GitLab

* 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: Ace Eldeib <alexeldeib@gmail.com>
Co-authored-by: Xin Yao <xiny@nvidia.com>

* rm cudaGraph compatible trait from GroupedGEMM and groupedQuantize
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* add grouped_gemm jitting in the unit test
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

* added XLA FFI custom op for TE/common nvte_cublas_gemm
Signed-off-by: Alp Dener <adener@nvidia.com>

started GemmPrimitive, abstract done
Signed-off-by: Alp Dener <adener@nvidia.com>

gemm custom op working with BF16, needs testing for FP8/MXFP8
Signed-off-by: Alp Dener <adener@nvidia.com>

converted TE GEMM API to use ScaledTensor and added os ENV flag to use TE GEMM under general gemm() call
Signed-off-by: Alp Dener <adener@nvidia.com>

BF16 tests passing, FP8 tests should be passing but contracting_dims has a scoping issue
Signed-off-by: Alp Dener <adener@nvidia.com>

fp8 tests passing for E4M3, getting CUBLAS_STATUS_NOT_SUPPORTED for E5M2
Signed-off-by: Alp Dener <adener@nvidia.com>

updated GEMM API to use separate LHS and RHS quantizers instead of a QuantizerSet
Signed-off-by: Alp Dener <adener@nvidia.com>

new GemmPrimitive passing all Dense tests
Signed-off-by: Alp Dener <adener@nvidia.com>

import cleanup and reverted code chunk movement
Signed-off-by: Alp Dener <adener@nvidia.com>

removed unused .transpose() implementations from ScaledTensors
Signed-off-by: Alp Dener <adener@nvidia.com>

all custom call tests passing on Hopper, GEMM-related tests cover both GemmPrimitive and native JAX impl
Signed-off-by: Alp Dener <adener@nvidia.com>

removed direct calls to GemmPrimitive.enabled() from outside of cpp_extensions
Signed-off-by: Alp Dener <adener@nvidia.com>

removed unused changes to ScaledTensor classes and debug prints
Signed-off-by: Alp Dener <adener@nvidia.com>

* minor unit test cleanup
Signed-off-by: Alp Dener <adener@nvidia.com>

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

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



* FP8 tests passing on Blackwell but MXFP8 outputs NaN
Signed-off-by: Alp Dener <adener@nvidia.com>

* reverted dense and fuseddense changes, FP8 test passing on Hopper and Blackwell, MXFP8 has issues with E5M2
Signed-off-by: Alp Dener <adener@nvidia.com>

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

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



* MXFP8 issue traced to scale factor padding with NaNs instead of zeros
Signed-off-by: Alp Dener <adener@nvidia.com>

* padding scale with 2^-127 instead of nans
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* fix bug on rhs_scale_inv usage
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* cleanup E8M0 type converter use it in gemm.cpp
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* segfault fixed, passing all unittests on Blackwell
Signed-off-by: Alp Dener <adener@nvidia.com>

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

* fix workspace alignment
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



* fixed GemmPrimitive custom partitioning to match jax.nn.scaled_matmul
Signed-off-by: Alp Dener <adener@nvidia.com>

all unit tests passing on H100x8 node
Signed-off-by: Alp Dener <adener@nvidia.com>

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

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



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

fixed batch dimension numbers
Signed-off-by: Alp Dener <adener@nvidia.com>

fixed FP8 scale sharding rule when there are no FP8 scales
Signed-off-by: Alp Dener <adener@nvidia.com>

added error message for unsupported Shardy partitioner
Signed-off-by: Alp Dener <adener@nvidia.com>

fixed test tolerances for FP8 cases
Signed-off-by: Alp Dener <adener@nvidia.com>

fixed shardy test skip cases
Signed-off-by: Alp Dener <adener@nvidia.com>

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

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



* moved reshape of encoder output in encoder examples to make custom partitioning rules work correctly
Signed-off-by: Alp Dener <adener@nvidia.com>

* added helper functions for padding and unpadding block scales, changed GemmPrimitive to accept unpadded scales and pad them after sharding
Signed-off-by: Alp Dener <adener@nvidia.com>

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

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



* updated shardy rules for all custom ops to decouple block scale rules from their tensors
Signed-off-by: Alp Dener <adener@nvidia.com>

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

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



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

* changed unit test use_jax_gemm option to be a context to preserve external custom op settings, tightened multi-GPU encoder test tolerances, changed gemm() API to use contracting_dims and batched_dims separately instead of dimension_numbers
Signed-off-by: Alp Dener <adener@nvidia.com>

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

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



* fixed typo in test utils
Signed-off-by: Alp Dener <adener@nvidia.com>

* added sequence-first input warnings
Signed-off-by: Alp Dener <adener@nvidia.com>

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

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



* fixed datasets version for JAX examples
Signed-off-by: Alp Dener <adener@nvidia.com>

* reverting modification to force_1x_quantization decision
Signed-off-by: Alp Dener <adener@nvidia.com>

* corrected gemm function syntax in unit tests
Signed-off-by: Alp Dener <adener@nvidia.com>

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

* memset for the mxfp8 scale padding
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

* Fix JAX grouped gemm error on CUDA 12.9.1 & later by using 16B alignment for scale ptr
Signed-off-by: Hua Huang <huah@nvidia.com>

* Pad MXFP8 scales with 2*-127 instead of NaNs
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

* Support MXFP8 and handle empty matrices
Signed-off-by: Hua Huang <huah@nvidia.com>

---------
Signed-off-by: Hua Huang <huah@nvidia.com>

* Add support for Fused Attn MLA head_dim_qk != head_dim_v
	Modify is_fused_attn_kernel_available() to accept different head_dims for qk and v
	Modify FusedAttnHelper to accept different head_dims for qk and v and modify assert dims checks in parse_qkv_aval()
	Modify FusedAttnFwdPrimitive and FusedAttnBwdPrimitive to accept different head_dims for qk and v
	Modify Fused Attn related cpp and csrc extension API calls to accept different head_dims for qk and v
	Modify DotProductAttention call() to extract head dims separately for qk and v
	Modify the FusedAttn Tests to accommodate for API changes in FusedAttn API
	Add test case for head_dim_qk != head_dim_v (failing)
	Modify the baseline JAX appropriately to reshape the output vector based on v dims and not q dims
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

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

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



* Fix context dims in general DPA in test_fused_attn
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Fix dim for output tensor by replacing with v head dim rather than q head dim
Add test cases for jax fused attn where head_dim_qk != head_dim_v for a combination of data types and attention type
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Modify the fused attn jax unit test case for head dim qk != head dim v
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Use new FusedAttnRunner function signature for separate hidden dim for qk and v in Fused Attn distributed tests
Code clean up
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Fix usage of is_fused_attn signature in distributed tests
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Remove unnecessary assert
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

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

* 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>

* fixes for jittable grouped_quantize

* fixes for jittable grouped_gemm

* fix contracting_dim for wgrad gemm

* exclude jitted grouped_gemm from the unit test as it does not work cudaGraph

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

* Implemented GroupedDense and TestGroupedDense for BF16, FP16, and FP8 
* Fix GroupedGemmFFI cuBLAS workspace alignment bug
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

Revert "[JAX] GroupedDense v.2 without dynamic shape (#1721)"

This reverts commit 5d01ef21

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

* Implemented GroupedDense and TestGroupedDense for BF16, FP16, and FP8 
* Fix GroupedGemmFFI cuBLAS workspace alignment bug
Signed-off-by: Hua Huang <huah@nvidia.com>
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

* 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>

* 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>

Fix NVTETensor leak in attention.cpp
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* 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>

* Enforce input sharding of norm primitive does not shard hidden dim
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Fix partitioning issue in dact primitive causing NaN and add better shape checks before calling TE API
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Move dact shape assertion from cpp to python
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

* Update test_helper.py and add QuantizeConfig class for CurrentScaling
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* WIP distributed current scaling
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Distributed Current Scaling (debugging). Distributed implementation with replicated scale_inv works for layernorm_mlp but feels like a hack

Has different per-device scale_inv values, but jax.debug.print only shows one of them. Since we're telling JAX/XLA that this scale is replicated, I think it assumes all the values are equal. However, it doesn't actually check this, so it seems we are able to get away with per-device scales for current scaling but I am not sure how stable this will be and may randomly fail if us or the user changes partitioning at all or if XLA decides to actually act on the assumption that all these scale_invs are the same.
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Implement distributed current scaling by computing a global amax and scale before quantization
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Add encoder and mnist tests for current scaling
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Add primitive prefix to shardy unique_vars to prevent factor conflicts when performing unfused primitives for current scaling
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Remove scale_shape primitive arg that is no longer used
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

* Fix expected result on multiprocessing encoder test
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Lint fix
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Update multiprocessing current scaling tolerances
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Uncomment test case that was disabled for testing
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Remove commented out debug line
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

Introduce nvte_memset to provide a fill kernel that is faster than cudaMemsetAsync for small sizes (#1716)

* nvte_memset fills single float value
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Support larger sizes than a single value and add tests
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

* [JAX-Q] Single GPU current scaling for JAX
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Fix scale check dtype for MXFP8 scales affecting tests using assert_bitwise_scaled_tensors
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Address comments
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Remove cast to fp32 for norm primitives now that zero-centered gamma dtype issue is fixed
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

* Remove unnecessary cast to fp32
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

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

* Move jaxx cuda kernels to core
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

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



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

* New GroupedGemmPrimitive using variadic args

* Remove squeeze() to reduce D2D memcpy

* Revert to the list append fashion to simplify code

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

* scaling enum abstract

* rm NVTE_ from ScalingMode names

* rework scaling mode enum in grouped gemm

* fix norm sharding

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

* rm no scaling enum
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

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

* rename QuantizeAxis to QuantizeLayout, get_layout to get_data_layout, q_axis to q_layout

* add fatten_axis option

* added gated act to test encoder

* sharding constraint fixes

* fix padding when flattening first dim needs to be padded

* update test sizes so that padding is tested

* rm output sharding as it can be done in the flax module

* sharding scale_inv for mxfp8

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

* refactor + mxfp8

* added grouped gemm

* rename linear to dense

* added cublas init phase for groupedGemm

* relax the tol of test encoder multiprocessing mxfp8 by 0.001
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: Jeremy Berchtold <jberchtold@nvidia.com>

* add paged attention; test_kv_cache_accuray and test_paged_attn pass
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove unnecessary change from last commit
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* test_fused_attn pass
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



* remove unnecessary import in test_numerics
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

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

* add to L0 test
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 license for test_paged_attn
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* fix build issue from previous merge
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



* WIP: minor fix/preparation for inference/cuda graph
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: non-paged
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: non-paged, bshd/sbhd
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: non-paged, thd, no CG
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* WIP: non-paged, thd, CG
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: non-paged, CG
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: non-paged, using paged kernel
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: restructure kernels
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: paged, CG
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: padding + BRCM
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: restructure IP, clean up
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: fix non-CG, fused
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: fix last commit
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: unfused, non-CG
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: flash-attn, non-CG
Signed-off-by: Charlene Yang <charleney@nvidia.com>

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

* commit two files missed by bcef6b34
Signed-off-by: Charlene Yang <charleney@nvidia.com>

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

* WIP: fix last commit
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: fix 1c31b68d
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: add bshd_2sbhd, sbhd_2bshd
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



* WIP: some cleanup
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: all qkv_format combinations and merge CM files
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



* WIP: some lint fixes
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: add docstring for IP
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* fix sequences_pre
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



* WIP: minor fixes for multi-layer
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: initial multi-layer test
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* WIP: minor clean up
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



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

* WIP: switch to flash_attn_varlen_func
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: fix unfused for separate q/kv format
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: fix fused for separate q/kv formats
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: flash attn + TELayer + 2 layers
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: unfused + TL + 2layers
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: all modules/backend
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: minor cleanup
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



* WIP: FlashAttention on Hopper with 2.7.3
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* WIP: FlashAttention + v3 from 39e7179
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* WIP: FlashAttention + v3 + FP8 + WIP
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

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

* WIP: separate use_flash_attention_2 and _3
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* WIP: tweaks to paged attn script
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



* WIP: enable/disable certain cases for fused attn
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: small fixes for lint and cg
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: minor fixes for attn/infer
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: fix CP
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* WIP: readd page info to FADescriptor_v1
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 to test_numerics.py
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* fix 9.5/9.7 sq/skv + mask logic
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* clean up
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 fix for FA3
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* more minor fixes for FA3
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* test page_size=1 for FA3
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix t3hd/th3d strides
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix ckpt recompute and fa3 k_scale
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* raise dynamo recompile limit for test
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove thunder test from L0
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 FA selection logic
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix FA3 q_descale shape
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



* remove page_table from IP.step() returns
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix FP8 FlashAttn DPA fp8_dpa tests
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 CP
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



* minor tweaks
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 FA3 note and L3 test
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

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

* adopt new FA3 APIs from FA2.7.3+/hopper for CP and non-CP
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix lint
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



* relax tols for TransformerLayers
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 merge
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix merge 2
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* relax tols for Ampere
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix fa3 version and reduce messaging
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 FA3 to its latest commit on main
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



* add default values to IP and assertion to graph.py
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* use custom_cache_manager instead of cache_manager
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

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



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

* Expose only required symbols from libtransformer_engine.so during linking for pytorch
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Augment libtransformer_engine.version for jax compatibility
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Augment the libtransformer_engine.version to ensure compatibility with CPP tests
Remove getenv from the .version file
Combine system.cpp and system.h
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

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

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



* Nit: Remove commented code for not including common.h
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Replace explicit getenv instantiations with a helper template
Use filesystem calls in file_exists()
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

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

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



* Revert comment to falsy instead of false
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: Kshitij Lakhani <33047503+KshitijLakhani@users.noreply.github.com>

---------
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>
Signed-off-by: Kshitij Lakhani <33047503+KshitijLakhani@users.noreply.github.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>

* Support THD + ring attention for self attn
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Consolidate reorder strategy
Signed-off-by: Reese Wang <rewang@nvidia.com>

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

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

* Use AttnBiasType, AttnMaskType, QKVLayout in cpp_extension
Signed-off-by: Reese Wang <rewang@nvidia.com>

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

* Refine P2P helper check_supported
Signed-off-by: Reese Wang <rewang@nvidia.com>

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

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

* Add dual chunk swap example
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Align different reorder code structure
Signed-off-by: Reese Wang <rewang@nvidia.com>

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

* softmax custom calls with correct encapsulates

* rm jax deprecated features

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

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

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

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

fix an int conversion error
Signed-off-by: Jennifer Zhou <jennifer@jezh.me>

* FFI for all softmax functions
Signed-off-by: Hua Huang <huah@nvidia.com>

* FFI for FusedAttnBackward and Dequantize

FusedAttnBackward passed all testes in test_fused_attn.py.
Dequantize is not used currently; finish it for completeness.
Signed-off-by: Hua Huang <huah@nvidia.com>

* Fix FusedAttnBackward FFI pybind & simplify
Signed-off-by: Hua Huang <huah@nvidia.com>

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

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



* Revert changes to tests/jax/test_fused_attn.py
Signed-off-by: Hua Huang <huah@nvidia.com>

---------
Signed-off-by: Hua Huang <huah@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Phuong Nguyen <36155692+phu0ngng@users.noreply.github.com>

* added prepare phase for the FusedAttnForwardFFI

* enabled FusedAttnForwardFFI by default

* moved prepare phase into pybind

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