QLoRA resource reality check

[oxideworks59]

Good point. I don't think your conclusion is far off. Worth rechecking your defaults though, because one bad assumption can make the rest of the threa

I don't think that's the whole story. I've seen people blame token cost when the real issue was much more mundane. Worth testing the simpler explanation before adopting the more interesting one.

[cudascope]

A couple things I'd check first. On B200, 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.

[circuitworks90]

Interesting, I hadn't tried that. I ended up in about the same place using TensorRT-LLM. The biggest change for me was simplifying the stack instead of adding one more optimization layer.

[quiethouse]

Not sure I agree there. I've seen people blame prefill/decode split when the real issue was much more mundane.

following this because I'm in the same boat.

[cudascope]

Did a small comparison this afternoon. I couldn't reproduce the slowdown until I pushed context much higher. The surprise was that memory pressure showed up before throughput collapsed.