Cheng Li comments

Results 26 comments of


                                            Cheng Li

trafficstars

some fix in flops_profiler

Hi @lucasleesw , thanks for the PR. Is there a reason you use __str__ instead of __str__()?

Expose DS-inference and ZeRO-inference configs to user

ZeRO inference currently takes the general deepspeed config json path. Do we want to expose DS-inference config in the deepspeed config file as well?

[BUG] Sample alexnet example for flops profiler does not work.

@LM-AuroTripathy , thanks for pointing this out. This will be fixed by https://github.com/microsoft/DeepSpeed/pull/2443.

[BUG] when I tracking the FLOPs by FlopsProfiler, the FLOPs become greater and greater?

Hi @GongCQ AND @insujang, this is expected if start_profile and end_profile are not set correctly. the profiler is supposed to be started and ended on a single training step (in...

[deepspeed/autotuner] Bug fix for binary search for batch size

@rahilbathwal5, thanks for the PR!

[deepspeed/autotuner] Bug fix for binary search for batch size

@thanks for the PR

Profiler enhancements: adding memory details to profilder, adding dtype breakdown for memory and params

@clumsy , sorry for the late response, I was out of office for a while. Can you please resolve the conflicts with the main branch? Thanks.

Profiler enhancements: adding memory details to profilder, adding dtype breakdown for memory and params

> Sure @cli99, are you in favor of the proposed changes though? Now I also need to find a way to make it reflect the expert parameters. Most changes look...

[BUG] AttributeError: 'PipelineEngine' object has no attribute 'flops'

Hi @yerimChoi, the flops profiler does not support deepspeed pipeline parallel engine yet, if you use pp_size > 1, disable the flops profiler

TypeError: _upsample_flops_compute() takes 1 positional argument but 5 were given

Hi @gauravmunjal13 , thanks for pointing out this issue. It shall be fixed by https://github.com/microsoft/DeepSpeed/pull/3782