Q-GaLore

This repo contains the pre-release version of Q-GaLore algorithm, proposed by Q-GaLore: Quantized GaLore with INT4 Projection and Layer-Adaptive Low-Rank Gradients.

Q-GaLore is a memory-efficient training methodology effective in both pre-training and fine-tuning scenarios. Q-GaLore incorporates two main components: (i) low precision training utilizing low-rank gradients, and (ii) lazy layer-wise subspace exploration. It enables full-parameter learning while requiring less memory, such as training a LLaMA-7B model from scratch on a single NVIDIA RTX 4060 Ti with only 16GB of memory.

Read this blog for more details!

Install Q-GaLore optimizer

Install via conda

conda env create -f environment.yml

or Install Q-GaLore optimizer and experiment dependencies

# install from pip
pip install q-galore-torch


# or install from source:
git clone https://github.com/VITA-Group/Q-GaLore.git
cd Q-GaLore
pip install -e .

pip install -r exp_requirements.txt

Usage

Pretraining LLaMA model on C4 dataset

We provide the command in scripts/pretrain_c4 for pretraining LLaMA model with sizes from 60M to 7B on C4 dataset. We also provide the simulation mode implementation of quantization with scripts in scripts/pretrain_c4/simulation. For example, training a LLaMA-60M with Q-GaLore-Adam8bit with the following scripts.

torchrun --standalone --nproc_per_node 1 run_pretrain.py \
    --model_config configs/llama_130m.json \
    --lr 0.015 \
    --galore_scale 0.25 \
    --rank 256 \
    --update_proj_gap 200 \
    --batch_size 256 \
    --total_batch_size 512 \
    --num_training_steps 20000 \
    --warmup_steps 2000 \
    --weight_decay 0 \
    --dtype bfloat16 \
    --eval_every 1000 \
    --optimizer q_galore_adamw8bit \
    --project 'g-galore-c4' \
    --weight_quant \
    --stochastic_round \
    --proj_quant \
    --name Q-Galore-Adam8bit-LLaMA-130M

Pretraining LLaMA-7B model within 16GB memory

The command of training LLaMA-7B model on single GPU as provided within scripts/pretrain_c4/single_gpu. With 16 batch size and activation checkpointing, the following scripts can pre-train a LLaMA-7B model with 15.26GB memory (tested on a single A6000 GPU)

# LLaMA-7B, 8-bit Q-GaLore-Adam, single GPU
# Memory cost: 15.26G, BSZ=16
torchrun --standalone --nproc_per_node 1 run_pretrain.py \
    --model_config configs/llama_7b.json \
    --lr 0.004 \
    --galore_scale 0.25 \
    --rank 1024 \
    --update_proj_gap 500 \
    --batch_size 16 \
    --total_batch_size 512 \
    --activation_checkpointing \
    --num_training_steps 150000 \
    --warmup_steps 15000 \
    --weight_decay 0 \
    --grad_clipping 1.0 \
    --dtype bfloat16 \
    --eval_every 1000 \
    --single_gpu \
    --proj_quant \
    --weight_quant \
    --stochastic_round \
    --optimizer q_galore_adamw8bit_per_layer

Citation

@misc{zhang2024qgalore,
      title={Q-GaLore: Quantized GaLore with INT4 Projection and Layer-Adaptive Low-Rank Gradients}, 
      author={Zhenyu Zhang and Ajay Jaiswal and Lu Yin and Shiwei Liu and Jiawei Zhao and Yuandong Tian and Zhangyang Wang},
      year={2024},
      eprint={2407.08296},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2407.08296}, 
}