GitLab

* refactor + mxfp8

* added grouped gemm

* rename linear to dense

* added cublas init phase for groupedGemm

* relax the tol of test encoder multiprocessing mxfp8 by 0.001
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>
Co-authored-by: Hua Huang <huah@nvidia.com>
Co-authored-by: Jeremy Berchtold <jberchtold@nvidia.com>

import te before te_jax
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

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

* switch order of dirs for finding so
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Use existing dir src build
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

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

* Add support for UB MNNVL
Signed-off-by: Nicolas Castet <ncastet@nvidia.com>

* Address review comments
Signed-off-by: Nicolas Castet <ncastet@nvidia.com>

* Fix lint
Signed-off-by: Nicolas Castet <ncastet@nvidia.com>

* Dlopen nvml lib since it comes with the cuda driver
Signed-off-by: Nicolas Castet <ncastet@nvidia.com>

* Add initial copyright date
Signed-off-by: Nicolas Castet <ncastet@nvidia.com>

---------
Signed-off-by: Nicolas Castet <ncastet@nvidia.com>

flax module with compute dtype inferred from the inputs
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* fixes L1 test

* fix test_multigpu_encoder

* fixes for other multi-encoder tests

* jax.extend.ffi to jax.ffi

* initialization with float32

* add init_dtype as an optional arg to all modules

* update use_scan query from xla flags

* relax threshold for test_encoder fp8

* relax the tols

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

Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

* add test_multiprocessing_encoder with processing spawning in bash

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

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

* Skip encoder tests on V100

* Fix mulitprocessing jax.distributed.init

* Remove XLA xla_gpu_deterministic_ops which causes segfault

---------
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Create README.md

added all PyT examples
Signed-off-by: Santosh Bhavani <santosh.bhavani@live.com>

* Update README.md

- Added JAX, PaddlePaddle, and third-party examples
- Fixed DL framework links
- Removed issue request for new PRs
Signed-off-by: Santosh Bhavani <sbhavani@nvidia.com>

---------
Signed-off-by: Santosh Bhavani <santosh.bhavani@live.com>
Signed-off-by: Santosh Bhavani <sbhavani@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Use jit instead of pjit

---------
Signed-off-by: Frederic Bastien <fbastien@nvidia.com>
Co-authored-by: Phuong Nguyen <36155692+phu0ngng@users.noreply.github.com>

* Switch to nltk>3.8.1 and new data
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

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

[C/PyTorch] Fixed incorrect use of `torch.distributed.new_group()` when creating intra-node group in `initialize_ub()` (#1087)

* updated initialize_ub() to use new_subgroups_by_enumeration() to generate intra-node groups, added new unit tests for TE layers with comm overlap
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>

[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>

* rm tensor check if the workspace is empty
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* add trust_remote=true for load_dataset() in the mnist test
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

* Reformat FP8 Meta

1. Reformat FP8 meta to be one-set-per-tensor.
2. Remove fp8_max and scale_inv.
3. Remove unused functions in fp8.py
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Fix unit-tests
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Remove ShardingType and MajorShardingType
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Fix lint errors
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Fixed unittests.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Rename few variables.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Add jit to update_amax_list
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Fixed naming error in LayernormMLP
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Fixed bugs in test_distributed_layernorm_mlp.py
Signed-off-by: Ming Huang <mingh@nvidia.com>

---------
Signed-off-by: Ming Huang <mingh@nvidia.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>

* Allow multi-dims for dgamma and dbeta in LN descriptor.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Fix the jit error in examples/jax
Signed-off-by: Ming Huang <mingh@nvidia.com>

---------
Signed-off-by: Ming Huang <mingh@nvidia.com>

* fixed static argnums for jax.jit in single gpu encoder test, changed warning filtering for pytest
Signed-off-by: Alp Dener <adener@nvidia.com>

* propagating the fix to the JAX mnist example
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed missing space ibetween flags i QAA scripts
Signed-off-by: Alp Dener <adener@nvidia.com>

* added TE warnings into the ignore list
Signed-off-by: Alp Dener <adener@nvidia.com>

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

* Adding support of sequence parallelism
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adding RoPE
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Fix wrong batch_logical_axes
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Rnaming FSDP outer env var
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Poring RoPE to Praxis layers.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Porting GeLU + [FP8 Cast].
Signed-off-by: Ming Huang <mingh@nvidia.com>

* WAR to make XLA successfully match FP8 GEMM on FFN1 with GeLU.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Allowing arbitrary dimension of NVShape for the workspace allocation
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adding checkpoint_name to fused functions of mlp.py to get better perf with nn.scan.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Modify with review feedback.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Fix bugs
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Fix typo.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Fixed for lint
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Follow review feedback to modify code.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Fix typo.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Port SP to Praxis
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Fix an issue when enabling both GQA and RoPE.
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Update docs
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

---------
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

Signed-off-by: hugo-syn <hugo.vincent@synacktiv.com>

* Implemented deferred initialization via `device='meta'` option for te.Linear and added new PyTorch example to demonstrate its use with FullyShardedDataParallel execution.
Signed-off-by: Alp Dener <adener@nvidia.com>

* correcting Float8Tensor initialization and fixing linting errors
Signed-off-by: Alp Dener <adener@nvidia.com>

* removed duplicate code from upstream rebase, local tests passing
Signed-off-by: Alp Dener <adener@nvidia.com>

* improved comments/documentation for FSDP example
Signed-off-by: Alp Dener <adener@nvidia.com>

* converted reset_parameters() into a base module function
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed Float8Tensor creation with deferred init, all tests passing locally
Signed-off-by: Alp Dener <adener@nvidia.com>

* extended deferred initialization to all TE modules
Signed-off-by: Alp Dener <adener@nvidia.com>

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

* removed unnecessary reference to the parent module of parameter, added clarifying comments in parameter reset
Signed-off-by: Alp Dener <adener@nvidia.com>

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

Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

* Refactor sharding.py for the further custom_partitioning migration
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Migrating both FWD and BWD of LayerNorm/RMSNorm from xmap to custom_partitioning.
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Migrating both FWD and BWD of all kinds of softmax from xmap to custom_partitioning.
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Fix the wrong order of parameters to LN/RMSN bwd in ln_mlp_fp8.
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* WAR to LN/RMSN_fp8 before migrating to CP.
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Fix the wrong order of parameters of bwd of LN/RMSN_fp8.
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Following review feedback to modify
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Force the hidden dim in Norm ops to no sharding and add warning msg.
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Reuse fwd_rule in VJP functions
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Migrating both FWD and BWD of self-fused-attn from xmap to custom_partitioning.
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Migrating both FWD and BWD of cross-fused-attn from xmap to custom_partitioning.
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* add gelu and dgelu.
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Reuse fwd_rule in VJP functions for attentions
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Apply native FP8 Dtypes to fp8.py
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Migrating cast_and_transpose from xmap to custom_partitioning
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Migrating transpose from xmap to custom_partitioning
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Apply XLA pattern match to perform FP8 GEMM.
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* migrate layernorm_fp8 to custom_partitioning.
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Unify code style
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Extend supported of Transpose with FP8
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Implementing layernorm_fp8_dot based on migrated custom calls.
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Renaming variables and publish NVTE_FP8_COLLECTION_NAME
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Replace Q/DQ custom calls with native XLA implementations
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* migrate gelu_fp to custom_partitioning.
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Miner fix
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Support custom calls with mutli-dims
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Support gerneral dot indices in _fp8_dot_impl
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Implementing layernrom_geglu_fp8_mlp
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Remove GEMM custom calls
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Remove xmap related code
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Fix typo and add query-function to FP8MetaPackage
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Fix some bugs of custom calls
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Fix CT's bugs
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Update UTs/eaxmaples to adapt to the API changes.
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Unify kernel initilization in MLP.
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Modifing with code review's feedback
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Update README and Add deprecating warning to *ShardingType
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>

* Canonicalize the dtype
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adding assertion for non-supported batch dims.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adding doc/examples to _multidim_transpose
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Set FP8 meta as WeightHParamsCollection.OVERWRITE_WITH_GRADIENT in Praxis modules.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Set FP8 meta as WeightHParamsCollection.OVERWRITE_WITH_GRADIENT in Praxis modules.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Apply dtype-based rtol/atol to UTs
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Deprecate QKV_INTERLEAVED enum
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Skip test_distributed_custom_ops.py
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Fix the wrong sharding of bias in SelfAttn
Signed-off-by: Ming Huang <mingh@nvidia.com>

* WAR to fix the wrong cu_seqlen of MHA when DP/FSDP enabled
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adding distributed ops unit-tests
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adding license to test_distributed_*
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Follow review feedback to modify
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Use total bytes involved in collective ops as criteria.
Signed-off-by: Ming Huang <mingh@nvidia.com>

---------
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Ming-Xu Huang <mingh@nvidia.com>
Co-authored-by: Donglin Yang <dongliny@nvidia.com>

* Use FP8 tolerances in JAX FP8 tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Programmatically compute expected floating point error
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Loosen tolerance for MNIST test
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

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

* [Paddle] Add TP, DP, PP, FSDP
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Minor fix
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Fix CI failure
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Remove set_nccl_overlap_warning_if_tp
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Improve variable naming
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Refactor FP8 Buffer
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Stylic changes
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Fix FP32 parallel training
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Fix numel performance issue
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Squashed commit of the following:

commit 79e2e5fd774e67dcdda9aae01a9f31a6479c5d70
Author: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>
Date:   Sun Aug 20 14:39:16 2023 +0000

    Add TP test
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

commit 1d40ad60540490f97ed82ba877cc6eda8902cbf6
Author: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>
Date:   Sun Aug 20 14:22:25 2023 +0000

    Fix tp_size when disabled
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

commit 6632f735a0c8251862355fc74622af59fae3a509
Author: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>
Date:   Sun Aug 20 05:52:18 2023 +0000

    Add TP for attention and transformer layer
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Add shape check
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Add FSDP check for stage 1,2,3
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Review changes
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Fix group_sharding test
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Support NVTE_FUSE_ATTN
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Fix CI errors
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

---------
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Cast Flax collections to FrozenDict as WAR to adapt Flax 0.7.1.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adding min version of flax to requirements.txt in examples
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Fix praxis tests and rename compare_frozen_dict to compare_dict
Signed-off-by: Ming Huang <mingh@nvidia.com>

* [Paddle] Refactor FP8 state (#350)

Refactor fp8 state
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Store FP8 checkpointing data in CPU (#351)
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Make test_layer be able to run on both Flax >=0.7.1 and <=0.7.0
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Update Flax version
Signed-off-by: Tim Moon <tmoon@nvidia.com>

---------
Signed-off-by: Ming Huang <mingh@nvidia.com>
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: Tian Zheng <tizheng@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>

* Add FP8 support

- Add FP8 recipe
- Add FP8 path for nn layers
- Add MNIST FP8 example
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Update README
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Fix LayerNormMLP FP8 backward
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Fix FP8 training in float32 accumulation
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Fix FP8 checkpointing for non forward execution cases (same as #323)
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Refactors and improvements for better code stype, readability and organization
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Remove unnecassary pylint override
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

---------
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Add Linear layer (FP16)
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

- Add BF16 training example
- Add fp8_autocast (only supports non-fp8 for now)
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Remove FP8 stuff
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Simplify Linear layer forward
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Add LayerNorm layer (BF16)
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Add LayerNormLinear layer
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Store weights in BF16
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Add LayerNormMLP layer
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Add BF16 MNIST example
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Remove in-place cast for compatibility with Paddle AMP mechanism
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* README correction
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Add Paddle op as a backend option
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Fix code format
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Fix dtype change between iterations
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Minor fixes
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Move forward function out of base layer
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Use Paddle nvtx bindings
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

---------
Signed-off-by: Tian Zheng (Engrg-Hardware 1) <tizheng@nvidia.com>

* Fully remove attn_type and set self_attn_mask_type default to 'causal'
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Fix tests with new arguments
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Explicit self_attn_mask_type for examples
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Update transformer_engine/jax/flax/transformer.py
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: zlsh80826 <rewang@nvidia.com>

* Update transformer_engine/jax/flax/transformer.py
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: zlsh80826 <rewang@nvidia.com>

---------
Signed-off-by: Reese Wang <rewang@nvidia.com>
Signed-off-by: zlsh80826 <rewang@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

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

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

* Enable fused attention dropout
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Cast the uint32 key/counter to int64
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Update dropout support in fused attention docs
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Revise devPtrCuSeqlen* to align the naming
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Support different Jax PRNG impls
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Revert CastAsync since it is not used
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Implement is_training for 16-bit fused attn
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add fused attn with dropout sanity unit tests
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Enhance the comments readability and rng_state checker
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Change the attention dropout shape to align other frameworks
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Make encoder tests deterministic
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Change the default seed for the jax encoder tests
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Maintain offset in TE
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Enhance the resource safety
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Revert rng_state type to allow only i64
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Handle the corner case for elts_per_threads calculation
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Populate rng state by kernels
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Rename rng_state as seed in cpp_extensions
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Update the attention dropout comment
Signed-off-by: Reese Wang <rewang@nvidia.com>

---------
Signed-off-by: Reese Wang <rewang@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Adding JAX/Praxis modules and dependencies.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adding UTs to JAX/Praxis modules.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Remove praxis as a dependency due to not strictly needed
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Repalce is_fp8_supported to is_fp8_available
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Make Praxis as an optional dependency.

1. Removed 'from . import praxis' in __init__.py.
    1.1 Noted, keep 'from . import flax' for deprecated warning.
2. Changed te.flax to te_flax in examples and README.rst.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adding a workaround to FP8 training on Praxis.
Signed-off-by: Ming Huang <mingh@nvidia.com>

---------
Signed-off-by: Ming Huang <mingh@nvidia.com>

bugfix for softmax lowering
Signed-off-by: Ryan Jeng <rjeng@nvidia.com>