ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

[LLM] fix chatglm native int4 emoji output (#8739)

Browse source
This commit is contained in:
Yishuo Wang 2023-08-11 15:38:41 +08:00 committed by GitHub
parent 77efcf7b1d
commit f91035c298
1 changed files with 6 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

6
python/llm/src/bigdl/llm/ggml/model/chatglm/chatglm.py
View file

@ -235,6 +235,7 @@ class ChatGLM(GenerationMixin):
text = self.detokenize(output_tokens) text = self.detokenize(output_tokens)
split_text = text[len(prompt):] split_text = text[len(prompt):]
split_text.rstrip('<EFBFBD>') # remove partial emoji
if stop != []: if stop != []:
for stop_word in stop: for stop_word in stop:
split_text = split_text.split(stop_word)[0] split_text = split_text.split(stop_word)[0]
@ -309,6 +310,11 @@ class ChatGLM(GenerationMixin):
print('\n') print('\n')
break break
text = self.detokenize(output_tokens) text = self.detokenize(output_tokens)
if text.endswith('<EFBFBD>'):
# generated new token is part of an emoji
# (some emoji consists of multiple tokens)
# continue to generate more tokens to decode this emoji
continue
text = text[len(history_text):] text = text[len(history_text):]
history_text += text history_text += text
yield { yield {