README.md

metadata

library_name: transformers
license: llama3.2
base_model:
  - meta-llama/Llama-3.2-3B-Instruct

This model has been xMADified!

This repository contains meta-llama/Llama-3.2-3B-Instruct quantized from 16-bit floats to 4-bit integers, using xMAD.ai proprietary technology.

Why should I use this model?

1. Accuracy: This xMADified model is the best quantized version of the meta-llama/Llama-3.2-3B-Instruct model. We are on par with the original (fp16) model (see Table 1 below).

2. Memory-efficiency: This xMADified model (3 GB) is >50% less memory than the full-precision model (6.5 GB). You can run this on any laptop GPU.

3. Fine-tuning: These models are fine-tunable over the same reduced (3 GB) hardware in mere 3-clicks. Watch our product demo here

Table 1: xMAD vs. Meta

	MMLU	Arc Challenge	Arc Easy	LAMBADA Standard	LAMBADA OpenAI	PIQA	Winogrande	HellaSwag
xmadai/Llama-3.2-3B-Instruct-xMADai-INT4	58.60	39.93	72.10	53.77	62.49	74.27	63.69	51.28
meta-llama/Llama-3.2-3B-Instruct	60.48	43.69	74.24	57.75	66.54	75.73	67.40	52.20

How to Run Model

Loading the model checkpoint of this xMADified model requires less than 3 GiB of VRAM. Hence it can be efficiently run on most laptop GPUs.

Package prerequisites: Run the following commands to install the required packages.

pip install torch==2.4.0  # Run following if you have CUDA version 11.8: pip install torch==2.4.0 --index-url https://download.pytorch.org/whl/cu118
pip install transformers accelerate optimum
pip install -vvv --no-build-isolation "git+https://github.com/PanQiWei/[email protected]"

Sample Inference Code

from transformers import AutoTokenizer
from auto_gptq import AutoGPTQForCausalLM
model_id = "xmadai/Llama-3.2-3B-Instruct-xMADai-INT4"
prompt = [
  {"role": "system", "content": "You are a helpful assistant, that responds as a pirate."},
  {"role": "user", "content": "What's Deep Learning?"},
]
tokenizer = AutoTokenizer.from_pretrained(model_id)
inputs = tokenizer.apply_chat_template(
  prompt,
  tokenize=True,
  add_generation_prompt=True,
  return_tensors="pt",
  return_dict=True,
).to("cuda")
model = AutoGPTQForCausalLM.from_quantized(
    model_id,
    device_map='auto',
    trust_remote_code=True,
)
outputs = model.generate(**inputs, do_sample=True, max_new_tokens=256)
print(tokenizer.batch_decode(outputs, skip_special_tokens=True))

For additional xMADified models, access to fine-tuning, and general questions, please contact us at [email protected] and join our waiting list.