GitLab

Recent work by [Microsoft](https://arxiv.org/abs/1910.02054) and [Google](https://arxiv.org/abs/2004.13336

) has shown that data parallel training can be made significantly more efficient by sharding the model parameters and optimizer state across data parallel workers. These ideas are encapsulated in the new **`FullyShardedDataParallel` (FSDP)** wrapper, which is a drop-in replacement for PyTorch's `DistributedDataParallel` (DDP) wrapper.

Compared to PyTorch DDP:
* FSDP shards parameters (FP16 + FP32) and optimizer state across data parallel GPUs
* FSDP with `reshard_after_forward=False` has the same communication cost as PyTorch DDP and is similar to ZeRO-2
* FSDP with `reshard_after_forward=True` increases total communication by 50% and is similar to ZeRO-3:
    * all-gather parameters at start of forward pass and start of backward pass
    * reduce-scatter grads at end of backward pass
Co-authored-by: Min Xu <24926999+min-xu-ai@users.noreply.github.com>
Co-authored-by: Sam Shleifer <sshleifer@gmail.com>

* test with and without buckets for all the shardedDDP unit tests
* parametrize all the things
* refactoring, adding even more  combinations at times
* handle hosts not having cuda

* Adding multiple groups support to ShardedDDP + unit test
* adding gloo to the backends tested for multiple groups

* [fix] ShardedDDP train/eval modes
* Update CHANGELOG.md

* initial implementation, with unit test and assert
* added changelog and better debug string

* Better unit testing
* Make it possible to refresh the DDP assumptions when the model has changed. Make it optional so that you save some time
* Enabling accumulation tests

* Add fairscale.utils.containers
Co-authored-by: Min Xu <24926999+min-xu-ai@users.noreply.github.com>

* Add fairscale.nn.misc.checkpoint_activations
Co-authored-by: Sam Shleifer <sshleifer@gmail.com>
Co-authored-by: Min Xu <24926999+min-xu-ai@users.noreply.github.com>
Co-authored-by: Sam Shleifer <sshleifer@gmail.com>

* Adding a proper ddp parity / AMP unit test, overdue
* catch non-AMP pytorch

* adding the .to(device) support + unit testing
* doc update

* no idea about the root issue, but it proved to be fairly narrowed (gloo+cpu+python3.8+no cuda installed) so I guess that's out of scope for fairscale

* Add AMPnet implementation (clean version)

* Move ampnet to experimental

* Move stuff around pipeline

* Address review comments and fix pre-commit errors

* Refactor and modify delegate functionality

* Modify header in pipe.py

* working around broken mypy

* minor, but ease of life, one less papercut

* tentatively fixing the cpu version of circleci jobs, now pipe tests are the last ones standing
* fixing oss backcompat, trying to fix rpc in old pytorch also
* fixing the file based init in torch 1.5

* adding the pytest timeout plugin to properly root out hanging tests
* removing redundant code, slightly more reasonable timeout, works on single cuda
* finding the root bug for some of the cpu hangs, rpc init
* propagating all the rpc init test changes to the pipe and model parallel tests

* fix typo, backend for CPU test

* Add function to add handle for sync BN
* Add test to ensure batch norm handles have been added

* catching properly a given test failing if not enough gpus

* file based dist init
* nicer handling of broken world sizes vs. number of available GPUs, do not break but warn out

[OSS] Getting rid of the "should bucket" hash table, just use a list + non trainable params fix (#259)

* Getting rid of the "should bucket" hash table, just use a list
Properly handle all params, with or without requires_grad

* make sure that this case is unit tested

* unit test checking ddp and sharded_ddp equivalence, reproducing the issue that Sean spotted
* fixing the issue, not counting requests in flight properly
* adding a multiple optimizers case

* proper unit testing, but no other solution than disabling bucketing for now, couple of options tested do not work

* fallback on internal pytorch numbering

* rewrite using autograd and Variable execution queue to make the reduce automatic
* share buckets with OSS to remove duplication
* some speed still likely on the table since the speed vs. bucketing does not match expectations, could be a follow up

Adds support for:
* Reused layers (e.g. for weight sharing)
* Lazily-constructed layers
* Single-process control via PipeRPCWrapper
* PipelineStyle.AsyncScheudle, which lays the foundation for asynchronous pipeline work by introducing an event loop for each rank/worker to process either activations or gradients as they arrive

Also added examples for multi-process and PipeRPCWrapper