fix tests

Signed-off-by: youkaichao <[email protected]>

tests/kernels/test_attention_selector.py

@@ -4,31 +4,34 @@
 import torch
 
 from tests.kernels.utils import override_backend_env_variable
-from vllm.attention.selector import which_attn_to_use
+from tests.utils import fork_new_process_for_each_test
 from vllm.utils import STR_FLASH_ATTN_VAL, STR_INVALID_VAL
 
 
 @pytest.mark.parametrize(
     "name", ["TORCH_SDPA", "ROCM_FLASH", "XFORMERS", "FLASHINFER", "OPENVINO"])
 @pytest.mark.parametrize("device", ["cpu", "openvino", "rocm", "cuda"])
+@fork_new_process_for_each_test
 def test_env(name: str, device: str, monkeypatch):
     """Test that the attention selector can be set via environment variable.
     Note that we do not test FlashAttn because it is the default backend.
     """
-
     override_backend_env_variable(monkeypatch, name)
 
     monkeypatch.setenv("VLLM_TEST_FORCE_PLATFORM", device)
+    from vllm.attention.selector import which_attn_to_use
+
     backend = which_attn_to_use(16, torch.float16, torch.float16, 16, False)
     assert backend.name == name
 
 
+@fork_new_process_for_each_test
 def test_flash_attn(monkeypatch):
     """Test FlashAttn validation."""
     # TODO: When testing for v1, pipe in `use_v1` as an argument to
     # which_attn_to_use
-
     override_backend_env_variable(monkeypatch, STR_FLASH_ATTN_VAL)
+    from vllm.attention.selector import which_attn_to_use
 
     # Unsupported CUDA arch
     with patch("torch.cuda.get_device_capability", return_value=(7, 5)):
@@ -61,8 +64,11 @@ def test_flash_attn(monkeypatch):
     assert backend.name != STR_FLASH_ATTN_VAL
 
 
+@fork_new_process_for_each_test
 def test_invalid_env(monkeypatch):
     """Throw an exception if the backend name is invalid."""
     override_backend_env_variable(monkeypatch, STR_INVALID_VAL)
+    from vllm.attention.selector import which_attn_to_use
+
     with pytest.raises(ValueError):
         which_attn_to_use(16, torch.float16, None, 16, False)