deepspeed ignoring steps_per_print

[ansar-ayan]

Hi everyone,
deepspeed is ignoring the step_per_print parameter in logging. is this a bug and is there a workaround to this?

[tjruwase]

@ansar-ayan, can you please share more details of the problem you are seeing. Below is a use of the parameter.
https://github.com/microsoft/DeepSpeed/blob/c27483933d50a693fef9c48418d2664cf6a6a6f8/deepspeed/runtime/engine.py#L2169

[ansar-ayan]

Hey @tjruwase, Its when I am passing the config
model, _, _, _ = deepspeed.initialize( model=model, model_parameters=model.parameters(), config={ "train_micro_batch_size_per_gpu": batch_size, "optimizer": { "type": "Adam", "params": {"lr": 1e-4}, }, "zero_optimization": { "stage": 1, "offload_optimizer": {"device": "cpu"}, }, "logging": { "steps_per_print": 10000, }, }, )
Even after specifying the logging, it is ignoring the param. I am using the 0.15.1 version. Will depreciating the version work?

[ansar-ayan]

@tjruwase I was incorrectly using step_per_print.
Its working correctly now. Thanks for your help :)

[tjruwase]

@ansar-ayan, I am glad you are unblocked. Sorry that you wasted time on this since ideally DeepSpeed should have alerted you on the typo. We have partially enabled configuration validation for ZeRO via #2004 to detect these kinds of typos. Unfortunately, we currently lack bandwidth to extend to other parts of the codebase. We would appreciate if you had bandwidth to help with this feature. Thanks!