Sync from v0.13

tests/config/test_config.yaml

@@ -0,0 +1,4 @@
+port: 12312
+served_model_name: mymodel
+tensor_parallel_size: 2
+trust_remote_code: true

tests/config/test_config_generation.py

@@ -0,0 +1,75 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import pytest
+
+from vllm.engine.arg_utils import EngineArgs
+from vllm.model_executor.layers.quantization.quark.utils import deep_compare
+
+
+def test_cuda_empty_vs_unset_configs(monkeypatch: pytest.MonkeyPatch):
+    """Test that configs created with normal (untouched) CUDA_VISIBLE_DEVICES
+    and CUDA_VISIBLE_DEVICES="" are equivalent. This ensures consistent
+    behavior regardless of whether GPU visibility is disabled via empty string
+    or left in its normal state.
+    """
+
+    def create_config():
+        engine_args = EngineArgs(
+            model="deepseek-ai/DeepSeek-V2-Lite", trust_remote_code=True
+        )
+        return engine_args.create_engine_config()
+
+    # Create config with CUDA_VISIBLE_DEVICES set normally
+    normal_config = create_config()
+
+    # Create config with CUDA_VISIBLE_DEVICES=""
+    with monkeypatch.context() as m:
+        m.setenv("CUDA_VISIBLE_DEVICES", "")
+        empty_config = create_config()
+
+    normal_config_dict = vars(normal_config)
+    empty_config_dict = vars(empty_config)
+
+    # Remove instance_id before comparison as it's expected to be different
+    normal_config_dict.pop("instance_id", None)
+    empty_config_dict.pop("instance_id", None)
+
+    assert deep_compare(normal_config_dict, empty_config_dict), (
+        'Configs with normal CUDA_VISIBLE_DEVICES and CUDA_VISIBLE_DEVICES=""'
+        " should be equivalent"
+    )
+
+
+def test_ray_runtime_env(monkeypatch: pytest.MonkeyPatch):
+    # In testing, this method needs to be nested inside as ray does not
+    # see the test module.
+    def create_config():
+        engine_args = EngineArgs(
+            model="deepseek-ai/DeepSeek-V2-Lite", trust_remote_code=True
+        )
+        return engine_args.create_engine_config()
+
+    config = create_config()
+    parallel_config = config.parallel_config
+    assert parallel_config.ray_runtime_env is None
+
+    import ray
+
+    ray.init()
+
+    runtime_env = {
+        "env_vars": {
+            "TEST_ENV_VAR": "test_value",
+        },
+    }
+
+    config_ref = ray.remote(create_config).options(runtime_env=runtime_env).remote()
+
+    config = ray.get(config_ref)
+    parallel_config = config.parallel_config
+    assert parallel_config.ray_runtime_env is not None
+    assert (
+        parallel_config.ray_runtime_env.env_vars().get("TEST_ENV_VAR") == "test_value"
+    )
+
+    ray.shutdown()

tests/config/test_config_utils.py

@@ -0,0 +1,166 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from dataclasses import dataclass
+from enum import Enum
+
+import pytest
+
+from vllm.config.utils import get_hash_factors, hash_factors, normalize_value
+
+# Helpers
+
+
+def endswith_fqname(obj, suffix: str) -> bool:
+    # normalize_value(type) returns fully-qualified name
+    # Compare suffix to avoid brittle import paths.
+    out = normalize_value(obj)
+    return isinstance(out, str) and out.endswith(suffix)
+
+
+def expected_path(p_str: str = ".") -> str:
+    import pathlib
+
+    p = pathlib.Path(p_str)
+    return p.expanduser().resolve().as_posix()
+
+
+# Minimal dataclass to test get_hash_factors.
+# Avoid importing heavy vLLM configs.
+@dataclass
+class SimpleConfig:
+    a: object
+    b: object | None = None
+
+
+class DummyLogprobsMode(Enum):
+    RAW_LOGITS = "raw_logits"
+
+
+def test_hash_factors_deterministic():
+    """Test that hash_factors produces consistent SHA-256 hashes"""
+    factors = {"a": 1, "b": "test"}
+    hash1 = hash_factors(factors)
+    hash2 = hash_factors(factors)
+
+    assert hash1 == hash2
+    # Dict key insertion order should not affect the hash.
+    factors_reordered = {"b": "test", "a": 1}
+    assert hash_factors(factors_reordered) == hash1
+    assert len(hash1) == 64
+    assert all(c in "0123456789abcdef" for c in hash1)
+
+
+@pytest.mark.parametrize(
+    "inp, expected",
+    [
+        (None, None),
+        (True, True),
+        (1, 1),
+        (1.0, 1.0),
+        ("x", "x"),
+        (b"ab", "6162"),
+        (bytearray(b"ab"), "6162"),
+        ([1, 2], (1, 2)),
+        ({"b": 2, "a": 1}, (("a", 1), ("b", 2))),
+    ],
+)
+def test_normalize_value_matrix(inp, expected):
+    """Parametric input→expected normalization table."""
+    assert normalize_value(inp) == expected
+
+
+def test_normalize_value_enum():
+    # Enums normalize to (module.QualName, value).
+    # DummyLogprobsMode uses a string payload.
+    out = normalize_value(DummyLogprobsMode.RAW_LOGITS)
+    assert isinstance(out, tuple)
+    assert out[0].endswith("DummyLogprobsMode")
+    # Expect string payload 'raw_logits'.
+    assert out[1] == "raw_logits"
+
+
+def test_normalize_value_set_order_insensitive():
+    # Sets are unordered; normalize_value sorts elements for determinism.
+    assert normalize_value({3, 1, 2}) == normalize_value({1, 2, 3})
+
+
+def test_normalize_value_path_normalization():
+    from pathlib import Path  # local import to avoid global dependency
+
+    # Paths expand/resolve to absolute strings.
+    # Stabilizes hashing across working dirs.
+    assert normalize_value(Path(".")) == expected_path(".")
+
+
+def test_normalize_value_uuid_and_to_json():
+    # Objects may normalize via uuid() or to_json_string().
+    class HasUUID:
+        def uuid(self):
+            return "test-uuid"
+
+    class ToJson:
+        def to_json_string(self):
+            return '{"x":1}'
+
+    assert normalize_value(HasUUID()) == "test-uuid"
+    assert normalize_value(ToJson()) == '{"x":1}'
+
+
+@pytest.mark.parametrize(
+    "bad",
+    [
+        (lambda x: x),
+        (type("CallableInstance", (), {"__call__": lambda self: 0}))(),
+        (lambda: (lambda: 0))(),  # nested function instance
+    ],
+)
+def test_error_cases(bad):
+    """Inputs expected to raise TypeError."""
+    # Reject functions/lambdas/callable instances
+    # to avoid under-hashing.
+    with pytest.raises(TypeError):
+        normalize_value(bad)
+
+
+def test_enum_vs_int_disambiguation():
+    # int stays primitive
+    nf_int = normalize_value(1)
+    assert nf_int == 1
+
+    # enum becomes ("module.QualName", value)
+    nf_enum = normalize_value(DummyLogprobsMode.RAW_LOGITS)
+    assert isinstance(nf_enum, tuple) and len(nf_enum) == 2
+    enum_type, enum_val = nf_enum
+    assert enum_type.endswith(".DummyLogprobsMode")
+    assert enum_val == "raw_logits"
+
+    # Build factor dicts from configs with int vs enum
+    f_int = get_hash_factors(SimpleConfig(1), set())
+    f_enum = get_hash_factors(SimpleConfig(DummyLogprobsMode.RAW_LOGITS), set())
+    # The int case remains a primitive value
+    assert f_int["a"] == 1
+    # The enum case becomes a tagged tuple ("module.QualName", "raw_logits")
+    assert isinstance(f_enum["a"], tuple) and f_enum["a"][1] == "raw_logits"
+    # Factor dicts must differ so we don't collide primitives with Enums.
+    assert f_int != f_enum
+    # Hash digests must differ correspondingly
+    assert hash_factors(f_int) != hash_factors(f_enum)
+
+    # Hash functions produce stable hex strings
+    h_int = hash_factors(f_int)
+    h_enum = hash_factors(f_enum)
+    assert isinstance(h_int, str) and len(h_int) == 64
+    assert isinstance(h_enum, str) and len(h_enum) == 64
+
+
+def test_classes_are_types():
+    """Types normalize to FQNs; include real vLLM types."""
+    # Only classes allowed; functions/lambdas are rejected.
+    # Canonical form is the fully-qualified name.
+    assert isinstance(normalize_value(str), str)
+
+    class LocalDummy:
+        pass
+
+    assert endswith_fqname(LocalDummy, ".LocalDummy")

tests/config/test_config_with_model.yaml

@@ -0,0 +1,6 @@
+# Same as test_config.yaml but with model specified
+model: config-model
+port: 12312
+served_model_name: mymodel
+tensor_parallel_size: 2
+trust_remote_code: true

tests/config/test_mp_reducer.py

@@ -0,0 +1,53 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import sys
+from unittest.mock import patch
+
+from vllm.config import VllmConfig
+from vllm.engine.arg_utils import AsyncEngineArgs
+from vllm.v1.engine.async_llm import AsyncLLM
+
+
+def test_mp_reducer():
+    """
+    Test that _reduce_config reducer is registered when AsyncLLM is instantiated
+    without transformers_modules. This is a regression test for
+    https://github.com/vllm-project/vllm/pull/18640.
+    """
+
+    # Ensure transformers_modules is not in sys.modules
+    if "transformers_modules" in sys.modules:
+        del sys.modules["transformers_modules"]
+
+    with patch("multiprocessing.reducer.register") as mock_register:
+        engine_args = AsyncEngineArgs(
+            model="facebook/opt-125m",
+            max_model_len=32,
+            gpu_memory_utilization=0.1,
+            disable_log_stats=True,
+        )
+
+        async_llm = AsyncLLM.from_engine_args(
+            engine_args,
+            start_engine_loop=False,
+        )
+
+        assert mock_register.called, (
+            "multiprocessing.reducer.register should have been called"
+        )
+
+        vllm_config_registered = False
+        for call_args in mock_register.call_args_list:
+            # Verify that a reducer for VllmConfig was registered
+            if len(call_args[0]) >= 2 and call_args[0][0] == VllmConfig:
+                vllm_config_registered = True
+
+                reducer_func = call_args[0][1]
+                assert callable(reducer_func), "Reducer function should be callable"
+                break
+
+        assert vllm_config_registered, (
+            "VllmConfig should have been registered to multiprocessing.reducer"
+        )
+
+        async_llm.shutdown()

tests/config/test_multimodal_config.py

@@ -0,0 +1,25 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pytest
+
+from vllm.attention.backends.registry import AttentionBackendEnum
+from vllm.config.multimodal import MultiModalConfig
+
+
+def test_mm_encoder_attn_backend_str_conversion():
+    config = MultiModalConfig(mm_encoder_attn_backend="FLASH_ATTN")
+    assert config.mm_encoder_attn_backend == AttentionBackendEnum.FLASH_ATTN
+
+
+def test_mm_encoder_attn_backend_invalid():
+    with pytest.raises(ValueError):
+        MultiModalConfig(mm_encoder_attn_backend="not_a_backend")
+
+
+def test_mm_encoder_attn_backend_hash_updates():
+    base_hash = MultiModalConfig().compute_hash()
+    overridden_hash = MultiModalConfig(
+        mm_encoder_attn_backend=AttentionBackendEnum.FLASH_ATTN
+    ).compute_hash()
+    assert base_hash != overridden_hash