open clip at slurm

open_clip_slurm.md

Install

git clone https://github.com/mlfoundations/open_clip.git
cd open_clip
python3.8 -m venv .env
source .env/bin/activate
pip install -U pip
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu113
pip install -e .
pip install braceexpand pandas webdataset

Get slurm script

Update: I rather advise to use https://github.com/mlfoundations/open_clip/blob/main/docs/script_examples/stability_example.sh

wget https://gist.githubusercontent.com/rom1504/0d6b7e4e49626109a5a8e1c59a4e1aa6/raw/c73aa74c65f42def14cfee6bf8b25438ebd4e11e/start_in_container.sh
wget https://gist.githubusercontent.com/rom1504/0d6b7e4e49626109a5a8e1c59a4e1aa6/raw/c73aa74c65f42def14cfee6bf8b25438ebd4e11e/start_openclip.sh

Run the training

sbatch start_openclip.sh

Check what is going on

squeue -u your_user

ls -lt | head -10 to find the log file then less thefile

find one host in squeue then ssh thehost, then nvidia-smi, htop

Datasets

laion2b

--train-data 'pipe:s3cmd get -q s3://s-datasets/laion5b/laion2B-data/{000000..231349}.tar -' \
--train-num-samples 2170337258 \

laion400m

--train-data="pipe:aws s3 cp s3://s-datasets/laion400m/laion400m-dat-release/{00000..41455}.tar -" \
--train-num-samples 413000000 \

Thanks to @rwightman for helping me fix this!

Auto resume instructions

#SBATCH --requeue

checkpoint_path=`ls -t /fsx/rom1504/open_clip/src/logs/*ViT-g-14*/checkpoints/* | head -1`

--resume $checkpoint_path \

 || sbatch /fsx/rom1504/open_clip/good.sh

start_openclip.sh

#!/bin/bash
#SBATCH --partition=compute-od-gpu
#SBATCH --job-name=openclip
#SBATCH --nodes 8
#SBATCH --ntasks-per-node 8
#SBATCH --cpus-per-gpu=6
#SBATCH --gres=gpu:8
#SBATCH --output=%x_%j.out
#SBATCH --exclusive

## Prefer using https://github.com/mlfoundations/open_clip/blob/main/docs/script_examples/stability_example.sh


module load intelmpi
source /opt/intel/mpi/latest/env/vars.sh
export LD_LIBRARY_PATH=/opt/aws-ofi-nccl/lib:/opt/amazon/efa/lib64:/usr/local/cuda-11.0/efa/lib:/usr/local/cuda-11.0/lib:/usr/local/cuda-11.0/lib64:/usr/local/cuda-11.0:/opt/nccl/build/lib:/opt/aws-ofi-nccl-install/lib:/opt/aws-ofi-nccl/lib:$LD_LIBRARY_PATH
export NCCL_PROTO=simple
export PATH=/opt/amazon/efa/bin:$PATH
export LD_PRELOAD="/opt/nccl/build/lib/libnccl.so"

export FI_EFA_FORK_SAFE=1
export FI_LOG_LEVEL=1
export FI_EFA_USE_DEVICE_RDMA=1 # use for p4dn

#export NCCL_ALGO=ring
export NCCL_DEBUG=info
#export NCCL_DEBUG_SUBSYS=INIT,ENV,GRAPH,COLL

export PYTHONFAULTHANDLER=1

export CUDA_LAUNCH_BLOCKING=0
export OMPI_MCA_mtl_base_verbose=1
export FI_EFA_ENABLE_SHM_TRANSFER=0
export FI_PROVIDER=efa
export FI_EFA_TX_MIN_CREDITS=64
export NCCL_TREE_THRESHOLD=0


#export NCCL_P2P_DISABLE=1
#export NCCL_IBEXT_DISABLE=1
#export NCCL_SOCKET_IFNAME="eth0,en,eth,em,bond"

# sent to sub script
export HOSTNAMES=`scontrol show hostnames "$SLURM_JOB_NODELIST"`
export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
export MASTER_PORT=12802
export COUNT_NODE=`scontrol show hostnames "$SLURM_JOB_NODELIST" | wc -l`

echo go $COUNT_NODE
echo $HOSTNAMES

source /fsx/rom1504/open_clip/.env/bin/activate
cd /fsx/rom1504/open_clip/src/

srun --cpu_bind=v --accel-bind=gn python -m training.main \
    --save-frequency 1 \
    --train-data="pipe:aws s3 cp s3://s-datasets/laion400m/laion400m-dat-release/{00000..41455}.tar -" \
    --train-num-samples 413000000 \
    --dataset-type webdataset \
    --warmup 2000 \
    --batch-size=384 \
    --epochs=32 \
    --lr 5e-4 \
    --workers=2 \
    --model ViT-B-32 \
    --seed 0 \
    --local-loss \
    --grad-checkpointing \
    --ddp-static-graph \
    --gather-with-grad \
    --report-to wandb

You can also use squeue -u your_user to see what's going on and it includes the header.

indeed, changed

prefer using https://github.com/mlfoundations/open_clip/blob/main/docs/script_examples/stability_example.sh