GitLab

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

* CPU perf optimization in linear autograd function

Avoid enable_grad context when possible in cast function. Cache distributed group properties.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* CPU perf optimization in prepare_forward function

Avoid torch.nn.Module impl of __setattr__.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Avoid module import in TE module forwards
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use fast getter for params
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Reuse tensor dims in linear autograd func
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



* Apply optimizations to grouped linear
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug test failures
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



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

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

* Avoid deepcopy in tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Move _fast_setattr logic to __setattr__ method
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



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

* Add base class for tensor proxies
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Move tensor detaching logic to tensor proxy base class
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use Python wrappers to PyTorch extensions
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Include transpose caching logic in proxy encode function
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug dimension mismatch with amax history
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Move dequantize logic to proxy_decode func
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Rename to "QuantizedTensor"
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Rename "proxy_detach" to "detach"
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Include transpose cache in detach and clone funcs
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix linter warnings
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 FP8 workspaces with QuantizedTensor functions
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Move logic for FP8 transpose cache in FP8 workspaces to base class
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove cast-transpose logic from linear op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove unnecessary args for Float8Tensor when using FP8 attr dict
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove __torch_function__ to QuantizedTensor
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Update tests/pytorch/test_float8tensor.py
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

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

* Debug cast functions
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: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Add basic infrastructure for Sequential module
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add linear op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add FP8 support in linear op

Runs, but need to validate. Runtime errors with non-FP8 params and FP8 compute, or FP8 params and non-FP8 compute.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add reshape op and unit test
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add bias op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add unfused linear op

Test does not pass with FP8.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug unfused linear op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add test for linear+bias op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add separate abstract classes for unfused and fused ops
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Consolidate unfused ops in submodule
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add linear-bias fused op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use fused cast-transpose in linear ops
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Disable GEMM+bias fusion with FP32 activations

Not supported by cuBLAS.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add parallel unit test for unfused linear op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Refactor parallel tests to reduce job launches
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add all-reduce, all-gather, and reduce-scatter ops
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove unused file
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Add support for FP8 scale updates

Still need to implement amax reductions.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add license boilerplate
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fuse GEMM+bias in row TP

Add documentation for unfused ops
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Rename pipeline to fuser

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

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

* Preserve cached FP8 transpose between ops
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add option for fused wgrad accumulation
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Directly output FP8 from linear if needed
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix cuDNN front-end commit
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Use updated API for FP8 scale updates
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add tests for non-default FP8 recipes
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Rename UnfusedOperation to BasicOperation
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add unit test to check amax reduction with fusable op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Operator autograd state no longer needs to be initialized
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Initial functional implementation of linear op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug fused linear+bias op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove autograd context from functional linear impl
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use functional linear impl in fused linear+bias op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Rename subdirectory from "fuser" to "ops"

Avoid confusion with kernel fusers and graph compilers.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update with Float8Tensor changes in #820
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Correctly pass FP8 metadata from next op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Add convenience functions to manipulate Sequential class
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update name of PyTorch extensions module
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Clear saved tensor data in linear op after bprop
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Update name of PyTorch extensions module
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix test name in QA script
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update name of PyTorch extensions module
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Run distributed tests even when only 1 GPU is available
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Only run distributed tests with 2 GPUs if there are >=2 GPUs
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



* Review suggestions from @sudhakarsingh27 and @ksivaman

Fix spelling of "fusible". Avoid "input" name in internal APIs.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update transformer_engine/pytorch/ops/__init__.py
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>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

cache was added in Python 3.9.
Signed-off-by: Tim Moon <tmoon@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>

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

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

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

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

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

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

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

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

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

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

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

* first draft for review
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

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

* minor fix of an SWA message
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* use subprocess instaed of os.sys
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* add example script and update notebook
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

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

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

* fix Jax/Paddle related comments
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* rerun H100 benchmark
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* restrict fp8 tests to sm90+
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* move get_cudnn_version from common to pytorch utils
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

Tweak error message for invalid FP8 GEMM dims
Signed-off-by: Tim Moon <tmoon@nvidia.com>

[PyTorch] Workaround for incorrect output from torch.cuda.is_bf16_compatible() on V100s and TU102s (#626)

* replaced torch.cuda.is_bf16_compatible() with explicit sm_80 check via torch.cuda.get_device_capability()
Signed-off-by: Alp Dener <adener@nvidia.com>

* implement te.utils.is_bf16_compatible() to replace torch.cuda counterpart
Signed-off-by: Alp Dener <adener@nvidia.com>

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

* [PyTorch] Linear: fix computation for wgrad if sequence_parallel=True
Signed-off-by: Markus Schnoes <markus.schnoes@gmx.de>

* Remove buggy gather_along_last_dim
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* [PyTorch] Linear: fix line length
Signed-off-by: Markus Schnoes <markus.schnoes@gmx.de>

* Simplify logic for saving input tensor for Linear backward
Signed-off-by: Tim Moon <tmoon@nvidia.com>

---------
Signed-off-by: Markus Schnoes <markus.schnoes@gmx.de>
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>

Improve PyTorch memory usage
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Debug PyTorch and Paddle tests on Ada
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Only run Paddle layer tests with cuDNN fMHA on supported archs
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Reduce JAX FP8 GEMM sizes

Avoid split-k kernels on Ada.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Disable JAX fused self-attention test on Ada
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Run supported fused attention tests on Ada
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Run supported fused attention JAX tests on Ada
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Enable Paddle fused attention on Ada
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update reference scale calculation in TensorFlow test
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Restore backend support to reference FP8 attention impl in PyT test

Review suggestion from @cyanguwa
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Debug Paddle tests on Ada
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Loosen tolerances for Paddle attention tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Assume causal mask implies equal seqlens in Paddle attention tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

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

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

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

* Clearer error messages for dtype and shape assertions
Signed-off-by: Carlos Mocholí <carlossmocholi@gmail.com>

* Update transformer_engine/pytorch/utils.py
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: Carlos Mocholí <carlossmocholi@gmail.com>

* Update transformer_engine/pytorch/utils.py
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: Carlos Mocholí <carlossmocholi@gmail.com>

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

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

* Catch FP8 modulo16 error before cublas and fp8 kernels
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

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

* ignore self attention mask for causal type
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* further relax checks to run FA, update docs
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* fix pytorch softmax path
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

* minimum ampere requirement for fa
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

* fix qkv weight unfused path
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* fix non FA non interleaved case
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

retain grad related attrs while casting
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>