Windows + 32 GB RAM: the low-RAM resident mode never engages - the fit check runs after the engine's own cache warm-up #467

Version: Strata 0.1.34 (ready-made strata-windows-x64-hip.zip engine), Windows
OS: Windows 11 Pro (build 26300), AMD Software: Adrenalin driver 32.0.31035.1003
GPU: AMD Radeon RX 7900 XTX 24 GB (gfx1100)
CPU: Ryzen 7 7800X3D (AVX-512) | RAM: 32 GB (31.1 GiB usable) | Disk: NVMe SSD
Model: ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF IQ2_XS (revision ed59f92082b1e93c0e96d60a8b11aab089b52f09), installed by setup itself (--family qwen --model IQ2_XS)

What happens

The low-RAM resident mode (--resident-experts, chosen by setup) never engages on this PC; every start falls back to --mmap-experts with:

strata generate: WARNING: the resident RAM mode does not fit (FileExpertSource: resident complement 16.20 GiB
exceeds available RAM (0.44 GiB) minus the 4 GiB safety headroom); the experts the GPU does not hold are read
from the model folder through the OS file cache (--mmap-experts), which is slower when the RAM cannot keep them

setup.py --check itself says this configuration fits ("IQ2_XS needs ~48 GB RAM: fits in the low-RAM mode (the GPU holds ~53% of its experts, the rest stays in RAM)"), and MODELS.md states "a 32 GB PC with a 24 GB GPU runs Q2_0, IQ2_XS and the Coder this way".

Measurements

I probed memory every 2 s from just before starting the server (setup-written config; STRATA_RESIDENT_PIN=0 as setup wrote it):

T-0     available 20.7 GiB                              (clean state, right before start)
T+2s    available 19.9 GiB    strata.exe WS  0.6 GiB
T+10s   available 14.5 GiB    strata.exe WS  5.9 GiB
T+20s   available  7.6 GiB    strata.exe WS 12.7 GiB
T+36s   available  0.5 GiB    strata.exe WS 19.9 GiB
then stable at ~0.5-0.7 GiB

The WARNING's position in the engine log shows when the fit check runs - after the engine has fully initialized:

strata generate: pre-filled 12551 of 12551 slots from the profile; slot 0 verified
strata generate: R4 hit path ON - resident experts are computed on the GPU
strata generate: 7 expert-pool workers + the host thread
strata generate: session is up (engine 0.1.34)
strata generate: token graph hit path: 12551 resident experts, decided on the device
strata generate: WARNING: the resident RAM mode does not fit (...)      <-- here
strata serve: prompt chunk auto: 8192 tokens

The math

• Needed for the resident complement: 16.20 GiB + 4 GiB headroom = 20.2 GiB
• Available in the clean state: 20.7 GiB - only 0.5 GiB of margin, before any of the engine's own allocations
• Available once the engine has warmed its own caches: ~0.5 GiB

So the check cannot pass at the moment it runs, and no value of STRATA_RESIDENT_HEADROOM_GIB helps (for any headroom >= 0: 0.5 GiB - headroom < 16.2 GiB). On this machine the docs' "32 GB + a 24 GB card runs IQ2_XS in low-RAM mode" is not reachable on Windows.

Impact

Not a crash: the mmap fallback works (server healthy, correct answers; an informal test measured roughly 47 tok/s including reasoning tokens). But the resident protection against memory pressure is not available - anything that shrinks the OS file cache sends the missing experts back to the SSD - and the documented 32 GB configuration silently degrades.

Possible fix direction

Read the available memory for the fit check before the engine warms its own caches (or reuse the reading taken at the start of initialization), similar in spirit to the #403 fix ("the safety check uses the same RAM reading"). The file-cache pages the engine itself has just created may be counted against available memory here; excluding the engine's own working set from the reading would also fix it.

Related: #403 (a clamped budget fails its own safety check), #250 (miscalculated limits despite theoretically available space), #286 (Windows low-RAM tier).