app.py

from __future__ import annotations

import torch
torch.jit.script = lambda f: f

import shlex
import spaces

import gradio as gr
from threading import Thread
from transformers import TextIteratorStreamer
import hashlib
import os

from transformers import AutoModel, AutoProcessor
import sys
import subprocess
from PIL import Image

import time

# install packages for mamba
def install():
    print("Install personal packages", flush=True)
    subprocess.run(shlex.split("pip install causal_conv1d-1.2.0.post1-cp310-cp310-linux_x86_64.whl"))
    subprocess.run(shlex.split("pip install mamba_ssm-1.2.0.post1-cp310-cp310-linux_x86_64.whl"))

install()

from cobra import load
vlm = load("cobra+3b")

if torch.cuda.is_available():
    DEVICE = "cuda"
    DTYPE = torch.bfloat16
else:
    DEVICE = "cpu"
    DTYPE = torch.float32
vlm.to(DEVICE, dtype=DTYPE)

prompt_builder = vlm.get_prompt_builder()

@spaces.GPU(duration=20)
def bot_streaming(message, history, temperature, top_k, max_new_tokens):
    streamer = TextIteratorStreamer(vlm.llm_backbone.tokenizer, skip_special_tokens=True)
    
    if len(history) == 0:
        prompt_builder.prompt, prompt_builder.turn_count = "", 0

    image = None
    
    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 not None:
        image = Image.open(image).convert("RGB")

    
    prompt_builder.add_turn(role="human", message=message['text'])
    prompt_text = prompt_builder.get_prompt()

    generation_kwargs = {
        "image": image,
        "prompt_text": prompt_text,
        "streamer": streamer,
        "max_new_tokens": max_new_tokens,
        "use_cache": True,
        "temperature": temperature,
        "do_sample": True,
        "top_k": top_k,
    }
    
    # Generate from the VLM
    thread = Thread(target=vlm.generate, kwargs=generation_kwargs)
    thread.start()

    buffer = ""
    output_started = False
    for new_text in streamer:
        if not output_started:
            if "<|assistant|>\n" in new_text:
                output_started = True
            continue
        buffer += new_text
        if len(buffer) > 1:
            yield buffer

    prompt_builder.add_turn(role="gpt", message=buffer)

    return buffer
    
    
demo = gr.ChatInterface(fn=bot_streaming, 
                        additional_inputs=[gr.Slider(0, 1, value=0.2, label="Temperature"),
                                           gr.Slider(1, 3, value=1, step=1, label="Top k"),
                                           gr.Slider(1, 2048, value=256, step=1, label="Max New Tokens")],
                        title="Cobra", 
                        description="Try [Cobra](https://huggingface.co/papers/2403.14520) in this demo. Upload an image and start chatting about it. Clear the history before asking questions related to new images",
                        stop_btn="Stop Generation", multimodal=True)
demo.launch(debug=True)