GitLab

* Minor, ease of life to debug and makes it possible to test a host of models with the same code

* adding a cpu option
* adjust the reference loss

* bugfix
* adjust default non-regression loss, not all_reduced now

More realistic benchmarks, comparing apples to apples. DDP/OSS+DDP/OSS+SDP

Same bucketing strategy for OSS and SDP:
sort everything ahead of time, per rank and per size, smaller tensors first. Bucket the smallest elements in a fixed buffer, send async, then send all the others async, and get back to the bucket. Once done then scatter the contents if needed

- adding the buffer broadcast option
- minor cleanup in shardedDDP

- small benchmark refactor, only one for all backends and ddp
- deterministic, enforce alignment with pytorch ddp

* Broadcasting grad-enabled tensors is forbidden in Gloo, because this is not differentiable. Workaround

* Doc extensions to some APIs
* FIx the benchmark and tutorial

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

- rename oss_ddp to ShardedDataParallel
- some refactoring
- ShardedDataParallel owns the sharded optimizer, exposed if need be
- some small perf bumps

Changes the structure of the returned state dict with respect to the param_groups to make it closer to what a vanilla optimizer would return (un-shard them). Shard again when loading

* Aligning the optimizer state dict with what PyTorch expects

* Adding a check on the dict keys, ensure that `state` and `param_groups` are there

* after installing the specific isort, black and all, one liner to please the linter..

* Adding some measurement of the memory consumption while training + checkpointing

* mandatory lintfix commit

* brainfart, reset the memory use counter at the beginning of the training in case two of them are run in a row

* move reset stats call, hotfix

* move the optimizer to rmsprop, more stateful and still used in CV

* trying to figure out a sigsev in circleci

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>

* Aligning the optimizer state dict with what PyTorch expects

* Adding a check on the dict keys, ensure that `state` and `param_groups` are there

* after installing the specific isort, black and all, one liner to please the linter..

* specify chunks for pipe/transformer benchmark

Set chunks to be equal to len(balance) for pipe/transformer benchmark. Will update words per second and memory usage checks in next commit (must test on CircleCI to find appropriate values)

* change benchmark words per second and memory usage

Did six runs for words-per-second, with results: 9144.40, 9163.91, 9993.01, 9082.82, 9155.09, 9000.67
Peak allocated bytes per device (which does not change between runs) were 193206272, 645632, 562688, 92688384 for devices 0, 1, 2 and 3, respectively

* increase batch size

batch size was small enough that the GPU's computing power was not the bottleneck, slowing training and specifically making more chunks slower. Increasing batch size has therefore increased training speed

* update benchmark numbers

ran six times, with wps 36917.44, 36797.65, 37006.03, 36872.84, 37129.31, 37003.31 and peak allocated bytes 4061909504, 4050944, 10427392, 2031824896 for devices 0,1,2 and 3 respectively.
Co-authored-by: Jun Ru Anderson <andersonic@fb.com>

Implement scaling of optimizer state when using pure-fp16 training to avoid underflow. Update benchmark to use pure-fp16. Modify state_dict methods to store and load the optimizer state scale.
Co-authored-by: Jun Ru Anderson <andersonic@fb.com>

* initial commit, dummy training loop, pure pytorch but not DDP

* probably slightly broken, but rough DDP benchmark run

* adding the torchvision requirement for testing

* brainfart

* reduce the loss, do something slightly distributed

* Some cleanup, distributing the training on two GPUs

* some cleanup + adding a vanilla run, still not good to go

* less silly defaults, gtg for a start I think

* smaller batch to fit the smaller gpus used in the circleci rigs

* Adding some options for the benchmark, and regression testing

* [test] set torch seed for Adam tests (#49)

Set the torch seed for tests. xfail mixed precision and memory-efficient mixed-precision state_dict tests due to their states being cast to FP16 and back to FP32 during load_state_dict.
Co-authored-by: Jun Ru Anderson <andersonic@fb.com>

* linting, I really need to automate this isort insanity
Co-authored-by: Jun Ru Anderson <33384298+andersonic@users.noreply.github.com>
Co-authored-by: Jun Ru Anderson <andersonic@fb.com>

Set the torch seed for tests. xfail mixed precision and memory-efficient mixed-precision state_dict tests due to their states being cast to FP16 and back to FP32 during load_state_dict.
Co-authored-by: Jun Ru Anderson <andersonic@fb.com>

Allow training with optimizer state in fp16. Use an enum to select from full-precision, mixed precision, memory efficient mixed precision and pure fp16. Improve clarity of testing code
Co-authored-by: Jun Ru Anderson <andersonic@fb.com>

Add support for mixed-precision (half precision params, full precision gradients) and memory-efficient (half precision params and half precision gradients) training with Adam
Co-authored-by: Jun Ru Anderson <andersonic@fb.com>

Add FusedAdam, update benchmark and add tests.
Co-authored-by: Jun Ru Anderson <andersonic@fb.com>