fix ci by removing nvidia-cutlass-dsl-libs-base and force reinstall n… (#20380)

docker/Dockerfile

@@ -321,6 +321,11 @@ RUN --mount=type=cache,target=/root/.cache/pip \
     && rm -rf /root/.cargo /root/.rustup target dist ~/.cargo \
     && sed -i '/\.cargo\/env/d' /root/.profile /root/.bashrc 2>/dev/null || true
 
+# Pin nvidia-cutlass-dsl to 4.3.5
+RUN --mount=type=cache,target=/root/.cache/pip \
+   python3 -m pip uninstall nvidia-cutlass-dsl-libs-base; \
+   python3 -m pip install nvidia-cutlass-dsl==4.3.5 --force-reinstall;
+
 # Patching packages for CUDA 12/13 compatibility
 # TODO: Remove when torch version covers these packages
 RUN --mount=type=cache,target=/root/.cache/pip if [ "${CUDA_VERSION%%.*}" = "12" ]; then \

scripts/ci/cuda/ci_install_dependency.sh

@@ -330,10 +330,12 @@ bash "${SCRIPT_DIR}/ci_download_flashinfer_cubin.sh"
 
 # Clean nvidia-cutlass-dsl-libs-base for cutedsl lower than 0.4.4
 $PIP_UNINSTALL_CMD nvidia-cutlass-dsl-libs-base $PIP_UNINSTALL_SUFFIX || true
+$PIP_CMD install nvidia-cutlass-dsl==4.3.5 --force-reinstall $PIP_INSTALL_SUFFIX || true
 
 # Show current packages
 $PIP_CMD list
 python3 -c "import torch; print(torch.version.cuda)"
+python3 -c "import cutlass; import cutlass.cute;"
 
 # Prepare the CI runner (cleanup HuggingFace cache, etc.)
 bash "${SCRIPT_DIR}/prepare_runner.sh"