app.py

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

MODEL_NAME = "Declan1/llava-v1.6-mistral-7b-sydneyfish-a100"

# Load processor (handles image + text formatting)
processor = AutoProcessor.from_pretrained(MODEL_NAME)

# Load model (Vision + Language fused)
model = LlavaForConditionalGeneration.from_pretrained(
    MODEL_NAME,
    torch_dtype=torch.float16,
    device_map="auto"
)

def chat(image, prompt):
    # Process inputs
    inputs = processor(prompt, images=image, return_tensors="pt").to(model.device)
    
    # Generate output
    output = model.generate(**inputs, max_new_tokens=300)
    
    # Decode to readable text
    response = processor.decode(output[0], skip_special_tokens=True)
    return response

# Simple UI
interface = gr.Interface(
    fn=chat,
    inputs=[gr.Image(type="pil", label="Image"), gr.Textbox(label="Prompt")],
    outputs=gr.Textbox(label="Response"),
    title="SydneyFish LLaVA (Mistral 7B)",
    description="Upload an image and ask it something."
)

interface.launch()