| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """The `get_language_model` cache must not own the models it keys on. |
| 4 | |
| 5 | A strong-keyed module-level dict pins every model ever loaded in the process |
| 6 | for the life of the interpreter. In-process teardown |
| 7 | (``VLLM_ENABLE_V1_MULTIPROCESSING=0``) then keeps the weights -- and the KV |
| 8 | cache aliased onto the model's layers -- resident after engine shutdown. |
| 9 | """ |
| 10 | |
| 11 | import gc |
| 12 | import weakref |
| 13 | |
| 14 | import pytest |
| 15 | import torch.nn as nn |
| 16 | |
| 17 | from vllm.model_executor.models.interfaces import _language_model_by_module |
| 18 | |
| 19 | pytestmark = pytest.mark.cpu_test |
| 20 | |
| 21 | |
| 22 | class _LanguageModel(nn.Module): |
| 23 | def embed_input_ids(self, input_ids): # pragma: no cover - never called |
| 24 | raise NotImplementedError |
| 25 | |
| 26 | |
| 27 | class _MultiModalModel(nn.Module): |
| 28 | def __init__(self): |
| 29 | super().__init__() |
| 30 | self.language_model = _LanguageModel() |
| 31 | |
| 32 | |
| 33 | def test_cache_is_weak_keyed(): |
| 34 | assert isinstance(_language_model_by_module, weakref.WeakKeyDictionary) |
| 35 | |
| 36 | |
| 37 | def test_cached_entry_does_not_pin_the_model(): |
| 38 | model = _MultiModalModel() |
| 39 | _language_model_by_module[model] = model.language_model |
| 40 | ref = weakref.ref(model) |
| 41 | |
| 42 | del model |
| 43 | gc.collect() |
| 44 | |
| 45 | assert ref() is None, ( |
| 46 | "the get_language_model cache is keeping the model alive; in-process " |
| 47 | "engine shutdown will not release its weights or KV cache" |
| 48 | ) |
| 49 | |
| 50 | |
| 51 | def test_cached_entry_is_dropped_with_its_key(): |
| 52 | model = _MultiModalModel() |
| 53 | _language_model_by_module[model] = model.language_model |
| 54 | assert model in _language_model_by_module |
| 55 | |
| 56 | del model |
| 57 | gc.collect() |
| 58 | |
| 59 | assert not any(isinstance(k, _MultiModalModel) for k in _language_model_by_module) |