xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-07-10 03:27:09 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Core] Eliminate redundant is_encoder_decoder lookups (20-40us/step) (#29800)

Browse Source 
Signed-off-by: Wushi Dong <dongws@meta.com>
This commit is contained in:
Wushi Dong 2025-12-01 23:08:07 -08:00 committed by GitHub
parent f5b0846ba0
commit 0037b5746a
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 3 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/v1/worker/gpu_model_runner.py
View File

@ -2439,16 +2439,13 @@ class GPUModelRunner(
]: ]:
num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
is_first_rank = get_pp_group().is_first_rank is_first_rank = get_pp_group().is_first_rank
is_encoder_decoder = self.model_config.is_encoder_decoder
# _prepare_inputs may reorder the batch, so we must gather multi # _prepare_inputs may reorder the batch, so we must gather multi
# modal outputs after that to ensure the correct order # modal outputs after that to ensure the correct order
ec_connector_output = None ec_connector_output = None
if ( if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder:
self.supports_mm_inputs
and is_first_rank
and not self.model_config.is_encoder_decoder
):
# Run the multimodal encoder if any. # Run the multimodal encoder if any.
with self.maybe_get_ec_connector_output( with self.maybe_get_ec_connector_output(
scheduler_output, scheduler_output,
@ -2526,10 +2523,7 @@ class GPUModelRunner(
num_input_tokens, intermediate_tensors, True num_input_tokens, intermediate_tensors, True
) )
if ( if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
self.model_config.is_encoder_decoder
and scheduler_output.scheduled_encoder_inputs
):
# Run the encoder, just like we do with other multimodal inputs. # Run the encoder, just like we do with other multimodal inputs.
# For an encoder-decoder model, our processing here is a bit # For an encoder-decoder model, our processing here is a bit
# simpler, because the outputs are just passed to the decoder. # simpler, because the outputs are just passed to the decoder.