Commits · 5e4e0b2c378d2b1ec2ee65dfa85124e1dd805389 · OpenDAS / TransformerEngine

22 Sep, 2025 1 commit

[PyTorch] Add sink attention support from cuDNN (#2148) · 5e4e0b2c

Charlene Yang authored Sep 22, 2025



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

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

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

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

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

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

* Apply 1 suggestion(s) to 1 file(s)
Co-authored-by: Chen Cui <chcui@nvidia.com>
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* fix unfused grad; add softmax_type; add sink to bwd
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Apply 1 suggestion(s) to 1 file(s)
Co-authored-by: Chen Cui <chcui@nvidia.com>
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix padding mask; add swa tests; remove requires_grad for off-by-one
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* Apply 1 suggestion(s) to 1 file(s)
Co-authored-by: Chen Cui <chcui@nvidia.com>
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Apply 1 suggestion(s) to 1 file(s)
Co-authored-by: Chen Cui <chcui@nvidia.com>
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Apply 1 suggestion(s) to 1 file(s)
Co-authored-by: Chen Cui <chcui@nvidia.com>
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Apply 1 suggestion(s) to 1 file(s)
Co-authored-by: Chen Cui <chcui@nvidia.com>
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Apply 1 suggestion(s) to 1 file(s)
Co-authored-by: Chen Cui <chcui@nvidia.com>
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Apply 1 suggestion(s) to 1 file(s)
Co-authored-by: Chen Cui <chcui@nvidia.com>
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Apply 1 suggestion(s) to 1 file(s)
Co-authored-by: Chen Cui <chcui@nvidia.com>
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Apply 1 suggestion(s) to 1 file(s)
Co-authored-by: Chen Cui <chcui@nvidia.com>
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Apply 1 suggestion(s) to 1 file(s)
Co-authored-by: Chen Cui <chcui@nvidia.com>
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

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

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

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

* add CP A2A; dq/dk mismatches
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix CP A2A; need cleaner solution
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix CP A2A; pending cudnn kernel change
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* fix world size in unit test; avoid thd format
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix kernel_backend, dtype in unit test; fix head_dim for FP8 Hopper
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* fix fp8 context
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* allow no_mask/padding for SWA(left,0)
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Revert "allow no_mask/padding for SWA(left,0)"

This reverts commit 08b4ccc67a08b6882080b06aa715f541bb832aca.
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

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

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

* skip 9.13 for MLA, non 192/128
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* small cleanups and improvements
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* force sink/dsink to be float32
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* return to GH TE main FE commit
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



* update FE to 1.14.1
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* clean up before CI
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>

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

* add backend selection guard for unit tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add docstring for softmax type enums in C
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

5e4e0b2c

18 Sep, 2025 1 commit

[Pytorch] Add Cutlass Grouped GEMM Support for fine-grained MoE Model (#2045) · 8aee1bb7

alan yang authored Sep 18, 2025



* feat: add cutlass group gemm support
Signed-off-by: Min Yang <min.yang@shopee.com>

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

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



* refactor: refactor multi tensor gemm interface
Signed-off-by: Min Yang <min.yang@shopee.com>

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

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



* refactor: refactor nvte_multi_stream_cublas_gemm func and add license info
Signed-off-by: Min Yang <min.yang@shopee.com>

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

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



* feat: add unit test for cutlass group gemm
Signed-off-by: Min Yang <min.yang@shopee.com>

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

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



* feat: add cutlass support type protect
Signed-off-by: Min Yang <min.yang@shopee.com>

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

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

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



* feat: fix unit tests error
Signed-off-by: Min Yang <min.yang@shopee.com>

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

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



* feat: refactor host workspace malloc
Signed-off-by: Min Yang <min.yang@shopee.com>

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

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

* further relex threshold and add a env var to warn fall back
Signed-off-by: Xin Yao <xiny@nvidia.com>

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

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



---------
Signed-off-by: Min Yang <min.yang@shopee.com>
Signed-off-by: Xin Yao <xiny@nvidia.com>
Signed-off-by: alan yang <89962857+cassiewilliam@users.noreply.github.com>
Co-authored-by: Min Yang <min.yang@shopee.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Xin Yao <xiny@nvidia.com>
Co-authored-by: Phuong Nguyen <phuonguyen@nvidia.com>

8aee1bb7

17 Sep, 2025 2 commits

TE Gemma tutorial attempt#2 (#1839) · 7042d7ae

Sudhakar Singh authored Sep 16, 2025



* add tutorial files and other local changes
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* remove extraneous code for easy debu
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* make cuda graphs work with non-paged and paged attention
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



* perf imp for kv cache ops
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* add code for calibration
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



* optimize kv_cache reindex and copy kernels
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



* changes to make quantizers work with fp8_calibration
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* avoid reindexing from python side
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* rename variable from previous commit
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

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

* use quantizer only if needed
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* functionality of the tutorial tested and perf checked
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* remove files and update headers/licenses
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* update header/license
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



* update tutorial for review
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 weights downloadable on the fly; remove extra print statements
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



* fix lint and update comments
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 comma back, typo
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* sequence_start_positions should be None for training
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 paged attention numberes and update requirements.txt file
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* more fixes
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 tutorial work on blackwell
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* remove gemma FT tutorial for now
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* fixing the headings placement and rewording attention -> kv caching
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* fixes from comments
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



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

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

* add more comments to te_gemma.py and cleanup utils.py
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 more information about the hierarchy of the classes used in the tutorial
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 better cuda graphs picture
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* addd updated cuda graphs pictures
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* add illustrated cuda graphs
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

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

* small fixes in documentation
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* add torch.no_grad() to force reduced memory usage
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* some fixes from recent comments
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

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

* add te_rope_emb to class desc
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

* fix tutorial wording; add calibration fix to grouped_linear.py
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>

---------
Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>
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>
Co-authored-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

7042d7ae

FP8 Output Quantization for GEMM (#2123) · ba37529c

vthumbe1503 authored Sep 16, 2025



* Test working as I think it should work
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>

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

Restrict the number of cases for unfused quantization, some fp8->fp8 cases are handled by cublas
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>

fix merge conflict
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

bug: missed a } in the code
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>

Add cuBLASMp-backed GEMM-like API to TE common (#1824)

* Pick up cuBLASMp during build
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

* Change lib order to fix link error
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

* Context creation, incomplete...
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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

* A sanity AgGemm test, failing...
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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

* Take care of uneven distribution
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Use MPI to get position of local matrices
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

* Refactor & fixes
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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

* Gemm-AR, not working...
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

* Setting all-reduce epilogue for gemm-ar
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Use supported shapes for GEMM-AR
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

* First shot at fp8
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Use TensorHolder in tests
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* More test configs
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

* Parametrize dtypes for A, B and D separately
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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

* Flags parity with cublas_gemm, saving...
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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

* Fix bias test
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Aux, saving...
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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

* Use test::Tensor
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Set scale inv
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Remove unsupported test configs
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

* Replace libcal with NCCL
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Add NVTX markers to API functions
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Tweak GemmAr tests
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* More test config
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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

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

* Fix merge fallout
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Remove MPI dependency, comment API, add algo parameter
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Fix nvshmem dependency
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Fix nvshmem build
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Excluse CommGemm tests from L0_cppunittest
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Add cpp_distributed sh file for CI
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Adapt tp TensorAllocator
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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



* Skip GemmAr test on unsupported HW
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Oversibscribe is needed on some clusters
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Fix incomplete libcal removal
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Move CI tests to L1
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Rename context to include NVTE prefix
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Remove leftover code
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

* More detailed NVTE_CHECK diag
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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



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

* Include stdbool header for legacy C compilers
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Remove now unused argument
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Abstract away cuBLASMp algo behind our own enum
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* More detailed shape diag messages
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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



* Update transformer_engine/common/include/transformer_engine/comm_gemm.h
Co-authored-by: Przemyslaw Tredak <ptrendx@gmail.com>
Signed-off-by: Vladimir Cherepanov <56651474+mk-61@users.noreply.github.com>

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

---------
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>
Signed-off-by: Vladimir Cherepanov <56651474+mk-61@users.noreply.github.com>
Co-authored-by: Vladimir Cherepanov <vcherepanov@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Przemyslaw Tredak <ptrendx@gmail.com>
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

FP8 AllGather in FP8 GroupedGEMM + Fix Stream Usage Issue. (#2086)

* FP8 AllGather in FP8 GroupedGEMM

1. Support current scaling FP8 quantation with a given amax.
2. Support FP8 AG in fwd and BF16 RS in bwd.
3. The workflow is AR-max -> FP8 Quant -> FP8 AG -> FP8 GroupedGEMM.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Slightly refactor
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adding documents of new args.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adding unit-tests.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adding license.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Move unit-tests to L1.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Move quantizaer store/reset into FP8 only.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adding all layout support for Blackwell+
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adopt the feedback from code-review.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Fixed the wrong stream used by d2d in groupedGEMM FFI.
Signed-off-by: Ming Huang <mingh@nvidia.com>

---------
Signed-off-by: Ming Huang <mingh@nvidia.com>
Co-authored-by: Phuong Nguyen <phuonguyen@nvidia.com>
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

[JAX] Delay MeshResource validation until first usage (#2124)

Delay MeshResource validation until first usage
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>
Co-authored-by: Phuong Nguyen <phuonguyen@nvidia.com>
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

[JAX] Decouple Recipe and ScalingMode (#1728)

* Decouple recipe and scaling mode
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Expose global QuantizeConfig instance as a getter
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

* Merge branch 'main' into dev/jberchtold/jax-scaling-mode-and-recipe-decoupling
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Rename UsageType to TensorSource
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Update test_layer.py
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

---------
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>
Signed-off-by: jberchtold-nvidia <158520091+jberchtold-nvidia@users.noreply.github.com>
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

[JAX] `dot_1_output` sharding constraint + use AXIS_IS_UNSHARDED (#2128)

* add dot_1_output sharding constraint + use AXIS_IS_UNSHARDED
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

[JAX] Add amax input to DBiasQuantizePrimitive and FFI (#2118)

* add amax input to DBiasQuantizePrimitive and FFI
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



* make sure amax is init with zero
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

Further relax constraints to cuDNN 9.13 for disabling fused attn for kv caching (#2121)
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

Temporarily remove comm_gemm tests (#2133)
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

[PyTorch] Disable determinism for sm100 (#2130)

* disable determinism for sm100+ and cudnn<9.14
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix remaining CI failures
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



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

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

* remove sm100 from determinism table
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

[PyTorch] ONNX export of FP8 Current Scaling (#2068)

* Compute amax in normalization forward in current scaling in untuned kernels
Signed-off-by: Jan Bielak <jbielak@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



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



* 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>

* apply tims suggestions
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

[PyTorch][MOE] Tentative Fix For Replacing from_blob with empty for experts receiving zero tokens (#2134)

use torch empty for empty shape instead of from_blob
Signed-off-by: zhongboz <zhongboz@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

build: pull cached wheels (#2127)

* build: pull cached wheels
Signed-off-by: oliver könig <okoenig@nvidia.com>

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

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



* Update setup.py
Signed-off-by: oliver könig <okoenig@nvidia.com>

---------
Signed-off-by: oliver könig <okoenig@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

feat: Add support for multiple quantization modes in the UB communicators (#2043)
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

[Common] Add checks to CUDA kernel launch and CUDA API calls (#2074)

* add checks to cuda kernel launch and cuda API calls
Signed-off-by: Xin Yao <xiny@nvidia.com>

* Remove exceptions from destructors
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* fix weired dispatch in ln/rmsnorm
Signed-off-by: Xin Yao <xiny@nvidia.com>

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

[PyTorch] Support bf16+fp8 cudagraph (#2098)

* support bf16+fp8 model
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
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
Signed-off-by: Robin Zhang <robinz@nvidia.com>

---------
Signed-off-by: Robin Zhang <robinz@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>
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

Dropout with 8-bit RNG (#2014)

* Add dropout kernel with 8-bit RNG
Co-authored-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>
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



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

* Avoid ambiguous types
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Do not enforce dropout prob is representable in 8 bits
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Expand error message
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



* Fix small statistical bug from using less-equal instead of less-than

Refactor kernel implementations and add comments. Interpret masks as bytes rather than 16-bit uints.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix linter warning
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 helper function in PyTorch extensions
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: Tim Moon <tmoon@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

Create GPU reload buffers on main stream (#2131)

* Create GPU relaod buffers on main stream
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-ptyche01.ptyche.clusters.nvidia.com>

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

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



* Fixed typo
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-preos01.a51.clusters.nvidia.com>

* Fixed typo
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-preos01.a51.clusters.nvidia.com>

---------
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-ptyche01.ptyche.clusters.nvidia.com>
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-preos01.a51.clusters.nvidia.com>
Co-authored-by: Selvaraj Anandaraj <selvaraja@login-ptyche01.ptyche.clusters.nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Selvaraj Anandaraj <selvaraja@login-preos01.a51.clusters.nvidia.com>
Co-authored-by: Paweł Gadziński <62263673+pggPL@users.noreply.github.com>
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

mxfp8 unfused quant support, refined unit test, remove unecessary quantization code
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>

missed a quant code removal
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>

minor bug fix
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



Add cuBLASMp-backed GEMM-like API to TE common (#1824)

* Pick up cuBLASMp during build
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

* Change lib order to fix link error
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

* Context creation, incomplete...
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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

* A sanity AgGemm test, failing...
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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

* Take care of uneven distribution
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Use MPI to get position of local matrices
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

* Refactor & fixes
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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

* Gemm-AR, not working...
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

* Setting all-reduce epilogue for gemm-ar
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Use supported shapes for GEMM-AR
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

* First shot at fp8
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Use TensorHolder in tests
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* More test configs
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

* Parametrize dtypes for A, B and D separately
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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

* Flags parity with cublas_gemm, saving...
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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

* Fix bias test
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Aux, saving...
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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

* Use test::Tensor
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Set scale inv
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Remove unsupported test configs
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

* Replace libcal with NCCL
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Add NVTX markers to API functions
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Tweak GemmAr tests
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* More test config
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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

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

* Fix merge fallout
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Remove MPI dependency, comment API, add algo parameter
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Fix nvshmem dependency
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Fix nvshmem build
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Excluse CommGemm tests from L0_cppunittest
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Add cpp_distributed sh file for CI
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Adapt tp TensorAllocator
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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



* Skip GemmAr test on unsupported HW
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Oversibscribe is needed on some clusters
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Fix incomplete libcal removal
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Move CI tests to L1
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Rename context to include NVTE prefix
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Remove leftover code
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

* More detailed NVTE_CHECK diag
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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



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

* Include stdbool header for legacy C compilers
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Remove now unused argument
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* Abstract away cuBLASMp algo behind our own enum
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

* More detailed shape diag messages
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

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

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



* Update transformer_engine/common/include/transformer_engine/comm_gemm.h
Co-authored-by: Przemyslaw Tredak <ptrendx@gmail.com>
Signed-off-by: Vladimir Cherepanov <56651474+mk-61@users.noreply.github.com>

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

---------
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>
Signed-off-by: Vladimir Cherepanov <56651474+mk-61@users.noreply.github.com>
Co-authored-by: Vladimir Cherepanov <vcherepanov@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Przemyslaw Tredak <ptrendx@gmail.com>

FP8 AllGather in FP8 GroupedGEMM + Fix Stream Usage Issue. (#2086)

* FP8 AllGather in FP8 GroupedGEMM

1. Support current scaling FP8 quantation with a given amax.
2. Support FP8 AG in fwd and BF16 RS in bwd.
3. The workflow is AR-max -> FP8 Quant -> FP8 AG -> FP8 GroupedGEMM.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Slightly refactor
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adding documents of new args.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adding unit-tests.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adding license.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Move unit-tests to L1.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Move quantizaer store/reset into FP8 only.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adding all layout support for Blackwell+
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Adopt the feedback from code-review.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Fixed the wrong stream used by d2d in groupedGEMM FFI.
Signed-off-by: Ming Huang <mingh@nvidia.com>

---------
Signed-off-by: Ming Huang <mingh@nvidia.com>
Co-authored-by: Phuong Nguyen <phuonguyen@nvidia.com>

[JAX] Delay MeshResource validation until first usage (#2124)

Delay MeshResource validation until first usage
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>
Co-authored-by: Phuong Nguyen <phuonguyen@nvidia.com>

[JAX] Decouple Recipe and ScalingMode (#1728)

* Decouple recipe and scaling mode
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Expose global QuantizeConfig instance as a getter
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

* Merge branch 'main' into dev/jberchtold/jax-scaling-mode-and-recipe-decoupling
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Rename UsageType to TensorSource
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Update test_layer.py
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

---------
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>
Signed-off-by: jberchtold-nvidia <158520091+jberchtold-nvidia@users.noreply.github.com>

[JAX] `dot_1_output` sharding constraint + use AXIS_IS_UNSHARDED (#2128)

* add dot_1_output sharding constraint + use AXIS_IS_UNSHARDED
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

[JAX] Add amax input to DBiasQuantizePrimitive and FFI (#2118)

* add amax input to DBiasQuantizePrimitive and FFI
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



* make sure amax is init with zero
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* fix sharding rule
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>

Further relax constraints to cuDNN 9.13 for disabling fused attn for kv caching (#2121)
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

Temporarily remove comm_gemm tests (#2133)
Signed-off-by: Vladimir Cherepanov <vcherepanov@nvidia.com>

[PyTorch] Disable determinism for sm100 (#2130)

* disable determinism for sm100+ and cudnn<9.14
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix remaining CI failures
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



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

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

* remove sm100 from determinism table
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

---------
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>

[PyTorch] ONNX export of FP8 Current Scaling (#2068)

* Compute amax in normalization forward in current scaling in untuned kernels
Signed-off-by: Jan Bielak <jbielak@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



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



* 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>

* apply tims suggestions
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

[PyTorch][MOE] Tentative Fix For Replacing from_blob with empty for experts receiving zero tokens (#2134)

use torch empty for empty shape instead of from_blob
Signed-off-by: zhongboz <zhongboz@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

build: pull cached wheels (#2127)

* build: pull cached wheels
Signed-off-by: oliver könig <okoenig@nvidia.com>

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

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



* Update setup.py
Signed-off-by: oliver könig <okoenig@nvidia.com>

---------
Signed-off-by: oliver könig <okoenig@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

feat: Add support for multiple quantization modes in the UB communicators (#2043)

[Common] Add checks to CUDA kernel launch and CUDA API calls (#2074)

* add checks to cuda kernel launch and cuda API calls
Signed-off-by: Xin Yao <xiny@nvidia.com>

* Remove exceptions from destructors
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* fix weired dispatch in ln/rmsnorm
Signed-off-by: Xin Yao <xiny@nvidia.com>

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

[PyTorch] Support bf16+fp8 cudagraph (#2098)

* support bf16+fp8 model
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
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
Signed-off-by: Robin Zhang <robinz@nvidia.com>

---------
Signed-off-by: Robin Zhang <robinz@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>

Dropout with 8-bit RNG (#2014)

* Add dropout kernel with 8-bit RNG
Co-authored-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>
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



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

* Avoid ambiguous types
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Do not enforce dropout prob is representable in 8 bits
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Expand error message
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



* Fix small statistical bug from using less-equal instead of less-than

Refactor kernel implementations and add comments. Interpret masks as bytes rather than 16-bit uints.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix linter warning
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 helper function in PyTorch extensions
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: Tim Moon <tmoon@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

Create GPU reload buffers on main stream (#2131)

* Create GPU relaod buffers on main stream
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-ptyche01.ptyche.clusters.nvidia.com>

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

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



* Fixed typo
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-preos01.a51.clusters.nvidia.com>

* Fixed typo
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-preos01.a51.clusters.nvidia.com>

---------
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-ptyche01.ptyche.clusters.nvidia.com>
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-preos01.a51.clusters.nvidia.com>
Co-authored-by: Selvaraj Anandaraj <selvaraja@login-ptyche01.ptyche.clusters.nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Selvaraj Anandaraj <selvaraja@login-preos01.a51.clusters.nvidia.com>
Co-authored-by: Paweł Gadziński <62263673+pggPL@users.noreply.github.com>

minor code cleanup
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



minor cosmetics
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 comment
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



minor comment update
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

Fix CI failures for UB overlap changes (#2149)
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

minor bug: quantizer should not be none for unfused quantization
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

[JAX] Fix failing fused attn tests for dropout=0.1 and bias for sm100 (#2135)

* Fix failing tests for dropout=0.1 and bias for fused attn for blackwell
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



* Fix the skip message
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Assert in fused attn bwd pass for sm100
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

Add check for sm100
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 support to get all devs in the process for jax
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



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

* Make get_all_device_compute_capability more pythonic, thereby avoiding unnecessary type conversion
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Represent attn bias using enum instead of string
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

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

fix linting error
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

[PyTorch][CUDA Graph] Fix FP8 Weight Quantization Cache under CUDA Graph (#2119)

* add noop to comp amax
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* fix for fp8 blockwise recipe
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* resolve comments
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



---------
Signed-off-by: zhongboz <zhongboz@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>

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

* Update test_multi_process_distributed_grouped_gemm.py

change accidentally added while merging
Signed-off-by: vthumbe1503 <vthumbe@nvidia.com>

* Update dense.py

change accidentally added while merging
Signed-off-by: vthumbe1503 <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>

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

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



* address revie comments
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



* Bug solved: delayed scaling quantization with mxfp8 inputs didnt work
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 unit test error
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



* just to trigger ci
Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

* address review comments: quantization inside gemm and outside both should exactly match for fp32 accumulation
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>

* [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>

* fix merge conflict
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>

address review comments: quantization inside gemm and outside both should exactly match for fp32 accumulation

[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>

[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



---------
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>

ba37529c

16 Sep, 2025 1 commit

Do not use normalization forward + amax fusion if cuDNN backend is requested (#2174) · 258d0842

Jan Bielak authored Sep 16, 2025



* Do not use norm fwd + amax fusion if cudnn backend is requested
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

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

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



* Read envirornment vairable directly to avoid include error
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

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

258d0842

08 Sep, 2025 1 commit

Fused RoPE with combined QKV input. (#2122) · 84fa28d2

vasunvidia authored Sep 08, 2025



* Fused RoPE with combined QKV input.

Initial commit for Dropout with 8-bit RNG

Fix documentation

Initial commit for Fused QKV RoPE

WIP

Initial tests passing

Enable rotary percent and margin

Enable CP2, start_positions, interleaved

Cleanup test

Revert "Fix documentation"

This reverts commit 53df10044e7769982bd4af2ae2628e6b7717e715.

Revert "Initial commit for Dropout with 8-bit RNG"

This reverts commit 301505e24031cbcd679069e1c2cd4d00eedf2dca.

Cleanup.

Minor cleanup
Signed-off-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>

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

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

Signed-off-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>

* Optimize kernels
Signed-off-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>

* Misc. Cleanup
Signed-off-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>

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

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

Signed-off-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>

* Optimize kernel performance
Signed-off-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>

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

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

Signed-off-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>

* Move fused_qkv_rope test to test_fused_rope.py
Signed-off-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>

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

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



* apply shared memory optimization to separate fused rope kernels
Signed-off-by: Xin Yao <xiny@nvidia.com>

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

---------
Signed-off-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>
Signed-off-by: Xin Yao <xiny@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Xin Yao <xiny@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

84fa28d2

04 Sep, 2025 1 commit

[PyTorch][CUDA Graph] Fix FP8 Weight Quantization Cache under CUDA Graph (#2119) · 0f68f7b2

Zhongbo Zhu authored Sep 04, 2025



* add noop to comp amax
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* fix for fp8 blockwise recipe
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* resolve comments
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



---------
Signed-off-by: zhongboz <zhongboz@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>

0f68f7b2

31 Aug, 2025 1 commit

Dropout with 8-bit RNG (#2014) · e0e3d123

vasunvidia authored Aug 31, 2025



* Add dropout kernel with 8-bit RNG
Co-authored-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>
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



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

* Avoid ambiguous types
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Do not enforce dropout prob is representable in 8 bits
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Expand error message
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



* Fix small statistical bug from using less-equal instead of less-than

Refactor kernel implementations and add comments. Interpret masks as bytes rather than 16-bit uints.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix linter warning
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 helper function in PyTorch extensions
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: Tim Moon <tmoon@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

e0e3d123

28 Aug, 2025 1 commit

[PyTorch][MOE] Tentative Fix For Replacing from_blob with empty for experts... · c449c6cf

Zhongbo Zhu authored Aug 28, 2025


[PyTorch][MOE] Tentative Fix For Replacing from_blob with empty for experts receiving zero tokens (#2134)

use torch empty for empty shape instead of from_blob
Signed-off-by: zhongboz <zhongboz@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

c449c6cf

26 Aug, 2025 1 commit
- [PyTorch] Expose more activation functions (#2106) · 07db17b5
  Xin Yao authored Aug 26, 2025
```
expose more activation functions
Signed-off-by: Xin Yao <xiny@nvidia.com>
```
  07db17b5
20 Aug, 2025 1 commit

Add backward RMSNorm+Add fusion (#2028) · 5b4d89c3

Jan Bielak authored Aug 19, 2025



* Add rmsnorm_bwd_add
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Add BackwardAddRMSNorm fused operation
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Try to optimize register usage in kernels
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Add separate BackwardAdd stage for the fused backward add
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

---------
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

5b4d89c3

15 Aug, 2025 1 commit

Fuse linear+scale+add (#2042) · c654e4fe

Jan Bielak authored Aug 15, 2025



* Add `nvte_cublas_gemm_scaled`
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Support use of `alpha` and `beta` in `tex.generic_gemm`
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Support use of `alpha` and `beta` in `general_gemm`
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Support use of `alpha` and `beta` in `BasicLinear._functional_forward` and `BasicLinear._functional_backward`
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Add `ForwardLinearScaleAdd` fusion
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Add `BackwardLinearScale` fusion
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Apply suggestions from code review
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Remove calls to `validate_gemm_scale` from `BasicLinear`
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

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

c654e4fe

14 Aug, 2025 1 commit

[PyTorch] Disable fused dbias-quantize kernel for unsupported recipes (#2007) · a169e9e7

Tim Moon authored Aug 13, 2025



* Unfused impl for dbias-quantize
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Unfused impl for dact-dbias-quantize
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Disable fused bgrad-quantize for unsupported recipes
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove unfused dbias-quantize impls

Not supported in the core lib.
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



* Support unfused impls in tex functions
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Tweaks
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 unused imports
Signed-off-by: Tim Moon <tmoon@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>

a169e9e7

09 Aug, 2025 1 commit

Use userbuffers for MXFP8 wgrad all-gather overlap (#1982) · 077e26c3

Daniel Stokes authored Aug 09, 2025



* fix: Add stream synchronization before destroying MPI communicator (#1979)
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* feat: Implement column-wise userbuffer overlap for comm+GEMM operations

Add support for overlapping column-wise allgather communication with GEMM
operations to improve training performance:

* **Core infrastructure changes:**
  - Update bulk_overlap_columnwise_ag() to accept explicit stream parameter
  - Modify userbuffers send/recv loops to use rank-ordered iteration
  - Add userbuffers_send_all/recv_all function declarations

* **Python integration:**
  - Add bulk_overlap_ag_with_external_gemm() C++ extension function
  - Expose new overlap function via pybind11 bindings
  - Update overlap method configurations to include more ring_exchange ops

* **LayerNorm MLP optimization:**
  - Enable column-wise quantization for FC2 gradient output
  - Implement overlap of allgather communication with FC2 DGRAD GEMM
  - Use fill_userbuffers_buffer_for_all_gather for efficient buffering

This optimization allows overlapping communication and computation phases
more effectively, reducing training wall-clock time by hiding allgather
latency behind GEMM execution.
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* fix: Working userbuffer overlapping API
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* fix: Fix overwriting bulk overlap UB object for layernormLinear
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* fix: Update external overlap to use tp size instead of nvsize to determine number of copies
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* fix: Fix linter error
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* fix: Explanatory comments of overlap logic
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* fix: Fix the UB fused ops tests
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* fix: Fix linter errors
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

---------
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

077e26c3

07 Aug, 2025 1 commit

[PyTorch] Fix numeric overflow caused by int-type parameters and return value... · dd083bdf

ldl authored Aug 07, 2025


[PyTorch] Fix numeric overflow caused by int-type parameters and return value in the roundup function (#2034)
Signed-off-by: lvdunlin <lvdunlin@xiaomi.com>
Co-authored-by: lvdunlin <lvdunlin@xiaomi.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

dd083bdf

06 Aug, 2025 2 commits

[PyTorch] fix input_quantizer usage for save_original_input; fix blockwise FP8... · de69ca0e

hx authored Aug 07, 2025


[PyTorch] fix input_quantizer usage for save_original_input; fix blockwise FP8 convert_and_update_tensor (#1978)

* fix input_quantizer in save_original_input bwd
Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com>

* fix get shape of blockwise tensor with only compact colwise data
Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com>

* fix blockwise FP8 convert_and_update_tensor
Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com>

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

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



---------
Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

de69ca0e

[PyTorch] Multi-tensor swizzle scaling factors for MXFP8 and fuse padding zeros (#2019) · c0d2f1a5

Xin Yao authored Aug 07, 2025



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

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

* multi-tensor swizzle
Signed-off-by: Xin Yao <xiny@nvidia.com>

* pad zeros in swizzle kernels
Signed-off-by: Xin Yao <xiny@nvidia.com>

* unify single- and multi-tensor swizzle
Signed-off-by: Xin Yao <xiny@nvidia.com>

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

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



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

* fix bug for col swizzle
Signed-off-by: Xin Yao <xiny@nvidia.com>

* check context & fix signifiers
Signed-off-by: Xin Yao <xiny@nvidia.com>

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

c0d2f1a5

04 Aug, 2025 1 commit

[Core] Kernel that swaps first two tensor dimensions (#1998) · c3f8a9f5

Tim Moon authored Aug 04, 2025



* Add basic kernel for swapping first two tensor dims
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add NVRTC kernel for swapping first dims
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add PyTorch extension for swap first dims kernel
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Tune kernel
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



* Make sure writes are contiguous
Signed-off-by: Tim Moon <tmoon@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>

c3f8a9f5

31 Jul, 2025 1 commit

Fuse amax computation into normalization kernel for current scaling (#2013) · 51eb6362

Jan Bielak authored Jul 31, 2025



* Compute amax in normalization kernels as long as the pointer is provided, even if using non quantized output
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Fuse amax computation into normalization forward
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Use TE lahyernorm kernel instead of raising error about unsupported cuDNN feature
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

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

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



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

51eb6362

30 Jul, 2025 1 commit

Refactor normalization.cpp to use quantizer logic introduced in #1952 (#2006) · 11ac24cf

Jan Bielak authored Jul 29, 2025



Refactor normalization.cpp to use quantizer logic introduced in #1952 instead of manual quantization
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

11ac24cf

29 Jul, 2025 3 commits

Fuse amax computation into activation kernel (#2004) · feda5b55

Jan Bielak authored Jul 29, 2025



* Compute amax in activation kernels when the output pointer is provided, even for non-fp8 outputs
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
(cherry picked from commit 9f13fe2fefc58cae93bc467d87d01ecf792a0381)

* Initialize metatensor values
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Fuse computation of amax into the activation kernel for fp8 current scaling
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
(cherry picked from commit 2b54327ac9c931a5340983a79e99de5caa0399dd)
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

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

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



* Zero out amax in `create_hp_tensor_with_amax` instead of relying on `Float8CurrentScalingQuantizer.__init__` to zero-initialize it
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

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

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



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

feda5b55

[PyTorch] Refactor C++ quantizer infrastructure (#1952) · cb5013bd

Tim Moon authored Jul 29, 2025



* remove reciprocal op
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* Refactor Quantizer::create_tensor function
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



* Fix bug when constructing FP8 tensor
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



* Add quantize function to C++ quantizers
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Prototype function to coerce Python quantized tensors to match quantizer
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use quantizer class in tex.quantize
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add FP8 current scaling support for activation backward
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Disable quantized GEMM output with FP8 current scaling
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add coerce_tensor functions for MXFP8 and DSv3
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



* Avoid quantizing empty tensors
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use consistent shapes for FP8 transposes
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* In attention impl, construct FP8 tensors with pre-initialized scale-invs
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Initialize MXFP8 scales to zero
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



* Store copy of quantizer when creating quantized tensors
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Make sure quantized tensors have private quantizer

Avoid problems with in-place ops after quantizer usages are changed externally.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Rename "coerce_tensor" to "convert_and_update_tensor"
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



* Make sure CUDA context is available when launching NVRTC kernel
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Expose CUDA context creation function externally
Signed-off-by: Tim Moon <tmoon@nvidia.com>

---------
Signed-off-by: zhongboz <zhongboz@nvidia.com>
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: zhongboz <zhongboz@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

cb5013bd

Fix the use-after-free bug in unfused normalization (#2002) · 5a495a39
Przemyslaw Tredak authored Jul 28, 2025
```
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>
```
5a495a39

25 Jul, 2025 1 commit

[C][PyTorch] Remove deprecated `device_id` arg for multi tensor API (#1994) · 1470116e

Kirthi Shankar Sivamani authored Jul 25, 2025



* Remove deprecated device arg
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

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

1470116e

17 Jul, 2025 1 commit

[PyTorch] Add save_original_input in Linear/GroupedLinear to save memory (#1865) · 07afda98

hx authored Jul 17, 2025



* save original input
Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com>

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

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



* fix input_quantizer usage in Linear bwd
Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com>

* minor fix
Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com>

* refine the docstring
Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com>

* Merge remote-tracking branch 'origin/main' into save_bf16_in_fp8_gemm
Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com>

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

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



* decouple linear bwd with save_original_input; clean up UTs
Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com>

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

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



---------
Signed-off-by: Hongxiao Bai <hongxiaob@nvidia.com>
Signed-off-by: hx <hongxiaob@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Xin Yao <xiny@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

07afda98

14 Jul, 2025 1 commit

[PyTorch] Fix bugs in router fusion (#1944) · 397c4be6

Autumn1998 authored Jul 14, 2025



* fix underterminsic problem in CI
Signed-off-by: tongliu <tongliu@nvidia.com>

* fix bug on mbs>1
Signed-off-by: tongliu <tongliu@nvidia.com>

* fix bug on sm dispatcher
Signed-off-by: tongliu <tongliu@nvidia.com>

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

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



* fix CI initial values
Signed-off-by: tongliu <tongliu@nvidia.com>

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

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



---------
Signed-off-by: tongliu <tongliu@nvidia.com>
Co-authored-by: tongliu <tongliu@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: Xin Yao <xiny@nvidia.com>

397c4be6

10 Jul, 2025 1 commit

[PyTorch][MoE] Kernels fusions for the MoE router (#1883) · 62acae05

Autumn1998 authored Jul 11, 2025



* add router fusion
Signed-off-by: tongliu <tongliu@nvidia.com>

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

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



* fix ci
Signed-off-by: tongliu <tongliu@nvidia.com>

* fix ci with cuda 12.3
Signed-off-by: tongliu <tongliu@nvidia.com>

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

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



* Review suggestions
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



* Fix CI sm89/80
Signed-off-by: tongliu <tongliu@nvidia.com>

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

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



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

62acae05

09 Jul, 2025 1 commit

[PyTorch][MoE] MXFP8 Support to Reduce CPU Overhead By Fuse Torch Empty Calls (#1934) · 96ee7173

Zhongbo Zhu authored Jul 09, 2025



* functional passed
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* before zero padding in mxfp8 swizzle, use torch zeros to malloc for now
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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

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

96ee7173

26 Jun, 2025 2 commits

[PyTorch|common] Optimize unpadding kernel for FP8 (#1866) · 23cf4ff9

xiaoxi-wangfj authored Jun 26, 2025



* [PyTorch|common] Implement unpadding kernel for FP8

1. Add multi-tensor unpadding kernel
2. Replace split+cat with unpadding kernel in Fp8Padding and Fp8Unpadding
3. Add unpadding with padding unit tests
Signed-off-by: xiaoxi-wangfj <690912414@qq.com>

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

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



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

* Update padding.cu
Signed-off-by: Xin Yao <xiny@nvidia.com>

---------
Signed-off-by: xiaoxi-wangfj <690912414@qq.com>
Signed-off-by: Xin Yao <xiny@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Xin Yao <xiny@nvidia.com>

23cf4ff9

[PyTorch][MoE] Reduce CPU Overhead By Fuse Torch Empty Calls (#1793) · c30e961f

Zhongbo Zhu authored Jun 25, 2025



* finish python ref impl for bulk alloc
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* c++ bulk alloc worked, still draft version
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

* resolve rebase conflict
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* add license
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* use shared_ptr to auto manage reference count
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* attempt to fix misc training error
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* attempt to handle case where experts get zero token
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* updated with fused C++ function calls
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

* experiment with reducing py object construction time
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* fix seg fault bug in inference mode
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

* fuse torch split into bulk alloc
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

* rebase to latest main
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* fix unit test failure
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* fix lint error
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* refactor create_tensor to use get_scale_shape
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* refactor quantize to call quantize_cpp
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* Implement separate functions for multi-tensor quantize and split + multi-tensor quantize
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update grouped linear module with fused split+quantize func
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Move multi-tensor quantize func to cast.cpp
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Do not expose quantizer helper function externally
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

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

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



* Revert cuDNN frontend commit
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* fix corner cases with zero tokens
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* add comments
Signed-off-by: zhongboz <zhongboz@nvidia.com>

---------
Signed-off-by: zhongboz <zhongboz@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>

c30e961f

13 Jun, 2025 3 commits

Add support for head_dim > 128 (#1797) · 71c76b6b

Charlene Yang authored Jun 14, 2025



* add support for head dim > 128
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* remove debugging
Signed-off-by: Charlene Yang <charleney@nvidia.com>

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

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



* raise tols slightly to tolerate 1/2048 mismatches
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* fix is_training for test_te_layer
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* add bprop support for blackwell
Signed-off-by: Charlene Yang <charleney@nvidia.com>

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

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



* minor tweak for format
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* fix backend selection results
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* bump sm100 to sm100+
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* add sq=1 test for MLA
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* enable sq=1 for bprop
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* minor tweak in comments
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* fix head_dim logic and remove pytest skip
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* add FE fix for d>128
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* update FE again to take in small fixes
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* add cuDNN version info in L0 tests
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* increase tols for Unfused + large dim
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* Revert "add cuDNN version info in L0 tests"

This reverts commit 3e1b426ca5319a2c0540b9e73bba7047d0e583e5.
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* fix tols for Unfused
Signed-off-by: Charlene Yang <charleney@nvidia.com>

---------
Signed-off-by: Charlene Yang <charleney@nvidia.com>
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>

71c76b6b

[common] Added support of FP4 data type (#1779) · 7b94bd99

Oleg Goncharov authored Jun 13, 2025



* Added support of FP4 data type
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Refactoring to BitsNum in progress
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Fixed compilation errors. All C++ tests passed
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Fixed a typo
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

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

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



* Added FP4 guard to TMA tensor descriptor data type
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

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

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



* Fixed errors in JAX C++ extensions
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

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

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



* Removed dummy NVFP4 C++ test file
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

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

* Refactored the code per the review notes. Fixed JAX build error.
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

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

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



* Removed unnecessary static casts
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Typo fix
Signed-off-by: Oleg Goncharov <64355998+Oleg-Goncharov@users.noreply.github.com>

* Pass correct num bits to create_2D_tensor_map; fixes CI
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

---------
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: Oleg Goncharov <64355998+Oleg-Goncharov@users.noreply.github.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

7b94bd99

Add support for overlapping wgrad NCCL AG with dgrad GEMM (#1849) · d90ced7c

Daniel Stokes authored Jun 13, 2025



* Add support for overlapping wgrad NCCL AG with dgrad GEMM
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* Remove unused wait on memcpy API from UB
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* Add better commenting to MXFP8 overlap
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

---------
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>
Co-authored-by: dastokes <dastokes@dastokes-dvt-01.nvidia.com>

d90ced7c

06 Jun, 2025 2 commits

[JAX] GroupedQuantizer and GroupedScaledTensor (#1666) · 7948779c

Phuong Nguyen authored Jun 06, 2025



* refactor the multi_stream utils + implement nvte_multi_tensor_quantize in TE/Common

* implement GroupedQuantizer and grouped_quantize in jaxx

* fix logical_axes_names for transpose tensor in ScaledTensor
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>
Co-authored-by: Hua Huang <huah@nvidia.com>
Co-authored-by: Ming Huang <mingh@nvidia.com>

7948779c

[PyTorch] FP8 Subchannel Recipe With FP8 Gather And Configurable Scaling... · 9985b02c

Zhongbo Zhu authored Jun 05, 2025


[PyTorch] FP8 Subchannel Recipe With FP8 Gather And Configurable Scaling Factor Tensor Swizzling (#1707)

* functional kernel for columnwise + no-transpose option, still hacky
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* pass all quantizer unit tests
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* refactor, add gemm ready api
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* make format options private members, simplify api
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* swizzle scales right before gemm
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* bug fix of single layer test
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* attempt to fix lint issue
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* fp8 gather pass, need minor refine
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* fix return_layernorm_output_gathered case
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* remove special cases, add sanity check before gemm
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

* lint ungrouped imports
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* Implement dequantize for compact 1D blocks.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* add more unit test with dequantize compact supported
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

* make ag for subchannel respect async
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* zero tolerance in distributed test
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* fix zero tolerance test
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* resolve rebase issues
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* lint & format
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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

* bug fix
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* relax rtol for fp32 distributed test
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* fix some ci issue
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* fix ci test failure in debug mode
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* Force row-wise and column-wise data to have same data format

Prototype "all-gather usage" in quantizer.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove dead logic for high-precision AGs
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

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



* Debug FP8 block-wise tensor tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Handle case where LayerNormLinear returns gathered norm output
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* fix debug mode
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

9985b02c

05 Jun, 2025 1 commit

Fix `NVTE_FRAMEWORK=all` installation (#1850) · 12af02f2

Kirthi Shankar Sivamani authored Jun 04, 2025



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

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

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

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

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

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

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

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

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

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

12af02f2

29 May, 2025 1 commit

Avoid memory allocations and deallocations when creating NVTETensor (#1813) · 4292653c

Przemyslaw Tredak authored May 29, 2025



* Changed the Tensor allocation strategy
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

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

* Disable debug flag
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

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

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



* Fix the double free error
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

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

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



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

* Fixed pyTorch recipe extension
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

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

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



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

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

* Hide TensorAllocator and fix the usage in LayerNorm
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

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

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



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

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

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



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

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

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

4292653c

14 May, 2025 1 commit

Cleanup pytorch extensions (#1781) · cfbbfb89

Kirthi Shankar Sivamani authored May 13, 2025



* rm unused swizzle extensions
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

* Consistent namespaces and first refactor
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

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

* revert accidental perm change
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

cfbbfb89

13 May, 2025 1 commit

Release GIL for PyTorch extensions (#1767) · 7e43feae

Kirthi Shankar Sivamani authored May 13, 2025



* Disallow kwargs for pybind extensions and release GIL if possible
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

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

7e43feae

11 May, 2025 1 commit

[C][PyTorch]Make pytorch extensions pure cpp (#1754) · 51cd4415

Kirthi Shankar Sivamani authored May 10, 2025



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

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

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

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

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

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

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

* Move grad outside autocast
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

* Address review comments
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



* Change src file name in cmake
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* move the kernels too
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

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

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

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

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

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

51cd4415