[Question]: Why does GIN require both crossNicSupported and !globalNicFused conditions?

[yesho-top]

Question

Question

I'm trying to understand the requirements for enabling GIN (GPU-Initiated Networking) support in NCCL and found the following conditions in init.cc:

comm->ginSupport = comm->sharedRes->ginState.ncclGin != nullptr 
                   && crossNicSupported 
                   && !globalNicFused;

Background
From the code analysis:

• crossNicSupported checks if NCCL_CROSS_NIC != 0 (multi-NIC topology)
• globalNicFused is set to true when any NIC has vProps.ndevs > 1 (fused/merged NICs)

Questions

1. Why is crossNicSupported required for GIN?

• Does GIN not provide benefits in single-NIC scenarios?
• s the overhead of GIN not worthwhile without multi-rail networks?

2. Why does !globalNicFused disable GIN?

• What technical limitations prevent GIN from working with fused NICs?
• Is this related to direct GPU-to-physical-NIC mapping requirements?

3. Are there plans to support GIN with fused NICs in the future?

• if use two-port, and it is not bonded, it will be considered as fused Nics, so gin cannot be used.