What is the expected behavior of NCCL allreduce with NaN input?

Accelerated Computing GPU-Accelerated Libraries
1

How is NCCL supposed to behave when a NaN value provided to allreduce in one or more of the ranks ?

2

I would imagine some evidence of NaN in the result.

3

That is reasonable to expect : NaN op number = NaN. Would not crash, hang , or do other nasty stuff ?

4

I’m not aware of any such expectation.