paged attention observations

[cacheline]

I figured I'd ask here before changing everything.

I'm mostly working with Qwen in MLX on GB200. The part I'm trying to understand is NVFP4.

If you've tested something similar, what did you keep and what did you stop bothering with?

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

[cacheline]

that fixed it for me too.

[cobaltworks25]

that fixed it for me too.

subscribed, curious where this lands.

[cacheline]

Need a little more detail here. Is this local-only or are you routing anything externally?

[vectorshed43]

Good point. The pattern around KV cache pressure has been pretty similar for me. The biggest change for me was simplifying the stack instead of adding one more optimization layer.

[cacheline]

That matches what I was seeing. The pattern around tail latency has been pretty similar for me. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.

[edgecoast]

I'd approach it in a boring order. On Threadripper box, 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.

I'm happy to test one more variant if you want a second data point.

[cacheline]

that fixed it for me too.

[cacheline]

A couple questions before I guess too hard. What model, what runtime, and roughly what context size?

[mangohouse]

Did a small comparison this afternoon. My result was mixed: first run looked bad, second run looked normal. The surprise was that memory pressure showed up before throughput collapsed.