There appears to be no reason why it wouldn't be possible to use the current support for tall-skinny array chunking in da.linalg.svd to also support short-fat array chunking.

Transposing inputs and outputs also appears to be a simpler solution than updating the sfqr implementation (see @mrocklin's comments at dask/dask-ml#731 (comment)).

While there are likely some other use cases, this came up in the context of PCA and it looks like some minor changes will also be necessary to support sign determinacy. Having a full_matrices argument (#3576) would be a good solution for dask/dask-ml#732 (svd_flip in dask-ml is broken without it).