add triton based Softmax and unit tests #475
Open
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary:
This PR is to add a Triton based softmax, which supports both fwd and bwd as well as multi-dim inputs.
This kernel outperforms the PT native softmax in the fwd pass and thus might be favored esp. for inference situations.
It's also a first opportunity to consider expanding the TorchMM offerings to include triton kernels.
Thus, I have listed this under a separate Triton directory in order to separate it from the rest of the codebase which is PyTorch base.
Test plan:
Tested it in end to end training (short run).
Added a number of unit tests which compare it directly with the torch.softmax and tests both forward and backward passes, as well as 2d and multi-dim inputs and fp32 and bf16 dtypes.