Closed davorchap closed 1 month ago
We support 1D row-major tensors. Tilized tensors have to be 2D
We support 1D row-major tensors. Tilized tensors have to be 2D
Thanks @arakhmati
Perhaps the original shape after tilizing+padding would still be 1D? We change the layout but the original would still be 1D
@razorback3 fyi
What our suggestion is: 1D row-major tensor after padding and tilizing -> 2D tensor But, if we call _shape_withoutpadding() it returns 1D tensor
@davorchap Would you please share any update regarding @razorback3's comment?
@davorchap Would you please share any update regarding @razorback3's comment?
Yes, 1D original tensor shape will be preserved.
@nsmithtt , @arakhmati and team are working through a proposal to uplift the Tensor class to support this cleanly
@arakhmati @eyonland @yan-zaretskiy for vis for Tensor layout conversation
We can't change the dimensionality of the shape when calling shape_without_padding
because it will be misleading in many other cases. And it really doesn't make sense to get tilized tensor's shape as 1D. I think Moreh should add a free function that does what's required and use it in appropriate places. This can't be a part of our infra as of right now
User should be able to specify a 1D tensor shape , such as
(10)
, today it has to be specified as 2D(1,10)