biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[ROCm][CI][Bugfix] Disable Flash/MemEfficient SDP on ROCm to avoid HF Transformers accuracy issues (#29909)

Browse Source 
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Andreas Karatzas
2025-12-02 20:36:49 -06:00
committed by GitHub
parent 4dd7978374
commit 506ed87e87
3 changed files with 22 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
docker/Dockerfile.rocm
View File

@@ -65,7 +65,6 @@ COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/tests /tests
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/examples /examples
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/docker/Dockerfile.rocm /docker/
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/.buildkite /.buildkite
# Centralized v1 package - copied to both test and final stages
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/vllm/v1 /vllm_v1
# -----------------------
@@ -98,7 +97,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --system hf_transfer
ENV HF_HUB_ENABLE_HF_TRANSFER=1
# Copy in the v1 package
# Copy in the v1 package (for python-only install test group)
COPY --from=export_vllm /vllm_v1 /usr/local/lib/python${PYTHON_VERSION}/dist-packages/vllm/v1
# Source code is used in the `python_only_compile.sh` test
@@ -130,9 +129,6 @@ RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
&& pip uninstall -y vllm \
&& uv pip install --system *.whl
# Copy in the v1 package
COPY --from=export_vllm /vllm_v1 /usr/local/lib/python${PYTHON_VERSION}/dist-packages/vllm/v1
ARG COMMON_WORKDIR
# Copy over the benchmark scripts as well