Windows local LLM inference 2-3x slower when terminal unfocused; headless fix restores speed
koloved · reddit · 2026-09-10
A Reddit user running a 27B NVFP4 model via ninfer on an RTX 5090 found Windows inference dropped from 130-200 tok/s to 50-60 tok/s whenever the terminal window lost focus.
- Not GPU throttling: SM clocks were actually higher when slow (2550-2600 MHz), power stayed at 400W, no PCIe power-state drops, decode-host time unchanged
- The culprit was CPU-side scheduling: 'wait' ballooned from 17ms to 38-41ms when unfocused
- Fix: run the server detached/headless (docker run -d), which restores 130-200 tok/s regardless of focus
- Reproduced on a native Windows build, so it's not WSL2-specific
A Windows 11 foreground/background scheduling quirk every local inference user should know.
More from Infra
- You can now write CUDA kernels in plain Rust via cuda-oxide and cutile-rs — rjurney · 2026-09-10
- 2,400 experiments show layer dropout can match dense baselines in LLM pretraining — burkov · 2026-09-10
- Deep-Dive Speculative Decoding Blog Incoming: Drafter Training to vLLM Serving — auto_grad_ · 2026-09-10
- KV cache exposes agent economics: devs pay big for context re-reads that cost providers nothing — hackgoofer · 2026-09-10
- AMD details agent-native ROCm 10 and Hyperloom, auto-optimizing 14,000 models — AnushElangovan · 2026-09-10
- NASA chief backs orbital AI compute as SpaceX targets first space data center in 2027 — rohanpaul_ai · 2026-09-10