From b4d755ac050094cb903521ee035b034790136cb4 Mon Sep 17 00:00:00 2001 From: zhuhaoran Date: Sat, 13 Dec 2025 18:09:36 +0800 Subject: [PATCH] fix inductor tf32 setting error Signed-off-by: zhuhaoran --- tests/v1/e2e/test_async_scheduling.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/tests/v1/e2e/test_async_scheduling.py b/tests/v1/e2e/test_async_scheduling.py index 307b6e66682f6..5d64a41fbc9a5 100644 --- a/tests/v1/e2e/test_async_scheduling.py +++ b/tests/v1/e2e/test_async_scheduling.py @@ -4,6 +4,7 @@ from itertools import repeat from typing import Any import pytest +import torch import torch._dynamo.config as dynamo_config from vllm import SamplingParams @@ -158,6 +159,7 @@ def run_tests( m.setenv("VLLM_ATTENTION_BACKEND", "FLEX_ATTENTION") # lock matmul precision to full FP32 (IEEE) m.setenv("VLLM_FLOAT32_MATMUL_PRECISION", "ieee") + torch.backends.cuda.matmul.allow_tf32 = False # m.setenv("VLLM_BATCH_INVARIANT", "1") outputs: list[tuple[str, list, list]] = [] for n, (