np.array_api coverage

[ev-br]

Given the size of the numpy API, let's break its coverage into stages. The first reasonable milestone would be to have coverage of array_api analogs, modulo renames (acos vs arccos etc) and deliberate differences (unique etc). We will match the numpy namespace API, not array_api API, so this is just a rough guide for development.

Here's a dump of the set difference of the np.array_api namespace (numpy version 1.23.5) and the current torch_np namespace.

• all
• any
• arange
• argmin
• argsort
• can_cast
• expand_dims
• finfo
• flip
• from_dlpack
• iinfo
• max
• mean
• meshgrid
• min
• nonzero
• permute_dims
• result_type
• roll
• round
• sort
• stack
• std
• sum
• tensordot
• tril
• triu
• unique_all
• unique_counts
• unique_inverse
• unique_values
• var
• vecdot
• where

Renames:

• abs
• acos
• acosh
• atan
• atan2
• atanh
• asin
• asinh
• pow
• concat -- np.concatenate
• astype -- not a top-level numpy function
• bitwise_invert
• bitwise_left_shift
• bitwise_right_shift
• matrix_transpose

[ev-br]

With gh-65, this can finally be declared done, I think.

The main omission is from_dlpack , which seems to be a matter of forwarding to/from pytorch equivalent if needed.