[Bug]: Too much ram used

[EugeoSynthesisThirtyTwo]

Commit or version

d7ffdc4

Environment

WSL2 Ubuntu-26.04, Ryzen 9950x3D, 192GB RAM, 2x RTX 5090, 8TB samsung 9100 pro NVME ssd, gcc 15.2.0

Reproduction steps

git clone https://github.com/JustVugg/colibri
cd colibri/c
./setup.sh
COLI_MODEL=/home/me/docs/model/GLM-5.2-colibri-int4-with-int8-mtp ./coli chat --ram 96

Expected behavior

I should use 96GB of RAM, but instead it eats up my whole 192GB and it even writes on the disk (D:) as a swap file, as you can see in the screenshot.

Actual behavior and logs

```shell ▄▀▀▀▄ ▄ colibrì v1.0 ▄▄▄▄▀▀▀▀▄▀▀ piccolo motore, modello immenso ▀▀▀▀▀▀▀ GLM-5.2 · 744B MoE · int4 · streaming CPU ▀▀▀▀ chat · hf · ram 96GB · topp off ▀ ────────────────────────────────────────────────────────── [MTP] attiva: decodifica speculativa nativa (draft=3) [RAM_GB=96.0] cap ALZATO 8->54: il budget lo consente (proiezione picco 95.2 GB; CAP_RAISE=0 per disattivare) [KV] context slots: 1 x 4096 token, projected pool 0.75 GB scrivi e premi invio · :piu continua risposta · :reset memoria · :q esci

╭────────────────────────────────────────────────────────────────────────────────────────────────╮
│ › hi │
╰────────────────────────────────────────────────────────────────────────────────────────────────╯

◆ colibrì
He^C
interrotto
ciao — il colibrì torna al nido 🐦

[JustVugg]

Good report, and the diagnosis is reassuring: the engine is respecting your budget — what's ballooning is the WSL2 VM's page cache, not colibrì's memory. Walk through the evidence in your own log:

• proiezione picco 95.2 GB — the engine sized itself to your --ram 96 correctly, and its RSS stays there.
• But colibrì reads ~11 GB of experts per cold token through buffered I/O, and Linux keeps every read page in the page cache (that's normal and usually good — it's our free L2 cache). Inside WSL2, though, the VM doesn't return cached pages to Windows eagerly, so from Windows' side the VM grows to whatever it's allowed (default: most of your 192 GB) and Windows starts paging other apps to your D: swap. Engine 95 GB + page cache ~90 GB = the full-looking 192.

The fix is one file on the Windows side — %UserProfile%\.wslconfig:

[wsl2]
memory=120GB              # hard cap for the VM: engine budget + cache headroom
autoMemoryReclaim=gradual # give cached pages back to Windows over time

then wsl --shutdown and reopen. With that cap the page cache still works as our L2 inside the VM but can never push Windows into swapping.

Two extra notes for your (excellent) machine:

• If you'd rather not cap the VM: DROP=1 makes the engine advise the kernel to discard expert pages right after reading (fadvise DONTNEED) — page cache stays flat, at the cost of losing the free L2 re-hits. On 192 GB I'd cap the VM and keep the cache instead.
• Your --ram 96 on 192 GB is conservative — with the VM capped at, say, 160 GB you could give colibrì --ram 120+ and the auto-raised LRU would hold ~40% more experts.

Adding a WSL2 section to the README with exactly this, credited to your report. Keeping this open until that lands.