What happened:
apply_gufunc returns a tuple with the wrong number of outputs for multi-output signatures.

What you expected to happen:
apply_gufunc would return the same number of items as the signature implied.

Minimal Complete Verifiable Example:

import dask.array as da
from dask.array import apply_gufunc
import numpy

>>> len(apply_gufunc(lambda a, b: (a.max(axis=1), b.max()), '(i,I),(I)->(i),()', da.ones((10, 10), chunks=-1), da.ones(10, chunks=-1), output_dtypes=['float64', 'float64']))
1

This should return a tuple of 2 items (one array and one scalar), as the signature implies.

The issue is that meta_from_array encounters a "zero-size array to reduction operation" error. This error used to propagate (which would have taken us down a more useful codepath in apply_gufunc), but as of recently (#6736) it just selects the first element from args_meta:

@pentschev I wonder if this meta = args_meta[0] should be conditioned on len(args_meta) = 1? When there are multiple arguments, I think the only sensible behavior is to return None, since we don't know how the arguments will be combined. Even when there's only a single argument, I personally thing it's incorrect to assume in generality that the output type will be the same as the input type; an arbitrary user-defined function could do all sorts of other things before/after calling np.min.

Even if meta_from_array had raised an error, this still wouldn't have worked. This logic for generating meta from the output_dtypes

Overall, there is inconsistency throughout apply_gufunc as to whether nout, the type+length of output_dtypes, or the type+length of metas is the source of truth for how many items to return.

Anything else we need to know?:

Environment:

• Dask version: ac1bd05
• Python version: 3.8.8
• Operating System: macOS
• Install method (conda, pip, source): source