Server panic/reboot linked to Qwen3.5-27B

[zhangchn]

Hi,

My server running llama-server and Qwen3.5 (mostly for coding agents) rebooted several times a day since I switched from Qwen3.5-35B-A3B to Qwen3.5-27B. Any suggestions?

I tried to stop services like Ollama, and set a lower GPU power limit with nvidia-smi -pl 300 (which was 350W by default). But as long as 27B was run with CUDA, it would crash and the machine would reboot itself.

My server config:

• CPU: dual AMD EPYC 9255 24 core
• GPU: dual RTX 3090 with NVLink
• OS: Ubuntu 22.04.1 with kernel 6.8.0-52-generic
• driver version: 525.183.01
• llama-server: b8508, b8580 built with CUDA 12.4.0_550.54.14
  • arguments: -m Qwen3.5-27B-UD-IQ3_XXS.gguf --mmproj mmproj-F16.gguf --host 0.0.0.0 --fit on