paged attention observations with MLX

[southtinker]

I'd start with the simpler stack.

That matches what I was seeing. I don't think your conclusion is far off. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.