 1224 | int64_t max_num_reqs, |
| 1225 | int64_t max_context_len, |
| 1226 | int64_t max_total_num_tokens, |
| 1227 | int64_t buffer_size_per_thread, |
| 1228 | bool has_sink) { |
| 1229 | using Vec = at::vec::Vectorized<float>; |
| 1230 | |
| 1231 | // block length for heads |
| 1232 | const int64_t BLOCK_H = batches == 1 ? 6 : (batches > 16 ? 22 : 11); |
| 1233 | |
| 1234 | // strides |
| 1235 | const int64_t l_stride0 = num_heads * num_kv_splits * (head_size_v + 1); |
| 1236 | const int64_t l_stride1 = num_kv_splits * (head_size_v + 1); |
| 1237 | const int64_t l_stride2 = head_size_v + 1; |
| 1238 | |
| 1239 | TORCH_CHECK(logit_cap == 0.f, "decode MLA: expect no logit_cap."); |
| 1240 | |
| 1241 | // partition the heads into blocks for parallel |
| 1242 | const int64_t num_blocks = div_up(num_heads, BLOCK_H); |
| 1243 | |
| 1244 | // parallel on [batches, num_blocks, num_kv_splits] |
| 1245 | at::parallel_for(0, batches * num_blocks * num_kv_splits, 0, [&](int64_t begin, int64_t end) { |
| 1246 | int64_t bs{0}, block_id{0}, kv_id{0}; |
| 1247 | data_index_init(begin, bs, batches, block_id, num_blocks, kv_id, num_kv_splits); |
| 1248 | |
| 1249 | int tid = at::get_thread_num(); |
| 1250 | scalar_t* __restrict__ Btmp0 = buffer + tid * buffer_size_per_thread; |
| 1251 | scalar_t* __restrict__ Btmp1 = Btmp0 + BLOCK_N * head_size; |
| 1252 | |
| 1253 | // init Btmp1 just once for each thread to prevent NaN |
| 1254 | // Btmp0 is not needed as it computes full K every single time |
| 1255 | fill_stub(Btmp1, 0.f, BLOCK_N * head_size_v); |
| 1256 | |
| 1257 | alignas(64) float s_i[BLOCK_H * BLOCK_N]; |
| 1258 | float* __restrict__ s_delta = s_i; |
| 1259 | alignas(64) scalar_t s_delta2[BLOCK_H * BLOCK_N]; |
| 1260 | |
| 1261 | alignas(64) float s_prime[BLOCK_H]; |
| 1262 | alignas(64) float m_prime[BLOCK_H]; |
| 1263 | alignas(64) float m_delta[BLOCK_H]; |
| 1264 | |
| 1265 | for (int64_t i = begin; i < end; ++i) { |
| 1266 | const int64_t h_start = block_id * BLOCK_H; |
| 1267 | const int64_t h_end = std::min(block_id * BLOCK_H + BLOCK_H, num_heads); |
| 1268 | const int64_t h_size = h_end - h_start; |
| 1269 | |
| 1270 | // get query |
| 1271 | const scalar_t* __restrict__ q_ptr = query + bs * q_strideM + h_start * q_strideH; |
| 1272 | |
| 1273 | int64_t seq_len_kv = seq_lens[bs]; |
| 1274 | int64_t req_pool_id = req_pool_indices[bs]; |
| 1275 | TORCH_CHECK(seq_len_kv <= max_context_len, "seq_len_kv out of scope!"); |
| 1276 | TORCH_CHECK(req_pool_id < max_num_reqs, "req_pool_id out of scope!"); |
| 1277 | |
| 1278 | const int64_t SPLIT_SIZE = div_up(seq_len_kv, num_kv_splits); |
| 1279 | const int64_t kv_start = kv_id * SPLIT_SIZE; |
| 1280 | const int64_t kv_end = std::min(kv_start + SPLIT_SIZE, seq_len_kv); |
| 1281 | |
| 1282 | fill_stub(s_prime, 0.f, BLOCK_H); |
| 1283 | fill_stub(m_prime, -std::numeric_limits<float>::infinity(), BLOCK_H); |
| 1284 | |
| 1285 | // get v_prime, and init to zero |
| 1286 | float* __restrict__ v_prime = attn_logits + bs * l_stride0 + h_start * l_stride1 + kv_id * l_stride2; |
| 1287 | for (int64_t h = 0; h < h_size; ++h) { |
| 1288 | fill_stub(v_prime + h * l_stride1, 0.f, head_size_v); |
| 1289 | } |
| 1290 | |
| 1291 | // loop over K and V sequence with BLOCK_N |
| 1292 | for (int64_t n = kv_start; n < kv_end; n += BLOCK_N) { |
| 1293 | int64_t n_size = std::min(BLOCK_N, kv_end - n); |
| 1294 | const int64_t padded_n_size = div_up(int(n_size), TILE_K) * TILE_K; |
| 1295 | |
| 1296 | // get key and pack |
| 1297 | pack_vnni<scalar_t, index_t>( |
| 1298 | /* dst0 */ Btmp0, |
| 1299 | /* dst1 */ Btmp1, |
| 1300 | /* src */ k_buffer + /* head_kv_id */ 0 * k_strideH, |
| 1301 | /* ind */ req_to_token + req_pool_id * max_context_len + n, |
| 1302 | /* N */ n_size, |
| 1303 | /* K */ head_size, |
| 1304 | /* Kv */ head_size_v, |
| 1305 | /* ld_src */ k_strideN, |
| 1306 | /* ld_dst0 */ BLOCK_N, |
| 1307 | /* ld_dst1 */ head_size_v); |
| 1308 | |
| 1309 | // calculate s_i <- Q @ K |
| 1310 | at::native::cpublas::brgemm( |
| 1311 | /* M */ h_size, |
| 1312 | /* N */ n_size, |
| 1313 | /* K */ head_size, |
| 1314 | /* lda */ q_strideH, |
| 1315 | /* ldb */ BLOCK_N, |
| 1316 | /* ldc */ BLOCK_N, |
| 1317 | /* add_C */ false, |
| 1318 | /* A */ q_ptr, |
| 1319 | /* B */ Btmp0, |
| 1320 | /* C */ s_i); |
| 1321 | |
| 1322 | const Vec scale_vec = Vec(sm_scale); |
| 1323 | for (int64_t h = 0; h < h_size; ++h) { |
| 1324 | // s_i <- s_i * scale |
| 1325 | at::vec::map<float>( |
| 1326 | [scale_vec](Vec x) { return x * scale_vec; }, s_i + h * BLOCK_N, s_i + h * BLOCK_N, n_size); |
| 1327 | |
| 1328 | // m_i: max value per row |
| 1329 | float m_i = at::vec::reduce_all<float>( |
| 1330 | [](Vec& x, Vec& y) { return at::vec::maximum(x, y); }, s_i + h * BLOCK_N, n_size); |
| 1331 | m_i = std::max(m_i, m_prime[h]); |
| 1332 | |
| 1333 | // m_delta <- exp(m' - m_i) |
| 1334 | m_delta[h] = std::exp(m_prime[h] - m_i); |
| 1335 | |
| 1336 | // s_delta <- exp(s_i - m_i) |
| 1337 | at::vec::map<float>( |
| 1338 | [m_i](Vec x) { return (x - Vec(m_i)).exp_u20(); }, s_delta + h * BLOCK_N, s_i + h * BLOCK_N, n_size); |
| 1339 | |
| 1340 | // s' <- s' * m_delta + sum(s_delta) |
| 1341 | s_prime[h] *= m_delta[h]; |
| 1342 | s_prime[h] += at::vec::reduce_all<float>([](Vec& x, Vec& y) { return x + y; }, s_delta + h * BLOCK_N, n_size); |
| 1343 | |
| 1344 | m_prime[h] = m_i; |
| 1345 | |
| 1346 | // v' <- v' * m_delta |
| 1347 | float scale_m = m_delta[h]; |
| 1348 | at::vec::map<float>( |
| 1349 | [scale_m](Vec x) { return x * Vec(scale_m); }, |
| 1350 | v_prime + h * l_stride1, |
| 1351 | v_prime + h * l_stride1, |
| 1352 | head_size_v); |
| 1353 | |
| 1354 | // pad s_delta with 0 first and then convert to scalar_t |
| 1355 | fill_stub(s_delta + h * BLOCK_N + n_size, 0.f, padded_n_size - n_size); |
| 1356 | copy_stub<scalar_t, BLOCK_N>(s_delta2 + h * BLOCK_N, s_delta + h * BLOCK_N); |
| 1357 | } |
| 1358 | |
| 1359 | // calculate V' <- s_delta @ V + V' |
| 1360 | at::native::cpublas::brgemm( |
| 1361 | /* M */ h_size, |
| 1362 | /* N */ head_size_v, |
| 1363 | /* K */ padded_n_size, // n_size |
| 1364 | /* lda */ BLOCK_N, |
| 1365 | /* ldb */ head_size_v, |
| 1366 | /* ldc */ l_stride1, |
| 1367 | /* add_C */ true, |
| 1368 | /* A */ s_delta2, |
| 1369 | /* B */ Btmp1, |
| 1370 | /* C */ v_prime); |
| 1371 | } // loop with KV blocks |
| 1372 | |
| 1373 | // only update v' when kv_split_size > 0 |
| 1374 | if (kv_end > kv_start) { |
| 1375 | for (int64_t h = 0; h < h_size; ++h) { |
| 1376 | float s = 1 / s_prime[h]; |
| 1377 | at::vec::map<float>( |
| 1378 | [s](Vec out) { return out * Vec(s); }, v_prime + h * l_stride1, v_prime + h * l_stride1, head_size_v); |
| 1379 | (v_prime + h * l_stride1)[head_size_v] = m_prime[h] + std::log(s_prime[h]); |
| 1380 | } |
| 1381 | } else { |
| 1382 | for (int64_t h = 0; h < h_size; ++h) { |
| 1383 | (v_prime + h * l_stride1)[head_size_v] = -std::numeric_limits<float>::infinity(); |
| 1384 | } |
| 1385 | } |
| 1386 | |
| 1387 | // move to the next index |
| 1388 | data_index_step(bs, batches, block_id, num_blocks, kv_id, num_kv_splits); |
| 1389 | } |
| 1390 | at::native::cpublas::brgemm_release(); |
| 1391 | }); |
| 1392 | |
| 1393 | decode_accumulate_kv_splits( |
| 1394 | output, attn_logits, sinks, batches, num_heads, head_size_v, num_kv_splits, l_stride1, l_stride2, has_sink); |
| 1395 | } // MLA |
| 1396 | |
| 1397 | template <typename scalar_t, typename index_t, int64_t BLOCK_N> |
| 1398 | void decode_attention_grouped_kernel_impl( |
| 1399 | scalar_t* __restrict__ output, |
| 1400 | float* __restrict__ attn_logits, |
| 1401 | const scalar_t* __restrict__ query, |
| 1402 | const scalar_t* __restrict__ k_buffer, |
| 1403 | const scalar_t* __restrict__ v_buffer, |
| 1404 | const index_t* __restrict__ req_to_token, |
| 1405 | const int64_t* __restrict__ req_pool_indices, |
| 1406 | const int64_t* __restrict__ seq_lens, |
| 1407 | const int64_t* __restrict__ encoder_lens, |
| 1408 | const scalar_t* __restrict__ sinks, |
| 1409 | int64_t batches, |
| 1410 | int64_t num_heads, |
| 1411 | int64_t num_heads_kv, |
| 1412 | int64_t head_size, |
| 1413 | int64_t head_size_v, |
| 1414 | int64_t num_kv_splits, |
| 1415 | int64_t q_strideM, |
| 1416 | int64_t q_strideH, |
| 1417 | int64_t k_strideN, |
| 1418 | int64_t k_strideH, |
| 1419 | int64_t v_strideN, |
| 1420 | int64_t v_strideH, |
| 1421 | float sm_scale, |
| 1422 | float logit_cap, |
| 1423 | int64_t max_num_reqs, |
| 1424 | int64_t max_context_len, |
| 1425 | int64_t max_total_num_tokens, |
| 1426 | int64_t sliding_window_size, |
| 1427 | bool is_cross_attn, |
| 1428 | bool has_encoder_lens, |
| 1429 | bool has_sink) { |
| 1430 | using Vec = at::vec::Vectorized<float>; |
| 1431 | |
| 1432 | // block length for heads |
| 1433 | // we parallel on [batches, divup(num_heads, BLOCK_H), num_kv_splits] |
| 1434 | // use smaller BLOCK_H when batches is small to utilize all cores |
| 1435 | constexpr int64_t kBLOCK_H = 16; |
| 1436 | const int64_t BLOCK_H = std::min(4 * batches, kBLOCK_H); |
| 1437 | |
| 1438 | // strides |
| 1439 | const int64_t l_stride0 = num_heads * num_kv_splits * (head_size_v + 1); |
| 1440 | const int64_t l_stride1 = num_kv_splits * (head_size_v + 1); |
| 1441 | const int64_t l_stride2 = head_size_v + 1; |
| 1442 | |
| 1443 | const bool has_logit_cap = logit_cap > 0; |
| 1444 | float rlogit_cap = has_logit_cap ? 1 / logit_cap : 0.f; |
| 1445 | |
| 1446 | // partition the heads into blocks for parallel |
| 1447 | const int64_t num_groups = num_heads / num_heads_kv; |
| 1448 | const int64_t num_blocks = div_up(num_groups, BLOCK_H); |
| 1449 | |
| 1450 | // parallel on [batches, num_heads_kv, num_blocks, num_kv_splits] |
| 1451 | at::parallel_for(0, batches * num_heads_kv * num_blocks * num_kv_splits, 0, [&](int64_t begin, int64_t end) { |
| 1452 | int64_t bs{0}, head_kv_id{0}, block_id{0}, kv_id{0}; |
| 1453 | data_index_init(begin, bs, batches, head_kv_id, num_heads_kv, block_id, num_blocks, kv_id, num_kv_splits); |
| 1454 | |
| 1455 | alignas(64) float s_i[BLOCK_H * BLOCK_N]; |
| 1456 | float* __restrict__ s_delta = s_i; |
| 1457 | |
| 1458 | alignas(64) float s_prime[BLOCK_H]; |
| 1459 | alignas(64) float m_prime[BLOCK_H]; |
| 1460 | alignas(64) float m_delta[BLOCK_H]; |
| 1461 | |
| 1462 | for (int64_t i = begin; i < end; ++i) { |
| 1463 | const int64_t h_start = head_kv_id * num_groups + block_id * BLOCK_H; |
| 1464 | const int64_t h_end = head_kv_id * num_groups + std::min(block_id * BLOCK_H + BLOCK_H, num_groups); |
| 1465 | const int64_t h_size = h_end - h_start; |
| 1466 | |
| 1467 | // get query |
| 1468 | const scalar_t* __restrict__ q_ptr = query + bs * q_strideM + h_start * q_strideH; |
| 1469 | |
| 1470 | int64_t seq_len_kv = is_cross_attn ? encoder_lens[bs] : seq_lens[bs]; |
| 1471 | int64_t req_pool_id = req_pool_indices[bs]; |
| 1472 | int64_t kv_offset = (has_encoder_lens && (!is_cross_attn)) ? encoder_lens[bs] : 0; |
| 1473 | TORCH_CHECK(seq_len_kv <= max_context_len, "seq_len_kv out of scope!"); |
| 1474 | TORCH_CHECK(req_pool_id < max_num_reqs, "req_pool_id out of scope!"); |
| 1475 | if (sliding_window_size > 0 && seq_len_kv > sliding_window_size) { |
| 1476 | kv_offset = seq_len_kv - sliding_window_size; |
| 1477 | seq_len_kv = sliding_window_size; |
| 1478 | } |
| 1479 | const int64_t SPLIT_SIZE = div_up(seq_len_kv, num_kv_splits); |
| 1480 | const int64_t kv_start = kv_id * SPLIT_SIZE; |
| 1481 | const int64_t kv_end = std::min(kv_start + SPLIT_SIZE, seq_len_kv); |
| 1482 | |
| 1483 | fill_stub(s_prime, 0.f, BLOCK_H); |
| 1484 | fill_stub(m_prime, -std::numeric_limits<float>::infinity(), BLOCK_H); |
| 1485 | |
| 1486 | // get v_prime, and init to zero |
| 1487 | float* __restrict__ v_prime = attn_logits + bs * l_stride0 + h_start * l_stride1 + kv_id * l_stride2; |
| 1488 | for (int64_t h = 0; h < h_size; ++h) { |
| 1489 | fill_stub(v_prime + h * l_stride1, 0.f, head_size_v); |
| 1490 | } |
| 1491 | |
| 1492 | // loop over K and V sequence with BLOCK_N |
| 1493 | for (int64_t n = kv_start; n < kv_end; n += BLOCK_N) { |
| 1494 | int64_t n_size = std::min(BLOCK_N, kv_end - n); |
| 1495 | |
| 1496 | // calculate Q @ K |
| 1497 | index_gemm_kernel_nt<scalar_t, index_t>( |
| 1498 | /* A */ q_ptr, |
| 1499 | /* B */ k_buffer + head_kv_id * k_strideH, |
| 1500 | /* C */ s_i, |
| 1501 | /* ind */ req_to_token + req_pool_id * max_context_len + n + kv_offset, |
| 1502 | /* scl */ sm_scale, |
| 1503 | /* M */ h_size, |
| 1504 | /* N */ n_size, |
| 1505 | /* K */ head_size, |
| 1506 | /* lda */ q_strideH, |
| 1507 | /* ldb */ k_strideN, |
| 1508 | /* ldc */ BLOCK_N, |
| 1509 | /* mtt */ max_total_num_tokens); |
| 1510 | |
| 1511 | if (has_logit_cap) { |
| 1512 | at::vec::map<float>( |
| 1513 | [logit_cap, rlogit_cap](Vec x) { return Vec(logit_cap) * (x * Vec(rlogit_cap)).tanh(); }, |
| 1514 | s_i, |
| 1515 | s_i, |
| 1516 | BLOCK_H * BLOCK_N); |
| 1517 | } |
| 1518 | |
| 1519 | // update the sm_scale coefficients |
| 1520 | for (int64_t h = 0; h < h_size; ++h) { |
| 1521 | // m_i: max value per row |
| 1522 | float m_i = at::vec::reduce_all<float>( |
| 1523 | [](Vec& x, Vec& y) { return at::vec::maximum(x, y); }, s_i + h * BLOCK_N, n_size); |
| 1524 | m_i = std::max(m_i, m_prime[h]); |
| 1525 | |
| 1526 | // m_delta <- exp(m' - m_i) |
| 1527 | m_delta[h] = std::exp(m_prime[h] - m_i); |
| 1528 | |
| 1529 | // s_delta <- exp(s_i - m_i) |
| 1530 | at::vec::map<float>( |
| 1531 | [m_i](Vec x) { return (x - Vec(m_i)).exp_u20(); }, s_delta + h * BLOCK_N, s_i + h * BLOCK_N, n_size); |
| 1532 | |
| 1533 | // s' <- s' * m_delta + sum(s_delta) |
| 1534 | s_prime[h] *= m_delta[h]; |
| 1535 | s_prime[h] += at::vec::reduce_all<float>([](Vec& x, Vec& y) { return x + y; }, s_delta + h * BLOCK_N, n_size); |
| 1536 | |
| 1537 | m_prime[h] = m_i; |
| 1538 | } |
| 1539 | |
| 1540 | // calculate V' <- s_delta @ V + V' * m_delta |
| 1541 | index_gemm_kernel_nn<scalar_t, index_t>( |
| 1542 | /* A */ s_delta, |
| 1543 | /* B */ v_buffer + head_kv_id * v_strideH, |
| 1544 | /* C */ v_prime, |
| 1545 | /* ind */ req_to_token + req_pool_id * max_context_len + n + kv_offset, |
| 1546 | /* scl */ m_delta, |
| 1547 | /* M */ h_size, |
| 1548 | /* N */ head_size_v, |
| 1549 | /* K */ n_size, |
| 1550 | /* lda */ BLOCK_N, |
| 1551 | /* ldb */ v_strideN, |
| 1552 | /* ldc */ l_stride1, |
| 1553 | /* mtt */ max_total_num_tokens); |
| 1554 | } // loop with KV blocks |
| 1555 | |
| 1556 | // only update v' when kv_split_size > 0 |
| 1557 | if (kv_end > kv_start) { |
| 1558 | for (int64_t h = 0; h < h_size; ++h) { |
| 1559 | float s = 1 / s_prime[h]; |
| 1560 | at::vec::map<float>( |
| 1561 | [s](Vec out) { return out * Vec(s); }, v_prime + h * l_stride1, v_prime + h * l_stride1, head_size_v); |
| 1562 | (v_prime + h * l_stride1)[head_size_v] = m_prime[h] + std::log(s_prime[h]); |
| 1563 | } |
| 1564 | } else { |
| 1565 | for (int64_t h = 0; h < h_size; ++h) { |
| 1566 | (v_prime + h * l_stride1)[head_size_v] = -std::numeric_limits<float>::infinity(); |
| 1567 | } |
| 1568 | } |
| 1569 | |
| 1570 | // move to the next index |
| 1571 | data_index_step(bs, batches, head_kv_id, num_heads_kv, block_id, num_blocks, kv_id, num_kv_splits); |
| 1572 | } |
| 1573 | }); |
| 1574 | |
| 1575 | decode_accumulate_kv_splits( |
| 1576 | output, attn_logits, sinks, batches, num_heads, head_size_v, num_kv_splits, l_stride1, l_stride2, has_sink); |
| 1577 | } // GQA/MQA |
| 1578 | |
| 1579 | } // anonymous namespace |
| 1580 | |
| 1581 | // query: [num_tokens, num_heads, head_size] |
| 1582 | // output: [num_tokens, num_heads, head_size] |
| 1583 | // k_buffer: [max_total_num_tokens, num_heads, head_size] |
| 1584 | // v_buffer: [max_total_num_tokens, num_heads, head_size_v] |
| 1585 | // attn_logits: [num_seqs, num_heads, num_kv_splits, head_size_v + 1] |
| 1586 | // req_to_token: [max_num_reqs, max_context_len] int32 or int64 |
| 1587 | // req_pool_indices: [num_seqs] int64 |
| 1588 | // seq_lens: [num_seqs] int64 |
| 1589 | // encoder_lens: [num_seqs] int64 or None |
| 1590 | // sinks: [num_heads] or None |
| 1591 | void decode_attention_cpu( |
| 1592 | at::Tensor& query, |
| 1593 | at::Tensor& k_buffer, |
| 1594 | at::Tensor& v_buffer, |
| 1595 | at::Tensor& output, |
| 1596 | const std::optional<at::Tensor>& key, |
| 1597 | const std::optional<at::Tensor>& value, |
| 1598 | const std::optional<at::Tensor>& loc, |
| 1599 | at::Tensor& attn_logits, |
| 1600 | at::Tensor& req_to_token, |
| 1601 | at::Tensor& req_pool_indices, |
| 1602 | at::Tensor& seq_lens, |
| 1603 | double sm_scale, |
| 1604 | double logit_cap, |
| 1605 | bool is_cross_attn, |
| 1606 | int64_t sliding_window_size, |
| 1607 | std::optional<at::Tensor> encoder_lens, |
| 1608 | std::optional<at::Tensor> sinks) { |
| 1609 | CHECK_LAST_DIM_CONTIGUOUS_INPUT(query); |
| 1610 | CHECK_LAST_DIM_CONTIGUOUS_INPUT(k_buffer); |
| 1611 | CHECK_LAST_DIM_CONTIGUOUS_INPUT(v_buffer); |
| 1612 | CHECK_DIM(3, query); |
| 1613 | CHECK_DIM(3, k_buffer); |
| 1614 | CHECK_DIM(3, v_buffer); |
| 1615 | |
| 1616 | int64_t num_seqs = seq_lens.size(0); |
| 1617 | int64_t max_num_reqs = req_to_token.size(0); |
| 1618 | int64_t max_context_len = req_to_token.size(1); |
| 1619 | int64_t max_total_num_tokens = k_buffer.size(0); |
| 1620 | |
| 1621 | int64_t num_heads = query.size(1); |
| 1622 | int64_t num_heads_kv = k_buffer.size(1); |
| 1623 | int64_t head_size = query.size(2); |
| 1624 | int64_t head_size_v = v_buffer.size(2); |
| 1625 | |
| 1626 | int64_t num_kv_splits = attn_logits.size(2); |
| 1627 | |
| 1628 | CHECK_EQ(attn_logits.size(0), num_seqs); |
| 1629 | CHECK_EQ(attn_logits.size(1), num_heads); |
| 1630 | CHECK_EQ(attn_logits.size(3), head_size_v + 1); |
| 1631 | CHECK_EQ(attn_logits.scalar_type(), at::kFloat); |
| 1632 | |
| 1633 | // strides for query |
| 1634 | int64_t q_strideM = query.stride(0); |
| 1635 | int64_t q_strideH = query.stride(1); |
| 1636 | |
| 1637 | // strides for k_buffer and v_buffer |
| 1638 | int64_t k_strideN = k_buffer.stride(0); |
| 1639 | int64_t k_strideH = k_buffer.stride(1); |
| 1640 | int64_t v_strideN = v_buffer.stride(0); |
| 1641 | int64_t v_strideH = v_buffer.stride(1); |
| 1642 | |
| 1643 | // check index data types |
| 1644 | const auto index_dtype = req_to_token.scalar_type(); |
| 1645 | TORCH_CHECK( |
| 1646 | index_dtype == at::kInt || index_dtype == at::kLong, |
| 1647 | "decode: expect req_to_token to be int32 or int64, got ", |
| 1648 | index_dtype); |
| 1649 | TORCH_CHECK(seq_lens.scalar_type() == at::kLong, "decode: expect req_lens to be int64, got ", seq_lens.scalar_type()); |
| 1650 | TORCH_CHECK( |
| 1651 | req_pool_indices.scalar_type() == at::kLong, |
| 1652 | "decode: expect req_pool_indices to be int64, got ", |
| 1653 | req_pool_indices.scalar_type()); |
| 1654 | |
| 1655 | // check if we have MLA here |
| 1656 | void* k_buffer_data = k_buffer.data_ptr(); |
| 1657 | void* v_buffer_data = v_buffer.data_ptr(); |
| 1658 | const bool is_mla = (k_buffer_data == v_buffer_data) && (num_heads_kv == 1) && (head_size == head_size_v + 64); |
| 1659 | |
| 1660 | // block length for k_buffer and v_buffer |
| 1661 | constexpr int BLOCK_N = 256; |
| 1662 | |
| 1663 | // buffer for packing k_cache and v_cache |
| 1664 | int num_threads = at::get_num_threads(); |
| 1665 | int64_t size_per_thread = is_mla ? BLOCK_N * head_size + BLOCK_N * head_size_v : 0; |
| 1666 | auto buffer = at::empty({num_threads, size_per_thread}, k_buffer.options()); |
| 1667 | bool has_encoder_lens = encoder_lens.has_value(); |
| 1668 | // Since encoder_lens is not used when it is None, encoder_lens_t can be initialized as any tensor of int64_t dtype. |
| 1669 | at::Tensor encoder_lens_t = seq_lens; |
| 1670 | if (has_encoder_lens) { |
| 1671 | encoder_lens_t = encoder_lens.value(); |
| 1672 | CHECK_EQ(encoder_lens_t.size(0), num_seqs); |
| 1673 | } |
| 1674 | bool has_sink = sinks.has_value(); |
| 1675 | at::Tensor sinks_tensor = has_sink ? sinks.value() : at::empty({num_heads}, query.options()); |
| 1676 | CHECK_DIM(1, sinks_tensor); |
| 1677 | CHECK_EQ(sinks_tensor.size(0), num_heads); |
| 1678 | AT_DISPATCH_REDUCED_FLOATING_TYPES(query.scalar_type(), "decode_attention_kernel", [&] { |
| 1679 | AT_DISPATCH_INDEX_TYPES(index_dtype, "decode_attention_indices", [&] { |
| 1680 | if (key.has_value()) { |
| 1681 | TORCH_CHECK(value.has_value(), "key and value should have values at the same time") |
| 1682 | TORCH_CHECK(loc.has_value(), "loc must be given when key/value are given") |
| 1683 | auto loc_tensor = loc.value(); |
| 1684 | CHECK_DIM(1, loc_tensor); |
| 1685 | CHECK_EQ(loc_tensor.numel(), num_seqs); |
| 1686 | auto key_tensor = key.value(); |
| 1687 | auto value_tensor = value.value(); |
| 1688 | // for MLA, key and value shares the same storage and value could be non-contiguous |
| 1689 | CHECK_LAST_DIM_CONTIGUOUS_INPUT(key_tensor); |
| 1690 | CHECK_LAST_DIM_CONTIGUOUS_INPUT(value_tensor); |
| 1691 | CHECK_DIM(3, key_tensor); |
| 1692 | CHECK_DIM(3, value_tensor); |
| 1693 | // strides for new key and value |
| 1694 | int64_t nk_strideN = key_tensor.stride(0); |
| 1695 | int64_t nk_strideH = key_tensor.stride(1); |
| 1696 | int64_t nv_strideN = value_tensor.stride(0); |
| 1697 | int64_t nv_strideH = value_tensor.stride(1); |
| 1698 | // update the kv buffer |
| 1699 | decode_set_kv_buffer( |
| 1700 | (scalar_t*)k_buffer_data, |
| 1701 | (scalar_t*)v_buffer_data, |
| 1702 | key_tensor.data_ptr<scalar_t>(), |
| 1703 | value_tensor.data_ptr<scalar_t>(), |
| 1704 | loc_tensor.data_ptr<int64_t>(), |
| 1705 | num_seqs, |
| 1706 | num_heads_kv, |
| 1707 | head_size, |
| 1708 | head_size_v, |
| 1709 | k_strideN, |
| 1710 | k_strideH, |
| 1711 | v_strideN, |
| 1712 | v_strideH, |
| 1713 | nk_strideN, |
| 1714 | nk_strideH, |
| 1715 | nv_strideN, |
| 1716 | nv_strideH, |
| 1717 | is_mla); |
| 1718 | } |
| 1719 | |
| 1720 | if (num_heads == num_heads_kv) { |
| 1721 | // MHA |
| 1722 | decode_attention_kernel_impl<scalar_t, index_t, BLOCK_N>( |
| 1723 | output.data_ptr<scalar_t>(), |
| 1724 | attn_logits.data_ptr<float>(), |
| 1725 | query.data_ptr<scalar_t>(), |
| 1726 | (const scalar_t*)k_buffer_data, |
| 1727 | (const scalar_t*)v_buffer_data, |
| 1728 | req_to_token.data_ptr<index_t>(), |
| 1729 | req_pool_indices.data_ptr<int64_t>(), |
| 1730 | seq_lens.data_ptr<int64_t>(), |
| 1731 | encoder_lens_t.data_ptr<int64_t>(), |
| 1732 | sinks_tensor.data_ptr<scalar_t>(), |
| 1733 | num_seqs, |
| 1734 | num_heads, |
| 1735 | head_size, |
| 1736 | head_size_v, |
| 1737 | num_kv_splits, |
| 1738 | q_strideM, |
| 1739 | q_strideH, |
| 1740 | k_strideN, |
| 1741 | k_strideH, |
| 1742 | v_strideN, |
| 1743 | v_strideH, |
| 1744 | sm_scale, |
| 1745 | logit_cap, |
| 1746 | max_num_reqs, |
| 1747 | max_context_len, |
| 1748 | max_total_num_tokens, |
| 1749 | sliding_window_size, |
| 1750 | is_cross_attn, |
| 1751 | has_encoder_lens, |
| 1752 | has_sink); |
| 1753 | } else if (is_mla) { |
| 1754 | // MLA |
| 1755 | decode_attention_mla_kernel_impl<scalar_t, index_t, BLOCK_N>( |
| 1756 | output.data_ptr<scalar_t>(), |
| 1757 | attn_logits.data_ptr<float>(), |
| 1758 | query.data_ptr<scalar_t>(), |
| 1759 | (const scalar_t*)k_buffer_data, |
| 1760 | (const scalar_t*)v_buffer_data, |
| 1761 | req_to_token.data_ptr<index_t>(), |
| 1762 | req_pool_indices.data_ptr<int64_t>(), |
| 1763 | seq_lens.data_ptr<int64_t>(), |
| 1764 | buffer.data_ptr<scalar_t>(), |
| 1765 | sinks_tensor.data_ptr<scalar_t>(), |
| 1766 | num_seqs, |
| 1767 | num_heads, |
| 1768 | head_size, |
| 1769 | head_size_v, |
| 1770 | num_kv_splits, |
| 1771 | q_strideM, |
| 1772 | q_strideH, |
| 1773 | k_strideN, |
| 1774 | k_strideH, |
| 1775 | v_strideN, |
| 1776 | v_strideH, |
| 1777 | sm_scale, |
| 1778 | logit_cap, |
| 1779 | max_num_reqs, |
| 1780 | max_context_len, |
| 1781 | max_total_num_tokens, |
| 1782 | size_per_thread, |
| 1783 | has_sink); |
| 1784 | } else { |
| 1785 | // GQA/MQA |
| 1786 | decode_attention_grouped_kernel_impl<scalar_t, index_t, BLOCK_N>( |
| 1787 | output.data_ptr<scalar_t>(), |
| 1788 | attn_logits.data_ptr<float>(), |
| 1789 | query.data_ptr<scalar_t>(), |
| 1790 | (const scalar_t*)k_buffer_data, |
| 1791 | (const scalar_t*)v_buffer_data, |
| 1792 | req_to_token.data_ptr<index_t>(), |
| 1793 | req_pool_indices.data_ptr<int64_t>(), |
| 1794 | seq_lens.data_ptr<int64_t>(), |
| 1795 | encoder_lens_t.data_ptr<int64_t>(), |
| 1796 | sinks_tensor.data_ptr<scalar_t>(), |
| 1797 | num_seqs, |
| 1798 | num_heads, |
| 1799 | num_heads_kv, |
| 1800 | head_size, |
| 1801 | head_size_v, |
| 1802 | num_kv_splits, |
| 1803 | q_strideM, |
| 1804 | q_strideH, |
| 1805 | k_strideN, |
| 1806 | k_strideH, |
| 1807 | v_strideN, |
| 1808 | v_strideH, |
| 1809 | sm_scale, |
| 1810 | logit_cap, |
| 1811 | max_num_reqs, |
| 1812 | max_context_len, |
| 1813 | max_total_num_tokens, |
| 1814 | sliding_window_size, |
| 1815 | is_cross_attn, |
| 1816 | has_encoder_lens, |
| 1817 | has_sink); |
| 1818 | } |
| 1819 | }); |
| 1820 | }); |
| 1821 | } |