Closed starkhu closed 1 year ago
Thanks! But this is not a bug, because indices1_s[i] * capacity + locations1_s[i]
is always unique, so that =
and +=
actually do the same, you can change CPU's +=
to =
, although it doesn't make any differences in producing results.
The tutel/fasst_dispatch.py file calls the function
func_fwd
, but the cpu and cuda implementations of this function are inconsistent. Here is the code implementation of this function on cpu and cuda:The dispatched_input is computed differently on the two implementations. On the cpu, dispatched_input += gates1_sreshaped_input, and on cuda, dispatched_input=gates1_sreshaped_input.