[ROCm] Remove stale SDPA and skinny GEMM workarounds (#50907)

Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Andreas Karatzas
2026-08-11 14:19:16 -04:00
committed by GitHub
co-authored by OpenAI Codex
parent 4f2f31b82e
commit 1ab2801dde
56 changed files with 500 additions and 558 deletions
+2 -18
View File
@@ -2,8 +2,6 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Pytest configuration for vLLM language generation tests."""
import warnings
import torch
from vllm.platforms import current_platform
@@ -11,19 +9,5 @@ from vllm.platforms import current_platform
def pytest_sessionstart(session):
"""Configure ROCm-specific settings before test session starts."""
if not current_platform.is_rocm():
return
# Disable Flash/MemEfficient SDP on ROCm to avoid HF Transformers
# accuracy issues: https://github.com/vllm-project/vllm/issues/30167
# TODO: Remove once ROCm SDP accuracy issues are resolved on HuggingFace
torch.backends.cuda.enable_flash_sdp(False)
torch.backends.cuda.enable_mem_efficient_sdp(False)
torch.backends.cuda.enable_math_sdp(True)
torch.set_float32_matmul_precision("high")
warnings.warn(
"ROCm: Disabled flash_sdp and mem_efficient_sdp, enabled math_sdp "
"to avoid HuggingFace Transformers accuracy issues",
UserWarning,
stacklevel=1,
)
if current_platform.is_rocm():
torch.set_float32_matmul_precision("high")