GitLab

* [feat]: checkpoint and normalization

- added special handling of BN for track_running_stats and checkpointing
- we test BN/LN and checkpointing
- we test them with mixed precision

- cover them in terms of code path only
- numerically, AdaScale is different on SDP/FSDP than DDP, mainly
  due to partial view of the gradients.
- this doesn't mean it is definitely not useful but it is yet to
  be validated.
- not going to spend too much time until we have a real use case.

This adds a context manager that assists in making child modules with similar defaults.
Usage:
```
from fairscale.nn.misc import enable_wrap, wrap

with enable_wrap(**handleful_of_important_params):
    layer_1 = wrap(torch.nn.Linear(5, 5))
    layer_2 = wrap(torch.nn.Linear(5, 5), flatten_parameters=True) # Override parameters if you'd like

# without the context manager, creates Linear layer
layer_1 = wrap(torch.nn.Linear(5, 5))
```
If not within the FSDP context, this would be a no-op. This makes it easier to annotate layers without having to copy any changes in parameters.

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>

* 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>

* 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 a parity unit test
* code review, better testing, use torch defaults and check for the loss, log world size

* [doc]: AdaScale example and notes

* formatted notes correctly as suggested by Benjamin

* added feature and unit test to make sure lr_scheduler works

* update the example with lr_scheduler

* fixed doc with "make html"

* addressed Mike's suggestions

* 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

* adding a shard-aware GradScaler wrap, credits to Sean Naren for the idea
* adding stubs & explanations in the documentation

add a clip gradients util, equivalent to torch's but aware of the sharded states. Add a corresponding unit test

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

* small refactor, getting rid of the while loop

- Aurick noticed this bug and I ran into it yesterday
- after the fix, our cifar training shows same gain values from
  different replics now:

```
20-Oct-20 16:00:19 - DEBUG - rank1 - scale 2, gain ratio 1.3512124098087777
20-Oct-20 16:00:19 - DEBUG - rank0 - scale 2, gain ratio 1.3512124098087777
20-Oct-20 16:00:19 - DEBUG - rank1 - timing: data 0:00:00.000600 fwd 0:00:00.003678 loss 0:00:00.000086 bwd 0:00:00.314158 update 0:00:00.002132 rest 0:00:00.000399
20-Oct-20 16:00:19 - DEBUG - rank0 - timing: data 0:00:00.000643 fwd 0:00:00.003460 loss 0:00:00.000084 bwd 0:00:00.314678 update 0:00:00.002001 rest 0:00:00.000408
20-Oct-20 16:00:19 - DEBUG - rank1 - scale 2, gain ratio 1.3514997779980324
20-Oct-20 16:00:19 - DEBUG - rank0 - scale 2, gain ratio 1.3514997779980324
20-Oct-20 16:00:19 - DEBUG - rank1 - timing: data 0:00:00.000732 fwd 0:00:00.003689 loss 0:00:00.000086 bwd 0:00:00.314176 update 0:00:00.002146 rest 0:00:00.000397
20-Oct-20 16:00:19 - DEBUG - rank0 - timing: data 0:00:00.000646 fwd 0:00:00.003542 loss 0:00:00.000089 bwd 0:00:00.314549 update 0:00:00.001956 rest 0:00:00.000392
20-Oct-20 16:00:19 - DEBUG - rank1 - scale 2, gain ratio 1.352149646693932
20-Oct-20 16:00:19 - DEBUG - rank0 - scale 2, gain ratio 1.352149646693932
```

- fixed typing
- make it run less often to reduce CI time

testing: run it in a loop make sure it is run in the right frequency.

- close #143

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

Add GradScaler to Fairscale, subclassing PyTorch's GradScaler. Use GradScaler in the pipe benchmark; though it is not needed in this case, it is a good example of how to use gradient scaling for larger models that do require gradient scaling in order to converge.
Co-authored-by: Jun Ru Anderson <andersonic@fb.com>

Workaround PyTorch bug that casts state (pytorch/pytorch#43706).

Copied from https://github.com/pytorch/fairseq/blob/v0.9.0/fairseq/optim/fp16_optimizer.py#L251-L268

authored-by: Mandeep Singh Baines <msb@fb.com>

Co-authored-by: Jun Ru Anderson <andersonic@fb.com>