Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add 1-bit Adam support to DeepSpeed #380

Merged
merged 36 commits into from
Sep 9, 2020
Merged

Add 1-bit Adam support to DeepSpeed #380

merged 36 commits into from
Sep 9, 2020

Conversation

awan-10
Copy link
Contributor

@awan-10 awan-10 commented Sep 8, 2020

No description provided.

awan-10 and others added 8 commits September 8, 2020 10:23
* 1-bit Adam v1 (squash) (#346)

* testing_onebit

* test_passed

* test_passed

* updated compressed_allreduce

* 123

* cpu2gpu test added

* Add non-cuda-aware code path. Segfaults for > 2 procs.

* Works for 4 procs with numpy buffers now. TODO: cleanup, evalute perf.

* Fix gather. Cleanup.

* Add new tests.

* Reduce memory footprint. BERT large with BS=16 works.

* Revert "Reduce memory footprint. BERT large with BS=16 works."

This reverts commit e7f38fc.

* Update optim to support bert-large.

* with initialization added on bert_onebit_adam

* This works!!

* Force igather for cupy. Better performance. Need cleanup and reorg to
support TCP now.

* X

* testing the fintune task for FP32 training

* Added the fintune taks for FP32 training

* With the control flag of freeze_step added

* added the freeze_step inside fp32_onebitadam

* Seperate freeze_kernnel added

* Added the sanity test for the Compressed_Allreduce

* Test for Compressed_Allreduce passed, but AllGather need sync.

* add checks for finetuning.

* Running passed for finetune on EastUs

* Add one bit adam clean file.

* Refactor comms. code and move it to a new file.

* fix compile/run errors.

* Adding changes for onebit_adam from Hank

* Save memory by modifying in-place.

Co-authored-by: Your Name <[email protected]>
Co-authored-by: Ammar Ahmad Awan <[email protected]>
Co-authored-by: tanghl1994 <[email protected]>
Co-authored-by: Hank <[email protected]>
Co-authored-by: root <[email protected]>

* Staging 1bit adam v1 (#348)

* Refactor to correct locations.

* Deleted unused files.

* Fix imports for refactored codebase.

* update the com reduce test.

* Fix some errors.

* Fix optimizer name

* Delete unused tests.

* Fix formatting for pre-commit.

* Add cupy dependencies.

* add cupy for cuda 10add cupy for cuda 10

* Add mpi4py requirement.

Co-authored-by: Ammar Ahmad Awan <[email protected]>

* Use correct initialization for exp_avg.

* Cleanup onebit adam.

* minor wording fix.

* Cleanup custom collectives.

* Fixes for TCP support.

* fix formatting. fix formatting.

* move 1bit adam reqs

* delay importing 1bit adam unless it's used, this will ensure we delay importing mpi4py

* Fix cuda version parsing.

* Temporary tcp fix.

* Update install.sh

* Refactor code to properly support cuda-aware comm.

* Fix imports.

* add cuda_aware flag to tests.

* Cleanup. Add copyrights.

* Add 1-bit Adam tutorial v1.

* Minor fixes to copyright and print statements.

* Update utils.py

* Update utils.py

Co-authored-by: Jeff Rasley <[email protected]>
Co-authored-by: Your Name <[email protected]>
Co-authored-by: tanghl1994 <[email protected]>
Co-authored-by: Hank <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: Ammar Ahmad Awan <[email protected]>
* refactor pdsh and add openmpi launcher

* fix bugs/cleanup
@jeffra jeffra merged commit 01726ce into master Sep 9, 2020
@jeffra jeffra deleted the staging-1bit-adam-v2 branch November 19, 2020 23:27
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants