+- Inference (https://www.datasculpting.ai/inference)
+-- Forum: Evaluation & Testing (https://www.datasculpting.ai/inference/forumdisplay.php?fid=16)
+--- Forum: Benchmarks (https://www.datasculpting.ai/inference/forumdisplay.php?fid=65)
+--- Thread: comparing models without fooling yourself (/showthread.php?tid=986)
Pages:
1
2
comparing models without fooling yourself - kernelarch - 10-10-2026
Thought I'd post a small recipe because a lot of threads ask similar things.
Maybe useful as a starting point: vLLM on M3 Max, keeping INT4 in mind from the start.
Feel free to tear it apart if you see something dumb.
Re: comparing models without fooling yourself - cacheworks - 10-10-2026
What worked for me was pretty simple. Before rebuilding anything, I'd test one clean baseline with DeepSeek and compare from there. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.
Re: comparing models without fooling yourself - batchsize - 10-11-2026
Quote:What worked for me was pretty simple. Before rebuilding anything, I'd test one clean baseline with DeepSeek and compare from there. A lot of these thr
I'd approach it in a boring order. Before rebuilding anything, I'd test one clean baseline with DeepSeek and compare from there. Once you know the plain path works, then add the faster path back in piece by piece.
Re: comparing models without fooling yourself - benchplease - 10-13-2026
Ran a quick check on my box. I couldn't reproduce the slowdown until I pushed context much higher.
Re: comparing models without fooling yourself - kernelarch - 10-14-2026
What worked for me was pretty simple. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at continuous batching.
Re: comparing models without fooling yourself - localtinker - 10-14-2026
Quote:What worked for me was pretty simple. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at continuous batching.
Ran a quick check on my box. On 5090 with PyTorch, I saw basically the same behavior. The main thing that moved was latency, not quality.
Re: comparing models without fooling yourself - batchcraft - 10-16-2026
My first guess would be this: On 4090, the easiest win is often to back off the aggressive settings and see whether the problem survives. Once you know the plain path works, then add the faster path back in piece by piece.
Re: comparing models without fooling yourself - sparrowworks88 - 10-16-2026
following this because I'm in the same boat.
Re: comparing models without fooling yourself - benchplease - 10-17-2026
following this because I'm in the same boat.
Re: comparing models without fooling yourself - circuitworks90 - 10-18-2026
Thanks, that helps. The pattern around structured output has been pretty similar for me. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.