Topology XML file

[guntrogu]

Hi,

I would like to set the NCCL_TOPO_FILE to a valid value before running all_reduce_perf script.
What is the correct location for this file, or should I generate it ?

Thanks!

[sjeaugey]

Why would you like to set this file?

The NCCL_TOPO_FILE is usually generated by the VM provider so you should not need to generate it yourself, and the VM provider should document where to find this file.

Now if you're running your own VM, then indeed you may want to dump the XML topology on the baremetal system (NCCL_TOPO_DUMP_FILE=system.xml), keep only the <cpu> and <pci> tags (remove the <gpu> and <nic>), then inject it inside the VM using NCCL_TOPO_FILE), provided the PCI IDs are the same inside the VM and on baremetal, otherwise you'll have to adjust the PCI bus IDs to match what's inside the VM.

[AddyLaddy]

No further feedback so I'll close this issue. Please open another one if you are still seeing issues.