[CI] Force-reinstall nvidia-cutlass-dsl-libs-cu13 last to avoid wheel-mix TypeError (#25958)

Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>

python/pyproject.toml

@@ -38,7 +38,7 @@ 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[cu13]==4.5.0",
+  "nvidia-cutlass-dsl[cu13]==4.5.1",
   "nvidia-ml-py",
   "openai-harmony==0.0.4",
   "openai==2.6.1",

scripts/ci/cuda/ci_install_dependency.sh

@@ -335,6 +335,34 @@ download_flashinfer_cache() {
     mark_step_done "${FUNCNAME[0]}"
 }
 
+force_reinstall_cutlass_dsl_libs_cu13() {
+    # nvidia-cutlass-dsl[cu13] has additive PyPI extras: installing it pulls in
+    # both -libs-base and -libs-cu13. The two wheels ship intentionally-different
+    # content for the same paths (cutlass/_mlir/dialects/_gpu_ops_gen.py and
+    # cutlass/_mlir/_mlir_libs/_cutlass_ir.cpython-*.so) -- each Python wrapper
+    # is paired with a matching pybind11 .so. If install order leaves the .py
+    # from one wheel and the .so from the other, GPUModuleOp.__init__ raises
+    # TypeError: incompatible function arguments at kernel-compile time.
+    #
+    # Force-reinstall -libs-cu13 LAST so both files come from the same wheel
+    # (BOTH-cu13 state), eliminating the mismatch. The version is parsed from
+    # pyproject.toml so this stays in sync with whatever nvidia-cutlass-dsl
+    # version the project pins.
+    if [ "$CU_MAJOR" != "13" ]; then
+        return
+    fi
+
+    CUTLASS_DSL_VERSION=$(grep -Po -m1 'nvidia-cutlass-dsl(\[[^]]+\])?==\K[0-9A-Za-z\.\-]+' "${REPO_ROOT}/python/pyproject.toml" || echo "")
+    if [ -z "$CUTLASS_DSL_VERSION" ]; then
+        echo "WARNING: could not detect nvidia-cutlass-dsl version from pyproject.toml; skipping libs-cu13 force-reinstall"
+        return
+    fi
+
+    $PIP_CMD install --force-reinstall --no-deps "nvidia-cutlass-dsl-libs-cu13==${CUTLASS_DSL_VERSION}" $PIP_INSTALL_SUFFIX
+
+    mark_step_done "${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 +516,7 @@ main() {
     install_sglang_kernel
     install_sglang_router
     download_flashinfer_cache
+    force_reinstall_cutlass_dsl_libs_cu13
     stabilize_flashinfer_jit_paths
     install_extra_deps
     install_test_tools