biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[V1][Scheduler] Avoid calling _try_schedule_encoder_inputs for every request (#15778)

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon
2025-03-30 14:10:42 -07:00
committed by GitHub
parent 70fedd0f79
commit 9b459eca88
2 changed files with 30 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/v1/request.py
View File

@@ -59,6 +59,8 @@ class Request:
self.mm_positions = multi_modal_placeholders or []
self.mm_inputs = multi_modal_inputs or []
self.mm_hashes: list[str] = multi_modal_hashes or []
self.num_encoder_inputs = len(self.mm_inputs)
self.has_encoder_inputs = self.num_encoder_inputs > 0
# Sanity check
assert len(self.mm_inputs) == len(self.mm_positions)
@@ -117,13 +119,6 @@ class Request:
def get_finished_reason(self) -> Union[FinishReason, None]:
return RequestStatus.get_finished_reason(self.status)
def has_encoder_inputs(self) -> bool:
return len(self.mm_inputs) > 0
@property
def num_encoder_inputs(self) -> int:
return len(self.mm_positions)
def get_num_encoder_tokens(self, input_id: int) -> int:
assert input_id < len(self.mm_positions)
num_tokens = self.mm_positions[input_id]["length"]