| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import pytest |
| 5 | import torch |
| 6 | |
| 7 | from vllm.v1.core.kv_cache_utils import check_enough_kv_cache_memory |
| 8 | from vllm.v1.kv_cache_interface import FullAttentionSpec |
| 9 | |
| 10 | |
| 11 | def test_kv_cache_oom_no_memory(): |
| 12 | from unittest.mock import MagicMock |
| 13 | |
| 14 | config = MagicMock() |
| 15 | config.model_config.max_model_len = 2048 |
| 16 | |
| 17 | spec = { |
| 18 | "layer_0": FullAttentionSpec( |
| 19 | block_size=16, |
| 20 | num_kv_heads=8, |
| 21 | head_size=128, |
| 22 | dtype=torch.float16, |
| 23 | ) |
| 24 | } |
| 25 | |
| 26 | with pytest.raises(ValueError): |
| 27 | check_enough_kv_cache_memory(config, spec, 0) |
| 28 | |
| 29 | |
| 30 | def test_kv_cache_oom_insufficient_memory(monkeypatch): |
| 31 | from unittest.mock import MagicMock |
| 32 | |
| 33 | config = MagicMock() |
| 34 | config.model_config.max_model_len = 2048 |
| 35 | config.cache_config.block_size = 16 |
| 36 | config.parallel_config.tensor_parallel_size = 1 |
| 37 | config.parallel_config.pipeline_parallel_size = 1 |
| 38 | config.parallel_config.decode_context_parallel_size = 1 |
| 39 | |
| 40 | monkeypatch.setattr( |
| 41 | "vllm.v1.core.kv_cache_utils.max_memory_usage_bytes", |
| 42 | lambda c, s: 100 * 1024**3, # 100 GiB |
| 43 | ) |
| 44 | |
| 45 | spec = { |
| 46 | "layer_0": FullAttentionSpec( |
| 47 | block_size=16, |
| 48 | num_kv_heads=8, |
| 49 | head_size=128, |
| 50 | dtype=torch.float16, |
| 51 | ) |
| 52 | } |
| 53 | |
| 54 | with pytest.raises(ValueError): |
| 55 | check_enough_kv_cache_memory(config, spec, 1024**3) # 1 GiB |