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

[Misc][OpenAI] deprecate max_tokens in favor of new max_completion_tokens field for chat completion endpoint (#9837)

Browse Source
This commit is contained in:
Guillaume Calmettes
2024-10-31 02:15:56 +01:00
committed by GitHub
parent 64384bbcdf
commit abbfb6134d
14 changed files with 140 additions and 118 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
vllm/entrypoints/openai/serving_engine.py
View File

@@ -263,20 +263,26 @@ class OpenAIServing:
return TextTokensPrompt(prompt=input_text,
prompt_token_ids=input_ids)
if request.max_tokens is None:
# chat completion endpoint supports max_completion_tokens
if isinstance(request, ChatCompletionRequest):
# TODO(#9845): remove max_tokens when field dropped from OpenAI API
max_tokens = request.max_completion_tokens or request.max_tokens
else:
max_tokens = request.max_tokens
if max_tokens is None:
if token_num >= self.max_model_len:
raise ValueError(
f"This model's maximum context length is "
f"{self.max_model_len} tokens. However, you requested "
f"{token_num} tokens in the messages, "
f"Please reduce the length of the messages.")
elif token_num + request.max_tokens > self.max_model_len:
elif token_num + max_tokens > self.max_model_len:
raise ValueError(
f"This model's maximum context length is "
f"{self.max_model_len} tokens. However, you requested "
f"{request.max_tokens + token_num} tokens "
f"{max_tokens + token_num} tokens "
f"({token_num} in the messages, "
f"{request.max_tokens} in the completion). "
f"{max_tokens} in the completion). "
f"Please reduce the length of the messages or completion.")
return TextTokensPrompt(prompt=input_text, prompt_token_ids=input_ids)