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

Print the performance information more reliably

Browse Source
This commit is contained in:
oobabooga 2023-03-03 21:24:32 -03:00
parent 5a79863df3
commit 05e703b4a4
1 changed files with 8 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
modules/text_generation.py
View File

@ -86,12 +86,18 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
if not shared.args.cpu: if not shared.args.cpu:
torch.cuda.empty_cache() torch.cuda.empty_cache()
t0 = time.time()
# These models are not part of Hugging Face, so we handle them
# separately and terminate the function call earlier
if shared.is_RWKV or shared.is_LLaMA: if shared.is_RWKV or shared.is_LLaMA:
if shared.args.no_stream: if shared.args.no_stream:
reply = shared.model.generate(question, token_count=max_new_tokens, temperature=temperature, top_p=top_p) reply = shared.model.generate(question, token_count=max_new_tokens, temperature=temperature, top_p=top_p)
t1 = time.time()
print(f"Output generated in {(t1-t0):.2f} seconds.")
yield formatted_outputs(reply, shared.model_name) yield formatted_outputs(reply, shared.model_name)
else: else:
for i in range(max_new_tokens//8): for i in tqdm(range(max_new_tokens//8+1)):
reply = shared.model.generate(question, token_count=8, temperature=temperature, top_p=top_p) reply = shared.model.generate(question, token_count=8, temperature=temperature, top_p=top_p)
yield formatted_outputs(reply, shared.model_name) yield formatted_outputs(reply, shared.model_name)
question = reply question = reply
@ -160,7 +166,6 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
# Generate the entire reply at once # Generate the entire reply at once
if shared.args.no_stream: if shared.args.no_stream:
t0 = time.time()
with torch.no_grad(): with torch.no_grad():
output = eval(f"shared.model.generate({', '.join(generate_params)}){cuda}")[0] output = eval(f"shared.model.generate({', '.join(generate_params)}){cuda}")[0]
if shared.soft_prompt: if shared.soft_prompt:
@ -169,10 +174,10 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
reply = decode(output) reply = decode(output)
if not (shared.args.chat or shared.args.cai_chat): if not (shared.args.chat or shared.args.cai_chat):
reply = original_question + apply_extensions(reply[len(question):], "output") reply = original_question + apply_extensions(reply[len(question):], "output")
yield formatted_outputs(reply, shared.model_name)
t1 = time.time() t1 = time.time()
print(f"Output generated in {(t1-t0):.2f} seconds ({(len(output)-len(input_ids[0]))/(t1-t0)/8:.2f} it/s, {len(output)-len(input_ids[0])} tokens)") print(f"Output generated in {(t1-t0):.2f} seconds ({(len(output)-len(input_ids[0]))/(t1-t0)/8:.2f} it/s, {len(output)-len(input_ids[0])} tokens)")
yield formatted_outputs(reply, shared.model_name)
# Generate the reply 8 tokens at a time # Generate the reply 8 tokens at a time
else: else: