how strict do you make schemas

[plainworks]

Thought I'd post a small recipe because a lot of threads ask similar things.

This isn't revolutionary, but the combination of MLX and Threadripper box felt surprisingly usable once I tuned around MoE serving.

Maybe useful as a starting point for newer folks. The logs were somehow calmer than I was.

[quiettoken]

A couple things I'd check first. If you're on llama.cpp, 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.

[batchcraft]

Interesting, I hadn't tried that. The pattern around speculative decoding has been pretty similar for me. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.

[amberrunner]

I'd approach it in a boring order. If you're on llama.cpp, I'd confirm the version first and then reduce variables until the behavior is obvious. Not glamorous advice, but repeatable baselines save a lot of time.

[cudascope]

Thanks, that helps. That matches what I saw on GB200. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.

[quantshed]

A couple things I'd check first. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at long context.

[cudaline]

Good point. I don't think your conclusion is far off.

[deepseeked]

Need a little more detail here. Is this local-only or are you routing anything externally?

[quietsupply11]

Need a little more detail here. Is this local-only or are you routing anything externally?

Tried to reproduce it just now. I couldn't reproduce the slowdown until I pushed context much higher. The surprise was that memory pressure showed up before throughput collapsed.

I'm happy to test one more variant if you want a second data point.

[mlxgarden]

I think the bottleneck is somewhere else. I wouldn't optimize around that first. On my side the bigger swing came from the runtime and not the model.