[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
-6
View File
@@ -131,12 +131,6 @@ else:
VLLM_PATH = Path(__file__).parent.parent
"""Path to root of the vLLM repository."""
# ROCm: disable skinny GEMM to avoid non-deterministic results from
# atomic reductions in wvSplitKrc kernel.
# See: https://github.com/vllm-project/vllm/pull/33493#issuecomment-3906083975
ROCM_ENV_OVERRIDES = (
{"VLLM_ROCM_USE_SKINNY_GEMM": "0"} if current_platform.is_rocm() else {}
)
# ROCm: disable prefix caching and eliminate batch variance to reduce
# test flakiness.
ROCM_EXTRA_ARGS = (