app.py

import gradio as gr
import os, gc, copy, torch, re
from datetime import datetime
from huggingface_hub import hf_hub_download
from pynvml import *
nvmlInit()
gpu_h = nvmlDeviceGetHandleByIndex(0)
ctx_limit = 1024
title = "RWKV-x060-eng_single_round_test-1B6-20240427-ctx1024"

os.environ["RWKV_JIT_ON"] = '1'
os.environ["RWKV_CUDA_ON"] = '1' # if '1' then use CUDA kernel for seq mode (much faster)

from rwkv.model import RWKV
model_path = hf_hub_download(repo_id="BlinkDL/temp-latest-training-models", filename=f"{title}.pth")
model = RWKV(model=model_path, strategy='cuda fp16')
from rwkv.utils import PIPELINE, PIPELINE_ARGS
pipeline = PIPELINE(model, "rwkv_vocab_v20230424")

def generate_prompt(instruction):
    instruction = instruction.strip().replace('\r\n','\n')
    instruction = re.sub(r'\n+', '\n', instruction)
    return f"User: {instruction}\n\nAssistant:"""

def evaluate(
    ctx,
    token_count=500,
    temperature=1.0,
    top_p=0.3,
    presencePenalty = 0.3,
    countPenalty = 0.3,
):
    args = PIPELINE_ARGS(temperature = max(0.2, float(temperature)), top_p = float(top_p),
                     alpha_frequency = countPenalty,
                     alpha_presence = presencePenalty,
                     token_ban = [], # ban the generation of some tokens
                     token_stop = [0]) # stop generation whenever you see any token here
    ctx = generate_prompt(ctx)
    all_tokens = []
    out_last = 0
    out_str = ''
    occurrence = {}
    state = None
    for i in range(int(token_count)):
        out, state = model.forward(pipeline.encode(ctx)[-ctx_limit:] if i == 0 else [token], state)
        for n in occurrence:
            out[n] -= (args.alpha_presence + occurrence[n] * args.alpha_frequency)

        token = pipeline.sample_logits(out, temperature=args.temperature, top_p=args.top_p)
        if token in args.token_stop:
            break
        all_tokens += [token]
        for xxx in occurrence:
            occurrence[xxx] *= 0.996        
        if token not in occurrence:
            occurrence[token] = 1
        else:
            occurrence[token] += 1
        
        tmp = pipeline.decode(all_tokens[out_last:])
        if '\ufffd' not in tmp:
            out_str += tmp
            yield out_str.strip()
            out_last = i + 1

    gpu_info = nvmlDeviceGetMemoryInfo(gpu_h)
    timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
    print(f'{timestamp} - vram {gpu_info.total} used {gpu_info.used} free {gpu_info.free}')  
    del out
    del state
    gc.collect()
    torch.cuda.empty_cache()
    yield out_str.strip()

examples = [
    ["Build a rocket to land on moon.", 700, 1, 0.3, 0.3, 0.3],
    ["How can I craft an engaging story featuring vampires on Mars?", 700, 1, 0.3, 0.3, 0.3],
    ["How can I persuade Elon Musk to follow me on Twitter?", 700, 1, 0.3, 0.3, 0.3],
    ["Write a story using the following information: a man named Alex chops a tree down.", 700, 1, 0.3, 0.3, 0.3],
    ["Create a Python program to simulate moon landing.", 700, 1, 0.3, 0.3, 0.3],
    ["Write a simple website in HTML. When a user clicks the button, it shows a random joke from a list of 4 jokes.", 700, 1, 0.3, 0.3, 0.3],
]

##########################################################################

with gr.Blocks(title=title) as demo:
    gr.HTML(f"<div style=\"text-align: center;\">\n<h1>{title}</h1>\n</div>")
    with gr.Tab("Raw Generation"):
        gr.Markdown(f"This is [RWKV-6](https://huggingface.co/BlinkDL/temp-latest-training-models) with 1.6B params [state-tuned](https://twitter.com/BlinkDL_AI/status/1784496793075744966) on single-round English Q & A. RWKV is a 100% attention-free RNN [RWKV-LM](https://github.com/BlinkDL/RWKV-LM), and we have [300+ Github RWKV projects](https://github.com/search?o=desc&p=1&q=rwkv&s=updated&type=Repositories). Demo limited to ctxlen {ctx_limit}.")
        with gr.Row():
            with gr.Column():
                prompt = gr.Textbox(lines=2, label="Prompt", value="Build a rocket to land on moon.")
                token_count = gr.Slider(10, 700, label="Max Tokens", step=10, value=700)
                temperature = gr.Slider(0.2, 2.0, label="Temperature", step=0.1, value=1.0)
                top_p = gr.Slider(0.0, 1.0, label="Top P", step=0.05, value=0.3)
                presence_penalty = gr.Slider(0.0, 1.0, label="Presence Penalty", step=0.1, value=0)
                count_penalty = gr.Slider(0.0, 1.0, label="Count Penalty", step=0.1, value=1)
            with gr.Column():
                with gr.Row():
                    submit = gr.Button("Submit", variant="primary")
                    clear = gr.Button("Clear", variant="secondary")
                output = gr.Textbox(label="Output", lines=50)
        data = gr.Dataset(components=[prompt, token_count, temperature, top_p, presence_penalty, count_penalty], samples=examples, samples_per_page=50, label="Example Instructions", headers=["Prompt", "Max Tokens", "Temperature", "Top P", "Presence Penalty", "Count Penalty"])
        submit.click(evaluate, [prompt, token_count, temperature, top_p, presence_penalty, count_penalty], [output])
        clear.click(lambda: None, [], [output])
        data.click(lambda x: x, [data], [prompt, token_count, temperature, top_p, presence_penalty, count_penalty])

demo.queue(concurrency_count=1, max_size=10)
demo.launch(share=False)