Skip MM Encoder for non-first PP ranks (#24387)

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>

vllm/v1/worker/gpu_model_runner.py

@@ -1620,14 +1620,11 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
 
         # _prepare_inputs may reorder the batch, so we must gather multi
         # modal outputs after that to ensure the correct order
-        if self.supports_mm_inputs:
+        if self.supports_mm_inputs and get_pp_group().is_first_rank:
             # Run the multimodal encoder if any.
             self._execute_mm_encoder(scheduler_output)
             mm_embeds = self._gather_mm_embeddings(scheduler_output)
-        else:
-            mm_embeds = []
 
-        if self.supports_mm_inputs and get_pp_group().is_first_rank:
             # NOTE(woosuk): To unify token ids and soft tokens (vision
             # embeddings), we always use embeddings (rather than token ids)
             # as input to the multimodal model, even when the input is text.