I am running a AMD Athlon(tm) X4 860K Quad Core Processor with 32GiB of RAM, Radeon HD 7450, 16TiB of HDD storage and 256GiB SSD. The only upgrade I am considering is buying 4TiB SSD drives to replace the HDD drives, this is only because I've noticed SSD's have gotten really cheap.
I would plan for Docker and not Virtual Machines, as VM's emulate an entire computer and then you run an entire operating system within them and then the application, the result is they need far more resources to act as a host for an application. Server applications have been moving to Docker because its a defined way to sandbox applications, run them consistently and uses far less resources.
I then deployed Portainer Community Edition on to the server, this provides a Web UI to manage the docker contaners running on the server. I have 9 docker containers currently running on the server.
You mentioned Plex: Plex provide a docker image for running their application that supports NVidia GPU Acceleration and seems to run fine on AMD hardware. You will find almost every server application offers an official docker image.
With my business hat on, think how many docker containers you want and plan for that + 1 cores in your CPU, you can probably look up the applications you want to run and add up their recommended RAM usage, as a home rule of thumb 16 GiB of RAM is the minimum, 64GiB would be overkill.
I have a Dell R720XD as my home NAS. I bought it used off eBay. The XD version has 12x 3.5" drive bays. I have eight of them occupied with 8TB spinning disks. The server runs TrueNAS and the disks are part of a ZFS RAID. I spent a few hundred dollars on the server itself, a hundred or so on more RAM, and $2200 on the hard drives.
Not related to your initial issue.. but a quick question as it appears you might have already figured out a lot regarding hosting your own instance.
Do you know if you can lock it down so that only you have an account on the instance? I’d like to run my own but I’d want membership locked so I just use the instance to federate content.
I've been using a Synology NAS for years. I use file and photo sync and it works pretty well. I just set up Surveillance Station (their security camera recording software) and connected an inexpensive camera to it. It works really well although it took me a while to figure everything out.
If you're concerned about data loss, you can set up a backup to Amazon S3 or another backup/storage service or you can put a Synology NAS somewhere else and have it back up to that. You could also have it sync with a service like One Drive.
I like to tinker so I've set up a couple other things on it including a Plex server, pi hole, and a VPN Server. It's a pretty versatile device.
Synology in an offsite location because it just works, and I have no complaint about their mobile app. At home I’m running proxmox with truenas and raided drives. A nightly script will copy all new files from synology to a truenas share.
My network mostly uses NPC and summon names from Final Fantasy XI, because I played that game for many, many years and can associate the personalities of those characters with specific roles the host needs to have. I've also considered using Pokemon names for similar reasons, and with over 1000 current Pokemon species it'd be hard to max out in a home environment.
Sounds similar to my needs. My solution was a self hosted instance of own cloud on a raspberry pi 4 (nextcloud was good but too too many bells and whistles, and was unstable on my system). The owncloud android app automatically transfers my photos and videos, which are then automatically downloaded to my main PC. Important: This is not a backup solution, if I delete a photo from one instance, it will be be deleted on all instances. This system had physical redundancy, as all photos are on at least 2 separate devices at a time.
What was the rationale behind basically using a pi to just forward the files to your main PC? Wouldn't it be more efficient to just do everything on the main PC - using task scheduler or something?
I'm also concerned about putting all eggs in one basket. Ive experienced a shutdown ruining the formatting on a drive and losing everything on it. That, and the possibility of theft (PC is more attractive than a little hard drive) or water spilling on the comp and frying something, is what's preventing me from the simple option of just putting everything on the PC.
You need to investigate owncloud, it is a self hosted cloud drive, think Dropbox, the pi is the machine in my house which is already internet facing, and has a wordpress blog running, with a domain name attached, so it made sense to use that. The owncloud android app sends all my photos and videos to the pi, and then when I am at my pc, the owncloud windows app pulls the files from the pi,so all files are synced. Once set up correctly, all of this happens without any manual intervention, and files are stored in 3 physical discs, my phone, my pi ext hard drive,and my PC. I also have an off site backup on oracle S3 Archive.
Nothing wrong with practical. I'll often name a VM by what service is going on to it as a temporary measure until I'm sure it's going to work out and give it its final name.
There's been some work getting CLIP to run in pure C++ with quantization in GGML, and there's a curious FasterViT model I've seen months ago, so hopefully this can be made faster to inference and easier to host as one binary soon enough.
If you want to host a capable pretrained model, feel free to check out LLaMA, especially the LLaMA.cpp since it allows for speedy inference. For the front-end, there's text-generation-webui, official web UI, Serge, XInference, or chatbot-ui with LocalAI (a server that makes LLaMA.cpp use OpenAI's schema).
For the model fine-tunes, I'd personally recommend WizardLM. It's not perfect, far from it, but it seems the closest to GPT-3.5 in my experience. Be sure to never trust what it says though, it does hallucinate less then other fine-tunes I saw, but still does so frequently enough.
There isn't really much of a need to train a model on a particular community. If you need it to work with changing facts, just throwing results from the search engine into the context window. Most of these models were already trained on huge datasets including Reddit, so...
If you want to fine-tune it on most helpful comments to make sure it generates more consistent advise, I'd recommend QLoRa and a 1k instruction dataset like in LIMA paper. Though again, I'm not sure there's any use for that.
selfhosted
Popular
Esta revista es de un servidor federado y podría estar incompleta. Explorar más contenido en la instancia original.