GitLab

Make sure RoPE frequencies are in FP32
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fixed Llama tutorial. Changed batch size and added fused=True.
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>
Signed-off-by: root <root@ipp2-0037.nvidia.com>

* Tutorial updated but not complete yet.
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>
Signed-off-by: root <root@ipp2-0037.nvidia.com>

* Tutorial notebook reseted - removed fuse=true
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>
Signed-off-by: root <root@ipp2-0037.nvidia.com>

* Removed fused=true
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>
Signed-off-by: root <root@ipp2-0037.nvidia.com>

* Batch size back to 8
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>
Signed-off-by: root <root@ipp2-0037.nvidia.com>

* Typo and commented out line
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>
Signed-off-by: root <root@ipp2-0037.nvidia.com>

* fixed whitespace
Signed-off-by: root <root@ipp2-0037.nvidia.com>

* fixed whitespace
Signed-off-by: root <root@ipp2-0037.nvidia.com>

* Added comment to attention line. Fixed potential bug with loading weights - now loading works correctly, confirmed by the generation code.
Signed-off-by: root <root@ipp2-1661.nvidia.com>

* Comments
Signed-off-by: root <root@ipp2-1661.nvidia.com>

* Models cast added again
Signed-off-by: root <root@ipp2-1661.nvidia.com>

* Weight download info
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* Moved parameter gate_proj_size to config
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* gate_proj_size removed and put immediate_size instead
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* Llama 3 added to tutorial
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

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

* Fixed model loading
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

* Different dim for attention
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* Reversed other commit
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* Changed name to kv_channels
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* Fixed typo
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* Back to kv_channels in transformer layer
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* Back to kv_channels in transformer layer
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

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

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

* changed file modes
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* lint fix and resolved conflict
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* lint fix and resolved conflict
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

---------
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>
Signed-off-by: root <root@ipp2-0037.nvidia.com>
Signed-off-by: root <root@ipp2-1661.nvidia.com>
Co-authored-by: root <root@ipp2-2373.nvidia.com>
Co-authored-by: root <root@ipp2-1588.nvidia.com>
Co-authored-by: Pawel Gadzinski <pgadzinski@nvidia.com>
Co-authored-by: root <root@ipp2-0037.nvidia.com>
Co-authored-by: root <root@ipp2-1661.nvidia.com>
Co-authored-by: root <root@ipp2-2371.nvidia.com>
Co-authored-by: root <root@ipp2-1589.nvidia.com>
Co-authored-by: Sudhakar Singh <sudhakars@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

replaced deprecated pkg_resources with packaging
Signed-off-by: Alp Dener <adener@nvidia.com>

* fix inconsistency for attn mask; now True means participating in attn
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix sliding window window_size for decoder+padding combination
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* revert paddle changes regarding mask
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* revert softmax to 1-mask;0-keep
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* enforce 1-mask out; 0-keep rule for jax masks
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* revert pytorch mask changes; some kept in tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* revert to jax fused attn on main
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* inverse mask logic for get_cu_seqlens/_and_indices in PyTorch implementation and mask generation in unit tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* temporarily disable update_weight_scale_inv
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* enforce window_size for decoder
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add docstring for mask definition 1-mask out;0-keep
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* tweak make_decoder_mask and make_mask in jax tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* skip dBias for shapes other than 1HSS; otherwise dq/dk/dv NaNs
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* expand attn_biases from list to variables in save_for_backward
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix use of variable before assignment in jax dact_lu
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove window size definition for decoder
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add change notes in README for padding mask in PyTorch
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* tweak padding mask notes in README
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* expand list to tensors for save_for_backwards
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: cyanguwa <8636796+cyanguwa@users.noreply.github.com>

Signed-off-by: kunlunl <kunlunl@nvidia.com>
Co-authored-by: cyanguwa <8636796+cyanguwa@users.noreply.github.com>

Bump FA version to 2.5.8
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

* fix cuDNN version in unit tests for cuDNN v9
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add hook to ignore missing fused_attn._extra_states if training from old checkpoints
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove test and redundant implementation from last commit
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove warning message and replace with docstring
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove tp_size/tp_group in FusedAttention; amax reduction is handled with fp8_group
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* move core_attention.fused_attention._extra_state to core_attention._extra_state
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* simplify post_state_dict_hooks between FU and DPA
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* remove previous attempts to move core_attention.fused_attention to core_attention; keep the test
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* disable pylint self arg for hook which is required by hook
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: cyanguwa <8636796+cyanguwa@users.noreply.github.com>

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

* Update .gitignore
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
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: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

remove tp_size/tp_group as amax reduction is handled by fp8_group()
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* make FusedAttn with CP support bias
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* assert Alibi cannot work with CP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* syntax fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix variable name
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix tensor shapes
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* a typo fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix bias indexing for CP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* bug fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add attn bias tests
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* change dbias update location
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix CP test model configs
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* change CP test sequence length
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* make AttnFuncWithCP support qkv format of sbhd
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* make sure qkv are contiguous for CP in cuDNN fused attn
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* change assert message
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix code format
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

---------
Signed-off-by: Xiaowei Ren <xren@nvidia.com>
Co-authored-by: cyanguwa <8636796+cyanguwa@users.noreply.github.com>

* Try using global buffer for cu_seqlens
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Avoid using functools.lru_cache
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* fixes
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>

* 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>

* 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>

* 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>

This reverts commit 965803c9.

* 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>

[Pytorch] Update context parallel softmax lse correction func.
Signed-off-by: kitefang <kitefang@tencent.com>
Co-authored-by: kitefang <kitefang@tencent.com>

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

* Modified MHA and DPA logic to use causal softmax and FA for inference
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Adjusted unfused attention and softmax logic for inference
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Cleaned up the code per pylint
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Added test cases to evaluate numerics of incremental decoding
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Apply suggestions from code review
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: Oleg Goncharov <64355998+Oleg-Goncharov@users.noreply.github.com>

* Apply suggestions from code review [sequence start-end]
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: Oleg Goncharov <64355998+Oleg-Goncharov@users.noreply.github.com>

* Apply suggestions from code review [inference_params offset update]]
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: Oleg Goncharov <64355998+Oleg-Goncharov@users.noreply.github.com>

* Fixed bug in KV-cache indices and updated test suite
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Added inference_params description and applied suggestions from the code review
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Adjusted absolute tolerances in numerics tests
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Cleaned up the files per pylint
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

---------
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>
Signed-off-by: Oleg Goncharov <64355998+Oleg-Goncharov@users.noreply.github.com>
Co-authored-by: Przemyslaw Tredak <ptredak@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* added support for arbitrary bias shapes for fused_attn
Signed-off-by: Alp Dener <adener@nvidia.com>

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

* Add b1ss/bhss/11ss bias shapes when not requiring dBias
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* add bias_b/h to plan cache
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fixed compile errors after PR653 merge
Signed-off-by: Alp Dener <adener@nvidia.com>

* updated JAX unittests for new bias shapes
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed mismatched mask type checking
Signed-off-by: Alp Dener <adener@nvidia.com>

* corrected skip condition
Signed-off-by: Alp Dener <adener@nvidia.com>

* fix selection logic for A100s
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* corrected skip checks for bias shapes
Signed-off-by: Alp Dener <adener@nvidia.com>

* resolved test issues but neginf with float16 is still problematic with JAX
Signed-off-by: Alp Dener <adener@nvidia.com>

* new bias shapes passing TE JAX CI for seqlen <= 512, seq_q == seq_kv and h_q == h_kv conditions
Signed-off-by: Alp Dener <adener@nvidia.com>

* TE/JAX fused attn tests for new bias shapes passing with neg_inf=-2**27 for Bfloat16 and -2**15 for Float16
Signed-off-by: Alp Dener <adener@nvidia.com>

* code style fixes and test parameter ID cleanup
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed incorrect skip condition for backward fused attn test
Signed-off-by: Alp Dener <adener@nvidia.com>

---------
Signed-off-by: Alp Dener <adener@nvidia.com>
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: Alp Dener <adener@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>

* Use fused implementation of RoPE in MultiHeadAttention
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

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

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

* test alibi between fa and fu
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* move alibi slopes and bias to global to avoid repeating calculation
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix alibi slopes/bias generation
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix _is_flash_attention_supported to allow alibi type
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* disable padding mask when alibi is used for fused attn arbi backend
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add support for custom [n_heads] alibi_slopes in flash, fused, unfused attention
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* remove alibi_type=none tests as they are unnecessary
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* update cudnn-frontend to 1.0.2
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* change bias/dbias shape to allow b,1/1,h/b,h in arbi backend
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* tweak tests for arbi post_scale_bias [1,h,s,s] or alibi_slopes [n_heads]
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* change bias/dbias shape in max512 backend - incomplete
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove max512 changes from last commit and disable max512 (and arbi temporarily) for [b, h, s, s]; pending cuDNN backend support
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* replace || with () in docstring
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix bias shape for max512 backend
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* combine slopes/bias generation to one function get_alibi() and fix alibi tests
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 PR557 bugs
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Update transformer_engine/pytorch/attention.py
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: cyanguwa <8636796+cyanguwa@users.noreply.github.com>

* encapsulate global alibi tensors into a dict cache
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* reduce alibi slopes test size
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* update to cudnn-frontend 1.0.3
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* use dBias shape to define bias_b/bias_h because jax materializes dBias rather than Bias in bwd abstract
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: cyanguwa <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

Do not cache sequence lengths based on layer number
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Update cudnn frontend to 1.0.3 to fix cudnn v9 Nans
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* make d_out contiguous for bwd
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove cudnnDestroy to let torch handle it
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Update transformer_engine/pytorch/attention.py
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: cyanguwa <8636796+cyanguwa@users.noreply.github.com>

* Update transformer_engine/pytorch/attention.py
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: cyanguwa <8636796+cyanguwa@users.noreply.github.com>

* Update transformer_engine/pytorch/attention.py
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: cyanguwa <8636796+cyanguwa@users.noreply.github.com>

---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Signed-off-by: cyanguwa <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

Fixed offloading for PyT version/ Added Attention activation offloading support/ Native FP8 support (#632)

* Fixed offloading for PyT version/ Added Attention activation offloading support/ Native FP8 support
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>

* Removed activation offloading for fused attention
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>

* Fixed the illegal memory access issue for activation offloading of attention
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>

* Removed the version guard
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>

* Pipeline failures fix
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>

* Fixed lint erros
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>

* Lint error fix
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>

---------
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>
Co-authored-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>

Fix MHA docstring
Signed-off-by: Isaac Ong <isaacong.jw@gmail.com>

* fused apply rope
Signed-off-by: Xin Yao <xiny@nvidia.com>

* Apply suggestions from code review
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: Xin Yao <yaox12@outlook.com>

* resolve comments
Signed-off-by: Xin Yao <xiny@nvidia.com>

* make rotary_percent optional
Signed-off-by: Xin Yao <xiny@nvidia.com>

* fix ci
Signed-off-by: Xin Yao <xiny@nvidia.com>

* fix test
Signed-off-by: Xin Yao <xiny@nvidia.com>

* add rope test to qa
Signed-off-by: Xin Yao <xiny@nvidia.com>

* fix linting
Signed-off-by: Xin Yao <xiny@nvidia.com>

* sync apex: add transpose_output_memory
Signed-off-by: Xin Yao <xiny@nvidia.com>

* small fix
Signed-off-by: Xin Yao <xiny@nvidia.com>

* sync apex: fuse sin/cos
Signed-off-by: Xin Yao <xiny@nvidia.com>

* sync apex: fused rope for thd format
Signed-off-by: Xin Yao <xiny@nvidia.com>

* fix lint
Signed-off-by: Xin Yao <xiny@nvidia.com>

* Fix license headers
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

* add support for bshd format
Signed-off-by: Xin Yao <xiny@nvidia.com>

* support different seq length
Signed-off-by: Xin Yao <xiny@nvidia.com>

* update
Signed-off-by: Xin Yao <xiny@nvidia.com>

* update copyright
Signed-off-by: Xin Yao <xiny@nvidia.com>

* remove transpose_output_memory
Signed-off-by: Xin Yao <xiny@nvidia.com>

* Make outputs contiguous in SBHD case
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

---------
Signed-off-by: Xin Yao <xiny@nvidia.com>
Signed-off-by: Xin Yao <yaox12@outlook.com>
Signed-off-by: Przemek Tredak <ptredak@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: Przemyslaw Tredak <ptredak@nvidia.com>

[PyTorch] fix forward attention_type in MultiheadAttention
Signed-off-by: Markus Schnoes <markus.schnoes@gmx.de>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

fix failing tests due to PR #557
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>
Co-authored-by: cyanguwa <8636796+cyanguwa@users.noreply.github.com>

* make TransformerLayer accept a `bshd` or `sbhd` tensor format
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* Fixes from feedback
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* more feedback fixes
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* remove incorrect info from docstring
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* fix from feedback
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

---------
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* try to use cuDNN fused attention for context parallelism
Signed-off-by: xren <xren@nvidia.com>

* assert CP is only supported with NVTE_F16_arbitrary_seqlen
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* port fused attn api to context parallelism
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add one more assert
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* assert CP does not support padded tokens
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add qkv_format into CP implementation
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove qkv_format from CP function
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix qkv_for,at
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix bwd error with FA v2
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* make cp implementation support non-causal masking
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* bug fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove redundant asserts for CP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor assert information change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* assert core attn bias has not been supported with CP yet
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* make CP work with window_sizes of [-1, -1] and [-1, 0]
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add draft code for fa test with cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* move fused attn test to a specific folder
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add assert_close to flash attn cp test
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add more tests for CP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add optional arguments for FA v2.4+
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* minor change
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add skip condition for CP test
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* class and function naming fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* docstring fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* do not use fused attn if backend does not work with CP
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* create a separate folder for CP test as it needs multi-GPUs
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* add attn_mask_type check in attn_forwrad_func_with_cp
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* code format fix
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

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

* Refactor parameter split in Linear module

Remove module state from noop_cat. Support arbitrary names in parameter split. Handle tensor parallelism.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Make noop_cat a standalone operation
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update parameter splits in LayerNormLinear
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug case without bias

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

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

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

* Use jit_fuser for bias-dropout-add fusion
Signed-off-by: Jaemin Choi <jaeminc@nvidia.com>

* Use jit_fuser for CP FA kernel
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: Jaemin Choi <jaeminc@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Deterministic FA, bump minimum supported version
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

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

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

---------
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>

* Provide pre-computed max sequence to remove unnecessary kernels and D2H copies
Signed-off-by: Sangkug Lym <slym@nvidia.com>

* Tweak comments
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

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

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

avoid redundant computation for cu_seqlens
Signed-off-by: Hongbin Liu <hongbinl@nvidia.com>
Co-authored-by: Hongbin Liu <hongbinl@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>