Ollama runs gpt-oss on the CPU (GPU idle) while other models use the GPU

Context

Source(s): https://github.com/ollama/ollama/issues/11676 Signal: 31 reactions, 91 comments. Status: reported working by source, not reproduced by us.

Problem

gpt-oss:20b/120b load entirely or mostly on the CPU (ollama ps shows a CPU/GPU split such as 88% CPU) and are slow.

Not this when / boundary

The model is simply bigger than your VRAM; then partial CPU offload is expected (maintainer).

Root cause

Raising OLLAMA_NUM_PARALLEL or the context length increases the memory needed, so the model no longer fits in VRAM and is offloaded to the CPU (maintainer).

Fix

1. Remove OLLAMA_NUM_PARALLEL or set it back to 1, and restart the server. Users confirmed this.
2. Lower the context length (num_ctx). The reporter's GPUs worked at 4k; one user on an RTX 5090 could use 32k but not 64k.
3. To force layers onto the GPU, make a Modelfile with PARAMETER num_gpu 256 and your num_ctx, then ollama create (a maintainer posted this; a user thanked them).

How to verify

ollama ps shows 100% GPU (or mostly GPU).