Great write up, thanks. For video learners, Wolfgang does a good step-by-step on YouTube

I’d love you to check back later with your conclusions.

Guide to Self Hosting LLMs with Ollama.

• Download and run Ollama
• Open a terminal, type ollama run llama3.2

If it’s an M1, you def can and it will work great. With Ollama.

+1 for Forgejo. I started on Gogs, then gathered that there had been some drama with that and Gitea. Forgejo is FOSS, simple to get going, and comfortable to use if you’re coming from GitHub. It’s actively maintained, and communication with the project is great.

Thanks, I ended up going with Garage, but it has the same issue. I assumed I could just specify some buckets with their keys in the docker-compose or garage.toml, but no - they had to be done through the api or command line.

This is correct, I already installed the minio cli, but when I came back and read this, I tried it out and yes, once garage is running in the container, you can

alias garage="docker exec -ti <container name> /garage"

so you can do the cli things like garage bucket info test-bucket or whatever. The --help for the garage command is pretty great, which is good since they don’t write it up much in the docs.

Thanks. I ended up going with Garage (in Docker), and installed the minio client cli for these tasks.

One I’m writing. I use the host file system (as I have a strong preference for simple) for it’s storage, but I’m interested in adding Litestream for replicating the database onto AWS.

"Convert this text to make it sound like from a random person: "

Love the effort you’ve put into this question. You’ve clearly done some quality research and thinking.

When I asked myself this same question a couple of years ago, I ended up just buying a second hand Synology NAS to use alongside my mini-pc. That would meet your criteria, and avoids the (I’m not sure what magnitude) reliability risk of using disks connected over USB. It’s more proprietary than I’d like, but it’s battle tested and reliable for me.

starcoder2:latest       	f67ae0f64584	1.7 GB	3 days ago 	
phi3:latest             	d184c916657e	2.2 GB	3 weeks ago	
deepseek-coder-v2:latest	8577f96d693e	8.9 GB	3 weeks ago	
llama3:8b-instruct-q8_0 	1b8e49cece7f	8.5 GB	3 weeks ago	
dolphin-mistral:latest  	5dc8c5a2be65	4.1 GB	3 weeks ago	
codeqwen:latest         	df352abf55b1	4.2 GB	3 weeks ago	
llama3:latest           	365c0bd3c000	4.7 GB	4 weeks ago

I mostly use starcoder2 with Continue for code autocomplete, the big deepseek coder is a bit slow (I can feel it thinking), but it and the regular llama3 are good for chatbot type programming questions.

I don’t really have anything to compare the M1 performance to. I guess the 8GB models output text a little slower than the web versions of the same models, and the 4GB ones about the same. Using ollama in the terminal, there’s sometimes a 0.5-2 second pause before it starts outputting. Not with phi3 though - it’s surprisingly snappy for the quality of answers.

An M1 MacBook with 16GB cheerfully runs llama3:8b outputting about 5 words a second. A second hand MacBook like that probably costs half to a third of a secondhand RTX3090.

It must suck to be a bargain hunting gamer. First bitcoin, and now AI.

edit: a letter

I use the Continue VS Code plugin with Ollama to use a couple of different models (deepseek-coder-v2 & starcoder2) to recreate a local only Github Copilot type experience for coding. This is on an M1 Apple Silicon though. For autocomplete the generation needs to be pretty brisk - I’m not sure how that would go in a VM without a GPU.

Yep, I think there’s sound arguments for separating out your storage (NAS) and network (router/DNS/PiHole) infrastructure. After that, whatever suits your purpose. I virtualise all my serious services on one machine under Proxmox (mostly for ease of snapshots) then have another machine for things I’m fiddling with, usually again under Proxmox so they are easy to move to production when I’m happy with them.

My NAS and production server run 24/7, I’ve got a dev server that I turn off if I’m not expecting to use it for a week or so. Usually when I do that, I immediately need it for something and I’m away from home. I have chosen equipment to try and minimize energy use to allow for constant running.

My view on UPS is it’s a crucial part of getting your availability percentage up. As my home lab turned into crucial services I used to replace commercial cloud options, that became more important to me. Whether it is to you will depend on what you’re running and why.

I’ve heard that one of the most likely times for hard drives to fail is on power up, and it also makes sense to me that the heating/cooling cycles would be bad for the magnetic coating, so my NAS is configured to keep them spinning, and it hasn’t been turned off since I last did a drive change.

I agree. Get a domain name, point it to the internal address of your NGINX Proxy manager (or other reverse proxy that manages certificates that you are used to). A bit of work initially, then trivial to add services afterwards.

I didn’t really need encryption for my internal services (although I guess that’s good), but I kept getting papercuts with browser warnings, not being able to save passwords, and some services (eg container repository on Forgejo) just flat out refusing to trust a http connection.

Yeah na, put your home services in Tailscale, and for your VPS services set up the firewall for HTTP, HTTPS and SSH only, no root login, use keys, and run fail2ban to make hacking your SSH expensive. You’re a much smaller target than you think - really it’s just bots knocking on your door and they don’t have a profit motive for a DDOS.

From your description, I’d have the website on a VPS, and Immich at home behind TailScale. Job’s a goodun.

+1 for the main risk to my service reliability being me getting distracted by some other shiny thing and getting behind on maintenance.

I love this idea (of just picking something I’m loving each month), it would help me overcome my decision paralysis about who to support.