diff --git a/python/pyproject.toml b/python/pyproject.toml
index 7ec785be5..d5c2f3d6b 100755
--- a/python/pyproject.toml
+++ b/python/pyproject.toml
@@ -27,7 +27,7 @@ dependencies = [
   "datasets",
   "einops",
   "fastapi",
-  "flashinfer_python==0.6.11.post1", # keep it aligned with jit-cache version in Dockerfile
+  "flashinfer_python[cu13]==0.6.11.post1", # keep it aligned with jit-cache version in Dockerfile
   "flashinfer_cubin==0.6.11.post1",
   "gguf",
   "interegular",
@@ -37,7 +37,7 @@ dependencies = [
   "ninja",
   "easydict",  # Required by remote model code (e.g. DeepSeek-OCR) loaded via trust_remote_code; validated by transformers 5.4+ check_imports
   "numpy",
-  "nvidia-cutlass-dsl==4.5.0",
+  "nvidia-cutlass-dsl[cu13]==4.5.0",
   "nvidia-ml-py",
   "openai-harmony==0.0.4",
   "openai==2.6.1",
diff --git a/scripts/ci/cuda/ci_install_dependency.sh b/scripts/ci/cuda/ci_install_dependency.sh
index 398ba3993..0f2146d64 100755
--- a/scripts/ci/cuda/ci_install_dependency.sh
+++ b/scripts/ci/cuda/ci_install_dependency.sh
@@ -186,8 +186,8 @@ uninstall_stale_flashinfer() {
     # Keep flashinfer packages if version matches to avoid re-downloading:
     # - flashinfer-cubin: 150+ MB
     # - flashinfer-jit-cache: 1.2+ GB
-    FLASHINFER_PYTHON_REQUIRED=$(grep -Po -m1 '(?<=flashinfer_python==)[0-9A-Za-z\.\-]+' python/pyproject.toml || echo "")
-    FLASHINFER_CUBIN_REQUIRED=$(grep -Po -m1 '(?<=flashinfer_cubin==)[0-9A-Za-z\.\-]+' python/pyproject.toml || echo "")
+    FLASHINFER_PYTHON_REQUIRED=$(grep -Po -m1 'flashinfer_python(\[[^]]+\])?==\K[0-9A-Za-z\.\-]+' python/pyproject.toml || echo "")
+    FLASHINFER_CUBIN_REQUIRED=$(grep -Po -m1 'flashinfer_cubin(\[[^]]+\])?==\K[0-9A-Za-z\.\-]+' python/pyproject.toml || echo "")
     FLASHINFER_CUBIN_INSTALLED=$(pip show flashinfer-cubin 2>/dev/null | grep "^Version:" | awk '{print $2}' || echo "")
     FLASHINFER_JIT_INSTALLED=$(pip show flashinfer-jit-cache 2>/dev/null | grep "^Version:" | awk '{print $2}' | sed 's/+.*//' || echo "")
     FLASHINFER_JIT_CU_VERSION=$(pip show flashinfer-jit-cache 2>/dev/null | grep "^Version:" | awk '{print $2}' | sed -n 's/.*+//p' || echo "")
@@ -335,6 +335,22 @@ download_flashinfer_cache() {
     mark_step_done "${FUNCNAME[0]}"
 }
 
+purge_cutlass_libs_base() {
+    # nvidia-cutlass-dsl[cu13] extras are additive on PyPI: requires_dist always
+    # pulls -libs-base AND -libs-cu13 when [cu13] is requested. Both wheels write
+    # to the same site-packages paths with different content, leaving the wrapper
+    # (cutlass.py, cu13 style) mismatched with the binding (_gpu_ops_gen.py, base
+    # style) -> GPUModuleOp signature TypeError. See vllm-project/vllm#40082.
+    # Uninstall -libs-base, then force-reinstall -libs-cu13 so its files win.
+    $PIP_UNINSTALL_CMD nvidia-cutlass-dsl-libs-base $PIP_UNINSTALL_SUFFIX || true
+    CUTLASS_DSL_VERSION=$(grep -Po -m1 'nvidia-cutlass-dsl(\[[^]]+\])?==\K[0-9A-Za-z\.\-]+' python/pyproject.toml || echo "")
+    if [ -n "$CUTLASS_DSL_VERSION" ]; then
+        $PIP_CMD install --force-reinstall --no-deps "nvidia-cutlass-dsl-libs-cu13==${CUTLASS_DSL_VERSION}" $PIP_INSTALL_SUFFIX
+    fi
+
+    mark_step_done "${FUNCNAME[0]}"
+}
+
 stabilize_flashinfer_jit_paths() {
     # In venv mode, FlashInfer JIT writes build.ninja with hardcoded -isystem
     # paths. Per-job venvs get unique paths, but the JIT cache is shared on the
@@ -488,6 +504,7 @@ main() {
     install_sglang_kernel
     install_sglang_router
     download_flashinfer_cache
+    purge_cutlass_libs_base
     stabilize_flashinfer_jit_paths
     install_extra_deps
     install_test_tools