Ollama keeps unloading my models each request

[lubek-dc]

Issue

everything is loaded until i put in a prompt into the aider then it starts to load each gpu up again A

also keep in mind that i am the only user of this and i also use the openweb-ui and when i use it on there it works fine it dosent unload it it just works.

Version and model info

Aider: 0.77.1
Model: Ollama/Deepseek-r1:70b

[k0zakinio]

Also experiencing the same. Did you manage to fix it?

[k0zakinio]

I think I understand the issue, not sure how to resolve it. Ollama is restarting with a changed context length which triggers the reload. I am going to try llama.cpp instead and report back

[awaescher]

I had this too and found a solution. Ollama reloads the model if the context window is different to the one that's loaded. So I used a fixed context window to keep the same model instance in use:

- name: ollama_chat/qwen3:30b-a3b
  system_prompt_prefix: "Code has always be written in english. /no_think"
  use_temperature: 0.7
  reasoning_tag: think      # prevents reasoning thoughts in commit messages if reasoning is enabled
  extra_params:
    num_ctx: 30000          # fixed NumCtx prevents Ollama from un- & reloading models
    keep_alive: 30m         # prevents Ollama from unloading too soon
    max_tokens: 24000
    top_p: 0.8
    top_k: 20
    min_p: 0.0
    temperature: 0.7  

Two downsides:

• you need to find a matching context window size for your workflow/repository
• it's slowing down responses that would only require a smaller context window size