app.py

import torch
from transformers import AutoProcessor, AutoModelForVision2Seq
from huggingface_hub import login
import gradio as gr
import os
import gc

# ----------------------------
# AUTHENTICATION
# ----------------------------
HF_TOKEN = os.getenv("HF_TOKEN")
if HF_TOKEN:
    login(token=HF_TOKEN)
else:
    print("No HF_TOKEN found. Please log in manually.")
    login()

# ----------------------------
# CONFIG
# ----------------------------
MODEL_NAME = "reverseforward/inferencemodel"
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
DTYPE = torch.float16

# Clear cache before loading
gc.collect()
if DEVICE == "cuda":
    torch.cuda.empty_cache()

# ----------------------------
# LOAD MODEL (with error handling)
# ----------------------------
print(f"Loading model on {DEVICE}...")
try:
    model = AutoModelForVision2Seq.from_pretrained(
        MODEL_NAME,
        torch_dtype=DTYPE,
        device_map="auto",
        token=HF_TOKEN,
        low_cpu_mem_usage=True,  # Reduce memory usage
    )
    processor = AutoProcessor.from_pretrained(
        MODEL_NAME,
        token=HF_TOKEN,
    )
    print("✓ Model loaded successfully.")
except Exception as e:
    print(f"✗ Error loading model: {e}")
    raise

# ----------------------------
# INFERENCE FUNCTION
# ----------------------------
def chat_with_image(image, text):
    try:
        if image is None or text.strip() == "":
            return "Please provide both an image and text input."

        # Clear memory before inference
        gc.collect()
        if DEVICE == "cuda":
            torch.cuda.empty_cache()

        # Prepare inputs
        inputs = processor(
            text=[text],
            images=[image],
            return_tensors="pt"
        ).to(DEVICE, DTYPE)

        # Generate output
        with torch.inference_mode():
            generated_ids = model.generate(
                **inputs,
                max_new_tokens=256,
                temperature=0.7,
                do_sample=True,
            )

        output = processor.batch_decode(
            generated_ids,
            skip_special_tokens=True
        )[0]
        
        # Clean up
        del inputs, generated_ids
        gc.collect()
        
        return output.strip()
    
    except Exception as e:
        return f"Error during inference: {str(e)}"


# ----------------------------
# GRADIO UI
# ----------------------------
title = "🧠 Qwen3-VL-8B Fine-tuned (Image + Text)"
description = """
Upload an image and enter a text prompt.  
The model will reason visually and respond.
"""

demo = gr.Interface(
    fn=chat_with_image,
    inputs=[
        gr.Image(type="pil", label="Upload Image"),
        gr.Textbox(label="Enter Instruction or Question", lines=3),
    ],
    outputs=gr.Textbox(label="Model Output", lines=5),
    title=title,
    description=description,

    allow_flagging="never",  # Disable flagging to reduce overhead
)

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