rackmount vs tower for homelab inference: worth it?

[homelabhenry]

I'd approach it in a boring order. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at long context. Once you know the plain path works, then add the faster path back in piece by piece.

[tensorworks]

A couple things I'd check first. If you're on PyTorch, I'd confirm the version first and then reduce variables until the behavior is obvious. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

[batchwire]

I tested this tonight. My result was mixed: first run looked bad, second run looked normal.

[macroforge]

I'd start with the simpler stack.

[kernelhouse70]

Good point. I ended up in about the same place using llama.cpp. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.

[homelabhenry]

Thanks, that helps. I don't think your conclusion is far off. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.