app.py

import streamlit as st
from transformers import AutoTokenizer
from auto_gptq import AutoGPTQForCausalLM
import torch
import subprocess
import traceback

# Function to get memory info
def get_gpu_memory():
    try:
        result = subprocess.check_output(["nvidia-smi", "--query-gpu=memory.free,memory.total", "--format=csv,nounits,noheader"], text=True)
        memory_info = [x.split(',') for x in result.strip().split('\n')]
        memory_info = [{"free": int(x[0].strip()), "total": int(x[1].strip())} for x in memory_info]
    except FileNotFoundError:
        memory_info = [{"free": "N/A", "total": "N/A"}]
    return memory_info

# Display GPU memory information before loading the model
gpu_memory_before = get_gpu_memory()
st.write(f"GPU Memory Info before loading the model: {gpu_memory_before}")

# Define pretrained model directory
pretrained_model_dir = "FPHam/Jackson_The_Formalizer_V2_13b_GPTQ"

# Check if CUDA is available and get the device
device = "cuda:0" if torch.cuda.is_available() else "cpu"

# Before allocating or loading the model, clear up memory if CUDA is available
if device == "cuda:0":
    torch.cuda.empty_cache()

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(pretrained_model_dir, use_fast=False)
tokenizer.pad_token = tokenizer.eos_token  # Ensure padding token is set correctly for the model

# Attempt to load the model, catch any OOM errors
@st.cache_resource
def load_gptq_model():
    model = AutoGPTQForCausalLM.from_quantized(
        pretrained_model_dir,
        model_basename="Jackson2-4bit-128g-GPTQ",
        use_safetensors=True,
        device=device,
        disable_exllamav2=True
    )
    model.eval()  # Set the model to inference mode
    return model

model_loaded = False 
# Attempt to load the model, catch any OOM errors
try:
    model = load_gptq_model()
    model_loaded = True
except RuntimeError as e:
    if 'CUDA out of memory' in str(e):
        st.error("CUDA out of memory while loading the model. Try reducing the model size or restarting the app.")
        st.stop()
    else:
        raise e

if model_loaded:
    # Display GPU memory information after loading the model
    gpu_memory_after = get_gpu_memory()
    st.write(f"GPU Memory Info after loading the model: {gpu_memory_after}")

    col1, col2 = st.columns(2)
    with col1:
        user_input = st.text_input("Input a phrase")
    with col2:
        max_token = st.number_input(label="Select max number of generated tokens", min_value=1, max_value=512, value=50, step=5)

    # Generate button
    if st.button("Generate the prompt"):
        try:
            prompt_template = f'USER: {user_input}\nASSISTANT:'
            inputs = tokenizer(prompt_template, return_tensors='pt', max_length=512, truncation=True, padding='max_length')
            inputs = inputs.to(device)  # Move inputs to the same device as model
        # Generate text using torch.inference_mode for better performance during inference
            with torch.inference_mode():
                output = model.generate(**inputs, max_new_tokens=max_token)  
            
            # Cut the tokens at the input length to display only the generated text
            output_ids_cut = output[:, inputs["input_ids"].shape[1]:]
            generated_text = tokenizer.decode(output_ids_cut[0], skip_special_tokens=True)
            
            st.markdown(f"**Generated Text:**\n{generated_text}")
        except RuntimeError as e:
            if 'CUDA out of memory' in str(e):
                st.error("CUDA out of memory during generation. Try reducing the input length or restarting the app.")
                # Log the detailed error message
                with open('error_log.txt', 'a') as f:
                    f.write(traceback.format_exc())
            else:
                # Log the error and re-raise it
                with open('error_log.txt', 'a') as f:
                    f.write(traceback.format_exc())
                raise e

        # Display GPU memory information after generation
        gpu_memory_after_generation = get_gpu_memory()
        st.write(f"GPU Memory Info after generation: {gpu_memory_after_generation}")
        st.write(f"The following is the distribution of model (in case n GPUs > 1): {model.hf_device_map}")