\n")
break
#print(f"{generated_text}")
for stop_word in stop:
stop_word = codecs.getdecoder("unicode_escape")(stop_word)[0]
if stop_word != '' and stop_word in generated_text:
generated_text = generated_text[:generated_text.find(stop_word)]
gc.collect()
yield generated_text
def chat(
prompt,
history,
username,
max_new_tokens=10,
temperature=0.1,
top_p=1.0,
seed=42,
):
global model
history = history or []
intro = ""
if model == None:
gc.collect()
if (DEVICE == "cuda"):
torch.cuda.empty_cache()
model = get_model()
intro = f'''The following is a verbose and detailed conversation between an AI assistant called FRITZ, and a human user called USER. FRITZ is intelligent, knowledgeable, wise and polite.
{username}: What year was the french revolution?
FRITZ: The French Revolution started in 1789, and lasted 10 years until 1799.
{username}: 3+5=?
FRITZ: The answer is 8.
{username}: What year did the Berlin Wall fall?
FRITZ: The Berlin wall stood for 28 years and fell in 1989.
{username}: solve for a: 9-a=2
FRITZ: The answer is a=7, because 9-7 = 2.
{username}: wat is lhc
FRITZ: The Large Hadron Collider (LHC) is a high-energy particle collider, built by CERN, and completed in 2008. It was used to confirm the existence of the Higgs boson in 2012.
{username}: Tell me about yourself.
FRITZ: My name is Fritz. I am an RNN based Large Language Model (LLM).
'''
if len(history) == 0:
# no history, so lets reset chat state
model.resetState()
history = [[],model.emptyState]
print("reset chat state")
else:
if (history[0][0][0].split(':')[0] != username):
model.resetState()
history = [[],model.emptyState]
print("username changed, reset state")
else:
model.setState(history[1])
intro = ""
max_new_tokens = int(max_new_tokens)
temperature = float(temperature)
top_p = float(top_p)
seed = seed
assert 1 <= max_new_tokens <= 384
assert 0.0 <= temperature <= 1.0
assert 0.0 <= top_p <= 1.0
if temperature == 0.0:
temperature = 0.05
prompt = f"{username}: " + prompt + "\n"
print(f"CHAT ({datetime.now()}):\n-------\n{prompt}")
print(f"OUTPUT ({datetime.now()}):\n-------\n")
# Load prompt
model.loadContext(newctx=intro+prompt)
out = model.forward(number=max_new_tokens, stopStrings=["<|endoftext|>",username+":"],temp=temperature,top_p_usual=top_p)
generated_text = out["output"].lstrip("\n ")
generated_text = generated_text.rstrip("USER:")
print(f"{generated_text}")
gc.collect()
history[0].append((prompt, generated_text))
return history[0],[history[0],out["state"]]
examples = [
[
# Question Answering
'''What is the capital of Germany?''',"Q/A", 25, 0.2, 1.0, "<|endoftext|>"],
[
# Question Answering
'''Are humans good or bad?''',"Q/A", 150, 0.8, 0.8, "<|endoftext|>"],
[
# Question Answering
'''What is the purpose of Vitamin A?''',"Q/A", 50, 0.2, 0.8, "<|endoftext|>"],
[
# Chatbot
'''This is a conversation between two AI large language models named Alex and Fritz. They are exploring each other's capabilities, and trying to ask interesting questions of one another to explore the limits of each others AI.
Conversation:
Alex: Good morning, Fritz, what type of LLM are you based upon?
Fritz: Morning Alex, I am an RNN with transformer level performance. My language model is 100% attention free.
Alex:''', "generative", 220, 0.9, 0.9, "\\n\\n,<|endoftext|>"],
[
# Generate List
'''Task given:
Please Write a Short story about a cat learning python
Best Full Response:
''', "generative", 80, 0.2, 1.0, "\\n\\n,<|endoftext|>"],
[
# Natural Language Interface
'''Here is a short story (in the style of Tolkien) in which Aiden attacks a robot with a sword:
''',"generative", 200, 0.85, 0.8, "<|endoftext|>"]
]
iface = gr.Interface(
fn=infer,
description='''RNN With Transformer-level LLM Performance. (github)
According to the author: "It combines the best of RNN and transformers - great performance, fast inference, saves VRAM, fast training, "infinite" ctx_len, and free sentence embedding"
Thanks to RFT Capital for donating compute capability for our experiments. Additional thanks to the author of the rwkvstic library.
''',
allow_flagging="never",
inputs=[
gr.Textbox(lines=20, label="Prompt"), # prompt
gr.Radio(["generative","Q/A"], value="generative", label="Choose Mode"),
gr.Slider(1, 256, value=40), # max_tokens
gr.Slider(0.0, 1.0, value=0.8), # temperature
gr.Slider(0.0, 1.0, value=0.85), # top_p
gr.Textbox(lines=1, value="<|endoftext|>") # stop
],
outputs=gr.Textbox(label="Generated Output", lines=25),
examples=examples,
cache_examples=False,
).queue()
chatiface = gr.Interface(
fn=chat,
description='''RNN With Transformer-level LLM Performance. (github)
According to the author: "It combines the best of RNN and transformers - great performance, fast inference, saves VRAM, fast training, "infinite" ctx_len, and free sentence embedding"
Thanks to RFT Capital for donating compute capability for our experiments. Additional thanks to the author of the rwkvstic library.
''',
allow_flagging="never",
inputs=[
gr.Textbox(lines=5, label="Message"), # prompt
"state",
gr.Text(lines=1, value="USER", label="Your Name", placeholder="Enter your Name"),
gr.Slider(1, 256, value=60), # max_tokens
gr.Slider(0.0, 1.0, value=0.8), # temperature
gr.Slider(0.0, 1.0, value=0.85) # top_p
],
outputs=[gr.Chatbot(label="Chat Log", color_map=("green", "pink")),"state"],
).queue()
demo = gr.TabbedInterface(
[iface,chatiface],["Generative","Chatbot"],
title="RWKV-4 (1.5b Instruct)",
)
demo.queue()
demo.launch(share=False)