QSA host KV tier: give each layer its own slot map (decode at 231K 6.8 -> 15.6 tok/s) #621

Before you start

• I have read the Roadmap and this is not already planned there.
• I have read the FAQ.
• I have searched existing issues and found no duplicate.

Applies to

Engine

What problem does this solve

With --kv-host-pages (#499), every QSA layer shares one GPU slot map. At depth the 12 layers' selections (up to ~72 pages each) compete for the same ~515 resident pages, so they evict each other on every decode step. Every miss also copies the page's K/V rows for all layers over PCIe (~860 KB instead of the ~72 KB the layer needs). On Qwen3.8-Flash-Next this was the largest decode cost past 32K tokens: about 36 ms of a ~128 ms step at 64K.

Proposed solution

Give each layer its own residency: phys_of / logical_of / usage become [num_layers, ...], and the fused-copy descriptor is split per layer (that layer's K, V and fp8 scale banks only). The write pages are computed once per forward but ensured and translated per layer, before that layer's store; invalidate_pages clears every layer. GPU memory is unchanged. Reference implementation: gdevenyi@34d087eddf43

Measured on a GTX 1080 Ti, Qwen3.8-Flash-Next NVFP4, 32K tokens on the GPU plus 3584 host pages, together with the indexer change in the legacy-GPU issue:

• 231K-token prompt: time to first token 2 h 20 min -> 29.5 min, decode 6.8 -> 15.6 tok/s, recall 3/3
• a 2048-token prefill chunk at 64K depth: 41.8 s -> 13.4 s
• the "selected pages dropped" warnings near 250K are gone

Alternatives considered

A larger --kv-reserve-tokens on the GPU hides the thrash up to that depth, but costs expert-cache slots.