app.py

import gradio as gr
import spaces
from transformers import AutoModelForCausalLM, AutoProcessor, GPT2LMHeadModel, GPT2Tokenizer
import torch
from PIL import Image
import subprocess

# Install flash-attn with no CUDA build isolation
subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)

# Define models and processors with pinning to a stable revision
models = {
    "microsoft/Phi-3.5-vision-instruct": AutoModelForCausalLM.from_pretrained(
        "microsoft/Phi-3.5-vision-instruct",
        revision="specific-revision-hash",  # Pinning to a specific revision for stability
        trust_remote_code=True,
        torch_dtype="auto", 
        _attn_implementation="flash_attention_2"
    ).cuda().eval()
}

processors = {
    "microsoft/Phi-3.5-vision-instruct": AutoProcessor.from_pretrained(
        "microsoft/Phi-3.5-vision-instruct",
        revision="specific-revision-hash",  # Pinning to a specific revision for stability
        trust_remote_code=True
    )
}

# Fallback to GPT-2 for testing
def load_fallback_model():
    tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
    model = GPT2LMHeadModel.from_pretrained("gpt2").cuda().eval()
    return model, tokenizer

# Default description and prompt
DESCRIPTION = "[Phi-3.5-vision Demo](https://huggingface.co/microsoft/Phi-3.5-vision-instruct)"
default_question = "You are an image to prompt converter. Your work is to observe each and every detail of the image and craft a detailed prompt under 100 words."

# Gradio function for generating output from image input with error handling
@spaces.GPU
def run_example(image, text_input=default_question, model_id="microsoft/Phi-3.5-vision-instruct"):
    try:
        model = models[model_id]
        processor = processors[model_id]
    except KeyError as e:
        print(f"Error loading model: {e}. Falling back to GPT-2.")
        model, processor = load_fallback_model()
    
    user_prompt = '<|user|>\n'
    assistant_prompt = '<|assistant|>\n'
    prompt_suffix = "<|end|>\n"
    
    prompt = f"{user_prompt}<|image_1|>\n{text_input}{prompt_suffix}{assistant_prompt}"
    image = Image.fromarray(image).convert("RGB")

    inputs = processor(prompt, image, return_tensors="pt").to("cuda:0")
    generate_ids = model.generate(**inputs, max_new_tokens=1000, eos_token_id=processor.tokenizer.eos_token_id)
    generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
    response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
    return response

# Custom CSS for styling
css = """
  #output_text {
    height: 500px;
    overflow: auto;
    border: 1px solid #333;
  }
  #model_selector, #text_input {
    display: none !important;
  }
  #main_container {
    border: 2px solid black;
    padding: 20px;
    border-radius: 10px;
  }
"""

# Gradio interface with styling and layout improvements
with gr.Blocks(css=css) as demo:
    gr.Markdown(DESCRIPTION)
    with gr.Row(id="main_container"):
        with gr.Column():
            input_img = gr.Image(label="Input Image", interactive=True)
            model_selector = gr.Dropdown(choices=list(models.keys()), label="Model", value="microsoft/Phi-3.5-vision-instruct", visible=False)
            text_input = gr.Textbox(label="Question", value=default_question, visible=False)
            submit_btn = gr.Button(value="Generate Prompt")
        
        output_text = gr.Textbox(label="Output", id="output_text", interactive=False)

    # Link button action to function
    submit_btn.click(run_example, [input_img, text_input, model_selector], output_text)

# Launch Gradio interface
demo.queue(api_open=False)
demo.launch(debug=True, show_api=False)