From b160cab156106b212ee67beb16c9725d366b8857 Mon Sep 17 00:00:00 2001 From: hy123 Date: Wed, 19 Aug 2026 19:18:10 +0800 Subject: [PATCH] [Bugfix] Restore model info caching for package backends (#52690) Signed-off-by: haoyang.qian Co-authored-by: OpenAI Codex --- tests/models/test_registry.py | 30 ++++++++++++++++++++++++++ vllm/model_executor/models/registry.py | 3 ++- 2 files changed, 32 insertions(+), 1 deletion(-) diff --git a/tests/models/test_registry.py b/tests/models/test_registry.py index 70b8b18f76..81d6b31215 100644 --- a/tests/models/test_registry.py +++ b/tests/models/test_registry.py @@ -173,6 +173,36 @@ def test_lazy_modelinfo_package_hash_includes_submodules(tmp_path): assert first_hash != second_hash +def test_lazy_modelinfo_package_attempts_cache_load(monkeypatch): + cached_model_info = object() + loaded_hashes = [] + + def fake_load_cache(self, module_hash): + loaded_hashes.append(module_hash) + return cached_model_info + + monkeypatch.setattr( + _LazyRegisteredModel, + "_load_modelinfo_from_cache", + fake_load_cache, + ) + monkeypatch.setattr( + "vllm.model_executor.models.registry._run_in_subprocess", + lambda _: pytest.fail("Package-backed model should use the cache path"), + ) + + registered_model = _LazyRegisteredModel( + module_name="vllm.model_executor.models.transformers", + class_name="TransformersForCausalLM", + ) + + result = registered_model.inspect_model_cls() + + assert result is cached_model_info + assert len(loaded_hashes) == 1 + assert loaded_hashes[0] + + def test_hf_registry_coverage(): untested_archs = ( ModelRegistry.get_supported_archs() - HF_EXAMPLE_MODELS.get_supported_archs() diff --git a/vllm/model_executor/models/registry.py b/vllm/model_executor/models/registry.py index 027dc10e6d..78e1dae556 100644 --- a/vllm/model_executor/models/registry.py +++ b/vllm/model_executor/models/registry.py @@ -1004,9 +1004,10 @@ class _LazyRegisteredModel(_BaseRegisteredModel): # hardware-isolated ``vllm.models.`` layout) live outside # ``vllm/model_executor/models``. Resolve the module spec directly # so the file-hash cache stays warm for them. + model_path: Path | None = None if self.module_name.startswith("vllm.model_executor.models."): model_path = Path(__file__).parent / f"{self.module_name.split('.')[-1]}.py" - else: + if model_path is None or not model_path.exists(): try: spec = importlib.util.find_spec(self.module_name) except (ImportError, ValueError):