Open nandeeka opened 1 week ago
Thanks for filing the issue! We're looking into it actively and will get back with an answer for you shortly.
Hi Nandeeka! This requires us to implement the support for SundaMax8
in NKI, which we have added to our backlog of active tasks. We'll let you know when it's available.
@aws-qieqingy is there anything @nandeeka can do to bypass this limitation for short term ?
Also a note: NKICodegen is currently experimental and will come across unimplemented nki instructions until we've completed expressing each instruction in nki isa. Whenever this class of error occurs, it is most likely for this reason.
@AWSNB Unfortunately, we don't have any instruction that achieves the same functionality exposed through NKI right now.
I was curious about how the
torch.topk()
function is implemented in the Neuron compiler. After writing a kernel with this call and compiling it to NKI, the resulting code looks something like:I understand that the
NkiCodegen.codegenSundaMax8
andNkiCodegen.codegenSundaMaxIndex8
functions need to be black boxes because they require data-dependent control flow, which is currently not supported. My question is, is it possible for me to call these functions within my own NKI kernel?Thanks!