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

Refactor several function calls and the API

Browse source
This commit is contained in:
oobabooga 2023-04-06 01:22:15 -03:00 • committed by GitHub
parent 378d21e80c
commit 3f3e42e26c
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
8 changed files with 147 additions and 118 deletions
Download patch file Download diff file Expand all files Collapse all files

54
modules/text_generation.py
View file

@ -102,10 +102,11 @@ def set_manual_seed(seed):
def stop_everything_event():
shared.stop_everything = True
def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, seed, eos_token=None, stopping_strings=[]):
def generate_reply(question, generate_state, eos_token=None, stopping_strings=[]):
clear_torch_cache()
set_manual_seed(seed)
set_manual_seed(generate_state['seed'])
shared.stop_everything = False
generate_params = {}
t0 = time.time()
original_question = question
@ -117,9 +118,12 @@ 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 any((shared.is_RWKV, shared.is_llamacpp)):
for k in ['temperature', 'top_p', 'top_k', 'repetition_penalty']:
generate_params[k] = generate_state[k]
generate_params["token_count"] = generate_state["max_new_tokens"]
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, repetition_penalty=repetition_penalty)
reply = shared.model.generate(context=question, **generate_params)
output = original_question+reply
if not shared.is_chat():
reply = original_question + apply_extensions(reply, "output")
@ -130,7 +134,7 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
# RWKV has proper streaming, which is very nice.
# 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, repetition_penalty=repetition_penalty):
for reply in shared.model.generate_with_streaming(context=question, **generate_params):
output = original_question+reply
if not shared.is_chat():
reply = original_question + apply_extensions(reply, "output")
@ -145,7 +149,7 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
print(f"Output generated in {(t1-t0):.2f} seconds ({new_tokens/(t1-t0):.2f} tokens/s, {new_tokens} tokens, context {original_tokens})")
return
input_ids = encode(question, max_new_tokens)
input_ids = encode(question, generate_state['max_new_tokens'])
original_input_ids = input_ids
output = input_ids[0]
@ -158,33 +162,21 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
t = [encode(string, 0, add_special_tokens=False) for string in stopping_strings]
stopping_criteria_list.append(_SentinelTokenStoppingCriteria(sentinel_token_ids=t, starting_idx=len(input_ids[0])))
generate_params = {}
generate_params["max_new_tokens"] = generate_state['max_new_tokens']
if not shared.args.flexgen:
generate_params.update({
"max_new_tokens": max_new_tokens,
"eos_token_id": eos_token_ids,
"stopping_criteria": stopping_criteria_list,
"do_sample": do_sample,
"temperature": temperature,
"top_p": top_p,
"typical_p": typical_p,
"repetition_penalty": repetition_penalty,
"encoder_repetition_penalty": encoder_repetition_penalty,
"top_k": top_k,
"min_length": min_length if shared.args.no_stream else 0,
"no_repeat_ngram_size": no_repeat_ngram_size,
"num_beams": num_beams,
"penalty_alpha": penalty_alpha,
"length_penalty": length_penalty,
"early_stopping": early_stopping,
})
for k in ["do_sample", "temperature", "top_p", "typical_p", "repetition_penalty", "encoder_repetition_penalty", "top_k", "min_length", "no_repeat_ngram_size", "num_beams", "penalty_alpha", "length_penalty", "early_stopping"]:
generate_params[k] = generate_state[k]
generate_params["eos_token_id"] = eos_token_ids
generate_params["stopping_criteria"] = stopping_criteria_list
if shared.args.no_stream:
generate_params["min_length"] = 0
else:
generate_params.update({
"max_new_tokens": max_new_tokens if shared.args.no_stream else 8,
"do_sample": do_sample,
"temperature": temperature,
"stop": eos_token_ids[-1],
})
for k in ["do_sample", "temperature"]:
generate_params[k] = generate_state[k]
generate_params["stop"] = generate_state["eos_token_ids"][-1]
if not shared.args.no_stream:
generate_params["max_new_tokens"] = 8
if shared.args.no_cache:
generate_params.update({"use_cache": False})
if shared.args.deepspeed:
@ -244,7 +236,7 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
# Stream the output naively for FlexGen since it doesn't support 'stopping_criteria'
else:
for i in range(max_new_tokens//8+1):
for i in range(generate_state['max_new_tokens']//8+1):
clear_torch_cache()
with torch.no_grad():
output = shared.model.generate(**generate_params)[0]