app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM
import json
import torch
import requests
import time
import random
from PIL import Image
from typing import Union
import os
import base64
from together import Together
import pathlib
import gradio_client as grc
import spaces

global shrd
shrd = gr.JSON(visible=False)

device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using {device}" if device != "cpu" else "Using CPU")

def _load_model():
    tokenizer = AutoTokenizer.from_pretrained("vikhyatk/moondream2", trust_remote_code=True, revision="2024-05-08", torch_dtype=(torch.bfloat16 if device == 'cuda' else torch.float32))
    model = AutoModelForCausalLM.from_pretrained("vikhyatk/moondream2", device_map=device, trust_remote_code=True, revision="2024-05-08")
    return (model, tokenizer)

class MoonDream():
    def __init__(self, model=None, tokenizer=None):
        self.model, self.tokenizer = (model, tokenizer)
        if not model or model is None or not tokenizer or tokenizer is None:
            self.model, self.tokenizer = _load_model()
        self.device = device
        self.model.to(self.device)
    def __call__(self, question, imgs):
        imn = 0
        for img in imgs:
            img = self.model.encode_image(img)
            res = self.model.answer_question(question=question, image_embeds=img, tokenizer=self.tokenizer)
            yield res
        return

md = MoonDream()

SYSTEM_PROMPT = "You are Llama 3 70b. You have been given access to Moondream 2 for VQA when given images. When you have a question about an image, simple start your response with the text, '@question\\nMy question?'. When you do this, the request will be sent to Moondream 2. User can see this happening if they turn debug on, so be professional and stay on topic. Any chat from anyone starting with @answer is the answer to last question asked. If something appears out of sync, ask User to clear the chat."

@spaces.GPU
def _respond_one(question, img):
  txt = ""
  yield (txt := txt + MoonDream()(question, [img]))
  return txt

def respond_batch(question, **imgs):
  md = MoonDream()
  for img in imgs.values():
    res = md(question, img)
    for r in res:
      yield r
    yield "\n\n\n\n\n\n"
  return

def dual_images(img1: Image):
  # Ran once for each img to it's respective output. Output should be detailed str of description/feature extraction/interrogation.
  md = MoonDream()
  res = md("Describe the image in plain english ", [img1])
  txt = ""
  for r in res:
    yield (txt := txt + r)
  return

import os

def merge_descriptions_to_prompt(mi, d1, d2):
  from together import Together
  tog = Together(api_key=os.getenv("TOGETHER_KEY"))
  res = tog.completions.create(prompt=f"""Describe what would result if the following two descriptions were describing one thing.
### Description 1:
```text
{d1}
```
### Description 2:
```text
{d2}
```
Merge-Specific Instructions:
```text
{mi}
```
Ensure you end your output with ```\\n
---
Complete Description:
```text""", model="meta-llama/Meta-Llama-3-70B", stop=["```"], max_tokens=1024)
  return res.choices[0].text.split("```")[0]

def xform_image_description(img, inst):
  #md = MoonDream()
  from together import Together
  desc = dual_images(img)
  tog = Together(api_key=os.getenv("TOGETHER_KEY"))
  prompt=f"""Describe the image in aggressively verbose detail. I must know every freckle upon a man's brow and each blade of the grass intimately.\nDescription: ```text\n{desc}\n```\nInstructions:\n```text\n{inst}\n```\n\n\n---\nDetailed Description:\n```text"""
  res = tog.completions.create(prompt=prompt, model="meta-llama/Meta-Llama-3-70B", stop=["```"], max_tokens=1024)
  return res.choices[0].text[len(prompt):].split("```")[0]

def simple_desc(img, prompt):
  import base64
  gen = md(prompt, [img])
  total = ""
  for resp in gen:
    print(total := total + resp)
  img.resize((192,192)).save("tmp.png")
  bts = False
  with open("tmp.png", "rb") as f:
    bts = f.read()
  if bts:
    os.remove("tmp.png")
  res = {
    'image_b64': base64.b64encode(bts).decode('utf-8'),
    'description': total,
  }
  cl = grc.Client("http://127.0.0.1:7860/")
  result = cl.predict(
    message="Here's the description of your latest image, repeat any relevant details to keep them in context. Here's the description:\n```text\n" + total + "\n```\n\nAnd what the user wanted to begin with: `" + prompt + "`.",
    api_name="/chat"
  )
  print(result)
  return total, res, {**res, 'chat': result}

ifc_imgprompt2text = gr.Interface(simple_desc, inputs=[gr.Image(label="input", type="pil"), gr.Textbox(label="prompt")], outputs=[gr.Textbox(label="description"), gr.JSON(label="json")])

def chat(inpt, mess, desc):
  from together import Together
  print(inpt, mess)
  if mess is None:
    mess = []

  tog = Together(api_key=os.getenv("TOGETHER_KEY"))
  messages = [{
    'role': 'system',
    'content': SYSTEM_PROMPT
  }]
  if desc is not None and desc != "":
    messages.append({
      'role': 'system',
      'content': 'Here is a description of what you can see at the moment:\n```text\n' + desc + '\n```\nKeep this in mind when answering User\'s questions.'
    })
  messages.append({
    'role': 'user',
    'content': inpt
  })
  for cht in mess:
    print(cht)
  res = tog.chat.completions.create(
  messages=messages,
  model="meta-llama/Llama-3-70b-chat-hf", stop=["<|eot_id|>"], stream=True, safety_model="Meta-LLama/Llama-Guard-7b")
  txt = ""
  for pk in res:
    print(pk)
    txt += pk.choices[0].delta.content
    #mess[-1][-2] += pk.choices[0].delta.content
    yield txt #, json.dumps(messages)#mess#, json.dumps(messages)

chatbot = gr.Chatbot(
  [
    ["Hello?", "### Greetings\n\nWell, it seems I have a visitor! What can I do for you? &lt3;\n\n---"]
  ],
  elem_id="chatbot",
  bubble_full_width=False,
  sanitize_html=False,
  show_copy_button=True,
  avatar_images=[
    pathlib.Path("image.jpeg"), 
    pathlib.Path("image2.jpeg")
])

wizard_chatbot = gr.Chatbot(
  [
    ["Hello?", "### Greetings\n\nWell, it seems I have a visitor! What can I do for you? &lt3;\n\n---"]
  ],
  elem_id="chatbot_wizard",
  bubble_full_width=True,
  sanitize_html=False,
  show_copy_button=True,
  avatar_images=[
    pathlib.Path("image.png"),
    pathlib.Path("image2.jpeg")
  ]
)

def wizard_chat(inpt, mess):
  from together import Together
  print(inpt, mess)
  if mess is None:
    mess = []

  tog = Together(api_key=os.getenv("TOGETHER_KEY"))
  messages = []
  messages.append({
    'role': 'user',
    'content': "English; Please reply in English. " + inpt
  })
  for cht in mess:
    print(cht)
  res = tog.chat.completions.create(
  messages=messages,
  model="microsoft/WizardLM-2-8x22B", stop=["</s>"], stream=True, safety_model="Meta-LLama/Llama-Guard-7b")
  txt = ""
  for pk in res:
    print(pk)
    txt += pk.choices[0].delta.content
    #mess[-1][-2] += pk.choices[0].delta.content
    yield txt #, json.dumps(messages)#mess#, json.dumps(messages

botroom = None

def group_chat(room: str, **models):
  wzn = json.loads(wzn)
  lmn = json.loads(lmn)
  print(wzn, lmn)
  if not "replace_token" in wzn:
    wzn["replace_token"] = "<|wizard|>"
  if not "replace_token" in lmn:
    lmn["replace_token"] = "</Llama>"
  while room.find(lmn['replace_token']) != -1 or room.find(wzn['replace_token']) != -1:
    if not "prompt" in wzn and room.find(wzn['replace_token']) != -1:
      wzn["prompt"] = room[0:room.find(wzn['replace_token'])]
    if not "prompt" in lmn and room.find(lmn['replace_token']) != -1:
      lmn["prompt"] = room[0:room.find(lmn['replace_token'])]
    print(wzn, lmn)
    if "prompt" in wzn:
      print(wzn)
      res = wizard_chat(wzn['prompt'], [])
      tx = ""
      for r in res:
        yield cdd + r
        tx = r
  return cdd + txt
  # Let's make a more genetic model-merge with shadow config that has basic sane defaults for any model.
  # top_k 42
  # top_p 0.842
  # max_tokens 1536
  # temperature 0.693
  
  shadow_config = {
    "top_k": 42,
    "top_p": 0.842,
    "max_tokens": 1536,
    "temperature": 0.693,
    "repetition_penalty": 1.12
  }

  #models = {#

#  }

arch_room = None

def wizard_complete(cdd, wzs):
  tog = Together(api_key=os.getenv("TOGETHER_KEY"))
  if wzs.startswith("root="):
    wzs = wzs[5:]
  wzs = json.loads(wzs)
  print(wzs)
  if not "stop" in wzs:
    wzs["stop"] = ['###', '\n\n\n', '<|im_end|>', '<|im_start|>']
  if not "model" in wzs:
    wzs["model"] = "WizardLM/WizardCoder-Python-34B-V1.0"
  if not "prompt" in wzs:
    wzs["prompt"] = cdd
  res = tog.completions.create(prompt=wzs["prompt"], model=wzs["model"], stop=wzs["stop"], max_tokens=1024, stream=False)
  txt = cdd + res.choices[0].text
  return txt, txt

with gr.Blocks() as arch_room:
  with gr.Row():
    gr.Markdown(f"""
## Arcanistry


*POOF* -- You walk in, to a cloudy room filled with heavy smoke. In the center of the room rests a waist-height table. Upon the table, you see a... You don't understand... It's dark and light and cold and warm but... As you extend your hand, you hear the voice travel up your arm and into your ears...

---
""")
  with gr.Row():
    cdd = gr.Code("""### Human
I require a Python script that serves a simple file server in Python over MongoDB.

### Wizard
Sure! Here's the script:
```python""", language="markdown")
  with gr.Row():
    wzs = gr.Code(json.dumps({
      'token': '<|wizard|>',
      'model': 'WizardLM/WizardCoder-Python-34B-V1.0',
      'stop': ['###', '\n\n\n', '<|im_end|>', '<|im_start|>']
    }))
  with gr.Row():
    rnd = gr.Markdown("")
  with gr.Row():
    subm_prompt = gr.Button("Run Prompt")
    subm_prompt.click(wizard_complete, inputs=[cdd, wzs], outputs=[cdd, rnd])

with gr.TabbedInterface([ifc_imgprompt2text, c_ifc := gr.ChatInterface(chat, chatbot=chatbot, submit_btn=gr.Button(scale=1)), gr.ChatInterface(wizard_chat), arch_room], ["Prompt & Image 2 Text", "Chat w/ Llama 3 70b", "Chat w/ WizardLM 8x22B", "Arcanistry"]) as ifc:
  shrd = gr.JSON(visible=False)
  ifc.launch(share=False, debug=True, show_error=True)