mmarch/text-generation-webui
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

Implement a demo HF wrapper for exllama to utilize existing HF transformers decoding. (#2777)

Browse source
This commit is contained in:
LarryVRH 2023-06-22 02:31:42 +08:00 • committed by GitHub
parent a06acd6d09
commit 580c1ee748
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
7 changed files with 101 additions and 6 deletions
Download patch file Download diff file Expand all files Collapse all files

3
modules/text_generation.py
View file

@ -104,9 +104,8 @@ def get_reply_from_output_ids(output_ids, input_ids, original_question, state, i
else:
new_tokens = len(output_ids) - len(input_ids[0])
reply = decode(output_ids[-new_tokens:], state['skip_special_tokens'])
# Prevent LlamaTokenizer from skipping a space
if type(shared.tokenizer) is transformers.LlamaTokenizer and len(output_ids) > 0:
if type(shared.tokenizer) in [transformers.LlamaTokenizer, transformers.LlamaTokenizerFast] and len(output_ids) > 0:
if shared.tokenizer.convert_ids_to_tokens(int(output_ids[-new_tokens])).startswith('▁'):
reply = ' ' + reply