diff --git a/vllm/v1/worker/gpu/mm/encoder_cache.py b/vllm/v1/worker/gpu/mm/encoder_cache.py index 1fcbe64299..065df2975c 100644 --- a/vllm/v1/worker/gpu/mm/encoder_cache.py +++ b/vllm/v1/worker/gpu/mm/encoder_cache.py @@ -12,6 +12,9 @@ class EncoderCache: # MM hash -> encoder outputs self.encoder_outputs: dict[str, torch.Tensor] = {} + def __len__(self) -> int: + return len(self.encoder_outputs) + def add_request( self, req_id: str, mm_features: list[MultiModalFeatureSpec] ) -> None: diff --git a/vllm/v1/worker/gpu/model_states/mm_pruning.py b/vllm/v1/worker/gpu/model_states/mm_pruning.py index e1eb098792..781baa6d15 100644 --- a/vllm/v1/worker/gpu/model_states/mm_pruning.py +++ b/vllm/v1/worker/gpu/model_states/mm_pruning.py @@ -121,10 +121,10 @@ def maybe_create_mm_pruner( ) -> MultiModalPruner | None: """Create a MultiModalPruner if the model prunes embeddings and uses M-RoPE.""" if ( - not rope_state + rope_state is None or not rope_state.has_delta - or not encoder_cache - or not model_config.multimodal_config + or encoder_cache is None + or model_config.multimodal_config is None or not model_config.multimodal_config.is_multimodal_pruning_enabled() or not supports_multimodal_pruning(model) ):