app.py

import torch
from transformers import AutoProcessor, AutoModelForVision2Seq
import gradio as gr

# ----------------------------
# CONFIG
# ----------------------------
MODEL_NAME = "reverseforward/qwenmeasurement"  # change this to your repo name
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
DTYPE = torch.float16  # use float16 on A10G

# ----------------------------
# LOAD MODEL
# ----------------------------
print("Loading model...")
model = AutoModelForVision2Seq.from_pretrained(
    MODEL_NAME,
    torch_dtype=DTYPE,
    device_map="auto",
)
processor = AutoProcessor.from_pretrained(MODEL_NAME)
print("Model loaded successfully.")

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

    # Prepare inputs for Qwen3-VL
    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,
        )

    output = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
    return output.strip()


# ----------------------------
# 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"),
    ],
    outputs=gr.Textbox(label="Model Output"),
    title=title,
    description=description,
    examples=[
        ["examples/cat.jpg", "Describe this image."],
        ["examples/room.jpg", "How many chairs are visible?"],
    ],
)

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