Ollama uses CPU after 10k token input to gpt-oss:120b despite model loaded entirely on GPU. Only switches back to GPU after input is less than 10k tokens.

What version of Codex is running?

0.86.0

What subscription do you have?

ChatGPT Plus

Which model were you using?

gpt-oss:120b

What platform is your computer?

Windows 10

What terminal emulator and version are you using (if applicable)?

VSCode Powershell

What issue are you seeing?

<img width="727" height="125" alt="Image" src="https://github.com/user-attachments/assets/6aa5d2fe-a638-4c61-9701-7095797c660e" />
<img width="707" height="51" alt="Image" src="https://github.com/user-attachments/assets/6212d30b-6258-4cb2-8eba-c635f9c86df5" />

<img width="642" height="206" alt="Image" src="https://github.com/user-attachments/assets/ddf3bcdb-6d4b-44c2-b3c2-fc70294c93cc" />

What steps can reproduce the bug?

codex --oss -m gpt-oss:120b --search

Run the model locally. Get it to perform a lengthy task.

What is the expected behavior?

You'll start seeing CPU usage increase after 10K token input. The amount of CPU used varies but I have a Ryzen x7950 which elevates the usage around %30 when running in codex CLI. This doesn't happen with other scripts I run that involve Ollama's API locally. Not sure what's causing this behavior.

• I have an RTX Pro 6000 Blackwell MaxQ GPU on my MOBO set as the primary inference GPU.
• The Display adapter is an RTX 8000 Pro Quadro 48GB VRAM GPU.

CPU usage stops and GPU usage climbs when input token length is less than 10K tokens. I have to set auto-compact to 12K in the config.toml file to control temps (CPU switch cools down GPU, might as well exploit this bug) and to continue GPU utilization.

Any help would be appreciated kthx.

Additional information

_No response_