feat: add support for deepgemm sglang
This commit is contained in:
@@ -247,6 +247,12 @@ TORCH_LIBRARY(deep_gemm, m) {
|
||||
deep_gemm_wrappers::init_wrapper(library_root_path, cuda_home_path_by_torch);
|
||||
});
|
||||
|
||||
m.def("preload_kernels() -> ()");
|
||||
m.impl("preload_kernels", []() {
|
||||
const auto& cache_dir = deep_gemm::Compiler::get_cache_dir();
|
||||
deep_gemm::kernel_runtime_cache->preload_all(cache_dir);
|
||||
});
|
||||
|
||||
// layout APIs
|
||||
m.def("transform_sf_into_required_layout(Tensor sf, int mn, int k, int[] recipe, int? num_groups=None, bool is_sfa=False, bool disable_ue8m0_cast=False) -> Tensor", deep_gemm_wrappers::transform_sf_into_required_layout_wrapper);
|
||||
|
||||
@@ -540,4 +546,4 @@ int64_t deep_gemm_wrappers::get_mk_alignment_for_contiguous_layout_wrapper() {
|
||||
return static_cast<int64_t>(deep_gemm::get_mk_alignment_for_contiguous_layout());
|
||||
}
|
||||
|
||||
REGISTER_EXTENSION(deep_gemm_cpp)
|
||||
REGISTER_EXTENSION(deep_gemm_cpp)
|
||||
|
||||
Reference in New Issue
Block a user