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>

* rm batch_dim, sequence_dim, sequence_parallel_output
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* rm lhs_quantized_colwise and rhs_quantized_colwise
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* rm unnecessary transpose_batch_sequence arg from some modules
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

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

[JAX] Fixing GemmPrimitive partitioning rules to handle tensor-parallelism correctly for sequence-parallel inputs (#1980)

* updated GemmPrimitive partitioning rules to explicitly control all-reduce vs. reduce-scatter for sequence-parallelism
Signed-off-by: Alp Dener <adener@nvidia.com>

* corrected handling of FSDP sharding for the RHS operand
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



* use correct logical axes variable to identify sequence-parallel dim in LayerNormDenseGeneral
Signed-off-by: Alp Dener <adener@nvidia.com>

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

* added assert on sequence-parallel options when GemmPrimitive is disabled
Signed-off-by: Alp Dener <adener@nvidia.com>

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

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



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

* 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>

* TensorUsage + FP8 GEMM with all layouts handling on BW
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>


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

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

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

* [JAX] Update flax module param initialization to support logical partitioning axes
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Fix ffn1 intermediate result being replicated
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

* Add documentation and assert when logical_axes=None
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Fix bias in LayerNormMLP flax module
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Fix layer tests to not use nn_partitioning and instead use nn.with_logical_axes
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

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

* exclude GroupedGemm APIs
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>