AI/text-generation-webui
1
0
Fork 0
mirror of https://github.com/oobabooga/text-generation-webui.git synced 2024-10-01 01:26:03 -04:00
Code Issues Packages Projects Releases Wiki Activity

Use a list of eos_tokens instead of just a number

Browse Source 
This might be the cause of LLaMA ramblings that some people have experienced.
This commit is contained in:
oobabooga 2023-03-12 14:54:58 -03:00
parent 17210ff88f
commit c7aa51faa6
1 changed files with 7 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
modules/text_generation.py
View File

@ -119,7 +119,9 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
original_input_ids = input_ids
output = input_ids[0]
cuda = "" if (shared.args.cpu or shared.args.deepspeed or shared.args.flexgen) else ".cuda()"
n = shared.tokenizer.eos_token_id if eos_token is None else int(encode(eos_token)[0][-1])
eos_token_ids = [shared.tokenizer.eos_token_id]
if eos_token is not None:
eos_token_ids.append(int(encode(eos_token)[0][-1]))
stopping_criteria_list = transformers.StoppingCriteriaList()
if stopping_string is not None:
# Copied from https://github.com/PygmalionAI/gradio-ui/blob/master/src/model.py
@ -129,7 +131,7 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
if not shared.args.flexgen:
generate_params = [
f"max_new_tokens=max_new_tokens",
f"eos_token_id={n}",
f"eos_token_id={eos_token_ids}",
f"stopping_criteria=stopping_criteria_list",
f"do_sample={do_sample}",
f"temperature={temperature}",
@ -149,7 +151,7 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
f"max_new_tokens={max_new_tokens if shared.args.no_stream else 8}",
f"do_sample={do_sample}",
f"temperature={temperature}",
f"stop={n}",
f"stop={eos_token_ids[-1]}",
]
if shared.args.deepspeed:
generate_params.append("synced_gpus=True")
@ -198,7 +200,7 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
if not (shared.args.chat or shared.args.cai_chat):
reply = original_question + apply_extensions(reply[len(question):], "output")
if output[-1] == n:
if output[-1] in eos_token_ids:
break
yield formatted_outputs(reply, shared.model_name)
@ -219,7 +221,7 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
if not (shared.args.chat or shared.args.cai_chat):
reply = original_question + apply_extensions(reply[len(question):], "output")
if np.count_nonzero(input_ids[0] == n) < np.count_nonzero(output == n):
if np.count_nonzero(np.isin(input_ids[0], eos_token_ids)) < np.count_nonzero(np.isin(output, eos_token_ids)):
break
yield formatted_outputs(reply, shared.model_name)