[Chore] Fix torch precision warning (#30428)

Signed-off-by: yewentao256 <zhyanwentao@126.com>

tests/v1/e2e/test_async_scheduling.py

@@ -152,8 +152,8 @@ def run_tests(
                 m.setenv("VLLM_ATTENTION_BACKEND", "ROCM_AITER_FA")
         else:
             m.setenv("VLLM_ATTENTION_BACKEND", "FLEX_ATTENTION")
-        # lock matmul precision to full FP32
-        m.setenv("VLLM_FLOAT32_MATMUL_PRECISION", "highest")
+        # lock matmul precision to full FP32 (IEEE)
+        m.setenv("VLLM_FLOAT32_MATMUL_PRECISION", "ieee")
         # m.setenv("VLLM_BATCH_INVARIANT", "1")
         outputs: list[tuple[str, list, list]] = []
         for n, (

vllm/envs.py

@@ -74,7 +74,7 @@ if TYPE_CHECKING:
     VLLM_MEDIA_CONNECTOR: str = "http"
     VLLM_TARGET_DEVICE: str = "cuda"
     VLLM_MAIN_CUDA_VERSION: str = "12.9"
-    VLLM_FLOAT32_MATMUL_PRECISION: Literal["highest", "high", "medium"] = "highest"
+    VLLM_FLOAT32_MATMUL_PRECISION: Literal["ieee", "tf32"] = "ieee"
     MAX_JOBS: str | None = None
     NVCC_THREADS: str | None = None
     VLLM_USE_PRECOMPILED: bool = False
@@ -456,11 +456,13 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_MAIN_CUDA_VERSION": lambda: os.getenv("VLLM_MAIN_CUDA_VERSION", "").lower()
     or "12.9",
     # Controls PyTorch float32 matmul precision mode within vLLM workers.
-    # Valid options mirror torch.set_float32_matmul_precision
+    # Accepted values:
+    #   - "ieee" (default): force full IEEE FP32 matmul precision.
+    #   - "tf32": enable TensorFloat32-based fast matmul.
     "VLLM_FLOAT32_MATMUL_PRECISION": env_with_choices(
         "VLLM_FLOAT32_MATMUL_PRECISION",
-        "highest",
-        ["highest", "high", "medium"],
+        "ieee",
+        ["ieee", "tf32"],
         case_sensitive=False,
     ),
     # Maximum number of compilation jobs to run in parallel.

vllm/v1/worker/gpu_worker.py

@@ -81,7 +81,7 @@ class Worker(WorkerBase):
 
         # configure float32 matmul precision according to vLLM env.
         precision = envs.VLLM_FLOAT32_MATMUL_PRECISION
-        torch.set_float32_matmul_precision(precision)
+        torch.backends.cuda.matmul.fp32_precision = precision
 
         if self.model_config.trust_remote_code:
             # note: lazy import to avoid importing torch before initializing