app.py

import os
import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig, TextIteratorStreamer
import torch
from threading import Thread
from huggingface_hub import Repository
import json

theme = gr.themes.Monochrome(
    primary_hue="indigo",
    secondary_hue="blue",
    neutral_hue="slate",
    radius_size=gr.themes.sizes.radius_sm,
    font=[gr.themes.GoogleFont("Open Sans"), "ui-sans-serif", "system-ui", "sans-serif"],
)
os.environ["TOKENIZERS_PARALLELISM"] = "false"

# Load peft config for pre-trained checkpoint etc.
device = "cuda" if torch.cuda.is_available() else "cpu"
model_id = "SebastianSchramm/Cerebras-GPT-111M-instruction"
if device == "cpu":
    model = AutoModelForCausalLM.from_pretrained(model_id)
else:
    model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id)

prompt_template = "Below is an instruction that describes a task, paired with an input that provides further context.\n" \
                  "Write a response that appropriately completes the request.\n\n" \
                  "### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:"


def generate(instruction, input='', temperature=1.0, max_new_tokens=256, top_p=0.9, length_penalty=1.0):
    formatted_instruction = prompt_template.format(instruction=instruction, input=input)
    
    # make sure temperature top_p and length_penalty are floats
    temperature = float(temperature)
    top_p = float(top_p)
    length_penalty = float(length_penalty)
    
    # STREAMING BASED ON git+https://github.com/gante/transformers.git@streamer_iterator

    # streaming
    streamer = TextIteratorStreamer(tokenizer)
    model_inputs = tokenizer(formatted_instruction, return_tensors="pt", truncation=True, max_length=2048)
    # move to gpu
    model_inputs = {k: v.to(device) for k, v in model_inputs.items()}

    generate_kwargs = dict(
        top_p=top_p,
        top_k=0,
        temperature=temperature,
        do_sample=True,
        max_new_tokens=max_new_tokens,
        early_stopping=True,
        length_penalty=length_penalty,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.eos_token_id,
    )
    t = Thread(target=model.generate, kwargs={**dict(model_inputs, streamer=streamer), **generate_kwargs})
    t.start()

    output = ""
    hidden_output = ""
    for new_text in streamer:
        # skip streaming until new text is available
        if len(hidden_output) <= len(formatted_instruction):
            hidden_output += new_text
            continue
        # replace eos token
        if tokenizer.eos_token in new_text:
            new_text = new_text.replace(tokenizer.eos_token, "")
        output += new_text
        yield output
    return output

examples = []

def process_example(args):
    for x in generate(args):
        pass
    return x

with gr.Blocks(theme=theme) as demo:
    with gr.Column():
        gr.Markdown(
            """<h1><center>Instruction-tuned Cerebras GPT 111M Language Model for Text</center></h1>
            <p>
            Link to model: [Cerebras-GPT-111M-instruction](SebastianSchramm/Cerebras-GPT-111M-instruction)
            </p>
      """
        )
        with gr.Row():
            with gr.Column(scale=3):
                instruction = gr.Textbox(placeholder="Instruction...", label="Instruction")
                input = gr.Textbox(placeholder="Input...", label="Input")
                output = gr.Textbox(
                    interactive=False,
                    lines=8,
                    label="Response",
                    placeholder="Response will be shown here...",
                )
                submit = gr.Button("Generate", variant="primary")
                gr.Examples(
                    examples=examples,
                    inputs=[instruction, input],
                    cache_examples=True,
                    fn=process_example,
                    outputs=[output],
                )

            with gr.Column(scale=1):
                temperature = gr.Slider(
                    label="Temperature",
                    value=1.0,
                    minimum=0.01,
                    maximum=1.0,
                    step=0.1,
                    interactive=True,
                    info="The higher more random",
                )
                max_new_tokens = gr.Slider(
                    label="Max new tokens",
                    value=256,
                    minimum=0,
                    maximum=2048,
                    step=5,
                    interactive=True,
                    info="The maximum numbers of new tokens",
                )
                top_p = gr.Slider(
                    label="Top p",
                    value=0.9,
                    minimum=0.01,
                    maximum=1,
                    step=0.05,
                    interactive=True,
                    info="probabilities that add up are kept",
                )
                length_penalty = gr.Slider(
                    label="Length penalty",
                    value=1.0,
                    minimum=-10.0,
                    maximum=10.0,
                    step=0.1,
                    interactive=True,
                    info="> 0.0 longer, < 0.0 shorter",
                )

    submit.click(generate, inputs=[instruction, input, temperature, max_new_tokens, top_p, length_penalty], outputs=[output])
    instruction.submit(
        generate, inputs=[instruction, input, temperature, max_new_tokens, top_p, length_penalty], outputs=[output]
    )

demo.queue(concurrency_count=1)
demo.launch(enable_queue=True)