+- Inference (https://www.datasculpting.ai/inference)
+-- Forum: Inference Engines (https://www.datasculpting.ai/inference/forumdisplay.php?fid=9)
+--- Forum: llama.cpp / GGUF (https://www.datasculpting.ai/inference/forumdisplay.php?fid=35)
+--- Thread: speculative decoding in llama.cpp (/showthread.php?tid=385)
Pages:
1
2
speculative decoding in llama.cpp - quantghost - 09-20-2026
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?
Re: speculative decoding in llama.cpp - vrambandit - 09-23-2026
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.
Re: speculative decoding in llama.cpp - vectorport - 09-24-2026
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.
Re: speculative decoding in llama.cpp - macroforge - 09-25-2026
Quote: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.
Re: speculative decoding in llama.cpp - orbitgarden - 09-28-2026
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.
Re: speculative decoding in llama.cpp - vrambandit - 09-30-2026
Need a little more detail here. What model, what runtime, and roughly what context size?
Re: speculative decoding in llama.cpp - latentcraft - 09-30-2026
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.
Re: speculative decoding in llama.cpp - northcraft90 - 09-30-2026
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.
Re: speculative decoding in llama.cpp - vrambandit - 09-30-2026
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.
Re: speculative decoding in llama.cpp - pixelcrate82 - 09-30-2026
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.