llama-server "srv operator(): instance name=... exited with status 1"

Likely cause

This line comes from llama.cpp's llama-server running in router mode (started with no -m, serving models on demand), not from Hermes. When Hermes asks for a model such as ggml-org/gemma-4-31B-it-GGUF:Q8_0, the router starts a separate llama-server child process for it. That child crashed during load, and the router only reports its exit code. The most common reason is memory: Gemma 4 31B at Q8_0 needs about 33GB for the weights alone, before the KV cache for your context size, and any other model the router still has loaded holds memory too. Other causes: a llama.cpp build too old to recognise the gemma4 architecture, a download that failed or stopped partway, or a bad argument in your models preset file, which makes the child exit immediately.

The fix

1. 1 Scroll up in the router log to the lines just above this one. The router passes on the child's output, and the real error is there (failed to allocate, unknown model architecture, or a download error).
2. 2 Or run the model on its own in the foreground with an explicit context size so the error prints directly (see the command below).
3. 3 If memory is the problem, set --ctx-size explicitly instead of taking the model's default, move to a 4-bit quant, or reduce -ngl so fewer layers sit on the GPU. Hermes needs at least 64K context, so lower quantisation before you lower context.
4. 4 Start the router with --models-max 1 so the model it last used is unloaded before the 31B loads, instead of both competing for VRAM.
5. 5 If the log says unknown model architecture, update llama.cpp to a build with Gemma 4 support.
6. 6 Once the model loads on its own, point Hermes back at the router and retry.