A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 12, 03:30 UTC
[ROCm][CI][Bugfix] Disable Flash/MemEfficient SDP on ROCm to avoid HF Transformers accuracy issues (#29909)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Dec 2, 2025 at 20:36 UTC
506ed87e876e9dca3c64ac83e21051c02e9cb2e3
3 files changed
+22
-7
docker/Dockerfile.rocm
+1
-5
@@ -65,7 +65,6 @@ COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/tests /tests
65
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/examples /examples
66
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/docker/Dockerfile.rocm /docker/
67
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/.buildkite /.buildkite
68
-# Centralized v1 package - copied to both test and final stages
68
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/vllm/v1 /vllm_v1
69
70
# -----------------------
@@ -98,7 +97,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
97
uv pip install --system hf_transfer
98
ENV HF_HUB_ENABLE_HF_TRANSFER=1
99
101
-# Copy in the v1 package
100
+# Copy in the v1 package (for python-only install test group)
101
COPY --from=export_vllm /vllm_v1 /usr/local/lib/python${PYTHON_VERSION}/dist-packages/vllm/v1
102
103
# 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 \
129
&& pip uninstall -y vllm \
130
&& uv pip install --system *.whl
131
133
-# Copy in the v1 package
134
-COPY --from=export_vllm /vllm_v1 /usr/local/lib/python${PYTHON_VERSION}/dist-packages/vllm/v1
135
-
132
ARG COMMON_WORKDIR
133
134
# Copy over the benchmark scripts as well
requirements/rocm-test.txt
+2
-2
@@ -70,8 +70,8 @@ torchgeo==0.7.0
70
mteb==2.1.2
71
72
# Data processing
73
-xgrammar @ git+https://github.com/mlc-ai/xgrammar.git@eafd4db51b78acc64b3f0764ef27dfd206c28628
74
- # Test async scheduling
73
+xgrammar==0.1.27
74
+# Test async scheduling
75
76
# Utilities
77
num2words==0.5.14
tests/models/multimodal/generation/conftest.py
new
+19
@@ -0,0 +1,19 @@
1
+# SPDX-License-Identifier: Apache-2.0
2
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
+"""Pytest configuration for vLLM tests."""
4
+
5
+import torch
6
+
7
+from vllm.platforms import current_platform
8
+
9
+
10
+def pytest_configure(config):
11
+ """Disable Flash/MemEfficient SDP on ROCm to avoid HF
12
+ Transformers accuracy issues.
13
+ """
14
+ if not current_platform.is_rocm():
15
+ return
16
+
17
+ torch.backends.cuda.enable_flash_sdp(False)
18
+ torch.backends.cuda.enable_mem_efficient_sdp(False)
19
+ torch.backends.cuda.enable_math_sdp(True)