Shall we raise ValueError in the else branch of this condition? I am not even sure how we could meaningfully trigger this in tests or in a legit usage scenario.

Maybe we should always return X.shape[1] without the if isinstance(X.shape[1], numbers.Integral) condition.

I believe this code was inspired from the _num_samples case for .shape[0] where this is not always known for dask dataframes where you can get delayed first dimensions:

>>> df.shape
(Delayed('int-1d60df59-017f-49c3-80e5-f313b06e4c1d'), 120)

Furthermore, in this case _num_samples would fallback to calling len(df) which would naturally trigger the delayed computation.

For the _num_features case, we would fallback to doing len(df[0]). In the case of a dask dataframe df[0]) would be a series for the first column and therefore len(df[0]) would return the number of samples instead of the number of features.

Finally I don't see a valid common case where this would happen for the second dimension. I think we can remove this defensive coding condition.

Just for dask dataframes, should we get the first row with the following?

    if hasattr(X, 'iloc'):
        first_sample = X.iloc[0, :]
    else:
        first_sample = X[0]

(Rename to first_sample to be a little more clear)

I don't think that's necessary: I see no reason why X.shape[1] would be Delayed in a dask-dataframe. Only the first dimension is lazily depend on the result of index based partitioning, for instance after a groupby operation.