[Docker] Remove redundant flashinfer download-cubin step (#46517)
Signed-off-by: mgoin <mgoin64@gmail.com> Co-authored-by: Claude <noreply@anthropic.com>
Michael Goin committed
Jun 23, 2026 at 13:16 UTC
7c2e08451a474972ce5a409b88d937f9cc08a243
1 file changed
-7
docker/Dockerfile
-7
@@ -898,13 +898,6 @@ RUN --mount=type=cache,target=/opt/uv/cache \
898
fi; \
899
fi
900
901
-# Download FlashInfer precompiled cubins AFTER all pip installs are done.
902
-# This must run after the vLLM wheel and EP kernels installs above, because
903
-# those can reinstall/touch flashinfer packages. Downloading cubins earlier
904
-# (in the flashinfer-jit-cache layer) causes ~2.5 GB of layer duplication
905
-# when a later pip install overwrites flashinfer package files.
906
-RUN flashinfer show-config && flashinfer download-cubin
907
-
901
# CUDA image changed from /usr/local/nvidia to /usr/local/cuda in 12.8 but will
902
# return to /usr/local/nvidia in 13.0 to allow container providers to mount drivers
903
# consistently from the host (see https://github.com/vllm-project/vllm/issues/18859).