Merge tag 'v0.7.2' into v0.7.2-dev

tests/spec_decode/e2e/test_integration.py

+# SPDX-License-Identifier: Apache-2.0
 """Tests which cover integration of the speculative decoding framework with
 other features, e.g. cuda graphs.
 """

tests/spec_decode/e2e/test_integration_dist_tp2.py

+# SPDX-License-Identifier: Apache-2.0
 """Tests which cover integration of the speculative decoding framework with
 tensor parallelism.
 """

tests/spec_decode/e2e/test_integration_dist_tp4.py

+# SPDX-License-Identifier: Apache-2.0
 """Tests which cover integration of the speculative decoding framework with
 tensor parallelism.
 """

tests/spec_decode/e2e/test_logprobs.py

+# SPDX-License-Identifier: Apache-2.0
+
 from itertools import cycle
 
 import pytest

tests/spec_decode/e2e/test_medusa_correctness.py

+# SPDX-License-Identifier: Apache-2.0
 """This docstring details important information on the testing methodology.
 
 Most of the tests rely on "greedy equality", where we expect the output of

tests/spec_decode/e2e/test_mlp_correctness.py

+# SPDX-License-Identifier: Apache-2.0
 """This docstring details important information on the testing methodology.
 
 Most of the tests rely on "greedy equality", where we expect the output of

tests/spec_decode/e2e/test_multistep_correctness.py

+# SPDX-License-Identifier: Apache-2.0
 """The tests in this file verify end-to-end speculative decoding correctness.
 
 This docstring details important information on the testing methodology.

tests/spec_decode/e2e/test_ngram_correctness.py

+# SPDX-License-Identifier: Apache-2.0
 """This docstring details important information on the testing methodology.
 
 Most of the tests rely on "greedy equality", where we expect the output of

tests/spec_decode/e2e/test_seed.py

+# SPDX-License-Identifier: Apache-2.0
+
 import pytest
 import os
 

tests/spec_decode/test_batch_expansion.py

+# SPDX-License-Identifier: Apache-2.0
+
 from typing import List
 
 import pytest

tests/spec_decode/test_dynamic_spec_decode.py

+# SPDX-License-Identifier: Apache-2.0
+
 from unittest.mock import MagicMock, patch
 
 import pytest

tests/spec_decode/test_metrics.py

+# SPDX-License-Identifier: Apache-2.0
+
 import math
 from unittest.mock import MagicMock
 

tests/spec_decode/test_multi_step_worker.py

+# SPDX-License-Identifier: Apache-2.0
+
 import random
 from typing import Dict, List
 from unittest.mock import MagicMock

tests/spec_decode/test_ngram_worker.py

+# SPDX-License-Identifier: Apache-2.0
+
 import torch
 import os
 

tests/spec_decode/test_scorer.py

+# SPDX-License-Identifier: Apache-2.0
+
 import random
 from typing import List
 

tests/spec_decode/test_spec_decode_worker.py

+# SPDX-License-Identifier: Apache-2.0
+
 import random
 from collections import defaultdict
 from types import SimpleNamespace

tests/spec_decode/test_utils.py

+# SPDX-License-Identifier: Apache-2.0
+
 from unittest.mock import MagicMock
 
 import pytest

tests/spec_decode/utils.py

+# SPDX-License-Identifier: Apache-2.0
+
 from itertools import count
 from typing import Callable, Dict, List, Optional
 from typing import Sequence as GenericSequence

tests/standalone_tests/lazy_torch_compile.py → tests/standalone_tests/lazy_imports.py

+# SPDX-License-Identifier: Apache-2.0
+
 # Description: Test the lazy import module
 # The utility function cannot be placed in `vllm.utils`
 # this needs to be a standalone script
@@ -6,7 +8,17 @@ from contextlib import nullcontext
 
 from vllm_test_utils import BlameResult, blame
 
-module_name = "torch._inductor.async_compile"
+# List of modules that should not be imported too early.
+# Lazy import `torch._inductor.async_compile` to avoid creating
+# too many processes before we set the number of compiler threads.
+# Lazy import `cv2` to avoid bothering users who only use text models.
+# `cv2` can easily mess up the environment.
+module_names = ["torch._inductor.async_compile", "cv2"]
+
+
+def any_module_imported():
+    return any(module_name in sys.modules for module_name in module_names)
+
 
 # In CI, we only check finally if the module is imported.
 # If it is indeed imported, we can rerun the test with `use_blame=True`,
@@ -14,8 +26,7 @@ module_name = "torch._inductor.async_compile"
 # and help find the root cause.
 # We don't run it in CI by default because it is slow.
 use_blame = False
-context = blame(
-    lambda: module_name in sys.modules) if use_blame else nullcontext()
+context = blame(any_module_imported) if use_blame else nullcontext()
 with context as result:
     import vllm  # noqa
 
@@ -23,6 +34,6 @@ if use_blame:
     assert isinstance(result, BlameResult)
     print(f"the first import location is:\n{result.trace_stack}")
 
-assert module_name not in sys.modules, (
-    f"Module {module_name} is imported. To see the first"
+assert not any_module_imported(), (
+    f"Some the modules in {module_names} are imported. To see the first"
     f" import location, run the test with `use_blame=True`.")

tests/tensorizer_loader/conftest.py

+# SPDX-License-Identifier: Apache-2.0
+
 import functools
 import gc
 from typing import Callable, TypeVar