[MM] Cache common token sequences (#53560)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Signed-off-by: Cyrus Leung <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-08-24 07:48:10 -07:00
committed by GitHub
parent 4ca856b0b5
commit ecfa7bb373
14 changed files with 40 additions and 23 deletions
@@ -98,6 +98,7 @@ from vllm.multimodal.processing import (
PromptReplacement,
PromptUpdate,
PromptUpdateDetails,
cached_encode,
)
from vllm.multimodal.processing.dummy_inputs import BaseDummyInputsBuilder
from vllm.multimodal.video import (
@@ -1894,7 +1895,7 @@ class LlavaOnevision2MultiModalProcessor(
video_pad_id = vocab["<|video_pad|>"]
vision_start_id = vocab["<|vision_start|>"]
vision_end_id = vocab["<|vision_end|>"]
newline_ids = tokenizer.encode("\n", add_special_tokens=False)
newline_ids = cached_encode(tokenizer, "\n", add_special_tokens=False)
merge_length = image_processor.merge_size**2
decimals = int(hf_processor_mm_kwargs.get("timestamp_decimals", 1))