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

Performance regression fix

Browse Source
This commit is contained in:
Eugene Rakhmatulin
2026-05-08 13:40:55 -07:00
parent 29d5904b80
commit bca64f9a53
1 changed files with 12 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
Dockerfile
View File

@@ -233,18 +233,23 @@ RUN curl -fsL https://patch-diff.githubusercontent.com/raw/vllm-project/vllm/pul
# TEMPORARY PATCH: revert vLLM PR #41524 / commit c51df430, # TEMPORARY PATCH: revert vLLM PR #41524 / commit c51df430,
# which disables FlashInfer autotune and regresses DGX Spark throughput. # which disables FlashInfer autotune and regresses DGX Spark throughput.
RUN set -eux; \ RUN RUN set -eux; \
patch_commit="c51df43005726a09c6eb7348e8c1b00501c70a8e"; \ patch_commit="c51df43005726a09c6eb7348e8c1b00501c70a8e"; \
target="vllm/config/vllm.py"; \ target="vllm/config/vllm.py"; \
if grep -q '"enable_flashinfer_autotune": False' "$target"; then \ marker="https://github.com/flashinfer-ai/flashinfer/issues/3197"; \
echo "FlashInfer autotune default is disabled; reverting ${patch_commit}"; \ if grep -q "$marker" "$target"; then \
git revert --no-commit "$patch_commit"; \ echo "PR #41524 regression found; reverting ${patch_commit}"; \
if grep -q '"enable_flashinfer_autotune": False' "$target"; then \ if ! git revert --no-commit "$patch_commit"; then \
echo "ERROR: revert applied but FlashInfer autotune still appears disabled"; \ git revert --abort 2>/dev/null || true; \
echo "ERROR: PR #41524 appears present but could not be reverted"; \
exit 1; \
fi; \
if grep -q "$marker" "$target"; then \
echo "ERROR: revert completed but PR #41524 marker is still present"; \
exit 1; \ exit 1; \
fi; \ fi; \
else \ else \
echo "FlashInfer autotune regression not present; skipping revert"; \ echo "PR #41524 regression marker not present; skipping revert"; \
fi fi
# Prepare build requirements # Prepare build requirements