Training command line and deepspeed config for CodeGen 16B, 3xA100 GPUs

ds_config_AdamW_16B_reduce_mem.json

{
    "fp16": {
        "enabled": true,
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },
    "optimizer": {
        "type": "AdamW",
        "params": {
            "lr": "auto",
            "weight_decay": "auto"
        }
    },
    "scheduler": {
        "type": "WarmupDecayLR",
        "params": {
            "warmup_min_lr": "auto",
            "warmup_max_lr": "auto",
            "warmup_num_steps": "auto",
            "total_num_steps": "auto"
        }
    },
    "zero_optimization": {
        "stage": 3,
        "offload_optimizer": {
            "device": "cpu",
        "pin_memory": false
        },
        "offload_param": {
            "device": "cpu",
        "pin_memory": false
        },
        "overlap_comm": true,
        "contiguous_gradients": true,
        "reduce_bucket_size": "auto",
        "stage3_prefetch_bucket_size": "auto",
        "stage3_param_persistence_threshold": "auto",
        "sub_group_size": 1e9,
        "stage3_max_live_parameters": 1e9,
        "stage3_max_reuse_distance": 1e9,
        "stage3_gather_16bit_weights_on_model_save": true
    },
    "gradient_accumulation_steps": 32,
    "gradient_clipping": "auto",
    "steps_per_print": 2000,
    "train_batch_size": "auto",
    "train_micro_batch_size_per_gpu": "auto",
    "wall_clock_breakdown": false
}

train.sh

#!/bin/bash
  
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=4
#SBATCH --mem=400GB
#SBATCH --gres=gpu:a100:3
#SBATCH --job-name=codegenV16B
#SBATCH --time=166:00:00

module purge
singularity exec --nv \
      --overlay /scratch/st4920/pytorch-codegen-v2.ext3:ro \
      /scratch/work/public/singularity/cuda11.3.0-cudnn8-devel-ubuntu20.04.sif \
      /bin/bash -c "source /ext3/env.sh; conda activate codegen_deepspeed; deepspeed run_clm.py --model_name_or_path=Salesforce/codegen-16B-multi --save_steps=500 --per_device_train_batch_size=1 --num_train_epochs 1 --output_dir=CodeGen/codegen-16B-verilog-1-epochs --report_to 'wandb' --dataset_name code_segments_verilog_unfiltered --tokenizer_name Salesforce/codegen-16B-multi --learning_rate 2e-5 --block_size 2048 --do_train --fp16 --gradient_accumulation_steps 32 --overwrite_output_dir --deepspeed ds_config_AdamW_16B_reduce_mem.json"