I’ve read about this method in the GitHub issues, but to me it seemed impractical to have different models just to change the context size, and that was the point I started looking for alternatives

It was multiple models, mainly 32-70B

There are many projects out there optimizing the speed significantly. Ollama is unbeaten in the convenience though

Yeah, but there are many open issues on GitHub related to these settings not working right. I’m using the API, and just couldn’t get it to work. I used a request to generate a json file, and it never generated one longer than about 500 lines. With the same model on vllm, it worked instantly and generated about 2000 lines

Take a look at NVIDIA Project Digits. It’s supposed to release in May for 3k usd and will be kind of the only sensible way to host LLMs then:

https://www.nvidia.com/en-us/project-digits/

I’ve discovered it just a few days ago and now use it on all my machines

For anyone trying this, make sure you do not have “- TS_USERSPACE=false” in your yaml from previous experimentation. After removing this, it works for me too.

In the documentation they say to add sysctl entries, it is possible in docker compose like so:

tailscale:
    sysctls:
      - net.ipv4.ip_forward=1
      - net.ipv6.conf.all.forwarding=1

But it does not seem to make a difference for me. Does anyone know why these would not be required in this specific setup?

Thank you, really appreciate it!

Do you have any links/sources about this? I’m not saying you’re wrong, I’m just interested