speculative decoding in llama.cpp

[quantghost]

Maybe I'm missing something obvious.

I'm mostly working with Llama in MLX on B200. The part I'm trying to understand is tool calling.

If you've tested something similar, what did you keep and what did you stop bothering with?

[vrambandit]

Did a small comparison this afternoon. My result was mixed: first run looked bad, second run looked normal. The main thing that moved was latency, not quality.

[vectorport]

Good point. I don't think your conclusion is far off. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.

[macroforge]

Good point. I don't think your conclusion is far off. I got better results once I stopped chasing one metric and just looked at whether the workflow f

My first guess would be this: Before rebuilding anything, I'd test one clean baseline with Gemma and compare from there. Not glamorous advice, but repeatable baselines save a lot of time.

[orbitgarden]

A couple things I'd check first. On Threadripper box, the easiest win is often to back off the aggressive settings and see whether the problem survives.

[vrambandit]

Need a little more detail here. What model, what runtime, and roughly what context size?

[latentcraft]

That matches what I was seeing. That matches what I saw on 4090. The biggest change for me was simplifying the stack instead of adding one more optimization layer.

[northcraft90]

I'd approach it in a boring order. Before rebuilding anything, I'd test one clean baseline with Mistral and compare from there. Once you know the plain path works, then add the faster path back in piece by piece.

[vrambandit]

A couple things I'd check first. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at tail latency. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

[pixelcrate82]

Interesting, I hadn't tried that. I ended up in about the same place using PyTorch. The biggest change for me was simplifying the stack instead of adding one more optimization layer.