NVIDIA / TransformerEngine

A library for accelerating Transformer models on NVIDIA GPUs, including using 8-bit floating point (FP8) precision on Hopper and Ada GPUs, to provide better performance with lower memory utilization in both training and inference.

https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/index.html

Apache License 2.0

1.85k stars 309 forks source link

[PyTorch] Remove implicit padding and unpadding in `GroupedLinear` #984

Closed yaox12 closed 2 months ago

yaox12 commented 3 months ago

Description

This PR removes implicit padding and unpadding in GroupedLinear to:

Match other layers (Linear/LayerNormLinear/LayerNormMLP) that throw an error when dims are not supported for FP8 GEMM, instead of doing implicit padding.
Avoid redundant padding/unpadding between multiple GroupedLinear layers. Users can do the padding and unpadding manually before and after the whole module if needed.

Type of change

[ ] Documentation change (change only to the documentation, either a fix or a new content)
[ ] Bug fix (non-breaking change which fixes an issue)
[ ] New feature (non-breaking change which adds functionality)
[ ] Breaking change (fix or feature that would cause existing functionality to not work as expected)
[ ] Infra/Build change
[ ] Code refractor

Changes

Please list the changes introduced in this PR:

Remove implicit padding and unpadding in GroupedLinear

Checklist:

[ ] I have read and followed the contributing guidelines
[ ] The functionality is complete
[ ] I have commented my code, particularly in hard-to-understand areas
[ ] I have made corresponding changes to the documentation
[ ] My changes generate no new warnings
[ ] I have added tests that prove my fix is effective or that my feature works
[ ] New and existing unit tests pass locally with my changes

yaox12 commented 3 months ago

@phu0ngng Could you please take a look? This should be a minor fix. Thank you very much.

phu0ngng commented 3 months ago

/te-ci pytorch