onnxruntime

f8304035 - PagedAttention: make the latent split's occupancy target measurable

PagedAttention: make the latent split's occupancy target measurable ComputePagedLatentSplits stopped at two CTAs per SM with no record of why. The target is now one place, ORT_PAGED_LATENT_SPLIT_WAVES, and the comment carries the measured curve for DSV4 decode: two waves is the optimum, four is -0.9% target verify, eight is +1.6%, sixteen is +2.6%, and forcing fewer splits through the granularity costs +16% at three and +21% at one. Default unchanged, so the emitted launch is identical. The sweep is the point: this kernel is parallelism-bound, not bound by the selection-list rescan that each of its 48 CTAs performs, so hoisting that scan to a per-token precompute is worth far less than it looked.