GitLab

* add flag for DistributedAdam: step_support_amp_scaling
Co-authored-by: Kexin Yu <kexiny@nvidia.com>
Co-authored-by: Kexin Yu <kexinznzn›@gmail.com>

* fuse dropout into softmax in fprop for additive mask case

- resume() is a nested function and when it loads best_prec1
  it creates a local variable that hides the one from the parent
  function (which refers to the global one). This PR adds `global`
  to modify the global variable as intended
Signed-off-by: Janusz Lisiecki <jlisiecki@nvidia.com>

DistributedFusedAdam Model Parallelism Support (Megatron)
Co-authored-by: Kexin Yu <kexiny@nvidia.com>
Co-authored-by: Kexin Yu <kexinznzn›@gmail.com>

In this PR, we mainly tried to optimize the performance of Syncatchnorm and also fixed one potential issue in the welford_parallel kernel implementation.

For performance improvement, we batched the mean/var/count all_gather communication together and sent it once in the forward path
We also batch the all_reduce in backward path
We add the contiguous call on the input of welford_parallel kernel.
If there is any standard perf benchmark, I would be happy to run it.

Update asp readme

Update README for ASP

Added an outline to illustrate our recommended recipe to obtain a pruned model

Co-authored-by: pbialecki <pbialecki@nvidia.com>

* add device guards to the optimizers

* add untracked file

* set deviceGuard in multi_tensor_apply

* address review comments; fix lamb

* indent

* typo

Co-authored-by: Sukru Eryilmaz <seryilmaz@computelab-dgx1v-32.nvidia.com>

Asp sparse param dict update

Accept custom (layer type:param name) to include in sparse_parameter …

1. Support to include in sparse_parameter_list an user-supplied custom layer type and its parameter name. This is useful when users have their own implementation of nn.Linear or nn.Conv2D. For example, huggingface repo has a custom implementation of nn.Linear called LinearActivation.
2. Print info of layers in the model that are not pruned.

Merge pull request #917 from a-maci/master

Fixing the case when grads are None

Fixing mask multiplication with grad tensors

Grads can be None type. Adding this fix to skip multiplication with masks if this is the case.

Updating my repo

Fixed weight init for fused weight matrices in fused MHA by adding correct gain factor

Fixed variable name

* [sync BN]

support non-uniform batch size across process group.

TODO: test should be added once cleaned up.

* updating unit tests

* new unit tests for different inputs

* cleaning

* Only attempt to patch Tensor methods if defined

* syntax
Co-authored-by: Michael Carilli <mcarilli@nvidia.com>

add unit tests for FusedLAMB optimizer

2d masking and sparsity

Minor edit

Importance and usage is 2d masking

Update softmax.h

Fixing NULL stream launch in dispatch_additive_masked_softmax

add streaming support for softmax kernels