GitLab

* Skip the failing unit tests from the FusedRMSNorm PR

* Update test_lamb.py
Co-authored-by: Jithun Nair <37884920+jithunnair-amd@users.noreply.github.com>

* FusedRMSNorm/"T5LayerNorm" based on FusedLayerNorm (#1274)

* FusedRMSNorm based on FusedLayerNorm

* refactor duplicated kernels

* delete comments

* delete comments

* cleanup

* cleanup

* cleanup, fixed clobbering forward_affine_mixed_dtypes

* fix pybind naming and add MixedFused test

* undo skipping

* check elementwise_affine

* Update tests/L0/run_fused_layer_norm/test_fused_layer_norm.py

Oof, nice catch, thanks
Co-authored-by: Masaki Kozuki <masaki.kozuki.2014@gmail.com>
Co-authored-by: Masaki Kozuki <masaki.kozuki.2014@gmail.com>

* fix and generate docs for FusedRMSNorm (#1285)

* [FusedRMSNorm doc] document where epsilon is added (#1295)

* [FusedRMSNorm doc] add epsilon to formula

* correct

* better wording

* Fix some bugs

* Optimize HostRMSNormGradient and HostApplyRMSNorm for AMD GPUs

* Fix NaN issues in FusedRMSNorm

* Update test_fused_layer_norm.py

* Skip test_fused_layer_norm.TestAutocastFusedRMSNorm on ROCm

* Use at::cuda::warp_size() instead of at::cuda::getCurrentDeviceProperties()->warpSize
Co-authored-by: eqy <eddiey@nvidia.com>
Co-authored-by: Masaki Kozuki <masaki.kozuki.2014@gmail.com>
Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>

* use _all_gather_base
Signed-off-by: ericharper <complex451@gmail.com>

* use _reduce_scatter_base
Signed-off-by: ericharper <complex451@gmail.com>

* remove torch empty in backward
Signed-off-by: ericharper <complex451@gmail.com>

* check self.attn_mask_type
Signed-off-by: ericharper <complex451@gmail.com>

* remove extra arg
Signed-off-by: ericharper <complex451@gmail.com>

* update get_tensor_shapes logic
Signed-off-by: ericharper <complex451@gmail.com>

* Improvements in distributed Adam optimizer for Megatron

Add option to allocate gradient buckets out of one large buffer. Add option to initialize params in user-provided order. Perform communication when saving optimizer state. Support param sync with any dtype.

* Style fixes in distributed Adam helper classes

Review suggestions from @crcrpar

* Hack to enable training GPT-3

Seems to fix bug from #1416

* Add test to initialize model-parallelism for decoder-only Transformers

Namely GPT-3.

Bug fixes, perf improvements

* add test

* update batch sizes

* update batch sizes

* small updates

* delete comment

* add async comm

* add sync if needed

* update tests

* remove redundant imports

* code cleanup

* minor updates

* update dtype for comparison

* fix dtypes

* fix typo

* modify sizes and use common_utils.find_free_port

* fix typo and use double precision

* revert some changes, create test for profiling on L1

* remove redundant line

* revert UCC_TLS and add sync to fwd_bwd

* code clean up

* code clean up

* modify BERT test

* add comment

* follow the current signature
Signed-off-by: Masaki Kozuki <mkozuki@nvidia.com>

* call .backward on outputs
Signed-off-by: Masaki Kozuki <mkozuki@nvidia.com>

* update the other caller of _softmax_backward_data
Signed-off-by: Masaki Kozuki <mkozuki@nvidia.com>

* Time dimension shape check for fused scale mask softmax kernel
Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

* Add shape test
Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

* Fix mask shape
Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

* update: mpu for t5 rpe

* update: add rpe mpu group test

* fix semicolon bugs
Co-authored-by: Masaki Kozuki <masaki.kozuki.2014@gmail.com>

* fix semicolon bugs
Co-authored-by: Masaki Kozuki <masaki.kozuki.2014@gmail.com>
Co-authored-by: Masaki Kozuki <masaki.kozuki.2014@gmail.com>

* remove pyprof

* remove reparameterization

* remove pyprof test

* clean up

* Add features to distributed Adam for Megatron support

Support gradient clipping, gradient scaling, FP32 grad accumulation, and multiple dtypes and devices.

* Restore closure arg to distributed Adam

Review suggestion from @crcrpar

* Integer driver number comparison

* packaging

* it looks possible to remove this file

* add communication collectives

* update Column|RowParallelLinear

* update checkpoint function

* update function name

* parity between public and private collectives

* row parallel linear

* column parallel linear

* sequence parallel: p2p comm

fix typo

* sequence parallel: pipeline parallel

* fix typo

* add layernorm with sequence_parallel_enabled attr

* class variable -> member variable

* fix col parallel test with sequence parallel

* Initial test of `forward_backward_pipelining_without_interleaving` with `model_type=ModelType.encoder_and_decoder`

* add cases pretending to test sequence_parallel

* Apply 2 suggestion(s) to 1 file(s)

* update sequence_parallel_enabled docstring

* update docstring: order of tensor dimensions, sequence_parallel_enabled behavior

* Divide sequence_length if sequence parallel

tensor shape should be updated if sequence parallel is enabled.

* cherry-pick https://github.com/NVIDIA/Megatron-LM/commit/8474e6e54fcb9dfa37aea039352f9fb485fb6f61

* type annotation

* Fix matmul call in RowParallelLinear

Fix `sequence_parallel_enabled` to `False` as you can see in
https://github.com/NVIDIA/Megatron-LM/blob/d898a8991d1a08d29074f87819d1bf41517e35f5/megatron/mpu/layers.py#L511-L514

* update rowparallellinear test

* fix `loss_weight` is not defined in test_layers

* @eqy's comment

* mixed fused layer norm

* fix typo

* misc

* test_layers cleanup

* Skip Bert/GPT script

Since these two models haven't gotten updated for sequence parallle, e.g. the update of the order of dimension from (batch, sequence, feature) to (sequence, batch, feature) and global variables of arguments

* debug part 1/N: comment out `x.retain_grad`

* debug part 2/N: [ColumnParallelLinear] comment out overriding of sequence_parallel_enabled

* debug 3/N: add pipeline test with parallel mlp

* Fix handling `self.input_tensor` and argument

* tp2pp4 ModelType.encoder_or_decoder is failing, which can be at my fault because the backward is blaming the output and the grad_ouptut shape don't match

* revert debug 1/N

* defer tensor model parallel size > 1

* split tensor in sequence dim

* cosmetic

* cosmetic: remove archaic comment

* enable TP>1 for encoder_and_decoder as well

* set requires_grad=True always...

* Set `scatter_gather_tensors_in_pipeline` to :obj:`False`

for the sake of nemo megatron's GPT works with sequence parallel enabled.

* brush up comment of `requires_grad()`

There's a possibility that PyTorch DistributedDataParallel hangs
when some tensor (or parameter) doesn't require grad according to @ptrblck.
This forced `requires_grad` in my understanding is different from that.

* misc changes of scatter_gather_tensors_in_pipeline comment

* guard for torch_ucc

* cosmetic changes related to tests

* update command line arguments

* update TransformerLanguageModel

* rename

* move gpt to gpt.py

* update bert

* add all_gather for params in sequence parallel region

* misc. some diffs were lost during rebasing...

* updates for non sequence parallel execution

* gpt with sequence parallel

* Apply 2 suggestion(s) to 2 file(s)

* update tensor&pipeline parallel size

* why `sequence_parallel_enabled` is not supplied!? Did I messed up when rebasing?

* cosmetic fix

* correct key is sequence_parallel_enabled

* Increase default bucket size in distributed Adam

* Move distributed Adam unit test to contrib tests

Integrate into unit testing framework

* Tweak hyperparameters for dist Adam optimizer test

Improves numerical stability so we can keep tight tolerances. Adopting suggestions from @crcrpar.

* Use distributed test infrastructure in distributed Adam unit test

Suggestion from @crcrpar.

* add temporary dispatch of double, float, half, bfloat16

* fusedadam of bfloat16

* Add bfloat16 path to FusedAdam

* Gradient clipping routine with fused kernels

Identical API as PyTorch. Falls back to PyTorch impl when not computing L2 norm.

* Add unit test for gradient clipping

* Add fp16 case to gradient clipping unit test

* Tweaks to grad clipping unit test

Review suggestions from @crcrpar

* Debug gradient clipping tests

When checking that incorrect results produce assertion errors, make sure to generate a discrepancy outside the range of numerical error.

Remove legacy fuser usage from multihead attention in contrib in favor of the default which should be nvfuser.  Modify test scripts to activate fusion. (#1403)

ZeRO-2 support in DistributedFusedAdam

Adjust test options to have tighter tolerances.