Mirrors/text-generation-webui
1
0
Fork 0
mirror of https://github.com/oobabooga/text-generation-webui.git synced 2024-12-25 05:48:55 +01:00
Code Issues Packages Projects Releases Wiki Activity

Simplify generate() function

Browse Source
This commit is contained in:
oobabooga 2023-02-02 13:47:08 -03:00
parent 3f05cf5ddd
commit 638495b633
1 changed files with 8 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
server.py
View File

@ -215,10 +215,9 @@ def generate_reply(question, tokens, inference_settings, selected_model, eos_tok
preset = infile.read() preset = infile.read()
loaded_preset = inference_settings loaded_preset = inference_settings
cuda = "" if args.cpu else ".cuda()"
n = tokenizer.eos_token_id if eos_token is None else tokenizer.encode(eos_token, return_tensors='pt')[0][-1]
input_ids = encode(question, tokens) input_ids = encode(question, tokens)
cuda = "" if (args.cpu or args.deepspeed) else ".cuda()"
n = tokenizer.eos_token_id if eos_token is None else tokenizer.encode(eos_token, return_tensors='pt')[0][-1]
if stopping_string is not None: if stopping_string is not None:
# The stopping_criteria code below was copied from # The stopping_criteria code below was copied from
# https://github.com/PygmalionAI/gradio-ui/blob/master/src/model.py # https://github.com/PygmalionAI/gradio-ui/blob/master/src/model.py
@ -232,14 +231,15 @@ def generate_reply(question, tokens, inference_settings, selected_model, eos_tok
else: else:
stopping_criteria_list = None stopping_criteria_list = None
generate_params = [f"eos_token_id={n}", "stopping_criteria=stopping_criteria_list"]
if args.deepspeed:
generate_params.append("synced_gpus=True")
# Generate the entire reply at once # Generate the entire reply at once
if args.no_stream: if args.no_stream:
t0 = time.time() t0 = time.time()
with torch.no_grad(): with torch.no_grad():
if not args.deepspeed: output = eval(f"model.generate(input_ids, {','.join(generate_params)}, {preset}){cuda}")
output = eval(f"model.generate(input_ids, eos_token_id={n}, stopping_criteria=stopping_criteria_list, {preset}){cuda}")
else:
output = eval(f"model.generate(input_ids, synced_gpus=True, eos_token_id={n}, stopping_criteria=stopping_criteria_list, {preset})")
reply = decode(output[0]) reply = decode(output[0])
t1 = time.time() t1 = time.time()
print(f"Output generated in {(t1-t0):.2f} seconds ({(len(output[0])-len(input_ids[0]))/(t1-t0):.2f} it/s)") print(f"Output generated in {(t1-t0):.2f} seconds ({(len(output[0])-len(input_ids[0]))/(t1-t0):.2f} it/s)")
@ -253,10 +253,7 @@ def generate_reply(question, tokens, inference_settings, selected_model, eos_tok
preset = preset.replace('max_new_tokens=tokens', 'max_new_tokens=8') preset = preset.replace('max_new_tokens=tokens', 'max_new_tokens=8')
for i in tqdm(range(tokens//8+1)): for i in tqdm(range(tokens//8+1)):
with torch.no_grad(): with torch.no_grad():
if not args.deepspeed: output = eval(f"model.generate(input_ids, {','.join(generate_params)}, {preset}){cuda}")
output = eval(f"model.generate(input_ids, eos_token_id={n}, stopping_criteria=stopping_criteria_list, {preset}){cuda}")
else:
output = eval(f"model.generate(input_ids, synced_gpus=True, eos_token_id={n}, stopping_criteria=stopping_criteria_list, {preset})")
reply = decode(output[0]) reply = decode(output[0])
if not (args.chat or args.cai_chat): if not (args.chat or args.cai_chat):
reply = original_question + apply_extensions(reply[len(question):], "output") reply = original_question + apply_extensions(reply[len(question):], "output")