app.py

import gradio as gr
import os
import spaces
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
from threading import Thread
import torch

# Set an environment variable
HF_TOKEN = os.environ.get("HF_TOKEN", None)

DESCRIPTION = '''
<div>
<h1 style="text-align: center;">Mistral 8B Instruct</h1>
</div>
'''

LICENSE = """
<p/>
---
"""

PLACEHOLDER = """
<div style="padding: 30px; text-align: center; display: flex; flex-direction: column; align-items: center;">
   <h1 style="font-size: 28px; margin-bottom: 2px; opacity: 0.55;">Mistral-8B</h1>
   <p style="font-size: 18px; margin-bottom: 2px; opacity: 0.65;">Ask me anything...</p>
</div>
"""

css = """
h1 {
  text-align: center;
  display: block;
}

#duplicate-button {
  margin: auto;
  color: white;
  background: #1565c0;
  border-radius: 100vh;
}
"""

# Load the tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("mistralai/Ministral-8B-Instruct-2410")
model = AutoModelForCausalLM.from_pretrained("mistralai/Ministral-8B-Instruct-2410", device_map="auto")

# Ensure we have a pad token
if tokenizer.pad_token_id is None:
    tokenizer.pad_token_id = tokenizer.eos_token_id

terminators = [
    tokenizer.eos_token_id,
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

@spaces.GPU(duration=120)
def chat_mistral(message: str, 
                 history: list, 
                 temperature: float, 
                 top_p: float, 
                 max_new_tokens: int,
                 system_prompt: str) -> str:
    """
    Generate a streaming response using the Mistral-8B model.
    Args:
        message (str): The input message.
        history (list): The conversation history used by ChatInterface.
        temperature (float): The temperature for generating the response.
        top_p (float): The top-p (nucleus) sampling parameter.
        max_new_tokens (int): The maximum number of new tokens to generate.
        system_prompt (str): The system prompt to guide the assistant's behavior.
    Returns:
        str: The generated response.
    """
    conversation = []

    # Format system prompt correctly using [INST]
    if system_prompt:
        formatted_prompt = f"[INST] {system_prompt} [/INST]\n\n"
    else:
        formatted_prompt = ""

    # Modify first user message to include system prompt
    if history:
        first_user_msg = f"{formatted_prompt}{history[0][0]}" if formatted_prompt else history[0][0]
        conversation.append({"role": "user", "content": first_user_msg})
        conversation.append({"role": "assistant", "content": history[0][1]})

        for user, assistant in history[1:]:
            conversation.extend([{"role": "user", "content": user}, {"role": "assistant", "content": assistant}])
    else:
        # First message in a new conversation
        first_message = f"{formatted_prompt}{message}" if formatted_prompt else message
        conversation.append({"role": "user", "content": first_message})

    # Tokenize with padding and attention mask
    input_data = tokenizer.apply_chat_template(conversation, return_tensors="pt", padding=True, truncation=True)
    input_ids = input_data.to(model.device)

    attention_mask = input_ids.ne(tokenizer.pad_token_id).to(dtype=torch.long, device=model.device)

    streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)

    generate_kwargs = dict(
        input_ids=input_ids,
        attention_mask=attention_mask,  # Fixes the warning
        streamer=streamer,
        max_new_tokens=max_new_tokens,
        do_sample=True,
        temperature=temperature,
        top_p=top_p,
        pad_token_id=tokenizer.pad_token_id,  # Explicitly set
        eos_token_id=terminators,
    )

    if temperature == 0:
        generate_kwargs['do_sample'] = False
        
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()

    outputs = []
    for text in streamer:
        outputs.append(text)
        yield "".join(outputs)
        

# Gradio block
chatbot = gr.Chatbot(height=450, placeholder=PLACEHOLDER, label='Gradio ChatInterface')

with gr.Blocks(fill_height=True, css=css) as demo:
    
    gr.Markdown(DESCRIPTION)

    system_prompt_input = gr.Textbox(
        label="System Prompt",
        placeholder="Enter system instructions for the model...",
        lines=2
    )

    gr.ChatInterface(
        fn=chat_mistral,
        chatbot=chatbot,
        fill_height=True,
        additional_inputs_accordion=gr.Accordion(label="⚙️ Parameters", open=False, render=False),
        additional_inputs=[
            system_prompt_input,
            gr.Slider(minimum=0, maximum=1, step=0.1, value=0.8, label="Temperature", render=False),
            gr.Slider(minimum=0, maximum=1, step=0.1, value=0.9, label="Top-p", render=False),
            gr.Slider(minimum=128, maximum=4096, step=1, value=4096, label="Max new tokens", render=False),
        ],
        examples=[
            ['How to setup a human base on Mars? Give short answer.'],
            ['Explain theory of relativity to me like I’m 8 years old.'],
            ['What is 9,000 * 9,000?'],
            ['Write a pun-filled happy birthday message to my friend Alex.'],
            ['Justify why a penguin might make a good king of the jungle.']
        ],
        cache_examples=False
    )

if __name__ == "__main__":
    demo.launch()