when i use deepspeed in trainer,Set lr_scheduler_type=“constant_with_warmup” in TrainingArguments, set in deepspeed zero config is below:
{
"scheduler": {
"type": "WarmupDecayLR",
"params": {
"total_num_steps": "auto",
"warmup_min_lr": "auto",
"warmup_max_lr": "auto",
"warmup_num_steps": "auto"
}
}
}
or do not set scheduler in deepspeed zero config.
I want to know which strategy to use when training the model.