app.py

import gradio as gr

from transformers import AutoProcessor, LlavaForConditionalGeneration
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer, TextIteratorStreamer

from threading import Thread
import re
import time 
from PIL import Image
import torch
import spaces
import requests

model_id = "xtuner/llava-llama-3-8b-v1_1-transformers"

processor = AutoProcessor.from_pretrained(model_id)

model = LlavaForConditionalGeneration.from_pretrained(
    model_id, 
    torch_dtype=torch.float16, 
    low_cpu_mem_usage=True, 
)

model.to("cuda:0")
model.generation_config.eos_token_id = 128009

@spaces.GPU
def bot_streaming(message, history):
  print(message)
  if message["files"]:
    image = message["files"][-1]["path"]
  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]

  if image is None:
      gr.Error("You need to upload an image for LLaVA to work.")
  prompt=f"<|start_header_id|>user<|end_header_id|>\n\n<image>\n{message['text']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
  print(f"prompt: {prompt}")
  image = Image.open(image)
  inputs = processor(prompt, image, return_tensors='pt').to(0, torch.float16)

  streamer = TextIteratorStreamer(processor, **{"skip_special_tokens": True})
  generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024)
  generated_text = ""

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

  text_prompt =f"<|start_header_id|>user<|end_header_id|>\n\n{message['text']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
  print(f"text_prompt: {text_prompt}")

  buffer = ""
  for new_text in streamer:
    
    buffer += new_text
    
    generated_text_without_prompt = buffer[len(text_prompt):]
    time.sleep(0.04)
    yield generated_text_without_prompt


demo = gr.ChatInterface(fn=bot_streaming, fill_height=True, title="LLaVA Llama-3-8B", examples=[{"text": "What is on the flower?", "files":["./bee.jpg"]},
                                                                      {"text": "How to make this pastry?", "files":["./baklava.png"]}], 
                        description="Try [LLaVA Llama-3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers). 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)
demo.launch(debug=True)