app.py

import gradio as gr
import torch
from PIL import Image
from unsloth import FastVisionModel
from peft import PeftModel

# -----------------------------------------------------------------------------
# 1. Load Model Correctly (No Double Patching)
# -----------------------------------------------------------------------------


# Load model with 4-bit quantization
model, tokenizer = FastVisionModel.from_pretrained(
    "unsloth/Llama-3.2-11B-Vision-Instruct",
    load_in_4bit = True,
    device_map = "auto",
)

# -----------------------------------------------------------------------------
# 2. Load Your Fine-Tuned Adapter CORRECTLY
# -----------------------------------------------------------------------------

# Loading adapter WITHOUT get_peft_model()
model = PeftModel.from_pretrained(model, "fine_tuned_model")
model = model.merge_and_unload()  # Merge adapters into base model
model.to("cuda")
model.eval()

# -----------------------------------------------------------------------------
# 3. Data preprocessing step
# -----------------------------------------------------------------------------

def analyze(image, user_prompt):
    if image.mode != "RGB":
        image = image.convert("RGB")

    messages = [
        {"role": "user", "content": [
            {"type": "image", "image": image},
            {"type": "text", "text": user_prompt}
        ]}
    ]

    input_text = tokenizer.apply_chat_template(messages, add_generation_prompt=True)

    inputs = tokenizer(
        image,
        input_text,
        return_tensors = "pt",
        add_special_tokens = False,
    ).to("cuda")

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens = 512,
            use_cache = True,
            temperature = 1.0,
            min_p = 0.1,
        )

    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# -----------------------------------------------------------------------------
# 4. Launch Interface
# -----------------------------------------------------------------------------

gr.Interface(
    fn=analyze,
    inputs=[
        gr.Image(type="pil", label="Upload Medical Scan"),
        gr.Textbox(
            placeholder="Example: 'Describe any abnormalities in this chest X-ray'",
            label="Your Question",
            lines=2
        )
    ],
    outputs=gr.Textbox(label="Expert Analysis"),
    title=" DAS medhub Radiology AI Assistant (Fine-Tuned)",
    description="Upload a medical image and ask questions about it"
).launch(server_port=7860, debug=False)