Does the open-source code for FastPersist include the last two optimizations mentioned in the paper, "parallelizing checkpoint writes over DP ranks and pipelining checkpoint writes"?

[Buddingpopp]

I followed the guide to perform checkpoint testing, and FastPersist performed very well. https://github.com/deepspeedai/DeepSpeedExamples/tree/master/deepnvme/model_checkpoint
I'd like to know if the open-source version of FastPersist includes the last two optimizations mentioned in the paper(https://arxiv.org/abs/2406.13768). If not, are there any plans to open-source these optimizations in the near future? Thank you.

[tjruwase]

@Buddingpopp thanks for your interest.

1. Yes, the open-source includes the code for all the results in the paper, including parallelizing and decoupling optimizations.
2. Yes, we plan to properly release docs and blogs for using these optimizations, but currently lack the bandwidth.

[Buddingpopp]

Thank you, I'm looking forward to your docs and blogs. Do the micro-benchmarks currently support testing of parallelizing and decoupling optimizations?

[tjruwase]

No, the micro-benchmarks currently don't support testing of parallelizing and decoupling optimizations because they currently lack some advanced logic like process group management.