devatar/quantized_Llama-3.1-8B-Instruct

🚀 Quantized Llama-3.1-8B-Instruct Model

This is a 4-bit quantized version of the meta-llama/Llama-3.1-8B-Instruct model, optimized for efficient inference on resource-constrained environments like Google Colab's NVIDIA T4 GPU.

🧠 Model Description

The model was quantized using the bitsandbytes library to reduce memory usage while maintaining performance for instruction-following tasks.

🧮 Quantization Details

• Base Model: meta-llama/Llama-3.1-8B-Instruct
• Quantization Method: 4-bit (NormalFloat4, NF4) with double quantization
• Compute Dtype: float16
• Library: bitsandbytes==0.43.3
• Framework: transformers==4.45.1
• Hardware: NVIDIA T4 GPU (16GB VRAM) in Google Colab
• Date: Quantized on June 20, 2025

📦 Files Included

• README.md: This file
• config.json, pytorch_model.bin (or sharded checkpoints): Model weights
• special_tokens_map.json, tokenizer.json, tokenizer_config.json: Tokenizer files

Usage

To load and use the quantized model for inference:

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, pipeline
import torch

# Define quantization configuration
quant_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True
)

# Load the quantized model
model = AutoModelForCausalLM.from_pretrained(
    "your-username/quantized_Llama-3.1-8B-Instruct",  # Replace with your Hugging Face repo ID
    quantization_config=quant_config,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("your-username/quantized_Llama-3.1-8B-Instruct")

# Create a text generation pipeline
generator = pipeline("text-generation", model=model, tokenizer=tokenizer)

# Perform inference
prompt = "Hello, how can I assist you today?"
output = generator(prompt, max_length=50, num_return_sequences=1)
print(output)

Quantization Process

The model was quantized in Google Colab using the following script:

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch
from huggingface_hub import login

# Log in to Hugging Face
login()  # Requires a Hugging Face token

# Define quantization configuration
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True
)

# Load and quantize the model
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct",
    quantization_config=quantization_config,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
tokenizer.pad_token = tokenizer.eos_token if tokenizer.pad_token is None else tokenizer.pad_token

# Save the quantized model
quant_path = "/content/quantized_Llama-3.1-8B-Instruct"
model.save_pretrained(quant_path)
tokenizer.save_pretrained(quant_path)

Requirements

• Hardware: NVIDIA GPU with CUDA 11.4+ (e.g., T4, A100)
• Python: 3.10+
• Dependencies:
  • transformers==4.45.1
  • bitsandbytes==0.43.3
  • accelerate==0.33.0
  • torch (with CUDA support)

Notes

• The quantized model is stored in /content/quantized_Llama-3.1-8B-Instruct in the Colab environment.
• Due to Colab's ephemeral storage, consider pushing to Hugging Face Hub or saving to Google Drive for persistence.
• Access to the base model requires a Hugging Face token and approval from Meta AI.

License

This model inherits the license of the base model meta-llama/Llama-3.1-8B-Instruct. Refer to the original model card: Meta AI Llama 3.1-8B-Instruct.

Acknowledgments

• Created using Hugging Face Transformers and bitsandbytes for quantization.
• Quantized in Google Colab with a T4 GPU on June 20, 2025.