app.py

import gradio as gr
from huggingface_hub import InferenceClient
import spaces

import os
import warnings
import shutil
import time
from threading import Thread

from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig, AutoProcessor
from transformers import TextIteratorStreamer
import torch
from dc.model import *
from dc.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
from dc.conversation import conv_templates, SeparatorStyle
from PIL import Image
from dc.mm_utils import tokenizer_image_token, process_images, get_model_name_from_path




tokenizer = AutoTokenizer.from_pretrained('HuanjinYao/DenseConnector-v1.5-8B', use_fast=False)
model = LlavaLlamaForCausalLM.from_pretrained('HuanjinYao/DenseConnector-v1.5-8B', low_cpu_mem_usage=True,torch_dtype=torch.float16)

vision_tower = model.get_vision_tower()
if not vision_tower.is_loaded:
    vision_tower.load_model()
vision_tower.to(device='cuda', dtype=torch.float16)
image_processor = vision_tower.image_processor

model.to('cuda')

# model.generation_config.eos_token_id = 128009
tokenizer.unk_token = "<|reserved_special_token_0|>"
tokenizer.pad_token = tokenizer.unk_token
terminators = [
    tokenizer.eos_token_id,
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
]


@spaces.GPU
def bot_streaming(message, history):
    print(message)
    if message["files"]:
        # message["files"][-1] is a Dict or just a string
        if type(message["files"][-1]) == dict:
            image = message["files"][-1]["path"]
        else:
            image = message["files"][-1]
    else:
        # if there's no image uploaded for this turn, look for images in the past turns
        # kept inside tuples, take the last one
        for hist in history:
            if type(hist[0]) == tuple:
                image = hist[0][0]
    try:
        if image is None:
            # Handle the case where image is None
            gr.Error("You need to upload an image for LLaVA to work.")
    except NameError:
        # Handle the case where 'image' is not defined at all
        gr.Error("You need to upload an image for LLaVA to work.")

    print('process end')
    print('history', history)

    
    conv = conv_templates['llama_3'].copy()
    if len(history) == 0:
        message['text'] = DEFAULT_IMAGE_TOKEN + '\n' + message['text']
    else:
        for idx, (user, assistant) in enumerate(history[1:]):
            # conversation.extend([{"role": "user", "content": user}, {"role": "assistant", "content": assistant}])
            if idx == 0:
                user = DEFAULT_IMAGE_TOKEN + '\n' + user
            conv.append_message(conv.roles[0], user)
            conv.append_message(conv.roles[1], assistant)
    conv.append_message(conv.roles[0], message['text'])
    conv.append_message(conv.roles[1], None)
    prompt = conv.get_prompt()
    
    print(prompt)

    image = Image.open(image).convert('RGB')
    image_tensor = process_images([image], image_processor, model.config)[0]

    

    inputs = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0)

    image_tensor = image_tensor.unsqueeze(0)
    image_tensor = image_tensor.to(dtype=torch.float16, device='cuda', non_blocking=True)
    inputs = inputs.to(device='cuda', non_blocking=True)

    print('image', image_tensor.shape)
    print('inputs', inputs.shape)

    
    streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
    
    generation_kwargs = dict(inputs=inputs, images=image_tensor, streamer=streamer, max_new_tokens=1024, do_sample=False, eos_token_id = terminators)


    thread = Thread(target=model.generate, kwargs=generation_kwargs)

    print('here?')
    
    thread.start()

    print('start')

    buffer = ""
    # time.sleep(0.5)
    for new_text in streamer:
        print('new_text', new_text)
        if "<|eot_id|>" in new_text:
            new_text = new_text.split("<|eot_id|>")[0]
        buffer += new_text

        generated_text_without_prompt = buffer
        # time.sleep(0.06)
        yield generated_text_without_prompt


chatbot=gr.Chatbot(label=f"Chat with Dense Connector")
chat_input = gr.MultimodalTextbox(interactive=True, file_types=["image"], placeholder="Enter message or upload file...", show_label=False)
with gr.Blocks(fill_height=True, ) as demo:
    gr.ChatInterface(
    fn=bot_streaming,
    title="DenseConnector-v1.5-8B",
    # examples=[{"text": "What is on the flower?", "files": ["./bee.jpg"]},
    #           {"text": "How to make this pastry?", "files": ["./baklava.png"]}],
    description="Try [DenseConnector-v1.5-8B](https://huggingface.co/HuanjinYao/DenseConnector-v1.5-8B). Upload an image and start chatting about it, or simply try one of the examples below. If you don't upload an image, you will receive an error.",
    stop_btn="Stop Generation",
    multimodal=True,
    textbox=chat_input,
    chatbot=chatbot,
    )
    gr.Image(label="Upload an image to start")


if __name__ == "__main__":
    demo.launch()