It’s mind-bogglingly convenient, especially compared to the before times. Consider donating to them if you can.

No one’s mentioned Forgejo yet? Solid git and artifact repository.

+1 for the Seiko 5s. Love me a SNZG07J1

There’s lots of ways to skin this particular cat. My current approach is low powered Synology (j series?) for mass storage, then 1 litre PC’s running proxmox for my compute power using their NVME for storage, all backed up to the Synology.

Two good points here OP. Type docker image ls to see all the images you currently have locally - you’ll possibly be surprised how many. All the ones tagged <none> are old versions.

If you’re already using github, it includes an package repository you could push retagged images to, or for more self-hosty, a local instance of Forgejo would be a good option.

Build anything small into a container on your laptop, push it to DockerHub or the Github package registry then host it on fly.io for free.

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.

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.

Shoutout to Magic Earth, the (weirdly named) iOS app that uses OpenStreeMap data. Works on CarPlay, has reliable routing, and I get a buzz out of updating a changed a speed limit or something on OSM and then seeing the change implemented a few weeks later when I’m driving through there again.

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.