| 1 | #include <cuda.h> |
| 2 | #include <cudaTypedefs.h> |
| 3 | |
| 4 | #include <torch/csrc/stable/tensor.h> |
| 5 | |
| 6 | #include "libtorch_stable/torch_utils.h" |
| 7 | |
| 8 | #include "libtorch_stable/cutlass_extensions/common.hpp" |
| 9 | |
| 10 | void cutlass_scaled_mm_sm75(torch::stable::Tensor& c, |
| 11 | torch::stable::Tensor const& a, |
| 12 | torch::stable::Tensor const& b, |
| 13 | torch::stable::Tensor const& a_scales, |
| 14 | torch::stable::Tensor const& b_scales, |
| 15 | std::optional<torch::stable::Tensor> const& bias); |
| 16 | |
| 17 | void cutlass_scaled_mm_sm80(torch::stable::Tensor& c, |
| 18 | torch::stable::Tensor const& a, |
| 19 | torch::stable::Tensor const& b, |
| 20 | torch::stable::Tensor const& a_scales, |
| 21 | torch::stable::Tensor const& b_scales, |
| 22 | std::optional<torch::stable::Tensor> const& bias); |
| 23 | |
| 24 | void cutlass_scaled_mm_sm89(torch::stable::Tensor& c, |
| 25 | torch::stable::Tensor const& a, |
| 26 | torch::stable::Tensor const& b, |
| 27 | torch::stable::Tensor const& a_scales, |
| 28 | torch::stable::Tensor const& b_scales, |
| 29 | std::optional<torch::stable::Tensor> const& bias); |
| 30 | |
| 31 | #if defined ENABLE_SCALED_MM_SM90 && ENABLE_SCALED_MM_SM90 |
| 32 | void cutlass_scaled_mm_sm90(torch::stable::Tensor& c, |
| 33 | torch::stable::Tensor const& a, |
| 34 | torch::stable::Tensor const& b, |
| 35 | torch::stable::Tensor const& a_scales, |
| 36 | torch::stable::Tensor const& b_scales, |
| 37 | std::optional<torch::stable::Tensor> const& bias); |
| 38 | #endif |
| 39 | #if defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90 |
| 40 | void cutlass_moe_mm_sm90(torch::stable::Tensor& out_tensors, |
| 41 | torch::stable::Tensor const& a_tensors, |
| 42 | torch::stable::Tensor const& b_tensors, |
| 43 | torch::stable::Tensor const& a_scales, |
| 44 | torch::stable::Tensor const& b_scales, |
| 45 | torch::stable::Tensor const& expert_offsets, |
| 46 | torch::stable::Tensor const& problem_sizes, |
| 47 | torch::stable::Tensor const& a_strides, |
| 48 | torch::stable::Tensor const& b_strides, |
| 49 | torch::stable::Tensor const& c_strides, |
| 50 | bool per_act_token, bool per_out_ch); |
| 51 | |
| 52 | #endif |
| 53 | |
| 54 | #if defined ENABLE_CUTLASS_MOE_SM100 && ENABLE_CUTLASS_MOE_SM100 |
| 55 | void cutlass_moe_mm_sm100(torch::stable::Tensor& out_tensors, |
| 56 | torch::stable::Tensor const& a_tensors, |
| 57 | torch::stable::Tensor const& b_tensors, |
| 58 | torch::stable::Tensor const& a_scales, |
| 59 | torch::stable::Tensor const& b_scales, |
| 60 | torch::stable::Tensor const& expert_offsets, |
| 61 | torch::stable::Tensor const& problem_sizes, |
| 62 | torch::stable::Tensor const& a_strides, |
| 63 | torch::stable::Tensor const& b_strides, |
| 64 | torch::stable::Tensor const& c_strides, |
| 65 | bool per_act_token, bool per_out_ch); |
| 66 | #endif |
| 67 | |
| 68 | #if defined ENABLE_SCALED_MM_SM120 && ENABLE_SCALED_MM_SM120 |
| 69 | void cutlass_scaled_mm_sm120(torch::stable::Tensor& c, |
| 70 | torch::stable::Tensor const& a, |
| 71 | torch::stable::Tensor const& b, |
| 72 | torch::stable::Tensor const& a_scales, |
| 73 | torch::stable::Tensor const& b_scales, |
| 74 | std::optional<torch::stable::Tensor> const& bias); |
| 75 | #endif |
| 76 | |
| 77 | #if defined ENABLE_SCALED_MM_SM100 && ENABLE_SCALED_MM_SM100 |
| 78 | void cutlass_scaled_mm_sm100(torch::stable::Tensor& c, |
| 79 | torch::stable::Tensor const& a, |
| 80 | torch::stable::Tensor const& b, |
| 81 | torch::stable::Tensor const& a_scales, |
| 82 | torch::stable::Tensor const& b_scales, |
| 83 | std::optional<torch::stable::Tensor> const& bias); |
| 84 | #endif |
| 85 | |
| 86 | #if (defined(ENABLE_CUTLASS_MOE_SM90) && ENABLE_CUTLASS_MOE_SM90) || \ |
| 87 | (defined(ENABLE_CUTLASS_MOE_SM100) && ENABLE_CUTLASS_MOE_SM100) || \ |
| 88 | (defined(ENABLE_CUTLASS_MOE_SM120) && ENABLE_CUTLASS_MOE_SM120) |
| 89 | void get_cutlass_moe_mm_data_caller( |
| 90 | const torch::stable::Tensor& topk_ids, |
| 91 | torch::stable::Tensor& expert_offsets, |
| 92 | torch::stable::Tensor& problem_sizes1, |
| 93 | torch::stable::Tensor& problem_sizes2, |
| 94 | torch::stable::Tensor& input_permutation, |
| 95 | torch::stable::Tensor& output_permutation, const int64_t num_experts, |
| 96 | const int64_t n, const int64_t k, |
| 97 | const std::optional<torch::stable::Tensor>& blockscale_offsets, |
| 98 | const bool is_gated); |
| 99 | |
| 100 | void get_cutlass_moe_mm_problem_sizes_from_expert_offsets_caller( |
| 101 | const torch::stable::Tensor& expert_first_token_offset, |
| 102 | torch::stable::Tensor& problem_sizes1, |
| 103 | torch::stable::Tensor& problem_sizes2, const int64_t n, const int64_t k, |
| 104 | const bool swap_ab); |
| 105 | |
| 106 | void get_cutlass_batched_moe_mm_data_caller( |
| 107 | torch::stable::Tensor& expert_offsets, |
| 108 | torch::stable::Tensor& problem_sizes1, |
| 109 | torch::stable::Tensor& problem_sizes2, |
| 110 | const torch::stable::Tensor& expert_num_tokens, |
| 111 | const int64_t num_local_experts, const int64_t padded_m, const int64_t n, |
| 112 | const int64_t k); |
| 113 | #endif |
| 114 | |
| 115 | void cutlass_scaled_mm_azp_sm75( |
| 116 | torch::stable::Tensor& c, torch::stable::Tensor const& a, |
| 117 | torch::stable::Tensor const& b, torch::stable::Tensor const& a_scales, |
| 118 | torch::stable::Tensor const& b_scales, torch::stable::Tensor const& azp_adj, |
| 119 | std::optional<torch::stable::Tensor> const& azp, |
| 120 | std::optional<torch::stable::Tensor> const& bias); |
| 121 | |
| 122 | void cutlass_scaled_mm_azp_sm80( |
| 123 | torch::stable::Tensor& c, torch::stable::Tensor const& a, |
| 124 | torch::stable::Tensor const& b, torch::stable::Tensor const& a_scales, |
| 125 | torch::stable::Tensor const& b_scales, torch::stable::Tensor const& azp_adj, |
| 126 | std::optional<torch::stable::Tensor> const& azp, |
| 127 | std::optional<torch::stable::Tensor> const& bias); |
| 128 | |
| 129 | void cutlass_scaled_mm_azp_sm89( |
| 130 | torch::stable::Tensor& c, torch::stable::Tensor const& a, |
| 131 | torch::stable::Tensor const& b, torch::stable::Tensor const& a_scales, |
| 132 | torch::stable::Tensor const& b_scales, torch::stable::Tensor const& azp_adj, |
| 133 | std::optional<torch::stable::Tensor> const& azp, |
| 134 | std::optional<torch::stable::Tensor> const& bias); |
| 135 | |
| 136 | #if defined ENABLE_SCALED_MM_SM90 && ENABLE_SCALED_MM_SM90 |
| 137 | void cutlass_scaled_mm_azp_sm90( |
| 138 | torch::stable::Tensor& c, torch::stable::Tensor const& a, |
| 139 | torch::stable::Tensor const& b, torch::stable::Tensor const& a_scales, |
| 140 | torch::stable::Tensor const& b_scales, torch::stable::Tensor const& azp_adj, |
| 141 | std::optional<torch::stable::Tensor> const& azp, |
| 142 | std::optional<torch::stable::Tensor> const& bias); |
| 143 | #endif |
| 144 | |
| 145 | bool cutlass_scaled_mm_supports_fp8(int64_t cuda_device_capability) { |
| 146 | // CUTLASS FP8 kernels need at least |
| 147 | // CUDA 12.0 on SM90 systems (Hopper) |
| 148 | // CUDA 12.4 on SM89 systems (Lovelace) |
| 149 | |
| 150 | #if defined CUDA_VERSION |
| 151 | if (cuda_device_capability >= 90) { |
| 152 | return CUDA_VERSION >= 12000; |
| 153 | } else if (cuda_device_capability >= 89) { |
| 154 | return CUDA_VERSION >= 12040; |
| 155 | } |
| 156 | #endif |
| 157 | |
| 158 | return false; |
| 159 | } |
| 160 | |
| 161 | bool cutlass_scaled_mm_supports_block_fp8(int64_t cuda_device_capability) { |
| 162 | // CUTLASS block-quantized FP8 kernels need at least CUDA 12.0 |
| 163 | // and at least SM90 (Hopper) |
| 164 | |
| 165 | #if defined CUDA_VERSION |
| 166 | if (cuda_device_capability >= 100) { |
| 167 | return CUDA_VERSION >= 12080; |
| 168 | } else if (cuda_device_capability >= 90) { |
| 169 | return CUDA_VERSION >= 12000; |
| 170 | } |
| 171 | #endif |
| 172 | |
| 173 | return false; |
| 174 | } |
| 175 | |
| 176 | bool cutlass_group_gemm_supported(int64_t cuda_device_capability) { |
| 177 | // CUTLASS grouped FP8 kernels need at least CUDA 12.3 and SM90 (Hopper) |
| 178 | // or CUDA 12.8 and SM100 (Blackwell). Only report archs that have an |
| 179 | // actual cutlass_moe_mm dispatch compiled into this file. |
| 180 | |
| 181 | #if defined CUDA_VERSION |
| 182 | #if defined ENABLE_CUTLASS_MOE_SM100 && ENABLE_CUTLASS_MOE_SM100 |
| 183 | if (cuda_device_capability >= 100 && cuda_device_capability < 120) { |
| 184 | return CUDA_VERSION >= 12080; |
| 185 | } |
| 186 | #endif |
| 187 | #if defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90 |
| 188 | if (cuda_device_capability >= 90 && cuda_device_capability < 100) { |
| 189 | return CUDA_VERSION >= 12030; |
| 190 | } |
| 191 | #endif |
| 192 | #endif |
| 193 | |
| 194 | return false; |
| 195 | } |
| 196 | |
| 197 | void cutlass_scaled_mm(torch::stable::Tensor& c, torch::stable::Tensor const& a, |
| 198 | torch::stable::Tensor const& b, |
| 199 | torch::stable::Tensor const& a_scales, |
| 200 | torch::stable::Tensor const& b_scales, |
| 201 | std::optional<torch::stable::Tensor> const& bias) { |
| 202 | // Checks for conformality |
| 203 | STD_TORCH_CHECK(a.dim() == 2 && b.dim() == 2 && c.dim() == 2); |
| 204 | STD_TORCH_CHECK(c.size(0) == a.size(0) && a.size(1) == b.size(0) && |
| 205 | b.size(1) == c.size(1)); |
| 206 | |
| 207 | // Check for strides and alignment |
| 208 | STD_TORCH_CHECK(a.stride(1) == 1 && c.stride(1) == 1); // Row-major |
| 209 | STD_TORCH_CHECK(b.stride(0) == 1); // Column-major |
| 210 | STD_TORCH_CHECK(c.stride(0) % 16 == 0 && |
| 211 | b.stride(1) % 16 == 0); // 16 Byte Alignment |
| 212 | |
| 213 | if (bias) { |
| 214 | STD_TORCH_CHECK(bias->numel() == b.size(1) && bias->is_contiguous() && |
| 215 | bias->dim() == 1); |
| 216 | } |
| 217 | |
| 218 | const torch::stable::accelerator::DeviceGuard device_guard( |
| 219 | a.get_device_index()); |
| 220 | int32_t version_num = get_sm_version_num(); |
| 221 | |
| 222 | #if defined ENABLE_SCALED_MM_SM120 && ENABLE_SCALED_MM_SM120 |
| 223 | if (version_num >= 120) { |
| 224 | cutlass_scaled_mm_sm120(c, a, b, a_scales, b_scales, bias); |
| 225 | return; |
| 226 | } |
| 227 | #endif |
| 228 | |
| 229 | #if defined ENABLE_SCALED_MM_SM100 && ENABLE_SCALED_MM_SM100 |
| 230 | if (version_num >= 100 && version_num < 120) { |
| 231 | cutlass_scaled_mm_sm100(c, a, b, a_scales, b_scales, bias); |
| 232 | return; |
| 233 | } |
| 234 | #endif |
| 235 | |
| 236 | // Guard against compilation issues for sm90 kernels |
| 237 | #if defined ENABLE_SCALED_MM_SM90 && ENABLE_SCALED_MM_SM90 |
| 238 | if (version_num >= 90 && version_num < 100) { |
| 239 | // Hopper |
| 240 | cutlass_scaled_mm_sm90(c, a, b, a_scales, b_scales, bias); |
| 241 | return; |
| 242 | } |
| 243 | #endif |
| 244 | |
| 245 | #if defined ENABLE_SCALED_MM_C2X && ENABLE_SCALED_MM_C2X |
| 246 | if (version_num == 89) { |
| 247 | // Ada Lovelace |
| 248 | cutlass_scaled_mm_sm89(c, a, b, a_scales, b_scales, bias); |
| 249 | return; |
| 250 | } |
| 251 | |
| 252 | if (version_num >= 80) { |
| 253 | // Ampere |
| 254 | cutlass_scaled_mm_sm80(c, a, b, a_scales, b_scales, bias); |
| 255 | return; |
| 256 | } |
| 257 | |
| 258 | if (version_num >= 75) { |
| 259 | // Turing |
| 260 | cutlass_scaled_mm_sm75(c, a, b, a_scales, b_scales, bias); |
| 261 | return; |
| 262 | } |
| 263 | #endif |
| 264 | |
| 265 | STD_TORCH_CHECK_NOT_IMPLEMENTED( |
| 266 | false, |
| 267 | "No compiled cutlass_scaled_mm for a compute capability less than " |
| 268 | "CUDA device capability: ", |
| 269 | version_num); |
| 270 | } |
| 271 | |
| 272 | void cutlass_moe_mm(torch::stable::Tensor& out_tensors, |
| 273 | torch::stable::Tensor const& a_tensors, |
| 274 | torch::stable::Tensor const& b_tensors, |
| 275 | torch::stable::Tensor const& a_scales, |
| 276 | torch::stable::Tensor const& b_scales, |
| 277 | torch::stable::Tensor const& expert_offsets, |
| 278 | torch::stable::Tensor const& problem_sizes, |
| 279 | torch::stable::Tensor const& a_strides, |
| 280 | torch::stable::Tensor const& b_strides, |
| 281 | torch::stable::Tensor const& c_strides, bool per_act_token, |
| 282 | bool per_out_ch) { |
| 283 | int32_t version_num = get_sm_version_num(); |
| 284 | #if defined ENABLE_CUTLASS_MOE_SM100 && ENABLE_CUTLASS_MOE_SM100 |
| 285 | if (version_num >= 100 && version_num < 110) { |
| 286 | cutlass_moe_mm_sm100(out_tensors, a_tensors, b_tensors, a_scales, b_scales, |
| 287 | expert_offsets, problem_sizes, a_strides, b_strides, |
| 288 | c_strides, per_act_token, per_out_ch); |
| 289 | return; |
| 290 | } |
| 291 | #endif |
| 292 | #if defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90 |
| 293 | if (version_num >= 90 && version_num < 100) { |
| 294 | cutlass_moe_mm_sm90(out_tensors, a_tensors, b_tensors, a_scales, b_scales, |
| 295 | expert_offsets, problem_sizes, a_strides, b_strides, |
| 296 | c_strides, per_act_token, per_out_ch); |
| 297 | return; |
| 298 | } |
| 299 | #endif |
| 300 | STD_TORCH_CHECK_NOT_IMPLEMENTED( |
| 301 | false, |
| 302 | "No compiled cutlass_scaled_mm for CUDA device capability: ", version_num, |
| 303 | ". Required capability: 90 or 100"); |
| 304 | } |
| 305 | |
| 306 | void get_cutlass_moe_mm_data( |
| 307 | const torch::stable::Tensor& topk_ids, |
| 308 | torch::stable::Tensor& expert_offsets, |
| 309 | torch::stable::Tensor& problem_sizes1, |
| 310 | torch::stable::Tensor& problem_sizes2, |
| 311 | torch::stable::Tensor& input_permutation, |
| 312 | torch::stable::Tensor& output_permutation, const int64_t num_experts, |
| 313 | const int64_t n, const int64_t k, |
| 314 | const std::optional<torch::stable::Tensor>& blockscale_offsets, |
| 315 | const bool is_gated) { |
| 316 | // This function currently gets compiled only if we have a valid cutlass moe |
| 317 | // mm to run it for. |
| 318 | int32_t version_num = get_sm_version_num(); |
| 319 | #if (defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90) || \ |
| 320 | (defined ENABLE_CUTLASS_MOE_SM100 && ENABLE_CUTLASS_MOE_SM100) || \ |
| 321 | (defined ENABLE_CUTLASS_MOE_SM120 && ENABLE_CUTLASS_MOE_SM120) |
| 322 | get_cutlass_moe_mm_data_caller(topk_ids, expert_offsets, problem_sizes1, |
| 323 | problem_sizes2, input_permutation, |
| 324 | output_permutation, num_experts, n, k, |
| 325 | blockscale_offsets, is_gated); |
| 326 | return; |
| 327 | #endif |
| 328 | STD_TORCH_CHECK_NOT_IMPLEMENTED( |
| 329 | false, |
| 330 | "No compiled get_cutlass_moe_mm_data: no cutlass_scaled_mm kernel for " |
| 331 | "CUDA device capability: ", |
| 332 | version_num, ". Required capability: 90, 100, or 120"); |
| 333 | } |
| 334 | |
| 335 | void get_cutlass_moe_mm_problem_sizes_from_expert_offsets( |
| 336 | const torch::stable::Tensor& expert_first_token_offset, |
| 337 | torch::stable::Tensor& problem_sizes1, |
| 338 | torch::stable::Tensor& problem_sizes2, const int64_t n, const int64_t k, |
| 339 | const bool swap_ab) { |
| 340 | int32_t version_num = get_sm_version_num(); |
| 341 | #if (defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90) || \ |
| 342 | (defined ENABLE_CUTLASS_MOE_SM100 && ENABLE_CUTLASS_MOE_SM100) || \ |
| 343 | (defined ENABLE_CUTLASS_MOE_SM120 && ENABLE_CUTLASS_MOE_SM120) |
| 344 | get_cutlass_moe_mm_problem_sizes_from_expert_offsets_caller( |
| 345 | expert_first_token_offset, problem_sizes1, problem_sizes2, n, k, swap_ab); |
| 346 | return; |
| 347 | #endif |
| 348 | STD_TORCH_CHECK_NOT_IMPLEMENTED( |
| 349 | false, |
| 350 | "No compiled get_cutlass_moe_mm_problem_sizes_from_expert_offsets: " |
| 351 | "no cutlass_scaled_mm kernel for CUDA device capability: ", |
| 352 | version_num, ". Required capability: 90, 100, or 120"); |
| 353 | } |
| 354 | |
| 355 | void get_cutlass_batched_moe_mm_data( |
| 356 | torch::stable::Tensor& expert_offsets, |
| 357 | torch::stable::Tensor& problem_sizes1, |
| 358 | torch::stable::Tensor& problem_sizes2, |
| 359 | const torch::stable::Tensor& expert_num_tokens, |
| 360 | const int64_t num_local_experts, const int64_t padded_m, const int64_t n, |
| 361 | const int64_t k) { |
| 362 | // This function currently gets compiled only if we have a valid cutlass moe |
| 363 | // mm to run it for. |
| 364 | int32_t version_num = get_sm_version_num(); |
| 365 | #if (defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90) || \ |
| 366 | (defined ENABLE_CUTLASS_MOE_SM100 && ENABLE_CUTLASS_MOE_SM100) || \ |
| 367 | (defined ENABLE_CUTLASS_MOE_SM120 && ENABLE_CUTLASS_MOE_SM120) |
| 368 | get_cutlass_batched_moe_mm_data_caller(expert_offsets, problem_sizes1, |
| 369 | problem_sizes2, expert_num_tokens, |
| 370 | num_local_experts, padded_m, n, k); |
| 371 | return; |
| 372 | #endif |
| 373 | STD_TORCH_CHECK_NOT_IMPLEMENTED( |
| 374 | false, |
| 375 | "No compiled get_cutlass_batched_moe_mm_data: no " |
| 376 | "cutlass_scaled_mm kernel " |
| 377 | "for CUDA device capability: ", |
| 378 | version_num, ". Required capability: 90, 100, or 120"); |
| 379 | } |
| 380 | |
| 381 | void cutlass_scaled_mm_azp(torch::stable::Tensor& c, |
| 382 | torch::stable::Tensor const& a, |
| 383 | torch::stable::Tensor const& b, |
| 384 | torch::stable::Tensor const& a_scales, |
| 385 | torch::stable::Tensor const& b_scales, |
| 386 | torch::stable::Tensor const& azp_adj, |
| 387 | std::optional<torch::stable::Tensor> const& azp, |
| 388 | std::optional<torch::stable::Tensor> const& bias) { |
| 389 | // Checks for conformality |
| 390 | STD_TORCH_CHECK(a.dim() == 2 && b.dim() == 2 && c.dim() == 2); |
| 391 | STD_TORCH_CHECK(c.size(0) == a.size(0) && a.size(1) == b.size(0) && |
| 392 | b.size(1) == c.size(1)); |
| 393 | STD_TORCH_CHECK(a_scales.numel() == 1 || a_scales.numel() == a.size(0)); |
| 394 | STD_TORCH_CHECK(b_scales.numel() == 1 || b_scales.numel() == b.size(1)); |
| 395 | |
| 396 | // Check for strides and alignment |
| 397 | STD_TORCH_CHECK(a.stride(1) == 1 && c.stride(1) == 1); // Row-major |
| 398 | STD_TORCH_CHECK(b.stride(0) == 1); // Column-major |
| 399 | STD_TORCH_CHECK(c.stride(0) % 16 == 0 && |
| 400 | b.stride(1) % 16 == 0); // 16 Byte Alignment |
| 401 | STD_TORCH_CHECK(a_scales.is_contiguous() && b_scales.is_contiguous()); |
| 402 | |
| 403 | // bias, azp, azp_adj are all 1d |
| 404 | // bias and azp_adj have n elements, azp has m elements |
| 405 | if (bias) { |
| 406 | STD_TORCH_CHECK(bias->numel() == b.size(1) && bias->is_contiguous()); |
| 407 | } |
| 408 | if (azp) { |
| 409 | STD_TORCH_CHECK(azp->numel() == a.size(0) && azp->is_contiguous()); |
| 410 | } |
| 411 | STD_TORCH_CHECK(azp_adj.numel() == b.size(1) && azp_adj.is_contiguous()); |
| 412 | |
| 413 | // azp & bias types |
| 414 | STD_TORCH_CHECK(azp_adj.scalar_type() == torch::headeronly::ScalarType::Int); |
| 415 | STD_TORCH_CHECK(!azp || |
| 416 | azp->scalar_type() == torch::headeronly::ScalarType::Int); |
| 417 | STD_TORCH_CHECK(!bias || bias->scalar_type() == c.scalar_type(), |
| 418 | "currently bias dtype must match output dtype ", |
| 419 | c.scalar_type()); |
| 420 | |
| 421 | const torch::stable::accelerator::DeviceGuard device_guard( |
| 422 | a.get_device_index()); |
| 423 | |
| 424 | int32_t version_num = get_sm_version_num(); |
| 425 | |
| 426 | #if defined ENABLE_SCALED_MM_SM90 && ENABLE_SCALED_MM_SM90 |
| 427 | if (version_num >= 90) { |
| 428 | cutlass_scaled_mm_azp_sm90(c, a, b, a_scales, b_scales, azp_adj, azp, bias); |
| 429 | return; |
| 430 | } |
| 431 | #endif |
| 432 | |
| 433 | #if defined ENABLE_SCALED_MM_C2X && ENABLE_SCALED_MM_C2X |
| 434 | if (version_num == 89) { |
| 435 | // Ada Lovelace |
| 436 | cutlass_scaled_mm_azp_sm89(c, a, b, a_scales, b_scales, azp_adj, azp, bias); |
| 437 | return; |
| 438 | } |
| 439 | |
| 440 | if (version_num >= 80) { |
| 441 | // Ampere |
| 442 | cutlass_scaled_mm_azp_sm80(c, a, b, a_scales, b_scales, azp_adj, azp, bias); |
| 443 | return; |
| 444 | } |
| 445 | |
| 446 | // Turing |
| 447 | STD_TORCH_CHECK(version_num >= 75); |
| 448 | cutlass_scaled_mm_azp_sm75(c, a, b, a_scales, b_scales, azp_adj, azp, bias); |
| 449 | return; |
| 450 | #endif |
| 451 | |
| 452 | STD_TORCH_CHECK_NOT_IMPLEMENTED( |
| 453 | false, |
| 454 | "No compiled cutlass_scaled_mm_azp for a compute capability less than " |
| 455 | "CUDA device capability: ", |
| 456 | version_num); |
| 457 | } |