[known issue] kthvalue not available on CUDA

[vadimkantorov]

torch.arange(1, 10).cuda().kthvalue(5) will throw

A known issue (https://discuss.pytorch.org/t/missing-kthvalue-on-gpu/1798/2)

[apaszke]

A workaround for now would be to use topk. It should be very fast.

def kthvalue(tensor, k, dim=None):
    val, idx = tensor.topk(k, dim)
    return val[:1], idx[:1]

[vadimkantorov]

kthvalue returns k'th smallest element, while topk returns by default largest elements. So a workaround should be:

 # dim = None wouldn't work because topk requires a dim argument if largest = False is used
def kthvalue(tensor, k, dim=-1):
    val, idx = tensor.topk(k, largest = False, dim = dim)
    return val[-1:], idx[-1:]

[AnishShah]

Hi, can I work on this?

[soumith]

sure thing @AnishShah !

[vadimkantorov]

fix worked on in #8406

[vadimkantorov]

Fixed in #17544