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

[Bugfix] fix when skip tokenizer init (#21922)

Browse Source 
Signed-off-by: rongfu.leng <rongfu.leng@daocloud.io>
This commit is contained in:
rongfu.leng
2025-08-02 01:09:36 +08:00
committed by GitHub
parent 3f8e952179
commit b879ecd6e2
2 changed files with 33 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/v1/engine/processor.py
View File

@@ -89,6 +89,10 @@ class Processor:
return
if not params.allowed_token_ids:
raise ValueError("allowed_token_ids is not None and empty!")
if self.tokenizer is None:
# When skip_tokenizer_init=True, we can't validate token IDs
# Skip validation and let the model handle invalid tokens
return
tokenizer = self.tokenizer.get_lora_tokenizer(lora_request)
vocab_size = len(tokenizer)
if not all(0 <= tid < vocab_size for tid in params.allowed_token_ids):
@@ -283,8 +287,9 @@ class Processor:
len(decoder_inputs["prompt_token_ids"]))
sampling_params.update_from_generation_config(
self.generation_config_fields, eos_token_id)
sampling_params.update_from_tokenizer(
self.tokenizer.get_lora_tokenizer(lora_request))
if self.tokenizer is not None:
sampling_params.update_from_tokenizer(
self.tokenizer.get_lora_tokenizer(lora_request))
else:
pooling_params = params.clone()