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

[VLM] Avoid unnecessary tokenization (#12310)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-01-22 19:08:31 +08:00
committed by GitHub
parent 68ad4e3a8d
commit cd7b6f0857
9 changed files with 71 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
vllm/model_executor/models/chameleon.py
View File

@@ -122,8 +122,9 @@ class ChameleonMultiModalProcessor(
) -> list[int]:
# HF processor adds sep token for chat mode
tokenizer = self.info.get_tokenizer()
sep_token_id: int = \
tokenizer.vocab[tokenizer.sep_token] # type: ignore
vocab = tokenizer.get_vocab()
sep_token_id = vocab[tokenizer.sep_token] # type: ignore
return prompt_tokens + [sep_token_id]
@@ -141,18 +142,22 @@ class ChameleonMultiModalProcessor(
out_mm_kwargs: MultiModalKwargs,
) -> list[PromptReplacement]:
processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
image_tokens = processor.image_token * self.info.get_num_image_tokens()
tokenizer = self.info.get_tokenizer()
vocab = tokenizer.get_vocab()
image_start_id = vocab[processor.image_start_token]
image_token_id = vocab[processor.image_token]
image_end_id = vocab[processor.image_end_token]
num_image_tokens = self.info.get_num_image_tokens()
image_tokens = [image_token_id] * num_image_tokens
return [
PromptReplacement(
modality="image",
target="<image>",
target=[image_token_id],
replacement=PromptReplacementDetails(
full="".join([
processor.image_start_token,
image_tokens,
processor.image_end_token,
]),
full=([image_start_id] + image_tokens + [image_end_id]),
features=image_tokens,
),
)