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

Add proper streaming to RWKV

Browse source
This commit is contained in:
oobabooga 2023-03-07 18:17:56 -03:00
parent 8660227e1b
commit 19a34941ed
2 changed files with 52 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

14
modules/text_generation.py
View file

@ -92,17 +92,17 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
# separately and terminate the function call earlier
if shared.is_RWKV:
if shared.args.no_stream:
reply = shared.model.generate(question, token_count=max_new_tokens, temperature=temperature, top_p=top_p, top_k=top_k)
t1 = time.time()
print(f"Output generated in {(t1-t0):.2f} seconds.")
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)
else:
yield formatted_outputs(question, shared.model_name)
for i in tqdm(range(max_new_tokens//8+1)):
clear_torch_cache()
reply = shared.model.generate(question, token_count=8, temperature=temperature, top_p=top_p, top_k=top_k)
# 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):
yield formatted_outputs(reply, shared.model_name)
question = reply
t1 = time.time()
print(f"Output generated in {(t1-t0):.2f} seconds.")
return
original_question = question