gaopengff
gaopengff
In your CI, did you run parallel test cases in multi gpu scenario? May I know the launcher rank count of parallel test(2 or more)?
I got it, thanks.
Our developer team will work on it to provide more specific error message.
Now torch-ccl only support one rank do sending and another do receiving at the same time, if you change the code to ```python if my_rank == 0: o1 = dist.isend(A,1-my_rank)...
This issue does not exist. I'll close it.