comparing models without fooling yourself

[kernelarch]

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.

[cacheworks]

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.

[batchsize]

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.

[benchplease]

Ran a quick check on my box. I couldn't reproduce the slowdown until I pushed context much higher.

[kernelarch]

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.

[localtinker]

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.

[batchcraft]

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.

[sparrowworks88]

following this because I'm in the same boat.

[benchplease]

following this because I'm in the same boat.

[circuitworks90]

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.