| 1 | /* |
| 2 | * Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved. |
| 3 | * |
| 4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | * you may not use this file except in compliance with the License. |
| 6 | * You may obtain a copy of the License at |
| 7 | * |
| 8 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | * |
| 10 | * Unless required by applicable law or agreed to in writing, software |
| 11 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | * See the License for the specific language governing permissions and |
| 14 | * limitations under the License. |
| 15 | */ |
| 16 | |
| 17 | #include <torch/csrc/stable/tensor.h> |
| 18 | |
| 19 | #include "libtorch_stable/torch_utils.h" |
| 20 | |
| 21 | #include "libtorch_stable/cutlass_extensions/common.hpp" |
| 22 | |
| 23 | #if defined ENABLE_NVFP4_SM100 && ENABLE_NVFP4_SM100 |
| 24 | void cutlass_scaled_fp4_mm_sm100a(torch::stable::Tensor& D, |
| 25 | torch::stable::Tensor const& A, |
| 26 | torch::stable::Tensor const& B, |
| 27 | torch::stable::Tensor const& A_sf, |
| 28 | torch::stable::Tensor const& B_sf, |
| 29 | torch::stable::Tensor const& alpha); |
| 30 | #endif |
| 31 | |
| 32 | #if defined ENABLE_NVFP4_SM120 && ENABLE_NVFP4_SM120 |
| 33 | void cutlass_scaled_fp4_mm_sm120a(torch::stable::Tensor& D, |
| 34 | torch::stable::Tensor const& A, |
| 35 | torch::stable::Tensor const& B, |
| 36 | torch::stable::Tensor const& A_sf, |
| 37 | torch::stable::Tensor const& B_sf, |
| 38 | torch::stable::Tensor const& alpha); |
| 39 | #endif |
| 40 | |
| 41 | void cutlass_scaled_fp4_mm(torch::stable::Tensor& D, |
| 42 | const torch::stable::Tensor& A, |
| 43 | const torch::stable::Tensor& B, |
| 44 | const torch::stable::Tensor& A_sf, |
| 45 | const torch::stable::Tensor& B_sf, |
| 46 | const torch::stable::Tensor& alpha) { |
| 47 | // Make sure we're on A's device. |
| 48 | const torch::stable::accelerator::DeviceGuard device_guard( |
| 49 | A.get_device_index()); |
| 50 | const int32_t sm = get_sm_version_num(); |
| 51 | |
| 52 | #if defined(ENABLE_NVFP4_SM100) && ENABLE_NVFP4_SM100 |
| 53 | if (sm >= 100 && sm < 120) { |
| 54 | cutlass_scaled_fp4_mm_sm100a(D, A, B, A_sf, B_sf, alpha); |
| 55 | return; |
| 56 | } |
| 57 | #endif |
| 58 | |
| 59 | #if defined(ENABLE_NVFP4_SM120) && ENABLE_NVFP4_SM120 |
| 60 | if (sm >= 120 && sm < 130) { |
| 61 | cutlass_scaled_fp4_mm_sm120a(D, A, B, A_sf, B_sf, alpha); |
| 62 | return; |
| 63 | } |
| 64 | #endif |
| 65 | |
| 66 | STD_TORCH_CHECK_NOT_IMPLEMENTED( |
| 67 | false, "No compiled nvfp4 mm kernel for SM ", sm, |
| 68 | ". Recompile with CUDA >= 12.8 and CC >= 100."); |
| 69 | } |
| 70 | |
| 71 | bool cutlass_scaled_mm_supports_fp4(int64_t cuda_device_capability) { |
| 72 | int runtimeVersion; |
| 73 | cudaRuntimeGetVersion(&runtimeVersion); |
| 74 | if (runtimeVersion < 12080) return false; |
| 75 | // Only report support when the SM-specific kernel was actually compiled in, |
| 76 | // so the Python-side backend selector does not choose CUTLASS and then hit |
| 77 | // TORCH_CHECK_NOT_IMPLEMENTED (or worse, fall through to Marlin). |
| 78 | #if defined(ENABLE_NVFP4_SM100) && ENABLE_NVFP4_SM100 |
| 79 | if (cuda_device_capability >= 100 && cuda_device_capability < 120) |
| 80 | return true; |
| 81 | #endif |
| 82 | #if defined(ENABLE_NVFP4_SM120) && ENABLE_NVFP4_SM120 |
| 83 | if (cuda_device_capability >= 120 && cuda_device_capability < 130) |
| 84 | return true; |
| 85 | #endif |
| 86 | return false; |
| 87 | } |