#2295 16GB RTX 5080 + 96GB RAM: intermittent VirtualAlloc error 1450 on expert-arena large pages + long-context (150-200k) stability data

Descrição

Hi Niko,

First of all — thank you for building Strata. It's genuinely impressive engineering, and it turned a 16 GB consumer card into a daily driver for a 125B model. Much appreciated.

Hardware: RTX 5080 16 GB (TUF), Core Ultra 7 265K (8P+12E, AVX2, no AVX-512), 96 GB DDR5 (48×2), TUF Z890, Windows 11. Single user, single GPU.

Setup: IQ3_S pack, `--max-context 262144 --kv int8 --kv-resident 32768 --vram-reserve-mib 1212 --pcie-frac 1.00 --spec 4 --mtp --pool-workers 5 --adapt-every 1 --adapt-swaps 80 --adapt-decay 0.97`. Engine 0.1.42.

 1) Intermittent `VirtualAlloc error 1450` on the expert arena (large pages)

Since 10-12, the expert arena (50,295,996,416 B) is refused at engine start on some boots but not others:

| Boot | Large pages | Load speed |
|---|---|---|
| 10-12 06:16 | OK | 3.17 GiB/s |
| 10-12 12:42 | refused → 4KB | 3.79 GiB/s |
| 10-12 18:38 | refused → 4KB | 2.87 GiB/s |
| 10-13 17:02 | refused → 4KB | 1.55 GiB/s |
| 10-13 22:00 | OK | 1.89 GiB/s |

Same config, same machine. The only variable seems to be the system commit/working-set state at boot (log: "14.9 GiB left of a 107.1 GiB commit limit" on refused boots). Is this the lock-pages working-set quota (ERROR_WORKING_SET_QUOTA)? Would a pagefile tweak make large pages reliable here? Happy to run any diagnostic you suggest.

 2) Long-context stability (150k–200k tokens, zero errors)

This is why Strata is my daily driver:

- 150–200k ctx decode: mean 86.7–91.5 tok/s (median ~89), peak 110
- Max context reached: 200,248 tokens; 200k+ single sample 90.8 tok/s
- Cold prefill: ~3316–3515 tok/s at 150k–180k tokens
- Draft acceptance pooled: 85.1–89.8%
- KV streaming hit: 94.7–99.3%; PCIe re-fetch 22–40% (mostly 25–33%) at near-full context
- 326-request session: 779k generated tokens, instance avg 97.9 tok/s, zero errors

Happy to attach the full log if useful. Thanks again — the 0.1.42 TAIL_SKIP / PCIe-share work clearly pays off on this shape of machine.