Open ipdemes opened 2 years ago
Yes, unfortunately the tensor/BLAS libraries we use (tblis, cutensor, openblas, cublas) don't support integer datatypes. We would have to write custom implementations for integers.
I see, so this is a known issue. Should we then change the error message to "int 64 type is not currently supported by CuNumeric contract method used in dot/multiply" ?
Sure, please make a PR with your suggestion. Note that a number of functions go through module.py:_contract
(e.g. einsum
, tensordot
), so you wouldn't want the error message there to refer to specifically dot/multiply
.
Following test doesn't work with cuNumeric: a = [[1, 0], [0, 1]] b = [[4, 1], [2, 2]] c=np.dot(a, b)
It produces error:
It seems like
supported_dtypes
do not include integer types.