frozenleaves/vllm
1
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2025-10-20 23:03:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

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

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon
2025-09-07 09:38:35 -07:00
committed by GitHub
parent 0661cb9df3
commit 2e5d21378d
1 changed files with 1 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
vllm/v1/worker/gpu_model_runner.py
View File

@ -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.