Issues / #961

#961 Windows (WDDM): verify window wedges nvlddmkm — 0x141 engine timeout → 0x116 TDR failure (driver defect, not a Strata bug)

closed · @1593914054 · 1 Kommentare · Auf GitHub

Multi-GPUNVIDIA / CUDAModels & quantsWindows

Beschreibung

On Windows 11 (build 26200) with an RTX 4080 SUPER and driver 610.88, running the
speculative verify window (`--mmap-experts --spec 4`, Qwen3.8-Flash-Next IQ3_S) wedges
NVIDIA's kernel-mode driver nvlddmkm: the GPU display engine stops answering any driver
call, dxgmms2!VidSchiCheckHwProgress reports "no hardware progress" (live dump 0x141
VIDEO_ENGINE_TIMEOUT, responsible offset nvlddmkm+0x1959fa0), and the TDR reset then
fails with STATUS_INSUFFICIENT_RESOURCES (0xC000009A), bugchecking with 0x116.

This is a driver-level wedge, not a Strata kernel bug — proven by exhausting every
software lever: replacing the sampler with a trivial write, removing all mapped-host-
memory kernel writes, switching the doorbell seq from read-modify-write to store,
disabling the shared-expert stream, disabling the device-side plan, and disabling the
entire CUDA-graph capture all still wedge (the no-graph build merely wedges later).
The identical failure appears under unrelated workloads (2024 GPU overclocking, 2025
torch training), while llama.cpp is stable on the same hardware/driver at higher load.

Full root-cause analysis, the 0x141 dumps, and all reproduction logs are attached.

[strata-nvlddmkm-wedge-evidence.zip](https://github.com/user-attachments/files/33068316/strata-nvlddmkm-wedge-evidence.zip)

Mehr auf der Site

Links zu Install, Modellen, Releases.