feat: add support for deepgemm sglang

csrc/python_api.cpp

@@ -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)

deep_gemm/__init__.py

@@ -56,6 +56,7 @@ set_compile_mode = _wrap_op('set_compile_mode')
 get_compile_mode = _wrap_op('get_compile_mode')
 set_tc_util = _wrap_op('set_tc_util')
 get_tc_util = _wrap_op('get_tc_util')
+preload_kernels = _wrap_op('preload_kernels')
 
 fp8_gemm_nt = _wrap_op('fp8_gemm_nt')
 fp8_gemm_nn = _wrap_op('fp8_gemm_nn')
@@ -109,7 +110,7 @@ from .utils import *
 
 def _verify_ops_loaded():
     expected_ops = [
-        'init', 'set_num_sms', 'get_num_sms', 'set_tc_util', 'get_tc_util',
+        'init', 'preload_kernels', 'set_num_sms', 'get_num_sms', 'set_tc_util', 'get_tc_util',
         'fp8_gemm_nt', 'fp8_gemm_nn', 'fp8_gemm_tn', 'fp8_gemm_tt',
         'm_grouped_fp8_gemm_nt_contiguous', 'm_grouped_fp8_gemm_nn_contiguous',
         'm_grouped_fp8_gemm_nt_masked', 'k_grouped_fp8_gemm_nt_contiguous',
@@ -141,10 +142,10 @@ if __debug__:
 # Preload cached kernels if enabled via environment variable
 if os.environ.get('DG_PRELOAD_KERNELS', '0') == '1':
     try:
-        deep_gemm_cpp.preload_kernels()
+        preload_kernels()
     except Exception as e:
         # Preloading failure should not block initialization
         import warnings
         warnings.warn(f"Failed to preload DeepGEMM kernels: {e}")
 
-__version__ = '2.1.2'
+__version__ = '2.2.1'