Add activation sparsity (24 + fp8 dynamic quant) subclass #2213
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Summary: This PR adds in a kernel to accelerate $$xW^T$$ when x is sparse and we are memory bound. The idea here is that we can avoid loading the columns of $W$ that correspond to the zero elements of $x$. This lets us accelerate activation sparsity for bs=1 decode use cases. Test Plan: Reviewers: Subscribers: Tasks: Tags:
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/2213
Note: Links to docs will display an error until the docs builds have been completed. ❌ 6 New FailuresAs of commit 61aedfd with merge base 1017c7e ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
jcaip
changed the title
jcaip
changed the title
jerryzh168
reviewed
May 30, 2025
| ) | ||
|
|
||
| aten = torch.ops.aten | ||
|
|
||
|
|
||
| def _pad_dense_input(dense_input: torch.Tensor) -> torch.Tensor: |
There was a problem hiding this comment.
I think ideally we don't do padding at this level, since this interferes with aliasing semantics, and make some op implementation impossible, like slice. is it possible to move this to the kernel itself? or improve the kernel to be able to handle the inputs without padding?
jerryzh168
reviewed
May 30, 2025
| @@ -66,11 +97,12 @@ def __new__( | |||
| ) | |||
| kwargs["dtype"] = sparse.dtype | |||
| kwargs["requires_grad"] = False | |||
| shape = (sparse.shape[0], 2 * sparse.shape[-1]) | |||
| # shape = (sparse.shape[0], 2 * sparse.shape[-1]) | |||
jerryzh168
reviewed
May 30, 2025
| @@ -80,6 +112,7 @@ def __init__( | |||
| self.meta = meta | |||
| self.scale = scale | |||
| self._layout = _layout | |||
| self._shape = shape | |||
There was a problem hiding this comment.
why do we need a separate shape, instead of deriving from the sparse tensor?
There was a problem hiding this comment.
nit: also sparse naming might be a bit vague, renaming to sparse_data or something might be better
jerryzh168
reviewed
May 30, 2025
| from torchao.dtypes.floatx import Float8Layout | ||
|
|
||
| res = ( | ||
| isinstance(input_tensor, AffineQuantizedTensor) |
There was a problem hiding this comment.
just FYI, I feel we could simplify this list of checks if don't need to support a variety of kernels that all use the same subclassed tensor, it only needs to be specific enough so we can dispatch to the correct kernel. maybe this can be simplified if we split the AQT in the future.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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 adds in the following things:
Float8DynamicSemiSparseActivationFloat8WeightConfig, that works withLinearActivationQuantizedTensor.CutlassSemiSparseLayoutto support both weight and activation sparsity and adds a new impl + check.Test Plan:
pytest test/sparsity/test_activation.pyReviewers:
Subscribers:
Tasks:
Tags: