Playbook for 2 node RPC Clustering

[abdmalik-amd]

add playbook covering llama.cpp RPC-based distributed inference across two STX Halo systems, including VRAM configuration, build instructions, and model deployment.

update page.tsx to support h3 headers and in page section navigation

[eddierichter-amd]

This looks great @abdmalik-amd! I just had a couple small comments.

[danielholanda]

@eddierichter-amd were you able to actually reproduce the results here?

[eddierichter-amd]

@eddierichter-amd were you able to actually reproduce the results here?

You mean performance results? I don't have a 4-Strix Halo machine but I do infact have a 2-Strix Halo setup and following the same steps functionally works.

[adamlam2-amd]

Nice, I'm glad the functionality works. Some more minor UX considerations:

1. maybe specify what exact memory values are needed to run the model we are working with. Can mention to set the variable graphics memory to ~75% and also the TTM values to there to load larger models
2. i think rpc does not improve compute, but rather just allows for model offloading on the combined vram
3. im not too familiar with building llama.cpp from source, but ensure they work. What is Ninja for windows? is that installed. Do we need build-essential for Linux?
4. some windows v linux instructions - .\rpc-server.exe vs ./rpc-server
5. any images/screenshots/gifs you can add would be helpful!

Other than that, pretty good and should be able to pass onto QA

[abdmalik-amd]

Nice, I'm glad the functionality works. Some more minor UX considerations:

1. maybe specify what exact memory values are needed to run the model we are working with. Can mention to set the variable graphics memory to ~75% and also the TTM values to there to load larger models
2. i think rpc does not improve compute, but rather just allows for model offloading on the combined vram
3. im not too familiar with building llama.cpp from source, but ensure they work. What is Ninja for windows? is that installed. Do we need build-essential for Linux?
4. some windows v linux instructions - .\rpc-server.exe vs ./rpc-server
5. any images/screenshots/gifs you can add would be helpful!

Other than that, pretty good and should be able to pass onto QA

1. Sure I will update the memory text, the current memory text is using the memory macro
2. Correct
3. Ninja is included in VS Code Build Tools
4. Will fix this typo for the windows section
5. Sure I will look into adding some pictures/gifs of usage

[abdmalik-amd]

@eddierichter-amd @adamlam2-amd

Images for llama-cli & llama-server interface have been added to the playbook

[danielholanda]

@adamlam2-amd @eddierichter-amd Any additional comments or requirements before we merge this?

[adamlam2-amd]

lgtm - one thing we should keep note of is to differentiate between RPC clustering and RCCL clustering so that users know exactly which one to use.