Hello
I would like to fine-tune the EleutherAI/gpt-j-6B model on a HPC cluster. On the cluster I have only GPUs available with at most 24 GB memory. Thus, I’m running out of memory when fine-tuning on these GPUs. That’s why I would like to use DeepSpeed and fine-tune on 2 or 4 GPUs at the same time. I have the following TrainingArguments:
training_args = TrainingArguments(
disable_tqdm=True,
output_dir='/cluster/scratch/myUser/cache/models/checkpoints',
save_total_limit=10,
logging_dir='/cluster/scratch/myUser/cache/models/logs',
num_train_epochs=4,
evaluation_strategy='epoch',
save_strategy='steps',
save_steps=30,
logging_steps=10,
overwrite_output_dir=True,
per_device_train_batch_size=4,
per_device_eval_batch_size=4,
gradient_accumulation_steps=4,
eval_accumulation_steps=4,
gradient_checkpointing=True,
max_grad_norm=0.5,
lr_scheduler_type="cosine",
learning_rate=1e-4,
warmup_ratio=0.05,
weight_decay=0.1,
fp16_full_eval=True,
fp16=True,
fp16_opt_level='O1',
deepspeed="configs/ds_config.json",
report_to=['tensorboard']
)
In addition I have the following deepspeed configuration file ds_config.json:
{
"fp16": {
"enabled": "auto",
"loss_scale": 0,
"loss_scale_window": 1000,
"initial_scale_power": 16,
"hysteresis": 2,
"min_loss_scale": 1,
"opt_level": "O3"
},
"zero_optimization": {
"stage": 2,
"offload_param": {
"device": "none",
"buffer_count": 4,
"buffer_size": 1e8,
"max_in_cpu": 1e9,
"pin_memory": false
},
"offload_optimizer": {
"device": "none",
"buffer_count": 4,
"pin_memory": false,
"pipeline_read": false,
"pipeline_write": false,
"fast_init": false
},
"allgather_partitions": true,
"allgather_bucket_size": 2e8 ,
"reduce_bucket_size": 2e8,
"overlap_comm": true,
"reduce_scatter": true,
"contiguous_gradients": true,
"cpu_offload": false,
"cpu_offload_params" : false,
"sub_group_size": 1e7,
"stage3_prefetch_bucket_size": "auto",
"stage3_param_persistence_threshold": "auto",
"stage3_max_live_parameters": 1e7,
"stage3_max_reuse_distance": 1e7,
"stage3_gather_16bit_weights_on_model_save": true
},
"optimizer": {
"type": "AdamW",
"params": {
"lr": "auto",
"betas": "auto",
"eps": "auto",
"weight_decay": "auto"
}
},
"scheduler": {
"type": "WarmupLR",
"params": {
"warmup_min_lr": "auto",
"warmup_max_lr": "auto",
"warmup_num_steps": "auto"
}
},
"aio": {
"block_size": 1048576,
"queue_depth": 16,
"single_submit": false,
"overlap_events": true,
"thread_count": 1
},
"activation_checkpointing": {
"partitioned_activations":true,
"number_checkpoints": 100,
"contiguous_memory_optimization": true,
"cpu_checkpointing": true,
"profile": true,
"synchronize_checkpoint_boundary": true
},
"flops_profiler": {
"enabled": true,
"profile_step": 1,
"module_depth": -1,
"top_modules": 3,
"detailed": true
},
"tensorboard": {
"enabled": true,
"output_path": "./logs",
"job_name": "finetune_gpt_j_6b"
},
"steps_per_print": 100,
"zero_allow_untested_optimizer": true,
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"wall_clock_breakdown": false,
"memory_breakdown": false
}
I have also the following lines of code:
import torch.distributed as dist
model.resize_token_embeddings(len(model.tokenizer))
model.gradient_checkpointing_enable()
model.to(device=dist.get_rank())
optimizer = AdamW(model.parameters(), lr=1e-4)
scheduler = lr_scheduler.ExponentialLR(optimizer, gamma=0.66)
How can I reflect the values of TrainingArguments in the deepspeed config file? Especially, how can I use the scheduler with exponential learning rate?
Second, I’m starting the script with the command python -m torch.distributed.launch --nproc_per_node=4 --nnodes=1 train.py. In train.py I’m calling the train() method of the Trainer. I’m not sure if deepspeed is actually running as there is no output on the command line related to deepspeed. Should I start it in a different way? I’ve read that it could be started by using the deepspeed command on the command line but somehow this does not work (command not found). I have installed deepspeed over pip.