app_v3.py

################### WHATEVER IT TAKES PHASE ###################################
import streamlit as st
from transformers import AutoTokenizer, TextStreamer, pipeline
from auto_gptq import AutoGPTQForCausalLM
from huggingface_hub import snapshot_download
import os
import torch
import subprocess

# os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'max_split_size_mb:128' # => just makes sense with more than one GPU, since is trying to split 
#os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID" # => just makes sense when more GPUs
#os.environ["CUDA_VISIBLE_DEVICES"]="0,1" # => this is an example of numbers of devices


# Define pretrained and quantized model directories
pretrained_model_dir = "FPHam/Jackson_The_Formalizer_V2_13b_GPTQ"
#cwd = os.getcwd()

#quantized_model_dir = cwd + "/Jackson2-4bit-128g-GPTQ"

# Check if the model directory is empty (i.e., model not downloaded yet)
#if not os.path.exists(quantized_model_dir) or not os.listdir(quantized_model_dir):
    # Create the cache directory if it doesn't exist
#    os.makedirs(quantized_model_dir, exist_ok=True)
#    snapshot_download(repo_id=pretrained_model_dir, local_dir=quantized_model_dir, local_dir_use_symlinks=True)

#st.write(f'{os.listdir(quantized_model_dir)}')
#model_name_or_path = quantized_model_dir
model_basename = "Jackson2-4bit-128g-GPTQ"

# Before allocating or loading the model, clear up memory
#gc.collect()
#torch.cuda.empty_cache()

use_triton = False

if torch.cuda.is_available():
    torch.cuda.empty_cache()

#tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, use_fast=True, legacy=False)
tokenizer = AutoTokenizer.from_pretrained(pretrained_model_dir, use_fast=True)
device = "cuda:0" if torch.cuda.is_available() else "cpu" # best configuration besides the auto option
model = AutoGPTQForCausalLM.from_quantized(
    pretrained_model_dir,
    model_basename=model_basename,
    use_safetensors=True,
    device=device,
    max_memory={0: "10GIB"}
)


viz = torch.cuda.memory_summary()
st.write(viz)

def run():
    output: str = ""
    try:
        output = subprocess.check_output(["nvidia-smi"], text=True)
    except FileNotFoundError:
        output = subprocess.check_output(["ls", "-alh"], text=True)
    return f"#{output}"
st.write(run())
user_input = st.text_input("Input a phrase")

prompt_template = f'USER: {user_input}\nASSISTANT:'

if st.button("Generate the prompt"):
    inputs = tokenizer(prompt_template, return_tensors='pt', max_length=512, truncation=True, padding='max_length', batch_size=4)
    #inputs = tokenizer(prompt_template, return_tensors='pt')
    #streamer = TextStreamer(tokenizer)
    #pipe = pipeline(
    #    "text-generation",
    #    model=model,
    #    tokenizer=tokenizer,
    #    streamer=streamer,
    #    max_new_tokens=512,
    #    temperature=0.2,
    #    top_p=0.95,
    #    repetition_penalty=1.15
    #)

    output = model.generate(**prompt_template)
    st.markdown(f"tokenizer.decode(output)")
    #st.write(output[0]['generated_text'])