Alltoall on DGX2 returns infeasible

[SivaKesava1]

Hi,

I am trying to generate schedules for DGX2 for Alltoall collective following the README.
When I run the command, taccl solve DGX2 Alltoall --topology-file ../taccl/examples/topo/topo-dgx2-1MB.json --sketch-file ../taccl/examples/sketch/sk1-dgx2-n2.json, the tool raises value error saying the model is infeasible. The same command with Allgather run quickly and produces the schedule. Is there some changes to be made to the topology or sketch for Alltoall? The paper mentions that same sketch was used for both collectives.

Thanks
Siva

[sabujlaskar]

Hi @SivaKesava1, were you able to generate the schedule for the 2-node DGX2 system?

[SivaKesava1]

Hello @sabuj7177,

When we ran experiments comparing it with our work, we contacted the authors to get the exact version they used for their paper results, and it worked fine. I recommend reaching out to them directly for the best results.