diff --git a/sgl-kernel/CMakeLists.txt b/sgl-kernel/CMakeLists.txt
index e41cf6f0a..880428161 100644
--- a/sgl-kernel/CMakeLists.txt
+++ b/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        54f99a8af537b3c6eb4819b69907ccbe2b600792
+    GIT_TAG        fc08e5775a5db960690a41248655168c55bef228
     GIT_SHALLOW    OFF
 )
 FetchContent_Populate(repo-deepgemm)
@@ -548,10 +548,21 @@ set(DEEPGEMM_SOURCES
     "${repo-deepgemm_SOURCE_DIR}/csrc/python_api.cpp"
 )
 
-Python_add_library(deep_gemm_cpp MODULE USE_SABI ${SKBUILD_SABI_VERSION} WITH_SOABI ${DEEPGEMM_SOURCES})
+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 ""
+)
 
 # Link against necessary libraries, including nvrtc for JIT compilation.
-target_link_libraries(deep_gemm_cpp PRIVATE ${TORCH_LIBRARIES} c10 cuda nvrtc mscclpp_static)
+find_library(TORCH_PYTHON_LIB
+    NAMES torch_python
+    PATHS "${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)
 
 # Add include directories needed by DeepGEMM.
 target_include_directories(deep_gemm_cpp PRIVATE
diff --git a/sgl-kernel/build.sh b/sgl-kernel/build.sh
index b29d4d64a..d8ad84965 100755
--- a/sgl-kernel/build.sh
+++ b/sgl-kernel/build.sh
@@ -46,6 +46,7 @@ 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)
@@ -61,6 +62,7 @@ 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 \
diff --git a/sgl-kernel/csrc/elementwise/concat_mla.cu b/sgl-kernel/csrc/elementwise/concat_mla.cu
index 7d5b8595c..32144aa8a 100644
--- a/sgl-kernel/csrc/elementwise/concat_mla.cu
+++ b/sgl-kernel/csrc/elementwise/concat_mla.cu
@@ -215,3 +215,4 @@ 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