#5635 [Engine] 0.1.58 regression: iq3_s + 256k OOM-killed and --no-tui no longer holds foreground, breaking systemd service

open · @vshome59 · 2 comments · View on GitHub

Server & APINVIDIA / CUDAModels & quantsLinux

Description

Hello all!

Problem:
After the stable-channel auto-update on 2026-11-08 (~13:30 UTC) that moved the engine and control from 0.1.57 to 0.1.58, an existing systemd user service that had run reliably for days stopped working in two ways from the same update.

Environment:

    Strata control: 0.2.12 (build 0.1.58); engine: 0.1.57 → 0.1.58 (auto-update, channel stable, feed github.com/Niko1221/Strata)
    OS: Linux (kernel 7.x), 121 GiB RAM, 8 GiB swap
    GPU: 2× NVIDIA RTX 3060 (12 GiB each), --gpu-roles 1=main,0=helper
    Model: Qwen3.8-Flash-Next ISTA IQ3_S, context 256k (n_ctx 262144), --kv int8, --kv-resident 32768, DFlash2 drafter on
    Service: systemd user unit, Type=simple, ExecStart=strata launch --family qwen --model iq3_s --context 256k --gpu-roles 1=main,0=helper --yes --no-tui --port 8090 --access lan --api-key ***

Reproduce:

    Run the ExecStart command above on engine 0.1.58.
    Watch journalctl --user -u <unit>.

Expected:
Model loads and the service keeps serving on :8090 (as it did on 0.1.57 for days, including alongside a ~31 GiB KVM guest).

Actual (two symptoms, both from 0.1.58):

    Engine OOM: during model load, systemd-oomd kills the process — Main process exited, code=killed, status=9/KILL, Failed with result 'oom-kill'. Consumed … 114.8G memory peak (the same profile peaks at ~67.8G on 0.1.57). Reproduced repeatedly (05:57, 08:57, 09:14 on separate boots; even with all VMs stopped and 117 GiB free before start).
    control --no-tui exits: on 0.1.58 strata launch --no-tui prints "Strata is serving" and then returns (its own --help now says "with --no-tui, prints the address and returns"). With Type=simple this terminates the unit ~2 s after "serving" — no error, just inactive. On 0.1.57 the same command held the process in the foreground.

Evidence (sanitized):

ноя 09 09:14:47 … The model is loaded and answering. / Strata is serving.
ноя 09 09:14:49 … strata.service: Consumed … 114.8G memory peak.
ноя 09 05:57:13 … systemd-oomd killed 5 process(es) in this unit.
ноя 09 05:57:13 … Main process exited, code=killed, status=9/KILL
ноя 09 05:57:14 … Failed with result 'oom-kill'.

(strata components engines after rollback shows 0.1.57 in use, 0.1.58 newest.)

Reproduction status: Reproduced locally (multiple boots on the same host).

Workaround (recovered):

    strata components use 0.1.57 --family qwen (roll back engine; RAM peak drops to ~68 GiB, no OOM).
    Replace ExecStart with foreground form: strata serve --profile <run_config.json> --port 8090, so Type=simple stays alive.

Question for maintainers: is the --no-tui return-on-address change intentional? If so, the foreground/service path should be documented (e.g. serve for systemd); if not, it is a regression that breaks Type=simple units.

Related on this site

Quick links to related pages.