fix

Signed-off-by: baonudesifeizhai <baonudesifeizhai@gmail.com>

tests/compile/fullgraph/test_full_graph.py

@@ -13,7 +13,10 @@ from vllm import LLM, SamplingParams
 from vllm.attention.backends.registry import AttentionBackendEnum
 from vllm.config import CompilationConfig, CompilationMode, CUDAGraphMode, PassConfig
 from vllm.platforms import current_platform
-from vllm.utils.torch_utils import is_torch_equal_or_newer
+from vllm.utils.torch_utils import (
+    is_torch_equal,
+    is_torch_equal_or_newer,
+)
 
 from ...utils import create_new_process_for_each_test
 
@@ -156,6 +159,20 @@ def test_full_graph(
         )
         for model_info in models_list(all=False)
         if is_torch_equal_or_newer("2.9.0.dev")
+    ]
+    + [
+        # Test get_raw_stream patch with compile_sizes
+        # This tests that TorchInductor autotune works correctly with get_raw_stream
+        # patch in torch 2.9 and without patch in torch 2.10+
+        (
+            CompilationConfig(
+                mode=CompilationMode.VLLM_COMPILE,
+                compile_sizes=[1, 2],  # Triggers autotune which uses get_raw_stream
+                cudagraph_mode=CUDAGraphMode.NONE,
+            ),
+            "facebook/opt-125m",
+            {},
+        ),
     ],
 )
 # only test some of the models
@@ -178,6 +195,26 @@ def test_custom_compile_config(
     ):
         pytest.skip("inductor graph partition is only available in PyTorch 2.9+")
 
+    # For get_raw_stream patch test: verify version and log context
+    if compilation_config.compile_sizes and len(compilation_config.compile_sizes) > 1:
+        # Verify torch version >= 2.9.0 for compile_sizes autotune
+        if not is_torch_equal_or_newer("2.9.0"):
+            pytest.skip(
+                f"compile_sizes autotune requires torch >= 2.9.0, "
+                f"got {torch.__version__}"
+            )
+
+        # Log version context for get_raw_stream patch testing
+        is_torch_2_9 = is_torch_equal("2.9.0") or is_torch_equal("2.9.1")
+        version_context = (
+            "2.9.x (get_raw_stream patch applied)"
+            if is_torch_2_9
+            else "2.10+ (get_raw_stream patch not needed)"
+        )
+        print(
+            f"Testing compile_sizes with torch {torch.__version__} ({version_context})"
+        )
+
     print(f"MODEL={model}")
     run_model(compilation_config, model, **model_kwargs)
 

tests/compile/test_config.py

@@ -16,7 +16,6 @@ from vllm.platforms import current_platform
 from vllm.utils.torch_utils import (
     _is_torch_equal_or_newer,
     is_torch_equal,
-    is_torch_equal_or_newer,
 )
 
 # This import automatically registers `torch.ops.silly.attention`
@@ -55,55 +54,6 @@ def test_get_raw_stream_patch():
         assert get_raw_stream is _cuda_getCurrentRawStream
 
 
-# forked needed to workaround https://github.com/vllm-project/vllm/issues/21073
-@pytest.mark.forked
-@pytest.mark.skipif(
-    not current_platform.is_cuda(),
-    reason="compile_sizes with autotune requires CUDA",
-)
-def test_get_raw_stream_patch_e2e(vllm_runner, monkeypatch):
-    """
-    E2E test to verify get_raw_stream patch works correctly with compile_sizes.
-    This test should run in both torch 2.9 and 2.10 to ensure no regression.
-    When compile_sizes > 1, TorchInductor autotune uses get_raw_stream().
-    In torch 2.9.0/2.9.1, this function needs to be patched,
-    but in 2.10+ it should work without patch.
-    """
-    import torch
-
-    # Verify torch version >= 2.9.0
-    if not is_torch_equal_or_newer("2.9.0"):
-        pytest.skip(f"Test requires torch >= 2.9.0, got {torch.__version__}")
-
-    # Determine version context for logging
-    is_torch_2_9 = is_torch_equal("2.9.0") or is_torch_equal("2.9.1")
-    version_context = (
-        "2.9.x (patch applied)" if is_torch_2_9 else "2.10+ (patch not needed)"
-    )
-    print(f"Running test with torch {torch.__version__} ({version_context})")
-
-    monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
-
-    compilation_config = {
-        "mode": CompilationMode.VLLM_COMPILE,
-        "compile_sizes": [1, 2],  # Triggers autotune which uses get_raw_stream
-        "cudagraph_mode": CUDAGraphMode.NONE,
-    }
-
-    with vllm_runner(
-        "facebook/opt-125m",
-        compilation_config=compilation_config,
-        gpu_memory_utilization=0.4,
-    ) as llm:
-        from vllm import SamplingParams
-
-        outputs = llm.generate(
-            ["Hello, my name is"], SamplingParams(temperature=0, max_tokens=5)
-        )
-        assert len(outputs) == 1
-        assert len(outputs[0].outputs) > 0
-
-
 def test_copy_pass():
     vllm_config = VllmConfig()
     inductor_pass = FixFunctionalizationPass(vllm_config)