+- Inference (https://www.datasculpting.ai/inference)
+-- Forum: Inference Engines (https://www.datasculpting.ai/inference/forumdisplay.php?fid=9)
+--- Forum: vLLM / SGLang (https://www.datasculpting.ai/inference/forumdisplay.php?fid=36)
+--- Thread: prefill/decode separation in practice (/showthread.php?tid=1126)
Pages:
1
2
prefill/decode separation in practice - thetaplane - 10-12-2026
Spent some time on this today and wanted to compare notes.
Two paths both seem viable, but they break down differently once long context becomes important.
Interested in tradeoffs more than fan loyalty.
This is mostly for local work rather than public serving. Turns out 'just one more test' was a lie.
Re: prefill/decode separation in practice - kernelgrid - 10-13-2026
A couple questions before I guess too hard. Are you testing this on B200 or something else?
Re: prefill/decode separation in practice - phaseport36 - 10-13-2026
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: prefill/decode separation in practice - cacheline - 10-13-2026
A couple things I'd check first. On dual 3090, the easiest win is often to back off the aggressive settings and see whether the problem survives. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.
Re: prefill/decode separation in practice - phasestream53 - 10-13-2026
My first guess would be this: 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: prefill/decode separation in practice - batchsize - 10-13-2026
Interesting, I hadn't tried that. That matches what I saw on A6000. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.
Re: prefill/decode separation in practice - amberrunner - 10-13-2026
Quote:Interesting, I hadn't tried that. That matches what I saw on A6000. I got better results once I stopped chasing one metric and just looked at whether
I'd approach it in a boring order. If you're on Ollama, 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.
There are a few different conversations hiding inside this thread, so being explicit about the goal helps.
Re: prefill/decode separation in practice - turboarch95 - 10-13-2026
That matches what I was seeing. The pattern around token cost has been pretty similar for me. The biggest change for me was simplifying the stack instead of adding one more optimization layer.
Re: prefill/decode separation in practice - ggufvault - 10-13-2026
Need a little more detail here. Do you see the same thing with a smaller batch or shorter prompt?
Re: prefill/decode separation in practice - kernelmike - 10-13-2026
Quote:Need a little more detail here. Do you see the same thing with a smaller batch or shorter prompt?
A couple questions before I guess too hard. Is this local-only or are you routing anything externally?