GitLab

[PyTorch] Fixing hang in `initialize_ub()` for multi-node runs after PR901 removal of MPI-dependence (#986)

* Re-implementing PR901 (removing MPI-dependence in Userbuffers) with multi-node fixes

* passing data-parallel rank/size info from torch.distributed to userbuffers
Signed-off-by: Alp Dener <adener@nvidia.com>

* multi-node example working with UB_SKIPMC=1 but not with multicast
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed multi-node hang in initialize_ub(), updated comm+GEMM overlap example to support multi-node mixed tensor/data parallelism, added README
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed use case when Userbuffers is asked to allocate the TP overlap buffer with UB_SKIPMC=1
Signed-off-by: Alp Dener <adener@nvidia.com>

* corrected example problem to set device by local ordinal instead of global process rank
Signed-off-by: Alp Dener <adener@nvidia.com>

* double-free fix in userbuffers destructor
Signed-off-by: Alp Dener <adener@nvidia.com>

* removed unnecessary and incorrect torch.cuda.set_device(...)
Signed-off-by: Alp Dener <adener@nvidia.com>

* corrected inter-node ranks logic
Signed-off-by: Alp Dener <adener@nvidia.com>

* generalized node ID logic in initialize_ub to handle arbitrary world rank layouts within node
Signed-off-by: Alp Dener <adener@nvidia.com>

* added single-node comm+GEMM overlap unit tests
Signed-off-by: Alp Dener <adener@nvidia.com>

* LayerNormMLP example confirmed working with 2 nodes on Eos
Signed-off-by: Alp Dener <adener@nvidia.com>

* unit test cleanup
Signed-off-by: Alp Dener <adener@nvidia.com>

* corrected DP group ranks logic in LNMLP comm+GEMM overlap example
Signed-off-by: Alp Dener <adener@nvidia.com>

* corrected enums in unit test
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed incorrect Ubuf object init signature
Signed-off-by: Alp Dener <adener@nvidia.com>

* switched default backend for Userbuffer bootstrapping to Gloo with MPI and NCCL fallbacks, and initialize_ub option to manually select backend
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed all comm+GEMM overlap unit tests
Signed-off-by: Alp Dener <adener@nvidia.com>

* corrected all_gather use for Gloo backend
Signed-off-by: Alp Dener <adener@nvidia.com>

* changed userbuffers allgather callback to always use all_gather() instead of all_gather_into_tensor()
Signed-off-by: Alp Dener <adener@nvidia.com>

* restored and verified old MPI-based bootstrapping via NVTE_UB_WITH_MPI=1 option at compile time
Signed-off-by: Alp Dener <adener@nvidia.com>

* disabled scoped GIL release for comm+GEMM overlap algorithms
Signed-off-by: Alp Dener <adener@nvidia.com>

* avoid dist.init_device_mesh in comm+GEMM overlap example to support older PyTorch versions
Signed-off-by: Alp Dener <adener@nvidia.com>

* applied RS overlap FP8 fix from PR1004
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed segfault in Userbuffers destructor
Signed-off-by: Alp Dener <adener@nvidia.com>

* corrected comm+GEMM overlap unit test arguments
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed unit test run command for when Userbuffers is compiled with MPI
Signed-off-by: Alp Dener <adener@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Refactored torch.distributed collectives into pure C++ callbacks
Signed-off-by: Alp Dener <adener@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Alp Dener <adener@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* removed unwanted memcpyDtoD/fixed weight parametrisation
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos02.eos.clusters.nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos02.eos.clusters.nvidia.com>
Co-authored-by: Selvaraj Anandaraj <selvaraja@login-eos02.eos.clusters.nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Fixed convergence issues
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos02.eos.clusters.nvidia.com>

* Update transformer_engine/pytorch/module/layernorm_linear.py
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Update transformer_engine/pytorch/module/layernorm_mlp.py
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Update transformer_engine/pytorch/module/linear.py
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos02.eos.clusters.nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Selvaraj Anandaraj <selvaraja@login-eos02.eos.clusters.nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* DGRAD_RS UB overlap Bug fixes
Signed-off-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* remove implicit padding and unpadding
Signed-off-by: Xin Yao <xiny@nvidia.com>
---------
Signed-off-by: Xin Yao <xiny@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Phuong Nguyen <36155692+phu0ngng@users.noreply.github.com>

* update FE to 1.5.2
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* enable unfused attn for cross attn
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* unify logging info
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* omit cudnn 9.1.1 and 9.2.1 due to bugs
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* set cu_seqlens_padded to cu_seqlens by default
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* replace variable name with ctx.variable
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Revert "enable unfused attn for cross attn"

This reverts commit bc49f14fca904217a711b4a86c45a4a739a17a14.
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* restrict cudnn version for fp8 tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove mha_fill for FP8
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Revert "remove mha_fill for FP8"

This reverts commit 83ffc44114dc6eb3d426d742b6c5a4d34805ec04.
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* lower cudnn version to >=9.2.1
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* GroupedGEMM via multi-stream cublas

* fix A/B is nullptr while D is not nullptr

* add fp8 grouped gemm

* register with TorchScript

* add the GroupedLinear layer

---------
Signed-off-by: Xin Yao <xiny@nvidia.com>
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>
Co-authored-by: Jiang Shao <jiangs@nvidia.com>
Co-authored-by: Qi Zhang <qizhang@nvidia.com>
Co-authored-by: Phuong Nguyen <phuonguyen@nvidia.com>

* Add the option to use SM for P2P comm in TP overlap
Signed-off-by: Sangkug Lym <slym@nvidia.com>

* cleanup
Signed-off-by: Sangkug Lym <slym@nvidia.com>

* Python formatting with black
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Format C++ with clang-format
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Update transformer_engine/pytorch/csrc/comm_gemm_overlap.h
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

---------
Signed-off-by: Sangkug Lym <slym@nvidia.com>
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* subclass DPA with BaseModule and test with test_gpt_checkpointing
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* test DPA only
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* test save and load
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove debug info
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor tweaks
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor tweak
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add hook in case core_attention._extra_state is missing
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* check named buffers in BaseModule; remove FP8 scratchpad override function; test FP8 for sm90+
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* minor fixes: test size, interval in recipe, named_buffer loop
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* move BaseModule from FusedAttention to DPA
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Apply formatting
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Apply formatting
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* added DL framework callbacks for bootstrapping userbuffers without MPI
Signed-off-by: Alp Dener <adener@nvidia.com>

* removed userbuffers availability check in TE modules since userbuffers is now always compiled
Signed-off-by: Alp Dener <adener@nvidia.com>

* added comm+GEMM overlap example with LayerNormMLP
Signed-off-by: Alp Dener <adener@nvidia.com>

* lintin and review fixes
Signed-off-by: Alp Dener <adener@nvidia.com>

* linting and review fixes
Signed-off-by: Alp Dener <adener@nvidia.com>

* added header guards
Signed-off-by: Alp Dener <adener@nvidia.com>

* removed defunct userbuffers checks in build_utils and setup.py
Signed-off-by: Alp Dener <adener@nvidia.com>

* added exposed API in modules/base.py to __all__
Signed-off-by: Alp Dener <adener@nvidia.com>

* removed transformer_engine/CMakeLists.txt and shifted all TE/common compile into transformer_engine/common/CmakeLists.txt
Signed-off-by: Alp Dener <adener@nvidia.com>

---------
Signed-off-by: Alp Dener <adener@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* New TE wrapper for PyTorch FullyShardedDataParallel to make TE modules distribute their activations after the forward pass and gather them before the backward pass
Signed-off-by: Alp Dener <adener@nvidia.com>

* simplified TE module setup for FSDP comms
Signed-off-by: Alp Dener <adener@nvidia.com>

* FSDP scatter/gather for tensors saved into autograd ctx now working for base TE modules
Signed-off-by: Alp Dener <adener@nvidia.com>

* make sure activation recompute disables FSDP scatter/gather
Signed-off-by: Alp Dener <adener@nvidia.com>

* make sure Fp8 weight buffers are sharded at the end of the backward pass and gathered before forward
Signed-off-by: Alp Dener <adener@nvidia.com>

* Fixed typo in attribute name
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed bug in finding FSDP-wrapped TE modules
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed typo in fp8 weight tensor name
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed incorrect # of gradients
Signed-off-by: Alp Dener <adener@nvidia.com>

* Added fp8 amax gradient hook tensor to the parameter reset
Signed-off-by: Alp Dener <adener@nvidia.com>

* get rid of erroneous dummy tensor leftover from incorrect rebase
Signed-off-by: Alp Dener <adener@nvidia.com>

* Linting fixes
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixing git snafu and removing debug statements
Signed-off-by: Alp Dener <adener@nvidia.com>

---------
Signed-off-by: Alp Dener <adener@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

Cleanup
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Initial refactor of FP8 workspaces in Linear module
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove extra kernel launch
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Minor perf optimizations

Tensor base class functions in Float8Tensor have significant overhead.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug FP8 recipe test
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Refactor FP8 workspaces in LayerNormLinear and LayerNormMLP
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Document FP8 workspace function
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Revert changes to FP8 recipe tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add support for lazy FP8 transpose caching

Previous caching behavior (always fill cache) incorrectly filled cache during CUDA graph warmup steps.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix Pylint warnings
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug ONNX export

ONNX FP8 cast ops assumed that FP8 scales were created during model export (i.e. not initialized during training).
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug fused attention tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Make sure Float8Tensor.transpose_2d is backward compatible
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Revert changes to ONNX export operations

Work around ONNX test failures by filling FP8 scale tensors instead of copying into them.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug scale factor update in Float8Tensor transpose_2d
Signed-off-by: Tim Moon <tmoon@nvidia.com>

---------
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Calibration fix
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* Lint fix
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

---------
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>
Co-authored-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* added squared relu in te-torch
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* Add support for MoE with FP8.
Signed-off-by: Dennis Liu <denliu@nvidia.com>

* Fix unittest.
Signed-off-by: Dennis Liu <denliu@nvidia.com>

* Fix error in linear backward.
Signed-off-by: Dennis Liu <denliu@nvidia.com>

---------
Signed-off-by: Dennis Liu <denliu@nvidia.com>
Co-authored-by: Przemyslaw Tredak <ptredak@nvidia.com>

* Remove unnecessary Pylint overrides
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fixes to lint
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Support noop concat without providing full tensor

Stop storing fused buffers in linear modules.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug noop cat func
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Construct TE modules in tests with correct dtypes
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add tolerances to numerical tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use plain PyTorch concat when exporting to ONNX
Signed-off-by: Tim Moon <tmoon@nvidia.com>

---------
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* fixes; docs
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Check for FP8
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fix LoRa-like use cases
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Reviews
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* WIP: fp8 v1 fprop integration
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* WIP: minor fixes
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add debug info
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add more debug info
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fprop working for h1; w/ debug info
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* WIP: add bprop
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* cleanup; bprop running but has mismatches
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add gitlab frontend as submodule
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* clean up and add back v0.9.2 FE support; fprop/bprop passing with 5e-2 tols
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix after merge; add bias_b/h to caching descriptor
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* distinguish fwd/bwd tensor types for bprop
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor fix for F16 cases; include added dqkv_type and d_scale_dp
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* adjust out shape for bwd in test
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add casting from/to FP8 to DPA module
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* WIP: bshd_bshd_bshd layout
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* WIP: support all sbhd/bshd layouts
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* clean up
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add qkvpacked and kvpacked support in both FusedAttnFunc and C levels
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove qkvpacked/kvpacked calls in DPA module (used for testing)
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove tp setup; add allow_non_contiguous; update FE; revert to sbh3d in tests; clean up
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add NVTE_FP8_DPA_BWD to control whether to use FP8 bwd or F16 bwd
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix MQA
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix MQA/GQA in FP8 v1 API
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* update FE to 705d8e3, with API change
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* test causal mask
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* restrict mha_fill for THD format
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix fused attn with CP and comment out is_alibi code
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* clean up FE0.9 vs FE1.0 FP8 implementations, and related unit tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* change NVTE_FP8_DPA_BWD default to 1, and fix its use in qkvpacked/kvpacked APIs
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix lint and self.tp_size/group in FusedAttention()
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* update FE to 6902c94
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add FP8 MHA support
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* update to FE v1.3.0
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor fixes for FP8 MHA with different configs
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* emit stats regardless of is_training
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix linear when input is not Float8Tensor
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix d_out type when f16 bprop
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix user buffer for layernorm_linear/linear and revert two FP8 casts in MHA
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add docstring for fp8_dpa/mha in recipe
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fixes
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* fix backend selection to avoid FA
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* replace transpose with transpose_2d
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* use RMSE for FP8 unit tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* replace two more transpose with transpose_2d
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add FP8 initialization to FusedAttention
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* rm docs
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Revert "add FP8 initialization to FusedAttention"

This reverts commit 15fffd825d6f23f31ea709b16ba01dfd61efabf8.
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Change order of ctxs
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fixes
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* minor fixes
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add back docs and mark as beta
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor fixes for tests and docs
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

Don't use autograd hook for bwd reduction
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Add LN margin to inference
Signed-off-by: Sangkug Lym <slym@nvidia.com>

* cleanup
Signed-off-by: Sangkug Lym <slym@nvidia.com>

* Fix symbolic func registration
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fix grads
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Sangkug Lym <slym@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* FP8 cuda graphs
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>
Co-authored-by: Charlene Yang <charleney@nvidia.com>

* Fix numerics
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* exclude torch compile from numerics tests
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* More numerics fixes
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fix tests
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fix CI
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* rm fusion from unfused path
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>
Co-authored-by: Charlene Yang <charleney@nvidia.com>

fix the default userbuffer communicator init settings
Signed-off-by: Sangkug Lym <slym@nvidia.com>

* Enable DGRAD RS overlap
Signed-off-by: Jaemin Choi <jaeminc@nvidia.com>

* fix lint; apply suggestions
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Jaemin Choi <jaeminc@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* userbuffer fp8 reduction support for individual overlap
Signed-off-by: Sangkug Lym <slym@nvidia.com>

* cleanup dict ub_cfg dict value load
Signed-off-by: Sangkug Lym <slym@nvidia.com>

* cleanup
Signed-off-by: Sangkug Lym <slym@nvidia.com>

* Remove unnecessary fence from producer

From @erhoo82 
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Sangkug Lym <slym@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Atomic gemm for TP-AR and TP-RS overlap with P2P exchanges
Signed-off-by: Sangkug Lym <slym@nvidia.com>

* FP8 reduction for atomic TP-RS with p2p exchange
Signed-off-by: Sangkug Lym <slym@nvidia.com>

* Fix
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Sangkug Lym <slym@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Do not store input activations when not computing weight gradients
Signed-off-by: Sangkug Lym <slym@nvidia.com>

* fix userbuffer tp comm overlap case
Signed-off-by: Sangkug Lym <slym@nvidia.com>

---------
Signed-off-by: Sangkug Lym <slym@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

Perform FP8 cast on gathered layernorm output in LayerNormLinear
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Enable TP-AG overlap with return_layernorm_output
Signed-off-by: Jaemin Choi <jaeminc@nvidia.com>

* Use ub_overlap_ag
Signed-off-by: Jaemin Choi <jaeminc@nvidia.com>

---------
Signed-off-by: Jaemin Choi <jaeminc@nvidia.com>
Co-authored-by: Jaemin Choi <jaeminc@nvidia.com>

* TP-RS overlap with send/recv

Atomic GEMM based TP-RS overlap with send/recv
Signed-off-by: Sangkug Lym <slym@nvidia.com>

Specify userbuffer overlap method of each overlap instance
Signed-off-by: Sangkug Lym <slym@nvidia.com>

P2P TP-RS overlap with fp8 GEMM outputs
Signed-off-by: Sangkug Lym <slym@nvidia.com>

Fix TP-RS overlap with send/recv
Signed-off-by: Sangkug Lym <slym@nvidia.com>

* cleanup
Signed-off-by: Sangkug Lym <slym@nvidia.com>

* cleanup
Signed-off-by: Sangkug Lym <slym@nvidia.com>

* linting
Signed-off-by: Sangkug Lym <slym@nvidia.com>

* fix typo
Signed-off-by: Sangkug Lym <slym@nvidia.com>

---------
Signed-off-by: Sangkug Lym <slym@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* add_dtype_for_userbuf
Signed-off-by: Hongbin Liu <hongbinl@nvidia.com>

* Update transformer_engine/pytorch/module/base.py
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Fix syntax
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fix lint
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Hongbin Liu <hongbinl@nvidia.com>
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Hongbin Liu <hongbinl@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* first draft of return_layernorm_output_gathered
Signed-off-by: Chen Cui <chcui@nvidia.com>

* explain use case more thoroughly in docstring
Signed-off-by: Chen Cui <chcui@nvidia.com>

* add same option in `LayerNormMLP`
Signed-off-by: Chen Cui <chcui@nvidia.com>

* Update transformer_engine/pytorch/module/layernorm_linear.py
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: Chen Cui <cxcui@alumni.cmu.edu>

* Update transformer_engine/pytorch/module/layernorm_linear.py
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: Chen Cui <cxcui@alumni.cmu.edu>

* address comments
Signed-off-by: Chen Cui <chcui@nvidia.com>

* add same option in LayerNormMLP
Signed-off-by: Chen Cui <chcui@nvidia.com>

* address linter errors
Signed-off-by: Chen Cui <chcui@nvidia.com>

---------
Signed-off-by: Chen Cui <chcui@nvidia.com>
Signed-off-by: Chen Cui <cxcui@alumni.cmu.edu>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

Signed-off-by: Jaemin Choi <jaeminc@nvidia.com>
Co-authored-by: Jaemin Choi <jaeminc@nvidia.com>

* Avoid updating real during param cast
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Review comments
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* added non-reentrant mode support to TE checkpoint
Signed-off-by: Alp Dener <adener@nvidia.com>

* updated get_cuda_rng_tracker kwarg to get_rng_state_tracker to remain consistent with other TE API
Signed-off-by: Alp Dener <adener@nvidia.com>

* docstring cleanup
Signed-off-by: Alp Dener <adener@nvidia.com>

* added mechanism to disable bias_gelu_nvfusion in LayerNormMLP when checkpointing in non-reentrant mode
Signed-off-by: Alp Dener <adener@nvidia.com>

* refactored checkpoint and recompute hook names to match PyTorch implementation
Signed-off-by: Alp Dener <adener@nvidia.com>

* Fixed incorrect reference before assignment
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed argument error in calling native PyTorch checkpoint
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed linting errors for missing docstrings
Signed-off-by: Alp Dener <adener@nvidia.com>

* Fix lint
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* bias GELU fusion consistency between checkpoint test and reference comparison
Signed-off-by: Alp Dener <adener@nvidia.com>

---------
Signed-off-by: Alp Dener <adener@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Added QuickGELUActivation from HuggingFace/Transformers to common and pytorch
Signed-off-by: Alp Dener <adener@nvidia.com>

* Removing 'qgelu' from double-size activations list in LayerNormMLP.
Signed-off-by: Alp Dener <adener@nvidia.com>

* indent fix
Signed-off-by: Alp Dener <adener@nvidia.com>

---------
Signed-off-by: Alp Dener <adener@nvidia.com>
Co-authored-by: Przemyslaw Tredak <ptredak@nvidia.com>

* Add option to avoid updating transpose cache when possible
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix typo
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use string kwarg for FP8 transpose caching
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove unused attr
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

---------
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Support GEMM-GELU fusion with split AG overlap
Signed-off-by: Jaemin Choi <jaeminc@nvidia.com>

* Fix linter complaints
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: Jaemin Choi <minitu77@gmail.com>

* Avoid code duplication
Signed-off-by: Jaemin Choi <jaeminc@nvidia.com>

* Fix issue with modifying tuple
Signed-off-by: Jaemin Choi <jaeminc@nvidia.com>

* Disable GEMM-GELU fusion when split AG overlap is not enabled
Signed-off-by: Jaemin Choi <jaeminc@nvidia.com>

* Add ub_split_ag parameter to LayerNormMLP unit test
Signed-off-by: Jaemin Choi <jaeminc@nvidia.com>

* Move knob into LayerNormMLP, auto-disable fusion when split AG overlap is not enabled
Signed-off-by: Jaemin Choi <jaeminc@nvidia.com>

* Revert changes to test_layernorm_mlp_accuracy
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: Jaemin Choi <minitu77@gmail.com>

---------
Signed-off-by: Jaemin Choi <jaeminc@nvidia.com>
Signed-off-by: Jaemin Choi <minitu77@gmail.com>
Co-authored-by: Jaemin Choi <jaeminc@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>