GitLab

* docs: Add comprehensive Getting Started guide with benchmarks

- Add new Getting Started documentation with PyTorch and JAX tutorials
- Include benchmark scripts demonstrating TE performance benefits
- Add CSS styling for code output and tabs
- Replace old quickstart notebooks with improved documentation
- Add transformer layer diagram (SVG)
- Update docs configuration and workflow
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

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



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

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

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



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

* 2026 in copyright
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

* Add tests for 2528 and 2529
Signed-off-by: Peter St. John <pstjohn@nvidia.com>

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

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

---------
Signed-off-by: Peter St. John <pstjohn@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

Signed-off-by: xiaoxi-wangfj <690912414@qq.com>
Co-authored-by: Teddy Do <tdophung@nvidia.com>

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

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

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

* Update docs/envvars.rst
Co-authored-by: Paweł Gadziński <62263673+pggPL@users.noreply.github.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Update docs/envvars.rst
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Paweł Gadziński <62263673+pggPL@users.noreply.github.com>

Update copyright to include 2026
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* replace autograd.grad with autograd.backward
Signed-off-by: Robin Zhang <robinz@nvidia.com>

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

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



* get/set graphable rng state
Signed-off-by: Robin Zhang <robinz@nvidia.com>

* fix lint
Signed-off-by: Robin Zhang <robinz@nvidia.com>

---------
Signed-off-by: Robin Zhang <robinz@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

Signed-off-by: fuyue.lj <fuyue.lj@antgroup.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fix incorrect calculation of segment pos from segment ids for thd cases and load balanced cases in from_segment_ids_and_pos. Enforce passing of segment_pos for THD cases and lod balanced cases
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

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

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



* Correct the assert condition
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Modify fused attn tests to pass new args to from_segment_ids_and_pos()
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

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

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



* Calculate seg ids before pos
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

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

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



* 1. Change the signature for from_segment_ids_and_pos()
2. Add support for THD in from_segment_ids_and_pos()
3. Assert if load balanced segment_ids is passed to generate a segment_pos
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

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

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



* Pass keyword-only args by name
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* nit: Fix typo to use seg_ids instead of segment_ids
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* nit: Fix comments
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* Modify the function call to differentiate between load balancing and actually reordered segment_ids and segment_pos
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

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

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



* Fix the is_segment_ids_reordered to be set only when CP and load balancing
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Fix comments for from_segment_ids_and_pos()
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Code clean up

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



Fix lint errors
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

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

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



---------
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* [PyTorch] Fuse permute+pad and unpermute+unpad ops for FP8 optimization

1.Fused `moe_permute_with_probs` + `Fp8Padding` and fused `moe_unpermute` + `Fp8Unpadding`,
  that can remove the explicit padding/unpadding of moe expert, improved performance and reduced peak gpu memory usage.
2.Add tests of fused permute/pad and unpermute/unpad.
Signed-off-by: xiaoxi-wangfj <690912414@qq.com>

* [PyTorch/Common] Fuse permute+pad and unpermute+unpad support with_merging_probs
Signed-off-by: xiaoxi-wangfj <690912414@qq.com>

* [PyTorch]format code
Signed-off-by: xiaoxi-wangfj <690912414@qq.com>

* [Common]perf expert_idx loaded once
Signed-off-by: xiaoxi-wangfj <690912414@qq.com>

* fix: pad_offsets can be None
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: xiaoxi-wangfj <690912414@qq.com>

* add padding + merging probs bwd support. Not tested
Signed-off-by: tdophung <tdophung@nvidia.com>

* Fix garbage initialized act grad
Signed-off-by: tdophung <tdophung@nvidia.com>

* all test passing for jax permutation + pad
Signed-off-by: tdophung <tdophung@nvidia.com>

* change tokens_per_experts APIs to num_out_tokens with conservative allocation of worst case padding for output buffer
Signed-off-by: tdophung <tdophung@nvidia.com>

* change test permutation to reduce test time
Signed-off-by: tdophung <tdophung@nvidia.com>

* triggering PR refresh
Signed-off-by: tdophung <tdophung@nvidia.com>

* format code
Signed-off-by: tdophung <tdophung@nvidia.com>

* Remove some tests cases from pytorch side. Add a separate toekn_dispatch test for sanity in case combine accidentally undo an error on dispatch in the roundtrip test. Add distinction between L0 and L2 in test cases in jax
Signed-off-by: tdophung <tdophung@nvidia.com>

* format code
Signed-off-by: tdophung <tdophung@nvidia.com>

* remove chance for inefficiency in moving between CPU and GPU, remove redundant primitive using a new static bool for padding, add assert for align size
Signed-off-by: tdophung <tdophung@nvidia.com>

* fix lint in jax
Signed-off-by: tdophung <tdophung@nvidia.com>

* account for both jax newer and older than version 0.8.2. Adjusted gpu triton binding accordingly
Signed-off-by: tdophung <tdophung@nvidia.com>

* format code
Signed-off-by: tdophung <tdophung@nvidia.com>

* fix typo
Signed-off-by: tdophung <tdophung@nvidia.com>

---------
Signed-off-by: xiaoxi-wangfj <690912414@qq.com>
Signed-off-by: tdophung <tdophung@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: tdophung <tdophung@nvidia.com>

* add triton ptxas path  for gb300 to find where it is to avoid compilation errors
Signed-off-by: tdophung <tdophung@nvidia.com>

* add these flags in advance to preven future breaks when ops are extended to multi gpus
Signed-off-by: tdophung <tdophung@nvidia.com>

* add this also to L1
Signed-off-by: tdophung <tdophung@nvidia.com>

---------
Signed-off-by: tdophung <tdophung@nvidia.com>

* rowwise colwise RHT group quant v1
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* remove local array RW
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* change wait_barrier
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* fast math options
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* use mult to replace div
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* format
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* bulk move random states
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* greptile
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* lint
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* revert to use divides
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* avoid fp32 bf16 round-trip in RHT cast fusion
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* trigger fastmath by toggle NVTE_RHT_CAST_FUSION_USE_FAST_MATH
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* integrate row col rht fusion, functional
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* numerics aligned
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* style
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* remove device sync
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* 128 padding
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* revert colwise rng state creation because of row-col fused kernel
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* fix CI, linter
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* refactor RS for generating two random values
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* Avoid invalid configs with templated kernel
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* fix acc pipeline init with 0 arrival count
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* restore rowwise-only mode
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* switch to dynamic atomic scheduler
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* Avoid instantiating group RHT+cast kernel without row-wise or col-wise output
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Include fast math option in quantization config
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix linter warnings and review nits
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



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

* Fix bug where kernel is always launched on stream
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



* Restore BF16 intermediate downcast in fused RHT-cast kernels
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* fix numerical test of grouped kernel
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* Make sure row-wise and col-wise quantization use different RNG seeds
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Restore autoformatter
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: Zhongbo Zhu <zhongboz@nvidia.com>
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

[JAX] Remove unused TE DPA module dtype which fixes cuDNN backend detection to properly use input dtypes (#2485)

* Remove unused TE DPA module dtype which fixes cuDNN backend detection to properly use input dtypes
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

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



* Warning fallback
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

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



* adjust test tolerances slightly for encoder tests due to change in backend
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

* Handle meshs set with jax.set_mesh
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

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



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

* add early return back (removed in 2427)
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

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

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



* Make sure Float8Tensor.contiguous supports autograd

Expand quantized tensor tests to check identity ops.
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: Sudhakar Singh <sudhakars@nvidia.com>
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>

Signed-off-by: oliver könig <okoenig@nvidia.com>

* Fix meta device check failure when passing torch.device objects
Signed-off-by: LucienXian <fl.xian@foxmail.com>

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

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



---------
Signed-off-by: LucienXian <fl.xian@foxmail.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>

* Tutorial for integration te/jax quantization into an existing framework
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* add todos
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* support nvfp4 sr rng key, move wrapper module into TE itself, fix bfloat16 cast
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* update docstrings
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Fix QKV proj and out proj in Flax example transformer layer
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Use fused attention in quickstart_jax example
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

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



* remat policy
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* add tutorial to docs
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* update title
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* remove unused dtype from TE DPA module
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Fix notebook title
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Fix lint
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Add explanation of flax module wrapper
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

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



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

* Add ccache support to TE and use it in GitHub actions
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

* Move to allowed action with sccache
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

* Properly handle sccache
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

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

* Removing ccache from the custom docker workflows where we can't run the
action in the container
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

* JAX already uses same cmake options to build the extension so there is
no need to set CXX too
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

* Removed the unnecessary env variables
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

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

reset weight ws cache for NVFP4TensorStorage
Signed-off-by: Jinhang Choi <jinhangc@nvidia.com>

* Use GEMM-AR fallback on newer cuBLASMp
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Remove test skip logic completely
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

---------
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Skip delayed wgrad tests in distributed numerics when debug mode is enabled
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

---------
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Check calling convention for amax switch.

Wgrad gemms with colwise x colwise require
rowwise data via general_gemm. Since dy
has both for dgrad and wgrad, the brittleness
has likely not affected results.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Clear rowwise data when applicable.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Update test with columnwise cases.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Check enum value rather than implicit cast.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

---------
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* fix ce loss with ignore idx
Signed-off-by: ykarnati <ykarnati@nvidia.com>

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

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

Signed-off-by: ykarnati <ykarnati@nvidia.com>

* remove fix comments
Signed-off-by: ykarnati <ykarnati@nvidia.com>

* fallback divisor to 1
Signed-off-by: ykarnati <ykarnati@nvidia.com>

* have arg for n_rows and n_non_ignore
Signed-off-by: ykarnati <ykarnati@nvidia.com>

* fuse n_non_ignore to softmax kernel
Signed-off-by: ykarnati <ykarnati@nvidia.com>

* fix incorrect arg
Signed-off-by: ykarnati <ykarnati@nvidia.com>

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

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

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

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

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



---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Teddy Do <tdophung@nvidia.com>

* Unset NVTE_FUSED_RING_ATTENTION_USE_SCAN by default
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Add TODO
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

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

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



* Change the warning check in P2P helper to warn against using scan loop
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

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

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



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

Convert sample tuple to list in reuse
Signed-off-by: Robin Zhang <robinz@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

set_all_rng_states in set_states
Signed-off-by: Robin Zhang <robinz@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Add separate RNG states for columnwise quantization with Stochastic Rounding
Signed-off-by: Evgeny <etsykunov@nvidia.com>

* Fix single tensor path
Signed-off-by: Evgeny <etsykunov@nvidia.com>

---------
Signed-off-by: Evgeny <etsykunov@nvidia.com>

* update FE; initial pass at thd
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* produce Stats+Max instead of Max+Sum_Exp
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Revert "produce Stats+Max instead of Max+Sum_Exp"

This reverts commit c7d2b77b2da9ff3f68344097284187ac427eeb6a.
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* code drop
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

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



---------
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>
Signed-off-by: Paweł Gadziński <62263673+pggPL@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Make softmax_type in FFI optional
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

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



* add warn message
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

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



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

* branch off of initial permutation jax-triton PR
Signed-off-by: tdophung <tdophung@nvidia.com>

* Set 0 as the size of dummy tensors to reduce memory usage.
Signed-off-by: tdophung <tdophung@nvidia.com>

* Correct setting of permuted_probs_stride_token, unpermuted_probs_stride_token and unpermuted_probs_stride_expert in unpermutation
Signed-off-by: tdophung <tdophung@nvidia.com>

* Implement primitives, wrapper, test for wrapper, edit trit
on binding to accomodate scalars
Signed-off-by: tdophung <tdophung@nvidia.com>

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

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



* Change implemementation of VJP functions to match correct pattern. Deduce some static scalar args from shapes of inputs. Accept B, S instead of num_tokens. Change test to use value_and_grad to test vjp funcs properly
Signed-off-by: tdophung <tdophung@nvidia.com>

* formatting
Signed-off-by: tdophung <tdophung@nvidia.com>

* fix pylint
Signed-off-by: tdophung <tdophung@nvidia.com>

* fix test to compare to the correct reference impl. relax 1 tol for grad compare, fix lint the rightway
Signed-off-by: tdophung <tdophung@nvidia.com>

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

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



* fix test_permutation to use value_and_grad for reference impl, tighten tols, and add unpermute with probs for token combine bwd rule
Signed-off-by: tdophung <tdophung@nvidia.com>

* added forgotten file in prev commit
Signed-off-by: tdophung <tdophung@nvidia.com>

* format
Signed-off-by: tdophung <tdophung@nvidia.com>

* merge with_probs to without_probs
Signed-off-by: tdophung <tdophung@nvidia.com>

* add aserts and fix lint
Signed-off-by: tdophung <tdophung@nvidia.com>

---------
Signed-off-by: tdophung <tdophung@nvidia.com>
Co-authored-by: Ming Huang <mingh@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

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

change order
Signed-off-by: tdophung <tdophung@nvidia.com>

Fixes to runtime loading logic and add missing deps
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* bug fixed, test added
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

* fix contigous
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

* revert unecessary change
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

* revert another change
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

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

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



* address review comments
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

* Update transformer_engine/pytorch/tensor/mxfp8_tensor.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Signed-off-by: vthumbe1503 <vthumbe@nvidia.com>

* address review comments
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

* missed adding renamed file
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

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

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

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

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



* fix minor issue
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

* fix ci issue
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

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

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



* fix the test for bfloat16
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

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

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



---------
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>
Signed-off-by: vthumbe1503 <vthumbe@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>

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

* Add generic stripe_height support for load balancing
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Fix imports in test for deprecated jax.experimental.pjit
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Add test case for stripe_height greater than 1. Add stripe_height arg to reordering methods
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* Add Striped 1 and 4 test cases. Refactor the Load Balancing test case. Fix the incorrect shape in striping inverser reordering
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* Modify test code for CP + AG + THD + stripe height greater than 1
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* Add stripe_height arg to fused attn and fused attn fwd API. Add appropriate mask checks for AG+THD+CP and pick BRCM to be executed per rank. Add Fused Attn Primitive for CP + THD +AG + Striping. Add a method to reorder and all gather segment ids and offsets for kv
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* TMP: Throwaway testing commit
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* Add comments in primitive registration process
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* TMP: Throwaway test commit
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Undoing incorrect rebase/merge leftovers
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* TMP: Throwaway test commits
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* Add support for calculating q and kv seqlens and offsets per rank for CP+THD+AG+SW+Striped>1 primitive
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* Augment jax primitive register code comments
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* Fix the array sizes and padding values returned for seqlens and offsets to fit what the fused attn primitive non cp computation
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Add support in new primitive for softmax_offset related changes. Put in missing primitive registering line in again. Increase the seqoffsets arrays lengths by 1
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Add new set of helper functions for seqlens and seqoffsets fo AG+THD+CP+Stripe>1 which accounts for batching and seq offsets size b+1
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Add backward primitive for CP+THD+AG+Striped>1
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Modify tests for backward primitive for CP+THD+AG+Striped>1
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Move stripe_height along with other static args in fused_attn_bwd rule. Fix typo in CP+AG+TH+Striped>1 primitive
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Code clean up: remove older version for calculating seqlens and offsets for CP+AG+THD+striped>1 primitive
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Add test for CP+THD+AG+Striped>1
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Fix missing var
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

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

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



* Add SWA tests for AG+Striped>1+CP+THD+SWA
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Restoring test code
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Remove assert preventing SWA code path in CP+AG+Striped primitive
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Parametrize num_segments_per_seq in tests
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

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

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



* Clean up test code
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

Clean up test code in TE common
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

Clean up debug statements
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Rename stripe_height to stripe_size
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Code clean up and add additional comments
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

nit: Apply suggestions from code review
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Signed-off-by: Kshitij Lakhani <33047503+KshitijLakhani@users.noreply.github.com>

Fix type on fused attn tests
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Fix seqoffsets length to be passed onto FusedAttn primitive as it is b and not b+1 needed by cuDNN
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Remove commented code
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Signed-off-by: Kshitij Lakhani <33047503+KshitijLakhani@users.noreply.github.com>

Fix linting issues
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

Fix incorrect greptile change
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

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

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



* Skip THD test cases for CP + AG + Dual chunk. Skip BSHD cases for CP + AG + Striped>1. Correct the layout and shapr parameters passed to the tests
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Pass stripe_size explicitly for ring attn tests for THD cases
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Remove TODO
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Explicitly fail if THD + AG is being used with a non padding causal mask
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

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

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



* nit: Correct the ID for the test dist fused attn tests to account for cp*2 which is done under the hood
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Set num_segments_per_seq defaults to None instead of 0
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Augment comments. Add ValueError for stripe_size=0
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Test only 1 num_segments_per_seq combination for CP+AG+THD+Striped>1+SWA instead of 2. Modify the num segments and window size to easily to debug values
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Default stripe_size to None instead of 0. Modify stripe_size check for <=0 instead of ==0
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

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

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



* Remove incorrectly added file
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Explicitly pass zero sized arrays for seg ids and pos in the CP + AG + Striped primitive rather than using the seqlens or the offsets as placeholders
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Fix linting errors
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

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

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



* Add a deep dive doc for CP+THD+AG+Stripe>1+SWA regarding design considerations and decisions
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Put docs and pngs into it's separate dir
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Replace png screenshots with markdown coe blocks for the attention patterns. Remove unecessary pngs
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Add doc file to index.rst. Fix grammatical errors
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

---------
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>
Co-authored-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>
Co-authored-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

Remve assumption in quantize/activation kernels that data buffer is initialized
Signed-off-by: Tim Moon <tmoon@nvidia.com>