| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Test that ``InputProcessor.inject_into_mm_cache()`` correctly injects |
| 4 | pre-processed mm_kwargs into the processor cache and reports MM cache |
| 5 | hit rate metrics accurately. |
| 6 | |
| 7 | This is used by frameworks like Dynamo that run the HF processor on a |
| 8 | frontend and transfer pre-processed mm_kwargs to the backend, avoiding |
| 9 | redundant processing. |
| 10 | """ |
| 11 | |
| 12 | import logging |
| 13 | |
| 14 | import pytest |
| 15 | import regex as re |
| 16 | |
| 17 | from tests.entrypoints.multimodal.conftest import TEST_IMAGE_ASSETS |
| 18 | from vllm import LLM, SamplingParams |
| 19 | from vllm.renderers.params import ChatParams |
| 20 | from vllm.v1.metrics import loggers as stat_loggers |
| 21 | from vllm.v1.metrics.reader import Counter, Metric |
| 22 | |
| 23 | |
| 24 | def _make_messages(image_url: str): |
| 25 | return [ |
| 26 | { |
| 27 | "role": "user", |
| 28 | "content": [ |
| 29 | { |
| 30 | "type": "image_url", |
| 31 | "image_url": {"url": image_url}, |
| 32 | }, |
| 33 | ], |
| 34 | } |
| 35 | ] |
| 36 | |
| 37 | |
| 38 | def _get_counter_value(metrics: list[Metric], name: str): |
| 39 | metric = next(m for m in metrics if m.name == name) |
| 40 | assert isinstance(metric, Counter) |
| 41 | return metric.value |
| 42 | |
| 43 | |
| 44 | def _get_mm_cache_stats(metrics: list[Metric]): |
| 45 | mm_cache_queries = _get_counter_value(metrics, "vllm:mm_cache_queries") |
| 46 | mm_cache_hits = _get_counter_value(metrics, "vllm:mm_cache_hits") |
| 47 | return mm_cache_queries, mm_cache_hits |
| 48 | |
| 49 | |
| 50 | def _get_mm_cache_log(llm: LLM, caplog_vllm: pytest.LogCaptureFixture) -> float: |
| 51 | caplog_vllm.clear() |
| 52 | with caplog_vllm.at_level(logging.INFO, logger=stat_loggers.__name__): |
| 53 | llm.llm_engine.do_log_stats() |
| 54 | |
| 55 | assert len(caplog_vllm.records) == 1 |
| 56 | msg = caplog_vllm.records[0].getMessage() |
| 57 | |
| 58 | assert "MM cache hit rate" in msg |
| 59 | match = re.search(r"MM cache hit rate: ([0-9.]+)%", msg) |
| 60 | assert match is not None |
| 61 | return float(match.group(1)) |
| 62 | |
| 63 | |
| 64 | @pytest.mark.parametrize("image_urls", [TEST_IMAGE_ASSETS[:2]], indirect=True) |
| 65 | @pytest.mark.parametrize("mm_processor_cache_type", ["lru", "shm"]) |
| 66 | def test_inject_into_mm_cache( |
| 67 | num_gpus_available, |
| 68 | image_urls, |
| 69 | mm_processor_cache_type, |
| 70 | caplog_vllm, |
| 71 | vllm_runner, |
| 72 | ): |
| 73 | """Test that inject_into_mm_cache() injects pre-processed mm_kwargs into |
| 74 | the processor cache and MM cache hit metrics are updated correctly. |
| 75 | |
| 76 | Steps: |
| 77 | 1. Two normal requests (same image) -> cache miss then hit (baseline) |
| 78 | 2. Extract cached kwargs, call inject_into_mm_cache with a new hash, |
| 79 | then generate with a pre-rendered input -> verifies injection works |
| 80 | """ |
| 81 | with vllm_runner( |
| 82 | "llava-hf/llava-1.5-7b-hf", |
| 83 | max_model_len=4096, |
| 84 | max_num_seqs=5, |
| 85 | enforce_eager=True, |
| 86 | disable_log_stats=False, |
| 87 | limit_mm_per_prompt={"image": 2}, |
| 88 | mm_processor_cache_type=mm_processor_cache_type, |
| 89 | ) as runner: |
| 90 | # Step 1: Normal requests to populate the cache |
| 91 | runner.llm.chat(_make_messages(image_urls[0])) |
| 92 | assert _get_mm_cache_stats(runner.llm.get_metrics()) == (1, 0) |
| 93 | |
| 94 | runner.llm.chat(_make_messages(image_urls[0])) |
| 95 | assert _get_mm_cache_stats(runner.llm.get_metrics()) == (2, 1) |
| 96 | assert _get_mm_cache_log(runner.llm, caplog_vllm) == pytest.approx(50.0) |
| 97 | |
| 98 | # Step 2: Use a second image to get valid expanded tokens and |
| 99 | # placeholder positions via the renderer. |
| 100 | runner.llm.chat(_make_messages(image_urls[1])) |
| 101 | queries_before = _get_mm_cache_stats(runner.llm.get_metrics())[0] # 3 |
| 102 | |
| 103 | renderer = runner.llm.llm_engine.renderer |
| 104 | cache = renderer.mm_processor_cache |
| 105 | assert cache is not None, "Processor cache should be enabled" |
| 106 | |
| 107 | _, eng_prompts = renderer.render_chat( |
| 108 | [_make_messages(image_urls[1])], |
| 109 | ChatParams(), |
| 110 | ) |
| 111 | eng_input = eng_prompts[0] |
| 112 | |
| 113 | # Inject pre-processed mm_kwargs with a NEW hash via public API |
| 114 | new_mm_hash = "deadbeef" * 8 |
| 115 | mm_hashes = {"image": [new_mm_hash]} |
| 116 | mm_kwargs = eng_input["mm_kwargs"] |
| 117 | |
| 118 | runner.llm.llm_engine.input_processor.inject_into_mm_cache(mm_hashes, mm_kwargs) |
| 119 | |
| 120 | # Build pre-rendered input (no externally_processed flag needed) |
| 121 | pre_rendered_input = { |
| 122 | "type": "multimodal", |
| 123 | "prompt_token_ids": eng_input["prompt_token_ids"], |
| 124 | "mm_kwargs": mm_kwargs, |
| 125 | "mm_hashes": mm_hashes, |
| 126 | "mm_placeholders": eng_input["mm_placeholders"], |
| 127 | } |
| 128 | |
| 129 | runner.llm.generate( |
| 130 | pre_rendered_input, |
| 131 | sampling_params=SamplingParams(max_tokens=1), |
| 132 | ) |
| 133 | |
| 134 | # Verify cache was queried and injection happened |
| 135 | queries_after = _get_mm_cache_stats(runner.llm.get_metrics())[0] |
| 136 | assert queries_after > queries_before, ( |
| 137 | "Cache should have been queried for the injected item" |
| 138 | ) |
| 139 | mm_rate = _get_mm_cache_log(runner.llm, caplog_vllm) |
| 140 | assert mm_rate >= 0.0, "MM cache hit rate should be reported" |
| 141 | |
| 142 | |
| 143 | @pytest.mark.parametrize("image_urls", [TEST_IMAGE_ASSETS[:1]], indirect=True) |
| 144 | def test_inject_into_mm_cache_without_cache( |
| 145 | num_gpus_available, |
| 146 | image_urls, |
| 147 | vllm_runner, |
| 148 | ): |
| 149 | """Test that inject_into_mm_cache works gracefully when processor cache |
| 150 | is disabled (mm_processor_cache_gb=0). Should not crash. |
| 151 | """ |
| 152 | with vllm_runner( |
| 153 | "llava-hf/llava-1.5-7b-hf", |
| 154 | max_model_len=4096, |
| 155 | max_num_seqs=5, |
| 156 | enforce_eager=True, |
| 157 | disable_log_stats=False, |
| 158 | limit_mm_per_prompt={"image": 2}, |
| 159 | mm_processor_cache_gb=0, |
| 160 | ) as runner: |
| 161 | # Run a normal chat request first to warm up the model. |
| 162 | runner.llm.chat(_make_messages(image_urls[0])) |
| 163 | |
| 164 | # Use the renderer to get a proper EngineInput with expanded tokens |
| 165 | renderer = runner.llm.llm_engine.renderer |
| 166 | _, eng_prompts = renderer.render_chat( |
| 167 | [_make_messages(image_urls[0])], |
| 168 | ChatParams(), |
| 169 | ) |
| 170 | eng_input = eng_prompts[0] |
| 171 | |
| 172 | mm_hashes = {"image": ["abcd1234" * 8]} |
| 173 | mm_kwargs = eng_input["mm_kwargs"] |
| 174 | |
| 175 | # inject_into_mm_cache should not crash even without cache |
| 176 | runner.llm.llm_engine.input_processor.inject_into_mm_cache(mm_hashes, mm_kwargs) |
| 177 | |
| 178 | # Build and generate with pre-rendered input |
| 179 | pre_rendered_input = { |
| 180 | "type": "multimodal", |
| 181 | "prompt_token_ids": eng_input["prompt_token_ids"], |
| 182 | "mm_kwargs": mm_kwargs, |
| 183 | "mm_hashes": mm_hashes, |
| 184 | "mm_placeholders": eng_input["mm_placeholders"], |
| 185 | } |
| 186 | |
| 187 | result = runner.llm.generate( |
| 188 | pre_rendered_input, |
| 189 | sampling_params=SamplingParams(max_tokens=1), |
| 190 | ) |
| 191 | assert len(result) == 1, "Should produce one output" |
| 192 | assert len(result[0].outputs) >= 1, "Should have at least one output sequence" |