[ROCm][CI][Bugfix] Disable Flash/MemEfficient SDP on ROCm to avoid HF...

[ROCm][CI][Bugfix] Disable Flash/MemEfficient SDP on ROCm to avoid HF Transformers accuracy issues (#29909)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>

docker/Dockerfile.rocm

@@ -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

requirements/rocm-test.txt

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

tests/models/multimodal/generation/conftest.py

+# 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)