Revert "[sgl-kernel] upgrade deepgemm" (#18562)

sgl-kernel/CMakeLists.txt

@@ -56,7 +56,7 @@ FetchContent_Populate(repo-cutlass)
 FetchContent_Declare(
     repo-deepgemm
     GIT_REPOSITORY https://github.com/sgl-project/DeepGEMM
-    GIT_TAG        fc08e5775a5db960690a41248655168c55bef228
+    GIT_TAG        ffe2b6b97420a9f8c58268ca55755168e6e2f360
     GIT_SHALLOW    OFF
 )
 FetchContent_Populate(repo-deepgemm)
@@ -548,21 +548,10 @@ set(DEEPGEMM_SOURCES
     "${repo-deepgemm_SOURCE_DIR}/csrc/python_api.cpp"
 )
 
-Python_add_library(deep_gemm_cpp MODULE WITH_SOABI ${DEEPGEMM_SOURCES})
-
-# DeepGEMM's Python package (`deep_gemm/__init__.py`) expects the extension module name to be `_C`,
-# i.e. it does `from . import _C`. Make sure the built module is importable as `deep_gemm._C`.
-set_target_properties(deep_gemm_cpp PROPERTIES
-    OUTPUT_NAME "_C"
-    PREFIX ""
-)
+Python_add_library(deep_gemm_cpp MODULE USE_SABI ${SKBUILD_SABI_VERSION} WITH_SOABI ${DEEPGEMM_SOURCES})
 
 # Link against necessary libraries, including nvrtc for JIT compilation.
-find_library(TORCH_PYTHON_LIB
-    NAMES torch_python
-    PATHS "${TORCH_INSTALL_PREFIX}/lib" "${TORCH_INSTALL_PREFIX}/lib64"
-)
-target_link_libraries(deep_gemm_cpp PRIVATE ${TORCH_LIBRARIES} "${TORCH_PYTHON_LIB}" c10 cuda nvrtc mscclpp_static)
+target_link_libraries(deep_gemm_cpp PRIVATE ${TORCH_LIBRARIES} c10 cuda nvrtc mscclpp_static)
 
 # Add include directories needed by DeepGEMM.
 target_include_directories(deep_gemm_cpp PRIVATE

sgl-kernel/build.sh

@@ -46,7 +46,6 @@ echo "PYTHON_TAG:     ${PY_TAG}"
 echo "Output:         ${DIST_DIR}/"
 echo "Buildx cache:   ${BUILDX_CACHE_DIR}"
 echo "Builder:        ${BUILDER_NAME}"
-echo "USE_CCACHE:     ${USE_CCACHE}"
 echo "----------------------------------------"
 
 # Optional profiling build-args (empty string disables)
@@ -62,7 +61,6 @@ docker buildx build \
   --build-arg ARCH="${ARCH}" \
   --build-arg PYTHON_VERSION="${PYTHON_VERSION}" \
   --build-arg PYTHON_TAG="${PY_TAG}" \
-  --build-arg USE_CCACHE="${USE_CCACHE}" \
   "${BUILD_ARGS[@]}" \
   --cache-from type=local,src=${BUILDX_CACHE_DIR} \
   --cache-to type=local,dest=${BUILDX_CACHE_DIR},mode=max \

sgl-kernel/csrc/elementwise/concat_mla.cu

@@ -215,4 +215,3 @@ void concat_mla_absorb_q(at::Tensor a, at::Tensor b, at::Tensor out) {
   cudaError_t err = cudaGetLastError();
   TORCH_CHECK(err == cudaSuccess, "CUDA kernel launch failed: ", cudaGetErrorString(err));
 }
-// For ci build