Mirrors
/
hf_text-generation-inference
mirror of https://github.com/huggingface/text-generation-inference.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

fix: slice stopping criteria buffer

This commit is contained in:
OlivierDehaene 2023-12-14 17:01:43 +01:00
parent 083c2de9f8
commit 80a69204c1
1 changed files with 5 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
server/text_generation_server/utils/tokens.py
View File

@ -112,7 +112,7 @@ class StoppingCriteria:
self.stop_sequence_criterias = stop_sequence_criterias
self.max_new_tokens = max_new_tokens
self.current_tokens = 0
self.current_output = ""
self.current_output = "test"
self.ignore_eos_token = ignore_eos_token
def __call__(self, last_token: int, last_output: str) -> Tuple[bool, Optional[str]]:
@ -124,6 +124,10 @@ class StoppingCriteria:
return True, FinishReason.FINISH_REASON_EOS_TOKEN
self.current_output += last_output
# There is no need to keep an output that is too long
if len(self.current_output) > 300:
# Slice to -200 to avoid doing it all the time
self.current_output = self.current_output[-200:]
for stop_sequence_criteria in self.stop_sequence_criterias:
if stop_sequence_criteria(self.current_output):
return True, FinishReason.FINISH_REASON_STOP_SEQUENCE