The array API specification stipulates the following operators that are currently not supported by PyTorch:

• __dlpack__
• __dlpack_device__
• __rand__ (Support __rand__, __ror__ and __rxor__ #58120)
• __rlshift__ (Support __rlshift__ and __rrshift__ #58121, Support torch.bitwise_{left/right}_shift and __rlshift__, __rrshift__ #59544)
• __imatmul__ - not doing __imatmul__ because matmul doesn't always preserve shape, and still has a resize bug (see Micro-optimisations for matmul #64387)
• __rmatmul__ (Some torch.Tensor operators don’t return NotImplemented for invalid inputs #57719, Fix some tensor operators to return NotImplemented for invalid inputs #58216)
• __rmod__ (Support __rmod__ #58035, Support __rmod__ #58476)
• __ror__ (Support __rand__, __ror__ and __rxor__ #58120)
• __rrshift__ (Support __rlshift__ and __rrshift__ #58121, Support torch.bitwise_{left/right}_shift and __rlshift__, __rrshift__ #59544)
• __rxor__ (Support __rand__, __ror__ and __rxor__ #58120)
• from_dlpack (implemented, but not exposed yet in the torch namespace in main namespace in 1.10.0)
• broadcast_arrays
• bitwise_left_shift (Support torch.bitwise_{left/right}_shift and __rlshift__, __rrshift__ #59544)
• bitwise_invert
• bitwise_right_shift (Support torch.bitwise_{left/right}_shift and __rlshift__, __rrshift__ #59544)
• vecdot ([Array API] Add linalg.vecdot #70542)
• concat (we should be sure torch.cat is compliant and alias concat to it) (Support concat alias to cat #61767)
• expand_dims (Support expand_dims #57116)
• __ipow__ (Connect Tensor.__ipow__ to pow_ method #76900)
• to_device
• astype
• unique_(all|counts|inverse|values) unique should be split into four partial functions #70920
• linalg.diagonal [Array API] Add linalg.diagonal #70599
• linalg.matrix_transpose, matrix_transpose
• linalg.outer Support torch.linalg.outer #63293
• linalg.tensordot Create linalg.tensordot #63478
• linalg.trace Support torch.linalg.trace #62714
• linalg.vecdot [Array API] Add linalg.vecdot #70542

cc @mruberry @rgommers @pmeier