biondizzle/DeepGEMM
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Multiple updates and refactorings (#280)

Browse Source
This commit is contained in:
Zhean Xu
2026-01-16 17:06:52 +08:00
committed by GitHub
parent 3ccf40c53a
commit 0f5f266202
55 changed files with 2706 additions and 891 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
csrc/utils/compatibility.hpp
View File

@@ -2,9 +2,16 @@
#include <torch/version.h>
#include <cuda.h>
#include <cuda_runtime.h>
// `torch::kFloat8_e4m3fn` is supported since PyTorch 2.1
#define DG_FP8_COMPATIBLE (TORCH_VERSION_MAJOR > 2 or (TORCH_VERSION_MAJOR == 2 and TORCH_VERSION_MINOR >= 1))
// `cuTensorMapEncodeTiled` is supported since CUDA Driver API 12.1
#define DG_TENSORMAP_COMPATIBLE (CUDA_VERSION >= 12010)
#define DG_TENSORMAP_COMPATIBLE (CUDA_VERSION >= 12010)
// `cublasGetErrorString` is supported since CUDA Runtime API 11.4.2
#define DG_CUBLAS_GET_ERROR_STRING_COMPATIBLE (CUDART_VERSION >= 11042)
// `CUBLASLT_MATMUL_DESC_FAST_ACCUM` and `CUBLASLT_MATMUL_DESC_SM_COUNT_TARGET` are supported since CUDA Runtime API 11.8
#define DG_CUBLASLT_ADVANCED_FEATURES_COMPATIBLE (CUDART_VERSION >= 11080)