| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """`LLMEngine`'s cleanup finalizer must not own the model. |
| 4 | |
| 5 | ``weakref.finalize`` keeps its ``args`` strongly in a class-level registry |
| 6 | until it fires, and it only fires once the ``LLMEngine`` itself is collected. |
| 7 | Handing it the model directly therefore pins the weights -- and the KV cache |
| 8 | aliased onto the model's layers -- for as long as the engine is reachable, |
| 9 | which in-process is past engine shutdown. |
| 10 | """ |
| 11 | |
| 12 | import gc |
| 13 | import weakref |
| 14 | |
| 15 | import pytest |
| 16 | import torch.nn as nn |
| 17 | |
| 18 | from vllm.v1.engine.llm_engine import LLMEngine |
| 19 | |
| 20 | pytestmark = pytest.mark.cpu_test |
| 21 | |
| 22 | |
| 23 | class _Model(nn.Module): |
| 24 | pass |
| 25 | |
| 26 | |
| 27 | def test_finalize_args_are_held_strongly(): |
| 28 | """The premise: this is why the model must be passed weakly.""" |
| 29 | |
| 30 | class Owner: |
| 31 | pass |
| 32 | |
| 33 | owner, model = Owner(), _Model() |
| 34 | weakref.finalize(owner, lambda _m: None, model) |
| 35 | ref = weakref.ref(model) |
| 36 | |
| 37 | del model |
| 38 | gc.collect() |
| 39 | |
| 40 | assert ref() is not None |
| 41 | del owner |
| 42 | gc.collect() |
| 43 | assert ref() is None |
| 44 | |
| 45 | |
| 46 | def test_cleanup_tolerates_a_dead_model(): |
| 47 | model = _Model() |
| 48 | ref = weakref.ref(model) |
| 49 | del model |
| 50 | gc.collect() |
| 51 | assert ref() is None |
| 52 | |
| 53 | LLMEngine._cleanup_instance_caches(ref) |
| 54 | |
| 55 | |
| 56 | def test_cleanup_still_walks_a_live_model(monkeypatch): |
| 57 | from vllm.compilation import wrapper |
| 58 | |
| 59 | cleaned = [] |
| 60 | |
| 61 | class _FakeWrapper(nn.Module): |
| 62 | def cleanup(self): |
| 63 | cleaned.append(self) |
| 64 | |
| 65 | monkeypatch.setattr( |
| 66 | wrapper, "TorchCompileWithNoGuardsWrapper", _FakeWrapper, raising=False |
| 67 | ) |
| 68 | |
| 69 | model = _Model() |
| 70 | model.inner = _FakeWrapper() |
| 71 | |
| 72 | LLMEngine._cleanup_instance_caches(weakref.ref(model)) |
| 73 | |
| 74 | assert cleaned == [model.inner] |