[Feature] Using the AMD Infinity Fabric in SMChannels

[ThomasNing]

I am trying to use the SMChannels way to connect AMD GPUs together with Infinity Fabric.

In the core.hpp, here is the transport enum class:
"/// Enumerates the available transport types.
enum class Transport {
Unknown, // Unknown transport type.
CudaIpc, // CUDA IPC transport type.
Nvls, // NVLS transport type.
IB0, // InfiniBand device 0 transport type.
IB1, // InfiniBand device 1 transport type.
IB2, // InfiniBand device 2 transport type.
IB3, // InfiniBand device 3 transport type.
IB4, // InfiniBand device 4 transport type.
IB5, // InfiniBand device 5 transport type.
IB6, // InfiniBand device 6 transport type.
IB7, // InfiniBand device 7 transport type.
Ethernet, // Ethernet transport type.
NumTransports, // The number of transports.
};
"

Could I know which transport type I should select in here?

[chhwang]

SmChannels can only work on the CudaIpc transport.

[ThomasNing]

So to make the AMD Infinity Fabric route happen should I switch to ProxyChannel and select Nvls as the type? Or I should stay in the SmChannel and select type as the CudaIpc

[chhwang]

Both SmChannel and ProxyChannel will work over AMD Infinity Fabric if you use the CudaIpc transport. Nvls is specific to NVIDIA H100 or later GPUs.

[ThomasNing]

Got it! Thank you for the help Changhao!