[ModelRunner V2] Fix whisper test (#46773)
Nick Hill committed
Jun 25, 2026 at 22:10 UTC
5b33041746b9b9ab45bdbd9b42cdd5d19357879a
2 files changed
+6
-3
vllm/v1/worker/gpu/mm/encoder_cache.py
+3
@@ -12,6 +12,9 @@ class EncoderCache:
12
# MM hash -> encoder outputs
13
self.encoder_outputs: dict[str, torch.Tensor] = {}
14
15
+ def __len__(self) -> int:
16
+ return len(self.encoder_outputs)
17
+
18
def add_request(
19
self, req_id: str, mm_features: list[MultiModalFeatureSpec]
20
) -> None:
vllm/v1/worker/gpu/model_states/mm_pruning.py
+3
-3
@@ -121,10 +121,10 @@ def maybe_create_mm_pruner(
121
) -> MultiModalPruner | None:
122
"""Create a MultiModalPruner if the model prunes embeddings and uses M-RoPE."""
123
if (
124
- not rope_state
124
+ rope_state is None
125
or not rope_state.has_delta
126
- or not encoder_cache
127
- or not model_config.multimodal_config
126
+ or encoder_cache is None
127
+ or model_config.multimodal_config is None
128
or not model_config.multimodal_config.is_multimodal_pruning_enabled()
129
or not supports_multimodal_pruning(model)
130
):