| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import logging |
| 5 | |
| 6 | import pytest |
| 7 | import regex as re |
| 8 | |
| 9 | from tests.entrypoints.multimodal.conftest import TEST_IMAGE_ASSETS |
| 10 | from vllm import LLM |
| 11 | from vllm.entrypoints.chat_utils import ChatCompletionMessageParam |
| 12 | from vllm.v1.metrics import loggers as stat_loggers |
| 13 | from vllm.v1.metrics.reader import Counter, Metric |
| 14 | |
| 15 | |
| 16 | def _make_messages(image_url: str) -> list[ChatCompletionMessageParam]: |
| 17 | return [ |
| 18 | { |
| 19 | "role": "user", |
| 20 | "content": [ |
| 21 | { |
| 22 | "type": "image_url", |
| 23 | "image_url": {"url": image_url}, |
| 24 | }, |
| 25 | ], |
| 26 | } |
| 27 | ] |
| 28 | |
| 29 | |
| 30 | def _get_counter_value(metrics: list[Metric], name: str): |
| 31 | metric = next(m for m in metrics if m.name == name) |
| 32 | assert isinstance(metric, Counter) |
| 33 | return metric.value |
| 34 | |
| 35 | |
| 36 | def _get_mm_cache_stats(metrics: list[Metric]): |
| 37 | mm_cache_queries = _get_counter_value(metrics, "vllm:mm_cache_queries") |
| 38 | mm_cache_hits = _get_counter_value(metrics, "vllm:mm_cache_hits") |
| 39 | |
| 40 | return mm_cache_queries, mm_cache_hits |
| 41 | |
| 42 | |
| 43 | def _get_mm_cache_log(llm: LLM, caplog_vllm: pytest.LogCaptureFixture) -> float: |
| 44 | caplog_vllm.clear() |
| 45 | with caplog_vllm.at_level(logging.INFO, logger=stat_loggers.__name__): |
| 46 | llm.llm_engine.do_log_stats() |
| 47 | |
| 48 | assert len(caplog_vllm.records) == 1 |
| 49 | msg = caplog_vllm.records[0].getMessage() |
| 50 | |
| 51 | assert "MM cache hit rate" in msg |
| 52 | match = re.search(r"MM cache hit rate: ([0-9.]+)%", msg) |
| 53 | assert match is not None |
| 54 | return float(match.group(1)) |
| 55 | |
| 56 | |
| 57 | @pytest.mark.parametrize("image_urls", [TEST_IMAGE_ASSETS[:2]], indirect=True) |
| 58 | @pytest.mark.parametrize("mm_processor_cache_type", ["lru", "shm"]) |
| 59 | def test_mm_cache_stats( |
| 60 | num_gpus_available, |
| 61 | image_urls, |
| 62 | mm_processor_cache_type, |
| 63 | caplog_vllm, |
| 64 | vllm_runner, |
| 65 | ): |
| 66 | with vllm_runner( |
| 67 | "llava-hf/llava-1.5-7b-hf", |
| 68 | max_model_len=4096, |
| 69 | max_num_seqs=5, |
| 70 | enforce_eager=True, |
| 71 | mm_processor_cache_type=mm_processor_cache_type, |
| 72 | disable_log_stats=False, |
| 73 | limit_mm_per_prompt={"image": 2}, |
| 74 | ) as runner: |
| 75 | runner.llm.chat(_make_messages(image_urls[0])) |
| 76 | assert _get_mm_cache_stats(runner.llm.get_metrics()) == (1, 0) |
| 77 | assert _get_mm_cache_log(runner.llm, caplog_vllm) == pytest.approx(0.0) |
| 78 | |
| 79 | runner.llm.chat(_make_messages(image_urls[1])) |
| 80 | assert _get_mm_cache_stats(runner.llm.get_metrics()) == (2, 0) |
| 81 | assert _get_mm_cache_log(runner.llm, caplog_vllm) == pytest.approx(0.0) |
| 82 | |
| 83 | runner.llm.chat(_make_messages(image_urls[0])) |
| 84 | assert _get_mm_cache_stats(runner.llm.get_metrics()) == (3, 1) |
| 85 | assert _get_mm_cache_log(runner.llm, caplog_vllm) == pytest.approx(33.3) |
| 86 | |
| 87 | # NOTE: This only resets hit rate stats in CachingMetrics |
| 88 | # The raw queries and hits counts remain unaffected |
| 89 | runner.llm.reset_mm_cache() |
| 90 | |
| 91 | runner.llm.chat(_make_messages(image_urls[0])) |
| 92 | assert _get_mm_cache_stats(runner.llm.get_metrics()) == (4, 1) |
| 93 | assert _get_mm_cache_log(runner.llm, caplog_vllm) == pytest.approx(0.0) |
| 94 | |
| 95 | runner.llm.chat(_make_messages(image_urls[1])) |
| 96 | assert _get_mm_cache_stats(runner.llm.get_metrics()) == (5, 1) |
| 97 | assert _get_mm_cache_log(runner.llm, caplog_vllm) == pytest.approx(0.0) |