+- Inference (https://www.datasculpting.ai/inference)
+-- Forum: Evaluation & Testing (https://www.datasculpting.ai/inference/forumdisplay.php?fid=16)
+--- Forum: Evals (https://www.datasculpting.ai/inference/forumdisplay.php?fid=66)
+--- Thread: benchmarks that map to real use (/showthread.php?tid=899)
Pages:
1
2
benchmarks that map to real use - routeline23 - 10-08-2026
Thought I'd post a small recipe because a lot of threads ask similar things.
Maybe useful as a starting point: Ollama on 7900 XTX, keeping reranking in mind from the start.
Maybe useful as a starting point for newer folks.
I care more about stability than squeezing the last possible tokens/sec out of it.
Re: benchmarks that map to real use - kernelhouse70 - 10-09-2026
Appreciate the detail. That matches what I saw on GB200. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.
Re: benchmarks that map to real use - mlxgarden - 10-09-2026
Thanks, that helps. I don't think your conclusion is far off.
Re: benchmarks that map to real use - promptfield - 10-10-2026
Good point. I ended up in about the same place using vLLM. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.
Re: benchmarks that map to real use - vramforge - 10-11-2026
I'd start with the simpler stack.
Re: benchmarks that map to real use - vrambandit - 10-11-2026
Thanks, that helps. That matches what I saw on M3 Max. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.
Re: benchmarks that map to real use - torchstation - 10-11-2026
My first guess would be this: Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at INT4. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.
Re: benchmarks that map to real use - vrambandit - 10-11-2026
Quote:My first guess would be this: Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at INT4. A lot of these threads
subscribed, curious where this lands.
Re: benchmarks that map to real use - cacheline - 10-11-2026
worth checking the driver first.
Re: benchmarks that map to real use - orbitline - 10-11-2026
What worked for me was pretty simple. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at tail latency. Once you know the plain path works, then add the faster path back in piece by piece.