add --rpc-layers flag to explicitly set RPC layers

[kallewoof]

The current setup does not allow for very precise control of how many layers to put on the local GPUs vs the remote RPC connected server(s). This adds an additional --rpc-layers flag (-nrl) which allows the user to explicitly set the number of layers to offload to RPC end.

[rgerganov]

In which cases you would need this instead of the using existing --tensor-split?

[kallewoof]

In which cases you would need this instead of the using existing --tensor-split?

This lets you specify layers without having to think about proportions, which is a lot simpler than --tensor-split. I can find the right amount for the RPC server and GPUs separately.

[abc-nix]

In which cases you would need this instead of the using existing --tensor-split?

This lets you specify layers without having to think about proportions, which is a lot simpler than --tensor-split. I can find the right amount for the RPC server and GPUs separately.

You don't need to think about the proportions when using --tensor_split. You can directly use the number of layers instead. Example:

$ llama.cpp/build/bin/llama-server --rpc <IP1>:<PORT1> --rpc <IP2>:<PORT2> \
--device RPC[IP1:PORT1],CUDA0,CUDA1,RPC[IP2:PORT2] \
-ngl 33 --tensor_split 3/20/10/0 --device-draft CUDA1,RPC[IP2:PORT2] -ngld 99 [...]

33 layers are being offloaded, and they are divided between 3 devices, (3 layers for the first RPC device, 20 for the first CUDA device, and 10 for the last CUDA device).

OT: The only thing I am missing is a tensor-split for the draft model.

[kallewoof]

Got it. The name should've been "layer-split" instead. It sounds like some low level voodoo shit so I never touched it. :)
Also, the -h output does not at all indicate you can use the number of layers. You just have to know that.

[abc-nix]

There are a lot of undocumented things on llama.cpp. I learned many tricks reading responses to issues, to merge requests, and trying things out. As an example, the recent device order merge request is another magic trick that makes llama.cpp so much more flexible when using RPC.