#5751 [Engine] Model load hard-powers-off the whole PC on RTX 5070 Ti (Kernel-Power 41, no dump): autotune variants and 0.1.58 crash; 0.1.46 + conservative VRAM settings stable

open · @w1563 · 7 comentários · No GitHub

AMD / HIPNVIDIA / CUDAModels & quantsWindows

Descrição

Problem: Loading the IQ3_XXS model hard-powers-off the whole PC (Kernel-Power 41, BugcheckCode 0, no minidump, no WHEA event) during the desktop native autotune stage, and during plain CLI model load on engine 0.1.58. Engine 0.1.46 with conservative settings loads and decodes the same model stably.

Environment:
- Strata desktop 0.2.12, Windows 11 build 26300
- Engines tested: 0.1.46 (`f-021-rel`, commit a7bdbfdabc9a) and 0.1.58 (`f-0212-rc`, commit 8a42a9457db6), CUDA 13.0, sm_120
- GPU: GeForce RTX 5070 Ti 16 GB, driver 621.01, PCIe 3.0 x16 (link negotiates 16x, BAR1 16 GB mapped)
- CPU/RAM: Ryzen 9 5900X, 32 GB DDR4-3000
- Model: Qwen3.8-Flash-Next IQ3_XXS, 47 GB two-shard GGUF + pack `iq3_xxs`, `--mmap-experts`, `--kv int8`, expert profile + `--expert-cache auto`

Reproduce:
1. In the desktop app, load the IQ3_XXS model. This starts the native autotune stage ("Finding the fastest settings", 6 planned runs). The machine powers off ~60-80 s after the tune starts (observed twice: `native-tune/tune-7665154f` started 00:54:46 → power-off 00:55:52; `tune-a1fb3c2d` started 08:34:14 → power-off 08:35:37). After reboot the tune status file is left frozen at `{"stage": "base", "state": "running"}`.
2. With engine 0.1.58, a plain CLI load also powers off the machine deterministically (3 runs, also reproduced with `--spec 2`):
   ```
   strata.exe --pack <pack>\iq3_xxs --native <shard1.gguf> --ple-gguf <shard2.gguf> ^
     --expert-profile <expert-profile.bin> --expert-cache auto --prefill auto ^
     --spec 4 --spec-min-p 0.5 --mtp <mtp-dir> --max-context 16384 --kv int8 ^
     --mmap-experts --vram-reserve-mib 1728 --tokens *** --max-new 8
   ```
   The last log lines before power-off (engine stderr):
   ```
   strata verify: capturing the 1-token window (1413 MiB of VRAM free)
   strata verify: captured the 1-token window (upload no error, sync no error) in 13.4 ms
   <machine powers off here>
   ```
3. With engine 0.1.46 and the app's generated aggressive run config (`--max-context 65536 --vram-reserve-mib 700 --pool-workers 19 --vision`), load also powers off (observed twice). Engine stderr before the crash showed VRAM oversubscription: `only 1445 MiB free once the slots are written (reserve 700 MiB); shrinking the expert cache`.

Not crashing on the same machine (control runs, each bracketed by boot-time checks):
- Engine 0.1.46, same CLI load with `--max-context 16384 --vram-reserve-mib 1728 --pool-workers 8`, no `--vision`: full load, expert pre-fill (~4700 slots), decode and generation all succeed; also stable with 19 CPU threads saturated in parallel.
- Baseline stress: multi-thread CPU burn, high-bandwidth memory stress, combined CPU+memory stress, large sustained NVMe reads — all clean, so the platform itself is stable under load.

Expected: A model load (and the autotune stage) must not be able to hard-power-off the host; at minimum the tuner should detect an abnormal termination (the frozen `state: running` status file after a reboot is a reliable signal) and stop retrying instead of re-crashing on the next load.

Actual: Instant whole-machine power-off. No bugcheck, no dump, no WHEA — only Kernel-Power 41 in the System log. The app re-runs the autotune after reboot and can crash again in a loop.

Evidence: Kernel-Power 41 event timestamps match the autotune start times and CLI test start times to within ~1 minute (5 power-off events total, all during engine GPU work). Engine stderr tails captured before each power-off are quoted above. Timestamps and paths are from local event-log and app-log review.

Reproduction status: reproduced locally (5 power-off events across autotune, 0.1.58 CLI, and 0.1.46 with the aggressive config; stable control runs on 0.1.46 with conservative settings).

Analysis notes:
- On 0.1.58 the deterministic death point is immediately after the verify-window CUDA-graph capture completes, when graph replay begins. A string-level diff vs 0.1.46 shows 0.1.58 adds the DFlash2 / KVA capture / STRATA_GRAPH_PRIO subsystems. Possibly related to #5559 (0.1.58 crashes on AMD) and #5608 (page faults on RX 9070 XT), same engine versions.
- On 0.1.46 the crash only occurs under VRAM oversubscription (700 MiB reserve + 65 K context, driver paging GPU memory to system RAM) or during autotune variants; a conservative config is stable even under simultaneous heavy CPU load.

Workaround in use: pin engine 0.1.46 via `engine.pinned` in `components/.updates/state.json`, `--max-context 16384 --vram-reserve-mib 1728 --pool-workers 8`, `--vision` off, and skip the native autotune stage. Stable since.

No site

Links install, modelos, releases.