paged attention observations

[cobalttinker]

Has anyone else run into this?

I'm trying to decide whether to keep the current stack or change course. Main question is how people here are handling long context.

Not looking for the perfect answer, mostly what worked in real use.

For context, I'm not chasing leaderboard numbers. I mostly want the system to feel consistent day to day.

[cacheline]

Did a small comparison this afternoon. My result was mixed: first run looked bad, second run looked normal. The main thing that moved was latency, not quality.

[dockriver]

Appreciate the detail. That matches what I saw on 7900 XTX. The biggest change for me was simplifying the stack instead of adding one more optimization layer.

[cacheline]

A couple things I'd check first. Before rebuilding anything, I'd test one clean baseline with Kimi and compare from there. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

[visionforge]

A couple things I'd check first. Before rebuilding anything, I'd test one clean baseline with Kimi and compare from there. A lot of these threads end

My first guess would be this: If you're on llama.cpp, I'd confirm the version first and then reduce variables until the behavior is obvious. Once you know the plain path works, then add the faster path back in piece by piece.

[cacheline]

Good point. The pattern around quant drift 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.

[cacheline]

A couple things I'd check first. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at small local stacks. Once you know the plain path works, then add the faster path back in piece by piece.

[circuitkite]

Interesting, I hadn't tried that. The pattern around continuous batching has been pretty similar for me.

[oxidemixer]

Ran a quick check on my box. I couldn't reproduce the slowdown until I pushed context much higher. I expected the model choice to matter more than it did.

[cacheline]

Maybe, but I had the opposite result. I've seen people blame NVFP4 when the real issue was much more mundane. Worth testing the simpler explanation before adopting the more interesting one.