software-engineering/spark-vllm-docker
2
0
Fork 0
Code Issues Actions Packages Projects Releases Wiki Activity

Fix performance regression

Browse Source
This commit is contained in:
Eugene Rakhmatulin
2026-05-08 12:56:28 -07:00
parent b87854fd4c
commit 29d5904b80
1 changed files with 16 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
Dockerfile
View File

@@ -231,6 +231,22 @@ RUN curl -fsL https://patch-diff.githubusercontent.com/raw/vllm-project/vllm/pul
fi \ fi \
&& rm pr35568.diff && rm pr35568.diff
# TEMPORARY PATCH: revert vLLM PR #41524 / commit c51df430,
# which disables FlashInfer autotune and regresses DGX Spark throughput.
RUN set -eux; \
patch_commit="c51df43005726a09c6eb7348e8c1b00501c70a8e"; \
target="vllm/config/vllm.py"; \
if grep -q '"enable_flashinfer_autotune": False' "$target"; then \
echo "FlashInfer autotune default is disabled; reverting ${patch_commit}"; \
git revert --no-commit "$patch_commit"; \
if grep -q '"enable_flashinfer_autotune": False' "$target"; then \
echo "ERROR: revert applied but FlashInfer autotune still appears disabled"; \
exit 1; \
fi; \
else \
echo "FlashInfer autotune regression not present; skipping revert"; \
fi
# Prepare build requirements # Prepare build requirements
RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \ RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
python3 use_existing_torch.py && \ python3 use_existing_torch.py && \