previous: >>108978069READ THE (temp)WIKI! & help by contributing:https://igwiki.lyci.de/wiki/Home_server/hsg/ is about learning and expanding your horizons. Know all about NAS? Learn virtualization. Spun up some VMs? Learn about networking by standing up a OPNsense/PFsense box and configuring some VLANs. There's always more to learn and chances to grow. Think you’re god-tier already? Setup OpenStack and report back.>What software should I run?Install Gentoo. Or whatever flavor of *nix is best for the job or most comfy for you. Jellyfin/Emby/Plex to replace Netflix, Nextcloud to replace Googlel, Ampache/Navidrome to replace Spotify, the list goes on. Look at the awesome self-hosted list and ask.>Why should I have a home server?De-botnet your life. Learn something new. Serving applications to yourself, your family, and your frens feels good. Put your tech skills to good use for yourself and those close to you. Store their data with proper availability redundancy and backups and serve it back to them with a /comfy/ easy to use interface.>Links & resourcesCool stuff to host: https://github.com/awesome-selfhosted/awesome-selfhostedhttps://reddit.com/r/datahoarderhttps://www.reddit.com/r/homelab/wiki/indexhttps://wiki.debian.org/FreedomBox/FeaturesARM-based SBCs: https://docs.google.com/spreadsheets/d/1PGaVu0sPBEy5GgLM8N-CvHB2FESdlfBOdQKqLziJLhQLow-power x86 systems: https://docs.google.com/spreadsheets/d/1LHvT2fRp7I6Hf18LcSzsNnjp10VI-odvwZpQZKv_NCISFF cases https://docs.google.com/spreadsheets/d/1AddRvGWJ_f4B6UC7_IftDiVudVc8CJ8sxLUqlxVsCz4/Cheap disks: https://shucks.top/ https://diskprices.com/PCIE info: https://files.catbox.moe/id6o0n.pdf>i226-V NICs are bad for servers>For more SATA ports, use PCIe SAS HBAs in IT modeWiFi fixing: pastebin.com/raw/vXJ2PZxnCockpit is nice for remote administrationRemember:RAID protects you from DOWNTIMEBACKUPS protect you from DATA LOSS
UDIMM prices are so ass that I want to get a SO-DIMM motherboard. SO-DIMMs seem to be priced slightly more reasonable here.
>>108985085Eww stinky niggers. Bad op.
anyone try setting up the browser based cockmail yet? kinda curious about it
>>108985291Mutt's Law outta nowhere
How do I decide what switch to buy?I want an outdoors switch with PoE in and PoE passthrough, to power some cameras and only have to pass in a single cable from the wall. Should I just get the first tp-link shitware with PoE passthrough that I can find? Is it better to find an outdoors/weatherproof switch, or find a normal one and stick it in a weatherproof enclosure?There's always so many consumer models that I always get decision paralysis.And in this case, grabbing a used enterprise switch for cheap is probably not going to work because I need a small-ish one to mount outside, I can't put a fuckoff huge rackmount one.
>>108985085weird-looking elf
>>108985291make another thread and link it here
>>108986683for poe cameras just buy whatever chinkshit meets your requirements, stick it in a box and caulk it shut. I got one from aliexpress but that was before the tariff crap. the only bad thing about it is a lot of EMI
Is there a solution for scheduled yt-dlp downloads of streams? something like https://github.com/Zibbp/ganymede but with a simple web ui that hooks into yt-dlp. I had to improvise and ssh+tmux in my server and just let do yt-dlp <url> do its thing but it's kind of an hassle setting it up with just one hand if you know what I mean
Im nu to selfhosting shit, is there any starter hardware I should consider using?
>>108988300start with the hardware you start withthat one is freenext one will cost you
>>108988346idm spending money
>>108988300No such thing as one size fits all.
>>108988300>>108988346>>108988425I want to self host Jellyfin, Nextcloud, Ampache, pihole, Immich, backups for now. Im thinking of mini PC + Synology 4-bay NAS. Thoughts?
>>108988775Pihole is useless. Nextcloud is jeetware
>>108988795well I was more asking about the hardware but thanks.what are u hosting?
>>108988775A four-bay Synology can do those things by itself. Just add some RAM.
Hi nerds! How do I into routers?
what the best software to copy 1TB of data from hdd to hdd? needs to have an integrity check and ideally OSS.
>>108991776rsync
TL:DR How to document network with vlansI work in IT and have homeservers and do selfhostingI know networking but never really played much with VLANsI started to play with them at homeI now decently understand but not really feel sureI want to document how I setup shitI did pic relatedBut it feels like SO MUCH FUCKING WORKand its like fucking few 8 port switchesI played with Netbox and phpipamnone of them feel like they are about documenting ports specificlybig overview which vlan is where - yeahdocumenting that port X is tagged or untagged with PVID that and those vlansnopeI am like WTF? am I doing vlans wrong?This home documentation is kinda test run as I want to deploy and document stuff at locations we manageHow do you document network and VLANsI feel like there should be some tools for this
The android TV jellyfin app is absolutely shit to navigate, any way to make it less grating?
>>108992199on my phone I use findroid
>>108992184documenting is really just a lot of data entry. You often see some neat graph that documents everything so well that even you who don't own it understand what's going on. The reality behind trhat is that someone probably worked on that documentation for months and months until it got that way
>>108992215Are you retarded?
>>108992222well thats fineif it felt like tools are in place and I am doing it correctlybut i go with enterprise level netbox with all its complexity and approach as source of truth...and it feels it lacks per port per switch documentationso I feel like people who actually do stuff dont really document those for vlanswhich feels weird and I am not sure if I should go my own approach with drawioor try some different way
>>108990868you need to specify what you want.a random chink router off amazon is sufficient for NPC itoddlers
>>108992199I use scyphopote on my phone to control the client on my tv. It's pretty new but does what I was (play/pause, check cast out, start playback)https://github.com/eiffelbeef/scyphomote
do antivirus and firewall niggas like .ovh tld?it's one of the cheapest i can get, they shit themselves over .xyz and .top
>>108992184I just got the exact switch in your picrel. Did i fuck up?>>108992199I use wholphin on my TV, I'm pretty sure there are other 3rd party clients in the same tier. Some people just run the jellyfin plugin in kodi. I think there's also DUNE I've heard good things about, been meaning to try that. >>108992729Wait why does tld matter?
>>108994596>Wait why does tld matter?Some TLDs are only used for crime, so they're blacklisted by default by some vendors. .su is infamous for this.
>>108994666Is that true of .xyz? It's where I was considering getting my domain.
>>108994688I don't think .xyz has such piss-poor reputation.
>>108994692That makes sense, but then what is >>108992729 talking about?
>>108994699How should I know?
>>108994703Sorry, not saying you should
>>108985291>Anon cannot stop thinking about bbc cuck pornEvery, fucking, time.
>>108994688i needed a more aggresive antivirus than win defender for parents and it has popup for .xyz and .top so i assume most antiviruses will do tooi talked about that because sometimes you may want to visit .top or .xyz in work/university or random ass hotel to save data when watching movies
>just learned readarr has been dead for like a year>only found out after trying to figure out why I couldn't pull a docker image in a fresh install>never realized because i use watchtower to keep everything updatedI mean it was never great to begin with (lidarr is pretty shitty too) but still surprising. I thought the *arr programs were fairly popular
looking for backup solution ideas (around 15gb)>bd-r drive: $50-$150>discs: $1 for 25gb, $2 for 50gb>cost for 10 years of monthly backups: $170-$390vs>32+gb flash drives: $10-$30 each, depending on brand, speed, etc>cost for 10 years of monthly backups: $1200-$3600>cost for 1 year of rotating monthly backups: $120-$360tapes are stupidly expensive and way more capacity than i need, so it seems like bluray is the way to go
>>108998469what are you backuping
Anyone else have local dns suffixes for everything?>>108986683personally i use: https://www.amazon.com/dp/B0CRB47VH7What you are describing is exactly what I do, i use a 90 watt PoE injector and the switch is kept in an outdoor electrical box enclosure, and it powers 4 IP cameras>>108987650>Is there a solution for scheduled yt-dlp downloads of streams?if its twitch streams, the channel page will contain "isLiveBroadcast", and you can programatically keep checking, and when its True, rip it with streamlink>>108992184i locally host nautobot, you can use it to document everything you are describing that you want to document>>108998469as another anon asked, what are you backing up? Personally, i use forgejo and minio but it seems like you arent asking for a server solution
>>108998548>>108998628>what are you backuping>as another anon asked, what are you backing up?I have most of my shit running in docker now, so the config/data from those containers since i spent a long time fiddling with them to get them the way i wantalso going to back up some personal files (pictures, encrypted financial stuff, etc)
>>108998653>so the config/data from those containers since i spent a long time fiddling with themthen forgejo (or gitea), because you will need to maintain source control of changes you make
>>108998662i absolutely do not want to be making commits for all the shit i'm doing
>>108998674it sounds annoying, but i maintain orgs and repos for quite literally everything because i think its worth it. Heck, i even use the packages feature to build from wheelhouse
>>108998815There are gigabytes of changes daily. There is zero chance I am reviewing any of that, so that level of versioning would be worthless
>>108998999>There are gigabytes of changes dailyis your docker file a gig? No? Use forgejo.Is a system backup a gig? Thats highly likely, but you still want version control to be able to restore it to checkpoints, and thats where minio comes inYou dont want to deal with that, fine, then just toss important things onto a NAS. When something fucks up badly you might panic because even your backup is jacked up and you never realized it.
>>108997353and the only replacement that isnt vibecoded to shit is paywalled and discord only
this is coolfirst time using docker, wanted a gui so using Dockhand to install immich and worked with no issueswanted to move away from google photos on my pixel to locally hosted since I wanted that stuff private.Anything I should add that could be useful?
>>108999041please figure out a robust backup solution before you lose your important photos
>>108999051I do plan to copy the library over to my other computer every now and then.Immich has a auto db dump already set by default so thats covered
is there a storage app/solution that does the following:You upload your data (encrypted) to other people's NASYou must then host data from other people to your NAS as well as the same storage amount you backed up (uploaded)This way you don't pay a subscription but pay with storage So the advantage is you only need a single large NAS and not another NAS in a different location for offsite backup
>>108999365>relying on randos for your offsite backup
should I use ansible for my LXC containersor install guix/nixos in all of themor just rawdog it by manually installing shit and then forgetting
>>108999365Not really. HexOS is supposedly gonna add "buddy backups" so that you can store your encrypted backups on a friend's NAS, but I doubt it'd add any hard requirements.
I want to build either a medium or full size server rack but have only ever had a Pi as my media server. Where do I even begin to find hardware that isn't cheap Chinese shit?
>>108999365Doesn't IPFS approach this problem? It's a bit less one-to-one because you get, essentially, a cryptocurrency for hosting things that you can spend to have other people host your things. It's like your idea but generalised and decentralized.No idea if it works well. Never used it. But your problem reminded me of their concept.
>>109000507Same anon.Looked into this a bit more and it won't work for your case. The overlay technology is called "Filecoin" and whlie it could solve your problem in theory, there is a minimum 10TiB storage for any payout (that you would need to pay for the storage you want to use), so my bad, you can forget that idea.
Rebuilding my NAS/home server box. I've not fucked around a lot with local AI (FOSS, local models of all types ) yet, but I've been wondering if its worth running them on my home server box, as opposed to my main PC with a more powerful (5090) GPU. In any case I have 2 possible GPUs to use for the server rebuild - I own them already. >Nvidia 3090 24GB>AMD 9070XT 16GBand I'm trying to pick which to keep and which to sell. 3090 prices are pretty high now but I wonder if its worth holding onto and using it for my own usage - both potentially AI and other general GPU usage for server related things. Or is it better to sell and just hang onto the 9070XT for hobbyist AI and other server use?>>108999365IPFS I think does something similar in theory, but the problem is that people basically put the monetization of Filecoin atop it which is insane for a variety of reasons. Have you tried >Cryptomator which seems to basically encrypt your data and then you can use any cloud (or local for that matter) source to host it, but anyone looking at it from the outside will only see encrypted nothing so you can basically take advantage of even FAGMAN hosting for offsite backup/sync etc.
>>109000627Really depends on what AI tasks you want to perform and when.LLM, TTS, STT, image?Interactively and when you are present? Non-interactively for summarising feeds, tagging files you generate?If you decide to put the GPU in the server, I would keep the 3090. Whether you should do that or just use the 5090 is the main question though.
>>108999027qrd?
>>109000235>should I use ansibleansible is literally the bane of my existenceand its not because scripts are written in yaml of all fucking things, i just hate it. period.>>109000504>Where do I even begin to find hardware that isn't cheap Chinese shit?ebay, you can find a half depth supermicro server for pretty cheap although the ram has probably already been pulled from it and it will cost a fortune even loading half of the 8 available ram slots with 4x 32gb lrdimm
>>109001037https://lemmy.world/post/43516286/22318332i was mistaken about the paywall, but it is violating gplv3
>>109000667Well that's the thing, I'm not sure yet. I may fuck around with comfyui generating images or some such, though creating a SillyTavern/CharacterAI style characters etc, an "agent/assistant" with speech to text or text to speech could be nice, but I've not fucked around with OpenClaw though I hear that's a big part of agent tasks though it can massively do stupid things (the pomf or catbox owner was complaining about openclaw agents basically shitting up the entire hosting somehow). Running a local LLM to create an "assistant" (building a presence for it, be it waifu or jarvis inspired as many suggest) is a neat idea, but I'm not sure how capable it will be. I guess both interactive and non-interactive tasks are on the table. Seeing how far things can do with local hosting, training/weights etc...though some say that being able to opt into enhancing it through trying to distill shit from the proprietary models like the OpenAI or Anthropic is worthwhile but all of this is theoretical at the moment as I'm just keeping it open to this becoming a hobby to fuck around with as I'm not really sure what tasks I'll "need" or discover as of yet and learning the capabilities is the first step. The 5090 is in my main PC and I don't have two of them; I'm not sure if its something where the storage, output etc, other AI stuff could be put on the home server box yet computationally heavy tasks run remotely from my main rig with the 5090 and the results/training etc..saved to the server/NAS. I'll have to put SOME sort of GPU in the server to do any computational heavy lifting of that sort required (and if I use my AM4 hardware then it will need a GPU to output video at all more or less as I don't think a 5950X has a proper iGPU as I recall. It wouldn't be until the Zen4 era that they showed up), so I may very well put the 3090 / 9070XT in it. Also, I have an offer for a >$530 64GB DDR5 6400mhz Corsair Vengeance RGB kit Decent?
>>108999365I think this would be cool to do, but doesn't exist. Best you can do is organise something like that with friends where you either host each other's backups if you both have NASes, or you just shove a server in his house>>109000090If it's encrypted at the source it's not that big of a deal. You could just periodically query checksums and random data samples to verify it's all online and stored properly, and if some guys's server disappears then you just get a new offsite somewhere else.
>>109001040You're the first person I know who says they hate ansible. Why do you hate it if not for the yaml?
>>109001094For LLMs, more VRAM is better and pretty much the main bottleneck you'll encounter. That's why I mention the 5090 as worth considering. 32GB makes a good LLM like Gemma4 have more headroom for context and it'll be a little faster. It's doable on 24GB though.If you'll only be actively using it (which I am interpretting comfyui and character interaction as active, though you can leave image gen on auto and set up a preference model to filter for the gens you're likely to like to slopmaxx) you can use your desktop with the better GPU for that. You're not going to be, say, gaming and image genning at the same time, I would guess, so that GPU is doing nothing at that point.OpenClaw and other agent frameworks sound neat but, in my opinion, are actually pretty useless. People struggle to even explain what they use them for. I've only been able to think up a very limited set of uses for them, but maybe you have ideas that would be useful to you. Think that one through before you set it up for nothing though.TTS and STT can both run on CPU these days, so you could do that on either your desktop or server. Doesn't really matter. I would just pair it up with whichever you end up using for your other AI stuff. Whisper 3 and Kokoro 82M are decent and lightweight. Of course, a GPU speeds them up, but that might not be something you really care about.Probably not feasible to do any meaningful training on LLMs with your hardware. It always takes more hardware to train than infer.Since this whole thing is so open ended, maybe reframe it like this: what do you really get for selling the 3090 and keeping the 9070XT? Figure out the difference in market value between each, ask yourself if you really care all that much about the ~£200 difference (and considering you can afford three good GPUs, I doubt it) then weigh that money against the potential loss in functionality.(1/2)
>>109001094>>109001264You can keep the storage on your server and use it on your desktop.Either of those cards is overkill for a video transcoder. If you were money-motivated in this you would sell both and buy a motherboard and CPU with an iGPU, but again, you probably aren't I would guess.>$530 64GB DDR5It's a lot more than I would want to pay, but things have gotten a lot worse since I last bought RAM.Only you can decide if you want to waitfag or fold to industry scalping. Both options suck because the market sucks.(2/2)
>>108999025>is your docker file a gig? No? Use forgejo.did you miss the "data" part? the docker backup is 15gb. plex alone is 5gb. sonarr, radarr, etc are a few gigabytes each. and so on. >Is a system backup a gig? Thats highly likely, but you still want version control to be able to restore it to checkpoints, and thats where minio comes inor i could just do a full backup and not have to deal with incremental backups>You dont want to deal with that, fine, then just toss important things onto a NAS. When something fucks up badly you might panic because even your backup is jacked up and you never realized it.that's the point of doing full backups instead of the incremental stuff you're proposing
Setting up local DNS namespaces actually bricked one of my servers to the point where it couldnt deploy and the solution was to have an empty DNS zone for the domain "cluster.local"Holy hell what a strange nightmare it was to figure that out, it literally took me at least 5 fucking hours
>>108988775An Optiplex 7050 micro is around $150'ish on ebay and will do what you need
>>109001552I ended up buying an MSI CUBI N ADL S-225BEU N100 16gb since optiplexes are more expensive where I am.
>>108985085I run a few servers. But even with my big gpu server down. I find my power bill is still in the mid 200s like 278 or so in ohio. I'm by myself and my home is okaish size like 1200sqft but I have made my thermostat always kick off when gone. I dont run lights 24/7 etc. The only constant is my 50tb md1200 and synology nas. I only have one psu in my md1200. It's idle 90% of the time. Fans are in the 20% rpm range. How can I make sure it's not sucking me dry, or is it just the way power is these days? I can afford it but I'd like to know for sure the wattage of everything.
>>109001714Not bad. It will serve you well. You doing to do SnapRaid?
>>109001967*Going to do
>>109001750buy a wall outlet power meter
>>109001264When you say there's no "meaningful training" does that include things lik LoRAs , weights, and the like? Or even improving a model's "remembering" and how it interacts with your data? Or do you just mean the core capabilities of the model as opposed to tweaks and other improvements? I don't expect to be able to make any massive improvements or build a new model that makes it to a higher AI benchmark score overall than the one that came before, but I was under the impression even stuff like the various Stable Diffusion customization (like teaching it whatever hyper specific art style of a mangaka for future generation etc) would be capable with far less than a 5090 from how people talk about it. I am rather curious about building an assistant of sorts both for conversation and for some tasks like >I have a bunch of drives with old installs having all kinds of content from pictures, videos, to installers . Put them into different folder, label them with subject matter and time period, DO NOT FUCKING DELETE ANYTHING and then show mebut am not sure if it can do that properly without being trained on say.. how a huge assortment of imageboard images all relate to each other despite being of disparate subject matter. Anyway that's probably a different sort of thing, but i guess I just want to figure out what kind of capabilities there are and what I'd be buying to put a card like a 3090 at all into the server (or even a 9070XT). At this point its not so much about the money in the short term but simply hardware costs and availability in the future and it being a good time to sell vs necessity to keep for capability.>RAM price Yeah, I know its not great. I was lucky enough to get 64GB DDR5 back when it was $220, but now the same kit is over $1200 retail and so I'm looking to anything even close and it seems $450-550 seems to be the common pricing for various user resale. I worry it could get even worse?
>>109002037That gives me total though not individual.
>>108985085Nextcloud works on any computer?
>>109001967Not yet, I just got one 4tb drive, they are SOOO expensive here rn
>>109002344Yeahz it's insanebright now. I ended up going with 4 2TB drives to save money. Not what I wanted but at the end of the day had enough room for about 500 movies and 50 TV shows plus Nextcloud storage.
>>109002258i don't understand what you mean by individual. individual drives and components? if you plug your md1200 and nas into a kill-a-watt you can see exactly how much it costs to run them
>>109002390Individual servers. I'll have the gpu server then my media server that controls the md1200 and then the nas. I don't want the total wattage but the amount individually. They are in the server power rail where I couldn't use those wall smart plugs.
>>109002228I'm talking about LoRAs and finetuning (the "weights" you mentioned).When people talk about a model "remembering" your past interactions, it's not finetuning they're talking about usually. You use something like mem0 (or sometimes it's built into the front-end) and it lets the model call on previous chats you've had or facts it took from those and insert them into the context for the chat. It's not "training" but I know some projects misuse the term to refer to that.On training image LoRAs, yeah, you can do that with any of those cards depending on the size of the base model. I'm just talking about LLM finetuning here as being more resource demanding.I had a look and actually you could get QLoRAs for LLMs training on either of the NVIDIA cards with quantization per Unsloth. Didn't read far enough to see what quant though.https://unsloth.ai/docs/models/gemma-4/trainI should say as well, you can train LoRAs for smaller models, but you'll probably be mostly interested in that 27B-35B region, because there are some really great models in that range right now that fit on your cards. Perhaps that's just my bias towards those models talking.On your file tagger idea, it's a whole subject unto itself. Could an agent do it? Maybe. That's a big task that many others want that, if solved, I would guess I'd have heard about the few times I've tried to find out what people have actually successfully done with these agent programs, but I haven't. That'd be the "killer feature" they'd advertise if it could actually do it. Might not hurt to try if you are thinking of this as a hobby.The image classifying component of that system probably won't need imageboard specific images to train on. Imageboard images aren't that different from normal images. It's only really when you get into very different images where it starts to become relevant.>I worry it could get even worse?Difficult or impossible to say. It's market speculation, which is famously difficult.
>>109001750your power bill is worthless. what is your energy usage? $250 could be 750 kWh or it could be 2500kwh>>109002523you can either>use the energy monitoring device on your whole rack, which will give you an idea if there's even a point in starting>unplug the servers temporarily and plug them into an energy monitoring device that is plugged directly into the wall>use short extension cords like pic rel to keep them on the rail but allow more space for energy monitoring device