+- 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: prefix caching in llama.cpp on 4090 (/showthread.php?tid=1257)
prefix caching in llama.cpp on 4090 - bytesupply76 - 10-15-2026
Did a bit of testing and thought I'd post the result here.
After a few runs on Threadripper box, I ended up with decent performance, but the setup only really made sense once I adjusted around prefix caching.
Curious if anyone else landed in the same place.
Re: prefix caching in llama.cpp on 4090 - vrambandit - 10-15-2026
A couple things I'd check first. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at tool calling. Not glamorous advice, but repeatable baselines save a lot of time. Turns out 'just one more test' was a lie.
Re: prefix caching in llama.cpp on 4090 - vectorport - 10-15-2026
What worked for me was pretty simple. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at reranking. Not glamorous advice, but repeatable baselines save a lot of time.
Re: prefix caching in llama.cpp on 4090 - vrambandit - 10-15-2026
My first guess would be this: Before rebuilding anything, I'd test one clean baseline with Qwen and compare from there. Once you know the plain path works, then add the faster path back in piece by piece.
Re: prefix caching in llama.cpp on 4090 - bytesupply76 - 10-16-2026
I don't think that's the whole story. I've seen people blame router policy when the real issue was much more mundane. Not saying you're wrong, just that I wouldn't lock that in as the lesson yet.
Re: prefix caching in llama.cpp on 4090 - deltastation - 10-16-2026
I'd approach it in a boring order. On 7900 XTX, the easiest win is often to back off the aggressive settings and see whether the problem survives.
Re: prefix caching in llama.cpp on 4090 - vrambandit - 10-16-2026
Thanks, that helps. I ended up in about the same place using llama.cpp. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.