app.py

import streamlit as st
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

base_model_name = "chaseharmon/Rap-Mistral-Big"


@st.cache_resource
def load_model():
    nf4_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_use_double_quant=False,
        bnb_4bit_compute_dtype="float16"
    )

    model = AutoModelForCausalLM.from_pretrained(
        base_model_name,
        device_map='auto',
    )
    model.config.use_cache = False
    model.config.pretraining_tp = 1

    return model

@st.cache_resource
def load_tokenizer(): 
    tokenizer = AutoTokenizer.from_pretrained(base_model_name)

    tokenizer.pad_token = tokenizer.eos_token
    tokenizer.padding_side = "right"

    return tokenizer

def build_prompt(question):
  prompt=f"[INST] {question} [/INST] "
  return prompt

model = load_model()
model.eval()

tokenizer = load_tokenizer

st.title("Rap Verse Generation V1 Demo")
st.header("Supported Artists")
st.write("Lupe Fiasco, Common, Jay-Z, Yasiin Bey, Ab-Soul, Rakim")

question = st.chat_input("Write a verse in the style of Lupe Fiasco")
if question:
    prompt = build_prompt(question)
    inputs = tokenizer(prompt, return_tensors="pt")
    model_inputs = inputs.to('cuda')
    generated_ids = model.generate(**model_inputs, max_new_tokens=300, do_sample=True, pad_token_id=tokenizer.eos_token_id)
    decoded_output = tokenizer.batch_decode(generated_ids)