From 4658bf882b881287fc85797a23037aa91740b7a7 Mon Sep 17 00:00:00 2001 From: Tobias Wasner Date: Fri, 22 May 2026 12:54:29 +0200 Subject: [PATCH] [Bugfix] Clear P0 mm sender cache on sleep/pause to fix mm_hash desync (#43001) Signed-off-by: Tobias Wasner --- tests/multimodal/test_cache.py | 39 ++++++++++++++++++++++++++++++++++ vllm/v1/engine/async_llm.py | 4 ++++ vllm/v1/engine/llm_engine.py | 2 ++ 3 files changed, 45 insertions(+) diff --git a/tests/multimodal/test_cache.py b/tests/multimodal/test_cache.py index d01b94ac9a..30df1d831a 100644 --- a/tests/multimodal/test_cache.py +++ b/tests/multimodal/test_cache.py @@ -549,3 +549,42 @@ def test_processor_cache_shared_across_loras(): receiver_cache.get_and_update_features([feature_lora_b]) assert feature_lora_b.data == item_data + + +_SLEEP_VISION_PROMPT = ( + "<|im_start|>system\nYou are a helpful assistant.<|im_end|>" + "\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>" + "What is in the image?<|im_end|>\n" + "<|im_start|>assistant\n" +) + + +@pytest.mark.skipif( + not torch.cuda.is_available(), + reason="sleep mode regression requires a CUDA GPU", +) +def test_sleep_wake_preserves_mm_cache_consistency(): + """Regression for vllm-project/vllm#42995.""" + from vllm import LLM, SamplingParams + from vllm.assets.image import ImageAsset + + image = ImageAsset("stop_sign").pil_image + prompt = { + "prompt": _SLEEP_VISION_PROMPT, + "multi_modal_data": {"image": image}, + } + sampling_params = SamplingParams(temperature=0, max_tokens=8) + + llm = LLM( + model="Qwen/Qwen2-VL-2B-Instruct", + enable_sleep_mode=True, + enforce_eager=True, + gpu_memory_utilization=0.5, + max_model_len=2048, + ) + + llm.generate([prompt], sampling_params) + llm.sleep(level=1) + llm.wake_up() + output2 = llm.generate([prompt], sampling_params) + assert output2[0].outputs[0].text diff --git a/vllm/v1/engine/async_llm.py b/vllm/v1/engine/async_llm.py index 32cc3c0d2d..160f148f5c 100644 --- a/vllm/v1/engine/async_llm.py +++ b/vllm/v1/engine/async_llm.py @@ -781,6 +781,8 @@ class AsyncLLM(EngineClient): stacklevel=2, ) mode = "wait" + if clear_cache: + await self.renderer.clear_mm_cache_async() await self.engine_core.pause_scheduler_async(mode=mode, clear_cache=clear_cache) # Small sleep to help ensure that final outputs from any in-flight requests are # returned prior to this method returning. These outputs come out of the engine @@ -927,6 +929,8 @@ class AsyncLLM(EngineClient): await self.engine_core.reset_encoder_cache_async() async def sleep(self, level: int = 1, mode: PauseMode = "abort") -> None: + if level >= 1: + await self.renderer.clear_mm_cache_async() await self.engine_core.sleep_async(level, mode) if self.logger_manager is not None: diff --git a/vllm/v1/engine/llm_engine.py b/vllm/v1/engine/llm_engine.py index 62c7d0e2a8..f3e8a95b0d 100644 --- a/vllm/v1/engine/llm_engine.py +++ b/vllm/v1/engine/llm_engine.py @@ -350,6 +350,8 @@ class LLMEngine: self.engine_core.reset_encoder_cache() def sleep(self, level: int = 1, mode: PauseMode = "abort"): + if level >= 1: + self.renderer.clear_mm_cache() self.engine_core.sleep(level, mode) if self.logger_manager is not None: