-
Bug
-
Resolution: Done
-
Critical
-
None
-
None
-
False
-
-
False
-
-
-
AIPCC Accelerators 9, AIPCC Accelerators 10
vLLM v0.9.1 defaults on V0 which has pretty bad accuracy regressions on ROCm.
https://github.com/neuralmagic/nm-cicd/actions/runs/15590090771
Setting VLLM_USE_V1=1 to force using the V1 engine solves the issue