GitLab

* [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>

* 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>

* 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>

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

* 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>

* 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>

* 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>

* Initialize empty tensors with shape=[0] instead of shape=[].
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix runtime crash in LayerNorm

Still seeing correctness issues.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Make sure norm workspace sizes are not zero
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove assumption in swizzle kernel that data is available.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove assumption in multi-swizzle kernel that data is available.
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



* Remove unnecessary explicit call to default constructor
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Avoid accessing tensor data pointer if tensor has no entries
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 suggestions from code review
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

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

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



* Update transformer_engine/common/swizzle/swizzle.cu
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

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

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



* Review suggestions from @ptrendx and @greptile-apps
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



* Prefer using row-wise/col-wise shape based on which has data
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix merge conflict, expand docs, fix inconsistency in dim function
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Change Tensor::has_data to check whether tensor is initialized, not whether pointer is valid.
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 suggestion from @greptile-apps
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Debug incorrect tensor initialization in tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Clarify comments that has_data does not guarantee safe pointer accesses
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug test failure when computing amaxes
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: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>

* Add primary weighs fp8 support for mxfp8
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Fix unit test and add better error log to unit test
Signed-off-by: kunlunl <kunlunl@nvidia.com>

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

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



* Move post all-gather processing out of for loop
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Add descriptions and ASCII diagrams for partial cast and partial amax functions
Signed-off-by: kunlunl <kunlunl@nvidia.com>

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

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



* Minor fix based on greptile bot
Signed-off-by: kunlunl <kunlunl@nvidia.com>

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

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



* Fix compilation errors due to arch-specific PTX instructions
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove unused noop flag from C API
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



* Expose test_partial_cast
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Skip mxfp8 partial cast test if mxfp8 is not available
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Fix pytest error
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* pylint ignore unused manual_post_all_gather_processing
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Fix error when using is_mxfp8_available
Signed-off-by: kunlunl <kunlunl@nvidia.com>

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

* add grouped_tensor classes and helpers
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* rm non-contiguous option and dptrs
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

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



* address comments + rework CheckIn/OutputGroupedTensor
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

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



* fix for compilation
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* make first_dims/last_dims optional + data.shape 2d
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

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

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



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

* add data.shape info
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* added logical shape field
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

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



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

* fixed issues raised by greptile
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

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



* return default dtype when grouped_tensor is empty
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

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



* use has_data() for dim queries
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

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



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

* fix index bound
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* Update transformer_engine/common/transformer_engine.cpp
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* Update transformer_engine/common/transformer_engine.cpp
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* restore Tensor.has_data() + add experimental marks
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

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



* restore Tensor::has_columnwise_data
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

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

* init triton binding with test case/example

* added Triton as TE-JAX test dependency

* grid with blocksize from autotune
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

* init
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>

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

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

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

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

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

* cross entropy _input -> input rename
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* cross entropy _input -> input rename
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>

* a lot of small fixes
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* torch_version() change
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



* add missing module and fix warnings
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>

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

* removed training whitespace:
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* Update docs/api/pytorch.rst
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Signed-off-by: Paweł Gadziński <62263673+pggPL@users.noreply.github.com>

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

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

* Fix NumPy docstring parameter spacing and indentation

- Standardize parameter documentation to use 'param : type' format (space before and after colon) per NumPy style guide
- Fix inconsistent indentation in cpu_offload.py docstring
- Modified 51 Python files across transformer_engine/pytorch
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>

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

Do not initialize recipe state in base op class

Op attrs may not be set. Move recipe state initialization to linear op constructor.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

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

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

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

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

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

* 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>

* [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>

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

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

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

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

* 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>

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

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

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

* 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>

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

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

* 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>

* [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>

* 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



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

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

* fix
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: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* fix backward_dw cuda graph order
Signed-off-by: Pingtian Li <pingtianl@nvidia.com>

* add validation for num_layers_per_chunk
Signed-off-by: Pingtian Li <pingtianl@nvidia.com>

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

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



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

* 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



* revert back testing changes
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

* remove quantizer copy + fused adam working
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 test
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 mxfp8 bug, god knows who created it
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



* Update transformer_engine/pytorch/optimizers/fused_adam.py
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: vthumbe1503 <vthumbe@nvidia.com>

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

---------
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>
Signed-off-by: vthumbe1503 <vthumbe@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 <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>

* minor fix of torch view dtype
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* multi-tensor RHT amax, compiles
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* setup multi_tensor_quantize_nvfp4_impl
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* wire things up and run without crash
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

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

* unit test passing
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* finish unit test of split quantize api
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* bump up padding to 64 for nvfp4 grouped quantize
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* fix stochastic rounding
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

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

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

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

* enable multi-amax without RHT
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* fix col-only quantize mode
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* improve benchmark script
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* add NCU example script
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* add larger test case
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* add contiguous_data_and_scale check to bulk allocator
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* unified naming and differentiate between group_ and multi_
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* move regular amax into multi_tensor.h
Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com>

* Disentangle logic for split-quantize and general multi-tensor quantize
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use size_t for split sections
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Suggestions from @greptile-apps
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>
Co-authored-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

remove linear redundant check
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

* allow dp + fsdp and fixed sr_rng_state partitioning
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

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



* cleanup for lint test
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

* Change order of arguments to make jax works
Signed-off-by: tdophung <tdophung@nvidia.com>

* make num_experts a tl.constepxr again
Signed-off-by: tdophung <tdophung@nvidia.com>

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

* Add support for THD+CP+SWA through A2A comms
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* unblock the `padding`+`THD`+`CP(A2A)` with SWA case in A2A forward
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* add proper support for thd
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



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

* enable thd+cp tests as essential
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* add cp+thd+a2a test to essential
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* fix comments from greptile
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



* add proper skip for flash attention
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* fix the test to create separate tensors for flash and fused attention backend scenarios
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* remove redundant compare
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

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

* add note for cu_seqlens_kv and cu_seqlens_kv_padded
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* Update tests/pytorch/attention/test_attention_with_cp.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* Update transformer_engine/pytorch/attention/dot_product_attention/context_parallel.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

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

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

* fix the argument name
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



---------
Signed-off-by: Sudhakar Singh <sudhakars@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>

* Remove unnecessary SWA calculation from _segment_ids_pos_to_seqlens_offsets
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>

* Make BSHD default for Unfused DPA, DPA and MHA in TE JAX
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Remove explicit transpose_batch set for BSHD for DPA in JAX quickstart
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Add warnings in DPA and MHA to warn users of change defaults to BSHD instead of SBHD
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Minimize the scope of when to trigger warnings for changed defaults for transpose_batch_sequence
Signed-off-by: Kshitij Janardan 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 Janardan Lakhani <klakhani@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

Signed-off-by: janbernloehr <jan@bernloehrs.de>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>