Skip to content

Accelerating multi-node Large Language Model training with per-layer selective quantization (FP32 -> FP16) of the transformer architecture.

License

Notifications You must be signed in to change notification settings

KastanDay/LLM-Distributed-Quantization

 
 

Repository files navigation

Installing

git clone --recurse-submodules git@github.com:KastanDay/LLM-Distributed-Quantization.git

Install strict dependencies (on x64):

conda env create -f ./utilities/environment.yml

Launch slurm jobs

Specify the number of nodes required via --nnodes=8 in slurm. Also set --ntasks to the same number as the number of nodes.

Specify the config path to use as the first parameter. I recommend using absolute paths.

# launch sbatch from login node of slurm cluster 
sbatch LATEST_auto_multinode_launch.sh <config_filepath>

Updating

Subsequently pull new changes using:

git submodule update --remote

Experiment Tracking

All experiments are tracked with Weights and Biases. View the live progress here: https://wandb.ai/kastan/LLM-Distributed-Quantization

For transparency, the first couple hundred experiments are saved here: https://wandb.ai/kastan/col_ai

Benchmark for Tuning Accuracy and Efficiency

Overview

The benchmark includes our efforts in using Colossal-AI to train different tasks to achieve SOTA results. We are interested in both validataion accuracy and training speed, and prefer larger batch size to take advantage of more GPU devices. For example, we trained vision transformer with batch size 512 on CIFAR10 and 4096 on ImageNet1k, which are basically not used in existing works. Some of the results in the benchmark trained with 8x A100 are shown below.

Task Model Training Time Top-1 Accuracy
CIFAR10 ViT-Lite-7/4 ~ 16 min ~ 90.5%
ImageNet1k ViT-S/16 ~ 16.5 h ~ 74.5%

The train.py script in each task runs training with the specific configuration script in configs/ for different parallelisms. Supported parallelisms include data parallel only (ends with vanilla), 1D (ends with 1d), 2D (ends with 2d), 2.5D (ends with 2p5d), 3D (ends with 3d).

Each configuration scripts basically includes the following elements, taking ImageNet1k task as example:

TOTAL_BATCH_SIZE = 4096
LEARNING_RATE = 3e-3
WEIGHT_DECAY = 0.3

NUM_EPOCHS = 300
WARMUP_EPOCHS = 32

# data parallel only
TENSOR_PARALLEL_SIZE = 1    
TENSOR_PARALLEL_MODE = None

# parallelism setting
parallel = dict(
    pipeline=1,
    tensor=dict(mode=TENSOR_PARALLEL_MODE, size=TENSOR_PARALLEL_SIZE),
)

fp16 = dict(mode=AMP_TYPE.TORCH, ) # amp setting

gradient_accumulation = 2 # accumulate 2 steps for gradient update

BATCH_SIZE = TOTAL_BATCH_SIZE // gradient_accumulation # actual batch size for dataloader

clip_grad_norm = 1.0 # clip gradient with norm 1.0

Upper case elements are basically what train.py needs, and lower case elements are what Colossal-AI needs to initialize the training.

Usage

To start training, use the following command to run each worker:

$ DATA=/path/to/dataset python train.py --world_size=WORLD_SIZE \
                                        --rank=RANK \
                                        --local_rank=LOCAL_RANK \
                                        --host=MASTER_IP_ADDRESS \
                                        --port=MASTER_PORT \
                                        --config=CONFIG_FILE

It is also recommended to start training with torchrun as:

$ DATA=/path/to/dataset torchrun --nproc_per_node=NUM_GPUS_PER_NODE \
                                 --nnodes=NUM_NODES \
                                 --node_rank=NODE_RANK \
                                 --master_addr=MASTER_IP_ADDRESS \
                                 --master_port=MASTER_PORT \
                                 train.py --config=CONFIG_FILE

About

Accelerating multi-node Large Language Model training with per-layer selective quantization (FP32 -> FP16) of the transformer architecture.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 93.5%
  • Shell 6.5%