Mirrors/text-generation-webui
1
0
Fork 0
mirror of https://github.com/oobabooga/text-generation-webui.git synced 2024-11-22 08:07:56 +01:00
Code Issues Packages Projects Releases Wiki Activity

Various fixes in chat mode

Browse Source
This commit is contained in:
oobabooga 2023-03-12 02:53:08 -03:00
parent b0e8cb8c88
commit 341e135036
3 changed files with 22 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
modules/callbacks.py
View File

@ -64,6 +64,7 @@ class Iteratorize:
ret = self.mfunc(callback=_callback, **self.kwargs)
except ValueError:
pass
clear_torch_cache()
self.q.put(self.sentinel)
if self.c_callback:
self.c_callback(ret)

16
modules/chat.py
View File

@ -115,18 +115,14 @@ def chatbot_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical
visible_text = visible_text.replace('\n', '<br>')
text = apply_extensions(text, "input")
if custom_generate_chat_prompt is None:
prompt = generate_chat_prompt(text, max_new_tokens, name1, name2, context, chat_prompt_size)
else:
prompt = custom_generate_chat_prompt(text, max_new_tokens, name1, name2, context, chat_prompt_size)
# Generate
reply = ''
for i in range(chat_generation_attempts):
# The prompt needs to be generated here because, as the reply
# grows, it may become necessary to remove more old messages to
# fit into the 2048 tokens window.
if custom_generate_chat_prompt is None:
prompt = generate_chat_prompt(text, max_new_tokens, name1, name2, context, chat_prompt_size-len(encode(' '+reply)[0]))
else:
prompt = custom_generate_chat_prompt(text, max_new_tokens, name1, name2, context, chat_prompt_size-len(encode(' '+reply)[0]))
for reply in generate_reply(f"{prompt}{' ' if len(reply) > 0 else ''}{reply}", max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, eos_token=eos_token, stopping_string=f"\n{name1}:"):
# Extracting the reply
@ -160,10 +156,10 @@ def impersonate_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typ
if 'pygmalion' in shared.model_name.lower():
name1 = "You"
prompt = generate_chat_prompt(text, max_new_tokens, name1, name2, context, chat_prompt_size, impersonate=True)
reply = ''
for i in range(chat_generation_attempts):
prompt = generate_chat_prompt(text, max_new_tokens, name1, name2, context, chat_prompt_size-len(encode(' '+reply)[0]), impersonate=True)
for reply in generate_reply(prompt+reply, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, eos_token=eos_token, stopping_string=f"\n{name2}:"):
reply, next_character_found, substring_found = extract_message_from_reply(prompt, reply, name1, name2, check, impersonate=True)
if not substring_found:

3
modules/text_generation.py
View File

@ -92,6 +92,7 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
# These models are not part of Hugging Face, so we handle them
# separately and terminate the function call earlier
if shared.is_RWKV:
try:
if shared.args.no_stream:
reply = shared.model.generate(context=question, token_count=max_new_tokens, temperature=temperature, top_p=top_p, top_k=top_k)
yield formatted_outputs(reply, shared.model_name)
@ -101,7 +102,7 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
# No need to generate 8 tokens at a time.
for reply in shared.model.generate_with_streaming(context=question, token_count=max_new_tokens, temperature=temperature, top_p=top_p, top_k=top_k):
yield formatted_outputs(reply, shared.model_name)
finally:
t1 = time.time()
output = encode(reply)[0]
input_ids = encode(question)