GitLab

replacing new_group with torch.distributed.group.WORLD, avoids creating new
group in every iteration.

This should resolve the issue in Training gets stuck when using SyncBN #105

torch.distributed.new_group and torch.distributed.reduce_op are deprecated on pytorch 1.0.0, this fix can avoid some errors for now.

supporting user specified process group

* test passes

* notes

* Using C++-side flatten and unflatten functions

* Adding csrc

* Persistent synchronization event so it doesn't need to be created and destroyed each time

* Interop with parameter flattening in SSD

* Added deterministic option to imagenet main.py

* Adding options to split gradient averaging and allreduce in pure fp32

* Fixing allreduce_maybe_retain call

* Fixing allreduce_fallback

* Also sync active_i_buckets from rank 0

* Making retain_allreduce_buffers compatible with/orthogonal to delay_allreduce=True|False

* Correcting syntax error, now all seems to work with SSD

* Optional cpp extension build

* Add mixed precision adam optimizer (#59)

* Add FusedAdam Optimizer to Apex that places all the math into a cuda kernel.

* Added fixes to fused_adam to get it to work with network.

* wip work on python interface for adam with options

* fix dispatch for halfs, add python options to handle optional half gradients and params

* cleanup, get rid of grid-stride loop

* [syncBN]
  added syncBN in native pure python apex
  added fused cuda kernels used for sync BN. Using welford for mean/var
    optional installation using 'python setup.py install --cuda_ext'
  added unit test with side to side comparison between apex sync BN with
    PyTorch BN. Notice that for pytorch BN implementation, because of
    numerical issue for mean/var, the output will be slightly off.

* [syncBN PR]
  added fp16 support
  addressing review comments on:
    1. updating last pow 2
    2. look for import error when importing syncBN kernel

* [syncBN PR]
  added convert function to insert SyncBatchNorm
  refactored some kernel code

* fixing type issue (fp16/fp32/fp64)
added Kahan summation
editing unit test to use pytorch primitive ops with double, passing reasonable tests now

* updating tensor creation calls

* fixing the all_reduce contiguous tensor

* transposed all reduce results

* [syncBN]
support fp16 input & fp32 layer for apex fp16
partially fixing launch configs
enabling imagenet example to run with --sync_bn

* [syncBN PR]
Documentation added

* adjusting README

* adjusting again

* added some doc to imagenet example

* [syncBN]
  warp-level reduction
  bug fix: warp reduction logic updated. check for dummy element to avoid nan.
  improved launch config for better reduction kernels. Further improvements
would be to increase grid size.

* [syncBN]
  fixing undefined behavior in __shfl_down_sync from divergent threads in warp
reduction.
  changing at::native::empty to at::empty (upstream comments)

Moving gradient division back to after the allreduce.  Empirically, it appears underflow is more of a danger than overflow.

This is consistent with upstream, and safer against overflow.

* beautiful

* IT'S WORKING

* Hopefully fix race condition for fallback hook

* Updating test

* shared_param -> delayed_allreduce

* Adding a safety check

* One more check

* syntax...

* Fix appears to work in Tomasz's example.

* Somehow shared_param got de-enabled again?

* Proper implementation of LARC clipping
 * Documentation of LARC class
 * Modification of FP16_Optimizer to absorb optimizer instance that's being wrapped instead of creating new optimizer instance of same class.

* More stringent check for distributed refresh