NVIDIA driver regression with vLLM

[audiopath66]

Spent some time on this today and wanted to compare notes.

Two paths both seem viable, but they break down differently once prefix caching becomes important.

Trying to avoid rebuilding the whole thing twice.

[mintforge]

A couple things I'd check first. Before rebuilding anything, I'd test one clean baseline with Mixtral and compare from there.

[orbitbench6]

A couple questions before I guess too hard. Are you testing this on B200 or something else?

[paperport64]

A couple questions before I guess too hard. Are you testing this on B200 or something else?

My first guess would be this: Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at tokens/sec. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

[phaseport36]

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. Not glamorous advice, but repeatable baselines save a lot of time.

[dockember]

I don't think that's the whole story. The reported numbers can look compelling, but in ordinary usage I didn't see the same benefit.

[relaynest82]

That matches what I was seeing. That matches what I saw on mini PC. The biggest change for me was simplifying the stack instead of adding one more optimization layer.

[cobaltspark]

I tested this tonight. On A6000 with PyTorch, I saw basically the same behavior. The main thing that moved was latency, not quality.

[phaseport36]

Thanks, that helps. I ended up in about the same place using MLX. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.

[audiopath66]

I'd approach it in a boring order. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at MoE serving. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.