app.py

import os

os.system('pip install llama-cpp-python transformers torch')

import gradio as gr
from llama_cpp import Llama
from transformers import AutoTokenizer
from huggingface_hub import upload_file 
import json
from uuid import uuid4

model_id = "Elijahbodden/eliGPTv1.3"

# MODEL
model = Llama.from_pretrained(
    repo_id=model_id,
    filename="model.gguf",
    verbose=True,
    n_threads = 2,
    n_threads_batch = 2,
    n_ctx=8192,
  )


# TOKENIZER AND TEMPLATE
tokenizer = AutoTokenizer.from_pretrained(model_id)

presets = {
    # Gaslight the model by adding sentence fragments to the start
    # It's weird but it works
    # If you're curious, default makes sure it doesn't hallucinate by showing that the next message is the start of a new convo
    # I also include "oh" and "shit" bc the model overuses them and this lets repetition penalties do their thing
    "Default" : [{"from": "human", "value": "shit good convo, bye"}, {"from": "gpt", "value": "Haha oh ok cool ttyl"}],
    # I swear this is for science 🗿
    "Rizz ????" : [{"from": "human", "value": "omg it's so hot when you flirt with me"}, {"from": "gpt", "value": "haha well you're lucky can even string a sentence together, the way you take my breath away 😘"}, {"from": "human", "value": "alright love you, gn!"}, {"from": "gpt", "value": "ttyl babe 💕"}],
    "Thinky" : [{"from": "human", "value": "Woah you just totally blew my mind\ngehh now the fermi paradox is going to be bugging me 24/7\nok ttyl"}, {"from": "gpt", "value": "nah our deep convos are always the best, we should talk again soon\nttyl"}],
}

def custom_lp_logits_processor(ids, logits, lp_start, lp_decay, prompt_tok_len):
    generated_tok_number = len(ids) - prompt_tok_len
    if (generated_tok_number > lp_start):
        print(len(ids), lp_start, pow(lp_decay, len(ids)-lp_start)) 
        logits[tokenizer.eos_token_id] *= pow(lp_decay, generated_tok_number-lp_start)
    return logits

def respond(
    message,
    history: list[tuple[str, str]],
    preset,
    min_p,
    temperature,
    lp_start,
    lp_decay,
    frequency_penalty,
    presence_penalty,
    max_tokens
):
    print(preset, temperature, min_p, lp_start, lp_decay, frequency_penalty, presence_penalty, max_tokens)
    
    messages = presets[preset].copy()
    for val in history:
        if val[0]:
            messages.append({"from": "human", "value": val[0]})
        if val[1]:
            messages.append({"from": "gpt", "value": val[1]})

    messages.append({"from": "human", "value": message})

    response = ""

    print(tokenizer.apply_chat_template(messages, tokenize=False))
    
    convo = tokenizer.apply_chat_template(messages, tokenize=True)
    for message in model.create_completion(
        convo,
        temperature=temperature,
        stream=True,
        stop=["<|im_end|>"],
        min_p=min_p,
        max_tokens=max_tokens,
        # Disable top-k pruning
        top_k=100000000,
        frequency_penalty=frequency_penalty,
        presence_penalty=presence_penalty,
        logits_processor=lambda ids, logits: custom_lp_logits_processor(ids, logits, lp_start, lp_decay, len(convo))
    ):
        token = message["choices"][0]["text"]

        response += token
        yield response
    print(response)

ci = gr.ChatInterface(
    respond,
    additional_inputs_accordion=gr.Accordion(label="Options", open=True),
    additional_inputs=[
        gr.Radio(presets.keys(), label="Personality preset", info="Slightly influence the model's personality [WARNING, IF YOU CHANGE THIS WHILE THERE ARE MESSAGES IN THE CHAT THE MODEL WILL BECOME VERY SLOW]", value="Default"),
        # ("The model will become slow" is bc this uncaches the prompt and prompt processing is a big part of the generation time)
        gr.Slider(minimum=0.0, maximum=1.0, value=0.1, step=0.01, label="Min_p", info="Lower values give it more \"personality\""),
        gr.Slider(minimum=0.1, maximum=4.0, value=1.5, step=0.1, label="Temperature", info="How chaotic should the model be?"),
        gr.Slider(minimum=0, maximum=512, value=5, step=1, label="Length penalty start", info='When should the model start being more likely to shut up?'),
        gr.Slider(minimum=0.5, maximum=1.5, value=1.015, step=0.001, label="Length penalty decay factor", info='How fast should that stop likelihood increase?'),
        gr.Slider(minimum=0.0, maximum=1.0, value=0.1, step=0.01, label="Frequency penalty", info='"Don\'repeat yourself"'),
        gr.Slider(minimum=0.0, maximum=1.0, value=0.1, step=0.01, label="Presence penalty", info='"Use lots of diverse words"'),
        gr.Slider(minimum=1, maximum=1024, value=1024, step=1, label="Max new tokens", info="How many words can the model generate at most?"),
    ],
  )


with gr.Blocks(css=".bubble-gap {gap: 6px !important}", theme="shivi/calm_seafoam") as demo:
  gr.Markdown("# EliGPT v1.3")
  gr.Markdown("Llama 3 8b finetuned on 2.5k of my discord messages. [Train your own clone!](https://gist.github.com/Elijah-Bodden/1964bd02fcd19efef65f6e0cd92881c4)")
  with gr.Accordion("Q&A:", open=False):
      gr.Markdown("""Q: Why is the model so fucking slow  
        A: The model might be slow if it hasn't run recently or a lot of people are using it (it's running on llama.cpp on a single a very slow cpu). You can duplicate the space to get your own (free) instance with no wait times.  
        Q: Why is the model so dumb  
        A: Llama 3 8b is impressive, but it's still tiny. This model is basically what you'd get if you shoved my brain into a toddler's head - it's just too small to be smart  
        Q: Either it just made something up or I don't know you at all  
        A: Probably the former. It's prone to hallucinating facts and opinions I don't hold. Take everything it says with a big grain of salt  
      """)
  ci.render()

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