app.py

import gradio as gr
from transformers import AutoModel, AutoTokenizer
import torch
import threading
import os

# caching the mode 
model_cache = {}
tokenizer_cache = {}
model_lock = threading.Lock()

from huggingface_hub import login
hf_token = os.environ.get('hf_token', None)


# Define the models and their paths
model_paths = {
    "H2OVL-Mississippi-2B":"h2oai/h2ovl-mississippi-2b",
    "H2OVL-Mississippi-0.8B":"h2oai/h2ovl-mississippi-800m",
    # Add more models as needed
}


example_prompts = [
    "Read the text and provide word by word ocr for the document. <doc>",
    "Extract the text from the image.",
    "Extract the text from the image and fill the following json {'license_number':'',\n'full_name':'',\n'date_of_birth':'',\n'address':'',\n'issue_date':'',\n'expiration_date':'',\n}",
    "Please extract the following fields, and return the result in JSON format: supplier_name, supplier_address, customer_name, customer_address, invoice_number, invoice_total_amount, invoice_tax_amount",     
]


def load_model_and_set_image_function(model_name):
    # Get the model path from the model_paths dictionary
    model_path = model_paths[model_name]
    
    
    with model_lock:
        if model_name in model_cache:
            # model is already loaded; retrieve it from the cache
            print(f"Model {model_name} is already loaded. Retrieving from cache.")
            
        else:
            # load the model and tokenizer
            print(f"Loading model {model_name}...")

            model = AutoModel.from_pretrained(
                model_path,
                torch_dtype=torch.bfloat16,
                low_cpu_mem_usage=True,
                trust_remote_code=True,
                use_auth_token=hf_token,
                # device_map="auto"
            ).eval().cuda()

            tokenizer = AutoTokenizer.from_pretrained(
                model_path,
                trust_remote_code=True,
                use_fast=False,
                use_auth_token=hf_token
            )
            
            # add the model and tokenizer to the cache
            model_cache[model_name] = model
            tokenizer_cache[model_name] = tokenizer
            print(f"Model {model_name} loaded successfully.")
            

    return model_name
    

def inference(image_input, 
              user_message,
              temperature, 
              top_p, 
              max_new_tokens, 
              tile_num,
              chatbot,
              state, 
              model_name):
    
    # Check if model_state is None
    if model_name is None:
        chatbot.append(("System", "Please select a model to start the conversation."))
        return chatbot, state, ""   
    
    with model_lock:
        if model_name not in model_cache:
            chatbot.append(("System", "Model not loaded. Please wait for the model to load."))
            return chatbot, state, ""
        model = model_cache[model_name]
        tokenizer = tokenizer_cache[model_name]

    # Check for empty or invalid user message
    if not user_message or user_message.strip() == '' or user_message.lower() == 'system':
        chatbot.append(("System", "Please enter a valid message to continue the conversation."))
        return chatbot, state, ""
    
    
    # if image is provided, store it in image_state:
    if chatbot is None:
        chatbot = []
        
    if image_input is None:
        chatbot.append(("System", "Please provide an image to start the conversation."))
        return chatbot, state, ""
        
    # Initialize history (state) if it's None
    if state is None:
        state = None  # model.chat function handles None as empty history        

    # Append user message to chatbot
    chatbot.append((user_message, None))

    # Set generation config
    do_sample = (float(temperature) != 0.0)    


    generation_config = dict(
        num_beams=1,
        max_new_tokens=int(max_new_tokens),
        do_sample=do_sample,
        temperature= float(temperature),
        top_p= float(top_p),
    )

    # Call model.chat with history
    response_text, new_state = model.chat(
        tokenizer,
        image_input,
        user_message,
        max_tiles = int(tile_num),
        generation_config=generation_config,
        history=state,
        return_history=True
    )
    
    # update the satet with new_state
    state = new_state
    # Update chatbot with the model's response
    chatbot[-1] = (user_message, response_text)    
    
    return chatbot, state, ""

def regenerate_response(chatbot, 
                        temperature, 
                        top_p, 
                        max_new_tokens, 
                        tile_num,
                        state, 
                        image_input,
                        model_name):
    
    # Check if model_state is None
    if model_name is None:
        chatbot.append(("System", "Please select a model to start the conversation."))
        return chatbot, state
    
    
    with model_lock:
        if model_name not in model_cache:
            chatbot.append(("System", "Model not loaded. Please wait for the model to load."))
            return chatbot, state
        model = model_cache[model_name]
        tokenizer = tokenizer_cache[model_name]
        
    # Check if there is a previous user message
    if chatbot is None or len(chatbot) == 0:
        chatbot = []
        chatbot.append(("System", "Nothing to regenerate. Please start a conversation first."))
        return chatbot, state, 
    
    # Get the last user message
    last_user_message, _ = chatbot[-1]
    
    # Check for empty or invalid last user message
    if not last_user_message or last_user_message.strip() == '' or last_user_message.lower() == 'system':
        chatbot.append(("System", "Cannot regenerate response for an empty or invalid message."))
        return chatbot, state
    
    # Remove last assistant's response from state
    if state is not None and len(state) > 0:
        state = state[:-1]  # Remove last assistant's response from history
        if len(state) == 0:
            state = None
    else:
        state = None
   
    # Set generation config
    do_sample = (float(temperature) != 0.0)    

    generation_config = dict(
        num_beams=1,
        max_new_tokens=int(max_new_tokens),
        do_sample=do_sample,
        temperature= float(temperature),
        top_p= float(top_p),
    )
    
    
    # Regenerate the response
    response_text, new_state = model.chat(
        tokenizer,
        image_input,
        last_user_message,
        max_tiles = int(tile_num),
        generation_config=generation_config,
        history=state,  # Exclude last assistant's response
        return_history=True
    )
    
    # Update the state with new_state
    state = new_state
    
    # Update chatbot with the regenerated response
    chatbot[-1] = (last_user_message, response_text)
       
    return chatbot, state


def clear_all():
    return [], None, None, ""  # Clear chatbot, state, reset image_input

# Build the Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("# **H2OVL-Mississippi**")
    
    state= gr.State()
    model_state = gr.State()
    # tokenizer_state = gr.State()
    # image_load_function_state = gr.State()

    with gr.Row():
        model_dropdown = gr.Dropdown(
            choices=list(model_paths.keys()),
            label="Select Model",
            value="H2OVL-Mississippi-2B"
        )

    
    with gr.Row(equal_height=True):
        # First column with image input
        with gr.Column(scale=1):
            image_input = gr.Image(type="filepath", label="Upload an Image")
             
     
        # Second column with chatbot and user input
        with gr.Column(scale=2):    
            chatbot = gr.Chatbot(label="Conversation")
            user_input = gr.Dropdown(label="What is your question", 
                                    choices = example_prompts,
                                    value=None,
                                    allow_custom_value=True,
                                    interactive=True)
            
            
    def reset_chatbot_state():
        # reset chatbot and state
        return [], None
    
    
    # When the model selection changes, load the new model
    model_dropdown.change(
        fn=load_model_and_set_image_function,
        inputs=[model_dropdown],
        outputs=[model_state]
    )
    
    model_dropdown.change(
        fn=reset_chatbot_state,
        inputs=None,
        outputs=[chatbot, state]
    )
    
    
    # Reset chatbot and state when image input changes
    image_input.change(
        fn=reset_chatbot_state,
        inputs=None,
        outputs=[chatbot, state]
    )
    
        # Load the default model when the app starts
    demo.load(
        fn=load_model_and_set_image_function,
        inputs=[model_dropdown],
        outputs=[model_state]
    )
    

        
    with gr.Accordion('Parameters', open=False):
        with gr.Row():
            temperature_input = gr.Slider(
                minimum=0.0, 
                maximum=1.0, 
                step=0.1, 
                value=0.2, 
                interactive=True,
                label="Temperature")
            top_p_input = gr.Slider(
                minimum=0.0, 
                maximum=1.0, 
                step=0.1, 
                value=0.9,
                interactive=True, 
                label="Top P")
            max_new_tokens_input = gr.Slider(
                minimum=0, 
                maximum=4096, 
                step=64, 
                value=1024, 
                interactive=True,
                label="Max New Tokens (default: 1024)")
            tile_num = gr.Slider(
                minimum=2, 
                maximum=12, 
                step=1, 
                value=6, 
                interactive=True,
                label="Tile Number (default: 6)"
            )
            
    with gr.Row():
        submit_button = gr.Button("Submit")
        regenerate_button = gr.Button("Regenerate")
        clear_button = gr.Button("Clear")        

    # When the submit button is clicked, call the inference function
    submit_button.click(
        fn=inference, 
        inputs=[
            image_input, 
            user_input, 
            temperature_input, 
            top_p_input, 
            max_new_tokens_input, 
            tile_num,
            chatbot, 
            state, 
            model_state
        ], 
        outputs=[chatbot, state, user_input]
    )
    # When the regenerate button is clicked, re-run the last inference
    regenerate_button.click(
        fn=regenerate_response,
        inputs=[
            chatbot, 
            temperature_input, 
            top_p_input,
            max_new_tokens_input, 
            tile_num,
            state,
            image_input, 
            model_state
            ],
        outputs=[chatbot, state]
    )

    clear_button.click(
        fn=clear_all, 
        inputs=None, 
        outputs=[chatbot, state, image_input, user_input]
    )  
    
    def example_clicked(image_value, user_input_value):
        chatbot_value, state_value = [], None
        return image_value, user_input_value, chatbot_value, state_value  # Reset chatbot and state
    
                  
    gr.Examples(
        examples=[
            ["assets/driver_license.png", "Extract the text from the image and fill the following json {'license_number':'',\n'full_name':'',\n'date_of_birth':'',\n'address':'',\n'issue_date':'',\n'expiration_date':'',\n}"],
            ["assets/handwritten-note-example.jpg", "Read the text on the image"],
            ["assets/invoice.png", "Please extract the following fields, and return the result in JSON format: supplier_name, supplier_address, customer_name, customer_address, invoice_number, invoice_total_amount, invoice_tax_amount"],
            ["assets/CBA-1H23-Results-Presentation_wheel.png", "What is the efficiency of H2O.AI in document processing?"],
        ],
        inputs = [image_input, user_input],
        outputs = [image_input, user_input, chatbot, state],
        fn=example_clicked,
        label = "examples",
    )
demo.queue()   
demo.launch(max_threads=10)