Add BSR subclass +torch.compile and clean up superblock #680
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Outdated suggestions cannot be applied.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.
This PR adds in torch.compile support for block sparsity.
In a custom op, we create the
sprase_bsr_tensor
from the explicitcrow_indices, col_indices, values
tensors that are passed in to the custom op.I also created a tensor subclass which holds these same values.
At dispatch, when we see a
torch.nn.functional.linear
call, we dispatch into our custom optorch.ops.blocksparse.linear
, using the tensors stored in the subclass.This will allow us to add a public API similar to
semi_sparse_weight()
, which I plan to do in a future PR.This PR also cleans up the superblock prototype implementation, as there was a lot of repeated code, and also adds in kernel tuning for BSR.
For bfloat16 I see the following numbers, for a 1.23x gain: