app.py

import gradio as gr
import torch
import threading
import spaces

from transformers import AutoTokenizer, TextIteratorStreamer
from auto_gptq import AutoGPTQForCausalLM

# Model identifier
model_id = "jncraton/SmolLM2-1.7B-Instruct-ct2-int8"

# Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False, trust_remote_code=True)

print("Is CUDA available?", torch.cuda.is_available())

class ModelWrapper:
    def __init__(self):
        self.model = None  # Model will be loaded when GPU is allocated

    @spaces.GPU
    def generate(self, prompt):
        if self.model is None:
            # Explicitly set device_map to 'cuda'
            self.model = AutoGPTQForCausalLM.from_quantized(
                model_id,
                device_map={'': 'cuda:0'},
                trust_remote_code=True,
            )
            self.model.eval()

        print("Model is on device:", next(self.model.parameters()).device)

        # Tokenize the input prompt
        inputs = tokenizer(prompt, return_tensors='pt').to('cuda')
        print("Inputs are on device:", inputs['input_ids'].device)

        # Set up the streamer
        streamer = TextIteratorStreamer(tokenizer, skip_special_tokens=True)

        # Prepare generation arguments
        generation_kwargs = dict(
            **inputs,
            streamer=streamer,
            do_sample=True,
            max_new_tokens=512,
        )

        # Start generation in a separate thread to enable streaming
        thread = threading.Thread(target=self.model.generate, kwargs=generation_kwargs)
        thread.start()

        # Yield generated text in real-time
        generated_text = ""
        for new_text in streamer:
            generated_text += new_text
            yield generated_text

# Instantiate the model wrapper
model_wrapper = ModelWrapper()

# Create the Gradio interface
interface = gr.Interface(
    fn=model_wrapper.generate,
    inputs=gr.Textbox(lines=5, label="Input Prompt"),
    outputs=gr.Textbox(label="Generated Text", lines=10),
    title="Mistral-Large-Instruct-2407 Text Completion",
    description="Enter a prompt and receive a text completion using the Mistral-Large-Instruct-2407 INT4 model.",
    allow_flagging='never',
    live=False,
    cache_examples=False
)

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