model.py

import os
import torch
from huggingface_hub import login
from transformers import AutoProcessor, AutoModelForImageTextToText

# Login with your secret token
login(token=os.environ["HF_TOKEN"])

MODEL_ID = "google/medgemma-1.5-4b-it"

print("Loading MedGemma... this may take a few minutes")

processor = AutoProcessor.from_pretrained(MODEL_ID)

use_cuda = torch.cuda.is_available()
dtype = torch.bfloat16 if use_cuda else torch.float32
device = "cuda:0" if use_cuda else "cpu"

model = AutoModelForImageTextToText.from_pretrained(
    MODEL_ID,
    torch_dtype=dtype,
    device_map=device
)

model.generation_config.pad_token_id = processor.tokenizer.eos_token_id
model.eval()

print("MedGemma loaded successfully!")
print(f"MedGemma loaded on: {device}")

def generate_report(image):
    """
    Takes a PIL Image, returns a generated radiology report string.
    """
    messages = [
        {
            "role": "user",
            "content": [
                {"type": "image", "image": image},
                {"type": "text", "text": """You are an experienced radiologist. 
                Analyze this brain MRI scan and generate a structured clinical report with the following sections:
                
                TECHNIQUE:
                FINDINGS:
                IMPRESSION:
                
                Be specific, professional and concise."""}
            ]
        }
    ]

    # Prepare inputs
    inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,        # ← returns a dict, not a raw Tensor
    return_tensors="pt"
    ).to(model.device)

    # Generate the report
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=512,
            do_sample=False        
        )

    # Decode only the newly generated tokens
    input_length = inputs["input_ids"].shape[1]
    generated_tokens = outputs[0][input_length:]
    report = processor.decode(generated_tokens, skip_special_tokens=True)

    # Clean up the report spacing
    report = report.replace("FINDINGS:", "\nFINDINGS:")
    report = report.replace("IMPRESSION:", "\nIMPRESSION:")
    report = report.replace("TECHNIQUE:", "\nTECHNIQUE:")

    return report