Add full mlx blockwise support#2233
Open
jessegrabowski wants to merge 1 commit into
Open
Conversation
jessegrabowski
requested review from
cetagostini,
drbenvincent,
ricardoV94 and
williambdean
ricardoV94
reviewed
Jun 16, 2026
|
|
||
|
|
||
| # Equivalent blockwise to matmul but with dumb signature | ||
| odd_matmul = Blockwise(Dot(), signature="(i00,i01),(i10,i11)->(o00,o01)") |
Member
There was a problem hiding this comment.
This had a purpose, it's the default signature of a a fallback Blockwise
| rng = np.random.default_rng(42) | ||
|
|
||
| # Create a blockwise matmul with no batch dimensions (core operation only) | ||
| x = pt.matrix("x") |
Member
There was a problem hiding this comment.
Looking at it I don't think we have a single test with non static shapes?
| for arg, batch_shape in zip(args, batch_shapes): | ||
| padded = (1,) * (batch_ndim - len(batch_shape)) + batch_shape | ||
| rev_filled.append(padded[::-1]) | ||
| squeeze_axes = tuple(i for i, s in enumerate(batch_shape) if s == 1) |
Member
There was a problem hiding this comment.
for consistency Blockwise broadcasting should depend on static shapes, or it will be allowed in some backends but not others and this is already one big source of confusion in PyTensor.
That means you know ahead of time what are None or 0 vmapped axis and the dispatch should only verify them, not reinfer
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
MLX blockwise operations fail in non-trivial cases because our implementation isn't a full vectorize, it's just a single naive application of
mlx.vmap. We need to align and broadcast all input shapes, then applyvmaponce per non-core dimension. This is whatjnp.vectorizedoes under the hood. I just went ahead and re-implementedjnp.vectorizein MLX, giving us full vectorize support.Closes #2092 , and also address additional unreported cases (e.g. when data has batch dim).