GitLab

* add mask-based moe permutation

* change moe_chunk_permute to moe_sort_chunks_by_indices

* fix __all__ in pytorch/permutation.py

* fix func/var names and typos; update tols in UT

---------
Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com>
Co-authored-by: Phuong Nguyen <phuonguyen@nvidia.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>

* POC for segment_ids/segment_pos
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Change segment_pos position
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Use RemainingArgs to solve number of parameters mismatches
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Test mask_descriptor for accomendating different mask representations
Signed-off-by: Reese Wang <rewang@nvidia.com>

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

* Use descriptor in bwd
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Primitives only accepts pure jnp array
Signed-off-by: Reese Wang <rewang@nvidia.com>

* segment_ids/pos support POC
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Move seqlens/offsets generation to mask descriptor
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Rename MaskDescriptor to SequenceDescriptor
Signed-off-by: Reese Wang <rewang@nvidia.com>

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

* Utilize sequence desc on FA bwd
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Migrate to new API
Signed-off-by: Reese Wang <rewang@nvidia.com>

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

* Remove small inputs and test different input format
Signed-off-by: Reese Wang <rewang@nvidia.com>

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

* Fix seed shardings
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Optimize sequence converting overhead
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Optimize seq_offsets calculation
Signed-off-by: Reese Wang <rewang@nvidia.com>

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

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

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

* Remove reduntant line
Signed-off-by: Reese Wang <rewang@nvidia.com>

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

* Avoid `parameters` function in op backward pass
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

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



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

only compare the recipe in AttentionParams.fp8_meta
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

Consolidate the distributed fused attention tests to shared input generation and execition logic.
Signed-off-by: Michael Goldfarb <mgoldfarb@nvidia.com>

* corrected RS overlap BF16 output clashing with Float8Tensor constructor
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed empty dgrad buffer dtype at initialization
Signed-off-by: Alp Dener <adener@nvidia.com>

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

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



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

* Add a compile option to compile activation kernels with fast math
Signed-off-by: Guyue Huang <guyueh@nvidia.com>

* Fix
Signed-off-by: Guyue Huang <guyueh@nvidia.com>

* Apply suggestions from code review
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: guyueh1 <140554423+guyueh1@users.noreply.github.com>

---------
Signed-off-by: Guyue Huang <guyueh@nvidia.com>
Signed-off-by: guyueh1 <140554423+guyueh1@users.noreply.github.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* support AG overlap in sequence-parallel Linear forward and RS overlap in sequence-parallel Linear backward
Signed-off-by: Alp Dener <adener@nvidia.com>

* implemented TP overlap support for column-parallel te.Linear
Signed-off-by: Alp Dener <adener@nvidia.com>

* fixed backward pass for te.Linear column-parallel with TP overlap, updated unit tests
Signed-off-by: Alp Dener <adener@nvidia.com>

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

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



* improved error messages for internal failure to infer TP overlap options in te.Linear
Signed-off-by: Alp Dener <adener@nvidia.com>

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

* fixed incorrect TP overlap option asserts
Signed-off-by: Alp Dener <adener@nvidia.com>

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

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



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

Signed-off-by: Sergii Dymchenko <sdym@meta.com>

* add test_multiprocessing_encoder with processing spawning in bash

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

Take token count quantization of fused attention into consideration for CP results correction (#1396)

* fix second half lse shape
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

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

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

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



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

* make pad_between_seqs check do not consider padding at the end
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* change CP THD test to make it consider 0-length sequence
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

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

* only use varlen func of flash attention while qkv_format is THD
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* try to converge code of flash and fused attentions
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* fix bwd compute with P2P
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove redundant out_per_step view
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* enable cudnn>9.6 and THD+GQA
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* enable CP with FusedAttn+SWA+All_Gather
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* enable CP with FusedAttn+SWA+All_Gather
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* code cleaning for cu_seqlens
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

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

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



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

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

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

* fix lse_seqlen in thd out correction
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

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

Correct fused attention output after each step to reduce intermediate memory use.
Signed-off-by: Michael Goldfarb <mgoldfarb@nvidia.com>

the current implementation would release the output of ln, leading to an error if setting `return_layernorm_output=True`.
Signed-off-by: Liyuan Liu <llychinalz@gmail.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Fix SWA mask for THD and forcing seqlen_kv >= seqlen_q for SWA
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Generalize sliding window mask
Signed-off-by: Reese Wang <rewang@nvidia.com>

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

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

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

* add swa (left,0) + padding + brcm support
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

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



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

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

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



* upgrade to FE 1.9-rc
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* skip thd + CP + fused attn tests for cuDNN 9.6+ due to different stats shapes
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

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



---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Move test distributed encoder to L0 distributed test suit

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>
Co-authored-by: Reese Wang <rewang@nvidia.com>

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

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

* fix the order of checking valid swa and changing mask type
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

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



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

* revamp to get full mask
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

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



---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

add weights_only=False for torch.load
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Add util functions to attn_mask_type
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add util functions to qkv_layout
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Fix THD cross reference code
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Remove explicit segment_pad, encoding it to segment_ids
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add jax.jit, replace _token with segment_ids, rename bias shape enum
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add comment for make_mask
Signed-off-by: Reese Wang <rewang@nvidia.com>

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

* Add doc strings for the added functions
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Remove cache for fa deterministic which causes UT failed
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Rename fixture to avoid conflict
Signed-off-by: Reese Wang <rewang@nvidia.com>

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

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

* draft implementation of fsdp2 fp8 all gather
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

* fix the convergence issue
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

* Add warning
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

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

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



* disable lint error
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

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

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



* fix the lint error
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

* fix lint error
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

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

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



* fix lint error
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

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

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



* fix lint error
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

* add comments
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

* add ref
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

* add related tests
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

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

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



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

* softmax custom calls with correct encapsulates

* rm jax deprecated features

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

Add Jeremy to ci users
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* fix ctx.aval_out indexing for workspace
* add cudnn init to prepare phase of norm custom calls
* add thread_local for norm registry instance
---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* Bug Fix: Use default factory for not sharing mutable default values
---------
Signed-off-by: Reese Wang <rewang@nvidia.com>
Co-authored-by: Phuong Nguyen <phuonguyen@nvidia.com>

* cuDNN normalization integration
* TE Norm refactor
* TE Norm APIs changes.

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

Debug Mcore integration test

Avoid FP8 on Ampere and older. Generate synthetic data instead of depending on external data.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* always have padding mask type for both flash and fused attentions
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove an redundant assert
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

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

Store module extra state in tensor
Signed-off-by: Tim Moon <tmoon@nvidia.com>

Debug jobs to deploy nightly docs
Signed-off-by: Tim Moon <tmoon@nvidia.com>

Scale sequence length in CP tests to avoid tiny sizes.
Signed-off-by: Michael Goldfarb <mgoldfarb@nvidia.com>

* draft implementation
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

* compile error fix
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

* fix compile error
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

* remove print
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

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

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



* Edit comments
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

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

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



* edit the bulk-overlap test case
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

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

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



* add version guard
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

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

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



* add runtime version guard
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

* fix the version guard
Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>

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

* Update list of CI users
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update list of CI users
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* retain_graph=True for grouped gemm
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* remove an unnecessary retain_graph=True
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

* make retain_graph in graph capture configurable
Signed-off-by: Xiaowei Ren <xren@nvidia.com>

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

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

Moved framework agnostic THD kernels to common.

---------
Signed-off-by: Michael Goldfarb <mgoldfarb@nvidia.com>

* Align RNG tracker with megatron
Signed-off-by: Robin Zhang <robinz@nvidia.com>
Co-authored-by: Yifei Song <yifeis@nvidia.com>

* Fix module_params order and warmup bug in cudagraph
Signed-off-by: Robin Zhang <robinz@nvidia.com>
Co-authored-by: Yifei Song <yifeis@nvidia.com>

* Add fp8_group argument and fix fp8 accuracy issue for cudagraph
Signed-off-by: Robin Zhang <robinz@nvidia.com>
Co-authored-by: Yifei Song <yifeis@nvidia.com>

* Add TE modules and weights filters to support MoE models
Signed-off-by: Robin Zhang <robinz@nvidia.com>
Co-authored-by: Yifei Song <yifeis@nvidia.com>

* Revert self.fp8
Signed-off-by: Robin Zhang <robinz@nvidia.com>

* Use hooks to filter module params
Signed-off-by: Robin Zhang <robinz@nvidia.com>

* Filter all TE modules in hooks
Signed-off-by: Robin Zhang <robinz@nvidia.com>
Co-authored-by: Yifei Song <yifeis@nvidia.com>

* Format code
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



* Update graph.py
Signed-off-by: Xin Yao <yaox12@outlook.com>

* Revert CudaRNGStatesTracker
Signed-off-by: Robin Zhang <robinz@nvidia.com>

* Format Update
Signed-off-by: Yifei Song <yifeis@nvidia.com>

* Revert "Use hooks to filter module params"

This reverts commit 73a22e2e8bcf43ec84c23bc844b8d16d06626e26.
Signed-off-by: Yifei Song <yifeis@nvidia.com>

* Remove filtering module params
Signed-off-by: Robin Zhang <robinz@nvidia.com>

---------
Signed-off-by: Robin Zhang <robinz@nvidia.com>
Signed-off-by: Xin Yao <yaox12@outlook.com>
Signed-off-by: Yifei Song <yifeis@nvidia.com>
Co-authored-by: Yifei Song <yifeis@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Xin Yao <yaox12@outlook.com>
Co-authored-by: Xin Yao <xiny@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Add helper function to convert C++ container to string
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

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



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