GitLab

Co-authored-by: Min Xu <min.xu.public@gmail.com>

[FSDP] ssd_offload fixing backward path (grad_fn) for SsdFlatParameter and SsdFlatParameterView (#974)

* [FSDP] fixing backward path for SsdFlatParameter and SsdFlatParameterView when overriding .data

* Get ssd_offload unit tests passing

* [FSDP] get all test_fsdp_offload tests passing w/ ssd_offload on

* Update changelog

Co-authored-by: Min Xu <min.xu.public@gmail.com>

Improvements to ssd_offload to support pickling/unpickling SsdTensorHandle (and derived classes) (#964)

Verified that FSDP wrapped models using ssd_offload checkpoint save and restore correctly

* update pytest versions

* [test] test related changes

- upgrade to newer pytorch versions
- added function to make test more deterministic on A100 and TF32
- fixed some tests so that they are correctly skipped on a single GPU system

* more fixes

* formatting overly long lines

* format

* better test without trigger a warning

* fix an optim state bug with newer pytorch

- adam optimizer seems to return "step" as a singleton tensor now in the
nightly build
- this fixes it assumeing non-tensor value can still be loaded back by
the optimizer

* improve oss.py

- use min_loss for regression checking is a bit more reliable
- also increased the num epochs from 10 to 12

* small oss.py fix

* Update fairscale/nn/data_parallel/fully_sharded_data_parallel.py
Co-authored-by: Min Xu <min.xu.public@gmail.com>

* [feat] add CosFace paper's LMCL to MEVO

- added baseline algorithm to the reference kernel
- added MEVO version of LMCL
- added unit test to verify it is correct with respect to the reference as well as its memory usage

* updated changelog
Co-authored-by: Min Xu <min.xu.public@gmail.com>

* enable reduce scatter overlap with other operations

* fixed unit tests and added docstrings for the new parameters for fsdp

* fixed more unit tests

* fixed unit tests

* avoided the pickle error on process_group_reduce_scatter

* removed an unnecessary parameter in unit tests

* remove unnecessary prints

* fixed the docstring

* skipped the test_offload unit test because this unit test failed in the main branch

* removed the enable_reduce_scatter_overlap API parameter

* added doc string for the defualt value of process_group_reduce_scatter parameter

* fixed a syntax bug

* fixed a bug which cause unitest failure

* removed the all_gather in the ProcessGroupName enum

* added more comment

* changed the default value of process_group_reduce_scatter from None to ProcessGroupName.reduce_scatter

* Enabling ssd_offload training and test via tests/nn/data_parallel/test_fsdp_offload.py.
* Removed unused classes: SsdBuffer, SsdTensorHandleView, SsdParameter, SsdTensor
* Enhance test coverage of test_ssd_offloading_train_flatten_params_wrapper
* Modifications from PR #887 review comments.
* Update Changelog

- During eval, we will fallback to just output projection without fusing
- added unit test to ensure the shape is correct

* adding pre-commit files

* applying pre-commit to all files

* adding no-strict-optional argument to mypy in circle ci config

* fix typo

* updating python versions

* [skip ci] remove extra args

* adding python 3.9

* [skip ci] set pre-commit version in requirements-dev.txt

* set CACHE_VERSION

* move linters from circleci to github actions

* update python version

* update python version in benchmarks_2

* moving to python 3.9.7

* update release notes

* initial commit

* lint cleanup etc.

* helper functions; lint errors

* lint errors

* lint errors

* add back the boolean for named_parameters

* address comments and fix lint

* remove unused functions and class

* remove unused state

Add SlowMo Distributed Data Parallel for clusters with slow interconnects
Co-authored-by: Vinayak Tantia <tantia.vinayak1@gmail.com>

* [feat] MEVO kernel

- initial import from min/softmax and min/testing branches
- need to rename and further cleanup

* only test with newer pytorch

* renamed and added comments and code cleanup

* rename and reduce test memory

* testing

* minor fixing

* fixing

* more fix

* changelog

* more 1.7 and 1.8 paper cuts

* remove dead code

* addressed Benjamin's comments

* addressed more comments
Co-authored-by: Min Xu <min.xu.public@gmail.com>

* add doc strings

* add lower level SSD APIs and tests

* add the test to the list to be run

* remove unused imports

* more doc string changes

* fix lint errors

Fixes #827.
Co-authored-by: Eugen Hotaj <ehotaj@fb.com>

auto_shard.py currently uses torch.fx to create a symbolic DAG of
operations and linearizes that DAG into an nn.Sequential so it can later
be used for model offloading. This works in most cases but runs into
issues for certain eager mode features, such as dynamic conditionals,
shape-dependent computation, etc.

This PR extends auto_shard.py to first run a preprocessing step which wraps
any nn.Module which cannot be traced through. It adds a test for dynamic
conditionals and updates existing failing test code.

There are some immediate extensions to this approach which are marked as
TODO in the code.

* update python version for cpu tess

* run CPU tests with updated PyTorch version

* update nightly and test PyTorch versions

* skip failing multiprocess pipe test

* always skip test

* always skip test

* always skip test

* lint error

* skip unsupported versions

* improve skip message

* lint errors

* [fix] FSDP intra-backwards gradient accumulation.

Ensure gradient reduction accumulates into the unsharded gradient tensor
within a backwards pass. This matters when an FSDP module is called
multiple times within a forward pass, and reduction is _not_ deferred
using activation checkpoint forward counters, bucketing or some other
mechanism.

Closes #780

* [refactor] Remove forward counters. Comments.

Removed forward counters from the activation checkpointing utility, now
that FSDP does not require them for correct operation. Add more detailed
comment about memory usage behaviour with gradient reduction.

* [refactor] Delete deprecated forward counter usage.

* [refactor] Add state assertion as end of pre-backward hook.

* fixing bug in setting dependancies in parition handler

* modifying unit test to need the fix

* black

* Preparing pipeline for newer versions of pytorch

* updated error message

* Update torch to 1.9.0.dev20210614+cu102

* Update config.yml

* Update config.yml

* Update setup.py

* Update config.yml

* Update config.yml

* Update config.yml

* Update config.yml

[Offload][feature] Add auto shard functionality to remove requirement of nn.Sequential models. (#695)

* auto wrap functionality

* lint and doc strings

* fix lint errors

* lint errors and version skips

* remove mypy checking and add conditional import

* another math.prod instance

* another import fix

* address comments

* lint errors

* address comments

* fix lint errors

* add placeholder nodes to tracker list

This change also ensure that we calculate running_{mean,var} correctly
when wrapped.

* [feat] experimental.nn.SyncBatchNorm: initial commit

Fast/simple re-implementation of SyncBatchNorm.

When profiling SSL Vision, I was seeing a majority of cycles spent in
SyncBatchNorm. With this change, I see a 10% to 20% speedup on the
model I was profiling.

When running benchmarks/experimental/sync_batchnorm.py on 8 x V100,
I get a 6x speedup:

<class 'torch.nn.modules.batchnorm.BatchNorm2d'>
Elapsed time is  0.08709120750427246
Elapsed time is  0.12632274627685547
Elapsed time is  0.14095258712768555
Elapsed time is  0.16529417037963867
Elapsed time is  0.1419970989227295
Elapsed time is  0.15166854858398438
Elapsed time is  0.12000870704650879
Elapsed time is  0.17534875869750977
<class 'torch.nn.modules.batchnorm.SyncBatchNorm'>
Elapsed time is  2.5087168216705322
Elapsed time is  2.497001886367798
Elapsed time is  2.5204885005950928
Elapsed time is  2.526789903640747
Elapsed time is  2.5080230236053467
Elapsed time is  2.524489641189575
Elapsed time is  2.513214588165283
Elapsed time is  2.5359973907470703
<class 'fairscale.experimental.nn.sync_batchnorm.SyncBatchNorm'>
Elapsed time is  0.4126114845275879
Elapsed time is  0.39051294326782227
Elapsed time is  0.40685415267944336
Elapsed time is  0.4159870147705078
Elapsed time is  0.42383885383605957
Elapsed time is  0.4080159664154053
Elapsed time is  0.41202712059020996
Elapsed time is  0.42400121688842773

* adding auto graph generation for distributed pipeline

* ignore trace.py for my for now, since it needs pytorch 1.8

* fixing tests

* simplifying graph api

* remove unused debug utilities

* use inspect to find argument lists

* use sharded linear layer

* flkae8

* comment

* polishing

* polishing

[fix] Revert change that removed the option to run OffloadModel with out activation checkpointing. (#608)

* revert change made

* add tests and revert sync shard changes

* add tests

* remove file checked in by error

* inine var

* fix lint errors

* add checkpoint activation

* fix mypy

* use a bigger model

* modify tests for now

* resolve conflicts
Co-authored-by: Anjali Sridhar <anj@devfair0443.h2.fair>

replacing multip-process pipe implementation with more flexible one

Initial implementation of proposal pytorch/pytorch#55256

* renaming/adding error messages

* address comments

* address comments

* add more comments

* add more comments

* Fix ampnet unit test by adding delegate object

* Remove comments