xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-06-06 11:02:16 +08:00
Code Issues 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
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 22 additions and 7 deletions
Show Stats 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/examples /examples
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/docker/Dockerfile.rocm /docker/ COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/docker/Dockerfile.rocm /docker/
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/.buildkite /.buildkite 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 COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/vllm/v1 /vllm_v1
# ----------------------- # -----------------------
@ -98,7 +97,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --system hf_transfer uv pip install --system hf_transfer
ENV HF_HUB_ENABLE_HF_TRANSFER=1 ENV HF_HUB_ENABLE_HF_TRANSFER=1
# Copy in the v1 package # Copy in the v1 package (for python-only install test group)
COPY --from=export_vllm /vllm_v1 /usr/local/lib/python${PYTHON_VERSION}/dist-packages/vllm/v1 COPY --from=export_vllm /vllm_v1 /usr/local/lib/python${PYTHON_VERSION}/dist-packages/vllm/v1
# Source code is used in the `python_only_compile.sh` test # Source code is used in the `python_only_compile.sh` test
@ -130,9 +129,6 @@ RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
&& pip uninstall -y vllm \ && pip uninstall -y vllm \
&& uv pip install --system *.whl && uv pip install --system *.whl
# Copy in the v1 package
COPY --from=export_vllm /vllm_v1 /usr/local/lib/python${PYTHON_VERSION}/dist-packages/vllm/v1
ARG COMMON_WORKDIR ARG COMMON_WORKDIR
# Copy over the benchmark scripts as well # Copy over the benchmark scripts as well

4
requirements/rocm-test.txt
View File

@ -70,8 +70,8 @@ torchgeo==0.7.0
mteb==2.1.2 mteb==2.1.2
# Data processing # Data processing
xgrammar @ git+https://github.com/mlc-ai/xgrammar.git@eafd4db51b78acc64b3f0764ef27dfd206c28628 xgrammar==0.1.27
# Test async scheduling # Test async scheduling
# Utilities # Utilities
num2words==0.5.14 num2words==0.5.14

19
tests/models/multimodal/generation/conftest.py Normal file
View File

@ -0,0 +1,19 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Pytest configuration for vLLM tests."""
import torch
from vllm.platforms import current_platform
def pytest_configure(config):
"""Disable Flash/MemEfficient SDP on ROCm to avoid HF
Transformers accuracy issues.
"""
if not current_platform.is_rocm():
return
torch.backends.cuda.enable_flash_sdp(False)
torch.backends.cuda.enable_mem_efficient_sdp(False)
torch.backends.cuda.enable_math_sdp(True)