[MM] Simplify prompt updates: replace PromptSeq with list[int] (#53372)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Co-authored-by: Qwen Code <noreply@alibaba-inc.com>
This commit is contained in:
Cyrus Leung
2026-08-23 14:50:32 +08:00
committed by GitHub
co-authored by Qwen Code
parent a3561ef8e4
commit ebcd606467
53 changed files with 831 additions and 855 deletions
+8 -3
View File
@@ -54,6 +54,7 @@ from vllm.multimodal.processing import (
PromptReplacement,
PromptUpdate,
PromptUpdateDetails,
cached_encode,
)
from vllm.sequence import IntermediateTensors
from vllm.tokenizers import cached_tokenizer_from_config
@@ -450,6 +451,10 @@ class IsaacMultiModalProcessor(BaseMultiModalProcessor):
out_mm_kwargs: MultiModalKwargsItems,
) -> Sequence[PromptUpdate]:
image_processor = self.info.get_image_processor(**hf_processor_mm_kwargs)
tokenizer = self.info.get_tokenizer()
image_pad_token_ids = cached_encode(
tokenizer, "<|image_pad|>", add_special_tokens=False
)
pixel_shuffle_scale = getattr(image_processor, "pixel_shuffle_scale", 2)
merge_length = pixel_shuffle_scale**2
@@ -460,13 +465,13 @@ class IsaacMultiModalProcessor(BaseMultiModalProcessor):
assert isinstance(grid_thw, torch.Tensor)
feature_size = int(grid_thw.prod()) // merge_length
repl_full = "<|image_pad|>" * feature_size
return PromptUpdateDetails.select_text(repl_full, "<|image_pad|>")
repl_full = image_pad_token_ids * feature_size
return PromptUpdateDetails.select_token_ids(repl_full, image_pad_token_ids)
return [
PromptReplacement(
modality="image",
target="<image>",
target=cached_encode(tokenizer, "<image>", add_special_tokens=False),
replacement=get_replacement_isaac,
)
]