app.py

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

# Update model configuration for Mistral-small-24B
MODEL_ID = "mistralai/Mistral-24B-v0.1"
CHAT_TEMPLATE = "mistral"  # Mistral uses its own chat template
MODEL_NAME = MODEL_ID.split("/")[-1]
CONTEXT_LENGTH = 32768  # Mistral supports longer context
COLOR = "black"
EMOJI = "🌪️"  # Mistral-themed emoji
DESCRIPTION = f"This is {MODEL_NAME} model, a powerful 24B parameter language model from Mistral AI."

def load_system_message():
    try:
        with open('system_message.txt', 'r', encoding='utf-8') as file:
            return file.read().strip()
    except FileNotFoundError:
        print("Warning: system_message.txt not found. Using default message.")
        return "You are a helpful assistant. First recognize the user request and then reply carefully with thinking."
    except Exception as e:
        print(f"Error loading system message: {e}")
        return "You are a helpful assistant. First recognize the user request and then reply carefully with thinking."

SYSTEM_MESSAGE = load_system_message()

@spaces.GPU()
def predict(message, history, system_prompt, temperature, max_new_tokens, top_k, repetition_penalty, top_p):
    # Format history using Mistral's chat template
    messages = [{"role": "system", "content": SYSTEM_MESSAGE}]
    
    for user, assistant in history:
        messages.append({"role": "user", "content": user})
        messages.append({"role": "assistant", "content": assistant})
    
    messages.append({"role": "user", "content": message})
    
    # Convert messages to Mistral format
    prompt = tokenizer.apply_chat_template(messages, tokenize=False)
    
    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
    enc = tokenizer(prompt, return_tensors="pt", padding=True, truncation=True)
    input_ids, attention_mask = enc.input_ids, enc.attention_mask

    if input_ids.shape[1] > CONTEXT_LENGTH:
        input_ids = input_ids[:, -CONTEXT_LENGTH:]
        attention_mask = attention_mask[:, -CONTEXT_LENGTH:]

    generate_kwargs = dict(
        input_ids=input_ids.to(device),
        attention_mask=attention_mask.to(device),
        streamer=streamer,
        do_sample=True,
        temperature=temperature,
        max_new_tokens=max_new_tokens,
        top_k=top_k,
        repetition_penalty=repetition_penalty,
        top_p=top_p
    )

    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()

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

# Load model with optimized settings for Mistral-24B
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    use_double_quant=True,  # Enable double quantization
    bnb_4bit_quant_type="nf4"  # Use normal float 4 for better precision
)

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    device_map="auto",
    quantization_config=quantization_config,
    use_flash_attention_2=True,  # Enable Flash Attention 2 for better performance
    torch_dtype=torch.bfloat16
)

# Create Gradio interface
gr.ChatInterface(
    predict,
    title=EMOJI + " " + MODEL_NAME,
    description=DESCRIPTION,
    examples=[
        ['What are the key differences between classical and quantum computing?'],
        ['Explain the concept of recursive neural networks in simple terms.'],
        ['How does transfer learning work in large language models?'],
        ['What are the ethical considerations in AI development?']
    ],
    additional_inputs_accordion=gr.Accordion(label="⚙️ Parameters", open=False),
    additional_inputs=[
        gr.Textbox(SYSTEM_MESSAGE, label="System prompt", visible=False),  # Hidden system prompt
        gr.Slider(0, 1, 0.7, label="Temperature"),  # Adjusted default for Mistral
        gr.Slider(0, 32768, 12000, label="Max new tokens"),  # Increased for longer context
        gr.Slider(1, 100, 50, label="Top K sampling"),
        gr.Slider(0, 2, 1.1, label="Repetition penalty"),
        gr.Slider(0, 1, 0.95, label="Top P sampling"),
    ],
).queue().launch()