[ROCm][CI] Disable skinny GEMMs in multimodal tests to fix non-deterministic results (vllm-project#35049)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>

Author: AndreasKaratzas

tests/models/multimodal/conftest.py

@@ -2,13 +2,31 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Pytest configuration for vLLM multimodal tests."""
 
+import os
 import warnings
 
 import torch
 
 from vllm.platforms import current_platform
 
 
+def pytest_configure(config):
+    """Early ROCm configuration that must happen before test collection."""
+    if not current_platform.is_rocm():
+        return
+
+    # Disable skinny GEMM on ROCm to avoid non-deterministic results
+    # from atomic reductions in wvSplitKrc kernel.
+    # See: https://github.com/vllm-project/vllm/pull/33493#issuecomment-3906083975
+    os.environ["VLLM_ROCM_USE_SKINNY_GEMM"] = "0"
+    warnings.warn(
+        "ROCm: Set VLLM_ROCM_USE_SKINNY_GEMM=0 to avoid non-deterministic "
+        "results from skinny GEMM atomic reductions",
+        UserWarning,
+        stacklevel=1,
+    )
+
+
 def pytest_collection_modifyitems(config, items):
     """Configure ROCm-specific settings based on collected tests."""
     if not current_platform.is_rocm():