GitLab

* Add log util

* replace all occurrences of print and logging

* address format

* disable propagate to avoid duplicate log

* Transformer kernels release
Co-authored-by: Shaden Smith <ShadenTSmith@gmail.com>
Co-authored-by: Elton Zheng <eltonz@microsoft.com>
Co-authored-by: Reza Yazdani <reyazda@microsoft.com>
Co-authored-by: RezaYazdaniAminabadi <44502768+RezaYazdaniAminabadi@users.noreply.github.com>
Co-authored-by: Tunji Ruwase <olruwase@microsoft.com>
Co-authored-by: Shaden Smith <ShadenTSmith@gmail.com>
Co-authored-by: Shaden Smith <Shaden.Smith@microsoft.com>
Co-authored-by: Samyam Rajbhandari <samyamr@microsoft.com>
Co-authored-by: Shaden Smith <ShadenTSmith@gmail.com>
Co-authored-by: Jeff Rasley <jerasley@microsoft.com>
Co-authored-by: Samyam Rajbhandari <samyamr@microsoft.com>
Co-authored-by: Shaden Smith <ShadenTSmith@gmail.com>
Co-authored-by: Elton Zheng <eltonz@microsoft.com>
Co-authored-by: Reza Yazdani <reyazda@microsoft.com>
Co-authored-by: RezaYazdaniAminabadi <44502768+RezaYazdaniAminabadi@users.noreply.github.com>
Co-authored-by: Tunji Ruwase <olruwase@microsoft.com>
Co-authored-by: Shaden Smith <Shaden.Smith@microsoft.com>
Co-authored-by: Samyam Rajbhandari <samyamr@microsoft.com>

* fix: typo in code docs

* more pythonic code

Co-authored-by: Shaden Smith <Shaden.Smith@microsoft.com>
Co-authored-by: Jeff Rasley <jerasley@microsoft.com>

* add support for predivide as a flag
* add predivide json config, remove allgather_disable (as it's not currently used anymore)

Contiguous Gradients should be set to false by default. Its not useful unless the model is very large

* Fix for CPU memory Bloating Issue caused by pyorch backward graph creation in allgather. Fixed by calling detach on tensors before calling all_gather

* Fix for CPU memory Bloating Issue caused by pyorch backward graph creation in allgather. Fixed by calling detach on tensors before calling all_gather

* Fix for CPU memory Bloating Issue caused by pyorch backward graph creation in allgather. Fixed by calling detach on tensors before calling all_gather

* updates to support fp32 grad clipping and disable max_grad_norm

Updates for ZeRO stage 2 + ZeRO stage 1 w. RS
Co-authored-by: Tunji Ruwase <olruwase@microsoft.com>
Co-authored-by: Samyam Rajbhandari <samyamr@microsoft.com>
Co-authored-by: Shaden Smith <ShadenTSmith@gmail.com>
Co-authored-by: Elton Zheng <eltonz@microsoft.com>
Co-authored-by: Shaden Smith <Shaden.Smith@microsoft.com>
Co-authored-by: yuxionghe <yuxhe@microsoft.com>
Co-authored-by: Arash Ashari <arashari@microsoft.com>

* Support dynamic loss scale args in fp16 optimizers

* Update names

* update apex version to feb 5th commit

* use gradient clipping instead of max grad norm in tests

* add warning when user provides max_grad_norm

* update examples commit

1) CSR parameter names should end with .weight. 
2) When using basic optimizer directly, DeepSpeed should handle zero_grad. Letting the basic optimizer do the zero_grad resulted in residual gradients in the embedding layer due to unknown reasons.

Co-authored-by: Shaden Smith <Shaden.Smith@microsoft.com>

* Push to remote

* Correctly handle multi output models by doing loss scaling in backward()
Unit tests for multi output models

* Fix formatting issues

* Formatting issues fix

* Fix formatting

* Update DeepSpeedExamples submodule
Enable Megatron model tests

* added zero_allow_untested_optimizer flag helpers

* add zero_allow_untested_optimizer config constants

* zero_allow_untested_optimizer logic with assertion

* Added unit test and CustomOptimizer helper class

* remove session_params in deepspeed_constants.py

* add constants info at README.md

* add support for torch 1.3+ builds inside a docker build environment
* remove apex imports

* Enhancement: Ability to load checkpoint without loading the optimizer states. Unittest testing saving and loading checkpoint with fused, unfused and zero optimizer. The unitest takes about 165s

* add tests cases for onecycle policy with fp16/zero

* Make lr schedulers support fp16 optimizers

* Fix formatting

* More specific naming
Co-authored-by: Jeff Rasley <jerasley@microsoft.com>

* add support for deepspeed env file to pass custom env values

* simplify deepspeed config example

* add mpirun support for openmpi 4.0

* add master addr support from args

* switch mpi detection to use mpi4py

* set constant for default distributed port

* Make sure deepspeed_mpi exits in args

* add auto-detect to torch dist init

* update tests to infer distributed init status

* prevent crash if dist_init_required is True but already initiliazed

* only init if safe to do so (forgot to add this file in prev commit)