GitLab

* [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.

* [fix] auto_wrap: support wrapping based on wrapper_config

- user can use this to avoid assert if auto_wrap is used multiple times on a module
- user can traverse the modules multiple times and assign a wrapper_config
  to the module and then use auto_wrap once to wrap them

fix #649
fix #585

* added changelog

* fix tests

* fix a test

* added an optional assert for collision based on discussions with Quentin

* added config_auto_wrap_policy

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

* fix saving and loading checkpoints with use_sharded_state=True

* mypy fix

* better fix of the infinite recursion

- we need to specifically call FSDP.state_dict from its local state_dict
- added unit test that fails without the fix and works with the fix
- fixed mypy for the overloaded functions

* make cpu-only fsdp work for state_dict at least
Co-authored-by: Min Xu <min.xu@acm.org>
Co-authored-by: Min Xu <min.xu.public@gmail.com>
Co-authored-by: Min Xu <m1n@fb.com>

* [fix]: let FSDP handle model with multiple forward pass and checkpoint

* try CI again

* save

* save

* fixed case with bn

* minor

* add the new file

* minor

* added test of a single case, runtime is about 50s

* enable all 8 test cases

* cleanup

* cleanup

* skip flatten case with 1.6 and 1.7

* minor
Co-authored-by: Min Xu <min.xu@acm.org>

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>

Adds support for distributing pipeline stages across multiple processes (and therefore multiple machines)
* Adds a style argument to the Pipe constructor, defaulting to PipelineStyle.SingleProcess, but also supporting PipelineStyle.MultiProcess
* Added support for lazy construction of modules (see lazy_construction for an example)
* Added two implementations of inter-process communication: one based on rpc with globally visible queues, one based on send/recv
* Copied all the relevant tests from tests/pipe to tests/pipe_process and modified them to exercise PipelineStyle.MultiProcess