XLA's Dot should follow broadcast semantics from np.matmul, not np.dot #5523
Assignees
Labels
type:feature
Feature requests
Comments
|
@cwhipkey, @andydavis1 and @prb12 , could you please comment on this? Thanks. |
|
@shoyer thanks for the suggestion -- we'll think about it |
|
Based on this input and other considerations, we've decided to restrict the semantics of XLA's Dot operation to 1D and 2D arrays in the initial release. We may consider expanding it to higher dimensions in the future, and at that point we'll be considering different possible semantics. For the time being, however, this issue can be closed. Thanks for the feedback! |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
I notice that the XLA Dot operation copies "outer-product style" broadcast semantics from
numpy.dot:dotarray [s x r x t]In brief, I think this is a mistake. It would be better to follow the "matmul style" style broadcasting semantics of Python's
@operation and NumPy'smatmul.matmul's broadcasting is much more general, and in my opinion, also easier to understand. For example, it can do batch matrix-multiplication, but also can still do outer product style broadcasting if you insert dummy dimensions of length 1 (the axes do end up in a different order), e.g.,
batch matmul: [p x q x r]
matmul[p x r x t] -> [p x q x t]outer product matmul: [p x 1 x q x r]
matmul[1 x s x r x t] -> [p x s x q x t]If we could go back in time as NumPy developers, we assuredly would change
dotto work this way (now we cannot, because of backwards compatibility concerns). So it would be nice to change this for XLA before we lock in this behavior.The text was updated successfully, but these errors were encountered: