mehdidc

"""
This is helper script to launch sbatch jobs and to handle two issues
we encountered:

- freezing/hanging
- limited maximum job time (24 hours in the best case, can be 6 hours when total compute budget is over)

The script automatically relaunch the sbatch script when the job either freezes
or is stopped/canceled.

mehdidc

image_retrieval_recall@1,text_acc,text_retrieval_recall@5,text_retrieval_recall@1,image_retrieval_recall@5,task,text_retrieval_recall@10,language,acc,acc5,image_retrieval_recall@10,model,dataset,mean_per_class_recall,image_acc,pretrained,acc1,model_fullname
0.448199987411499,,0.8410000205039978,0.5920000076293945,0.7206000089645386,zeroshot_retrieval,0.8980000019073486,en,,,0.8123999834060669,RN50,wds/flickr30k,,,cc12m,,RN50 cc12m
0.45840001106262207,,0.8500000238418579,0.5889999866485596,0.7305999994277954,zeroshot_retrieval,0.8980000019073486,en,,,0.8163999915122986,RN50-quickgelu,wds/flickr30k,,,cc12m,,RN50-quickgelu cc12m
0.5221999883651733,,0.878000020980835,0.6740000247955322,0.7821999788284302,zeroshot_retrieval,0.9309999942779541,en,,,0.864799976348877,ViT-B-16,wds/flickr30k,,,commonpool_l_basic_s1b_b8k,,ViT-B-16 commonpool_l_basic_s1b_b8k
0.5103999972343445,,0.8889999985694885,0.6830000281333923,0.7735999822616577,zeroshot_retrieval,0.9369999766349792,en,,,0.8521999716758728,ViT-B-16,wds/flickr30k,,,

mehdidc

1.Preparing the env:

export REPO=/p/project/laionize/cherti1/Megatron-LM # MODIFY!
git clone https://github.com/bigcode-project/Megatron-LM.git $REPO
cd $REPO
ml Stages/2024 GCC/12.3.0 OpenMPI CUDA/12 PyTorch/2.1.0-CUDA-12-ALPHA
python -m venv $REPO/env
source $REPO/env/bin/activate
pip install -U pip

mehdidc

import torch
import mobileclip
from PIL import Image
import numpy as np
import pandas as pd
from torch.utils.flop_counter import FlopCounterMode
import open_clip
import fvcore
import fvcore.nn

mehdidc

#!/bin/bash
#SBATCH --nodes=8
#SBATCH --time=00:20:00
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=48
#SBATCH --gres=gpu:4
#SBATCH --partition=booster
#SBATCH --account=transfernetx
#SBATCH --exclude=jwb[0059,0067,0069,0193,0284,0287,0294,0359,0418,0637,0829,0832,0838,0898,0907,0921,0971,1004,1023,1029,1213,1126]
#SBATCH --threads-per-core=1

mehdidc

#!/bin/bash
#SBATCH --nodes=8
#SBATCH --time=00:20:00
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=48
#SBATCH --gres=gpu:4
#SBATCH --partition=booster
#SBATCH --account=transfernetx
#SBATCH --exclude=jwb[0059,0067,0069,0193,0284,0287,0294,0359,0418,0637,0829,0832,0838,0898,0907,0921,0971,1004,1023,1029,1213,1126]
#SBATCH --threads-per-core=1

mehdidc

checkpoints:
  checkpoint_interval: 10
  checkpoints_path: checkpoints
  checkpoints_path_is_shared_file_system: false
  resume_checkpoint_path: null
  save_initial_state: false

data_stages:
  - name: Stable Training Stage
    start_training_step: 1

mehdidc

#!/bin/bash -x
#SBATCH --account={account}
#SBATCH --nodes={nodes}
#SBATCH --gres=gpu:4
#SBATCH --ntasks-per-node=4
#SBATCH --cpus-per-task=24
#SBATCH --time=06:00:00
#SBATCH --partition={partition}
#SBATCH --output={output_file}
echo "Job Id:$SLURM_JOB_ID"

mehdidc

job_id_regexp: "Job Id:(\\d+)"
cmd: "sbatch {sbatch_script}"
check_interval_secs: 600
partition: booster
account: laionize
experiments:
  small:
    model_scale:
        model: [ViT-B-32]
    samples_seen_scale:        

mehdidc

Disable Laptop Keyboard in Ubuntu

Taken From :- https://blog.hostonnet.com/laptop-keyboard-ubuntu

To disable laptop built in keyboard in Ubuntu, first you need to identify the device id for keyboard.

Now run xinput –list command to get list of all input devices

hostonnet@pc4:~$ xinput --list
⎡ Virtual core pointer                      id=2    [master pointer  (3)]