#760 524K via rope-scaling on consumer Blackwell (sm_120, no clusters): the histogram fallback works end to end — field data + two budget gotchas on 16GB cards

描述

Your README lists 384K/512K (experimental) via the rope-scaling CLI. Just filed field data for the corner nobody can test in CI: **consumer Blackwell (RTX 5060 Ti 16GB, sm_120) has no thread-block clusters**, so `qsa_block_topk` runs entirely through the histogram fallback at those contexts — and it works, end to end, past the trained window.

Rig: 4x 5060 Ti (gen5, x8/x4/x8/x4), 9950X, 128GB DDR5-5000. Engine: the eddoursul/Strata `custom` fork extended to 4-GPU expert tiers (our build: https://github.com/johnmosesventura06/strata-4gpu, PR https://github.com/eddoursul/Strata/pull/4 carries your `qsa_select.cu` + `rope_scaling` machinery onto that fork), `UD-Q4_K_XL`, int8 KV, MTP on.

At `--rope-scaling yarn --rope-scale 2.0 --max-context 524288`:

- boot READY, `rope scaling yarn, factor 2 (freq_scale 0.5, mscale 1.069315)` — mscale exactly `1 + 0.1*ln(2)`
- 281,535-token cold prompt: **1,418 t/s** prefill; decode at 225K–312K depth: **84–111 t/s**
- warm replay of the 312K prompt: 12 ms (312,134 tokens from checkpoints)
- identity gate: `--rope-scale 1.0` @196,608 reproduced the unscaled decode band (98.2 t/s) — the scaling path is bit-neutral when off

Two budget facts on 16GB cards that the CLI docs don't mention:

1. At 524K the KV pool (~3.7GB int8) lands before the auto expert cache sizes itself; main-card residency drains to ~202 slots and ranked experts fall to host/tier. Measured cost is modest (decode -5% per residency halving), but the *staging consequence* is hard: `--prefill 24576`'s ~2.4GB of device buffers can no longer borrow or allocate and the engine dies between load and READY. `--prefill 4096` boots; the chunk clamp itself costs 18% of prefill (A/B'd with residency held constant: 2,159 -> 1,778 t/s at 256K with a fat cache).
2. `--vram-reserve-mib` is not the escape hatch: raising it to buy the staging back starves the ~1.4GB weight arena first and dies before that. The real lever at 524K is KV-side bytes (k8v4 — we haven't gaunted it yet).

And one doc suggestion: 384K/512K "experimental" probably needs a footnote that on sm_120 (all consumer Blackwell) selection lands on the histogram fallback rather than the cluster kernel — users will otherwise read "unsupported geometry or cap"-adjacent failures as broken rather than as the intended fallback... which, to be fair, this run proves is not broken.

Happy to re-run any arm on this box if it helps you pin the fallback's geometry limits.