Toots for zrail@hachyderm.io account

Written by Pete Keen on 2025-02-01 at 23:17

Think I finally have a workable local LLM setup for #HomeAssistant. It involves:

• Nvidia Tesla P4 GPU (8gb memory, Pascal arch)
• Faster whisper (linuxserver.io docker image) STT
• Piper TTS
• Ollama running on bare metal
• Qwen 2.5 7B
• a custom RAG pipeline/proxy in front of Ollama that filters the exposed entities

This whole assembly lets me run tool commands in ~3 seconds on a cold cache and half that on a warm cache.

I want to figure out how to make the kv cache more persistent. Right now it seems to get cleared on every assist interaction. I'm also not sure if the rag proxy is actually necessary and I'm super hesitant to release it widely as it's a pile of janky python and #Ollama specific.

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-31 at 17:14

Follow up: I put an on/off zwave switch I had in stock on a single-switch light and it works great. Only problem is the 4-gang box is very tight so if I want to put more in there (which I do) I'm going to have to get some of the easy-fit switches.

[#]HomeAutomation

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-31 at 00:26

Ok #HomeAutomation folks, I would like some advice. I have some lighting circuits I want to smarten up. I have at my disposal Lutron Caseta, zwave, and zigbee.

A few of these circuits are single switch but most of them are multi-switch, either two or three each. I want to use accessory switches rather than the existing three/four ways because the four way switches are literally popping out of the front of the box.

What's my best option? Install the handful of zwave switches I have on hand as a way to move forward? Or should I drop $$$$ on Lutron gear? Or something else?

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-29 at 13:59

If not pillow why pillow shaped

[#]cats #catsofmastodon

=> View attached media

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-27 at 01:11

A recipe:

• 1 serving of quick oats prepared according to package instructions

• brown sugar, as much as your heart desires

• fresh strawberries

• candied walnuts or some other variety of cronch

• 1 skoosh of almond extract (approximately 1/2 a capful)

Mix. Enjoy while watching an episode of your favorite comfort show.

[#]cooking #oatmeal

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-25 at 22:31

me:

me:

her, out of absolutely nowhere, slightly annoyed: sigh I guess I can play Skyrim again.

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-25 at 18:41

Today's discovery: TCL Roku TVs lie about what codecs they support (that's not the discovery). Amongst other lies, they lie about supporting interlaced files.

The whole TV locks up and restarts when it tries to play a 1080i file.

Pretty cool. Pretty pretty pretty cool.

So now I'm spending my Saturday figuring out how to find all the interlaced files in my library and deinterlace them.

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-24 at 23:23

insatiable urge to plan

https://mas.to/@RikerGoogling/113885901774918713

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-23 at 15:15

Hot #zigbee protip: add more hardwired router nodes, even if you think you have enough. Your wireless nodes will thank you.

[#]HomeAssistant #HomeAutomation #homelab

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-22 at 00:42

This is a neat #3dprinting idea: model 1.8mm channels into the side of something and then slide different color filament in as decoration. Right up there with using filament as a hinge pin.

https://www.reddit.com/r/3Dprinting/s/wNP8Kcb1sD

=> View attached media

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-21 at 00:54

Never felt the need to use filters on mastodon but gosh now is the time I suppose.

I can get my doom and gloom on bluesky, mastodon needs to stay a happy place full of homelabs and joyful electronics and good books.

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-17 at 20:23

I wrote down my thoughts about #TikTok but it's too long for here.

https://www.petekeen.net/goodbye-tiktok/

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-16 at 14:40

Me: hey Jarvis, raktajino double strong double sweet

[#]HomeAssistant: I'll get the kettle on.

espresso dispenser starts warming up

^^ that is what all of this #LLM malarkey is building towards.

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-15 at 17:40

My #homelab #selfhosted server is back to fully alive!!!!

To recap, last week I tried to add a GPU to it and all of the sudden the drive array wouldn't work. ZFS helpfully stopped all IO so nothing would get damaged any further, but I didn't know where to start.

Here's what I did:

• Restored the machine to the state it was in prior to GPU, no change

• Ordered a new HBA because all of the drives failing at once is deeply suspicious

• Memory testing, found bad memory, ordered and installed new memory

• Waited FOREVER for the new HBA to show up

• New HBA showed up, installed and everything is back to normal, with 32GB more memory and a seemingly snappier drive array.

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-10 at 17:36

I'm too used to running #HomeAssistant across six cores of my i7-12700 and sitting on a PCIe 4.0 NVMe.

Now that it's on a plain SSD and two J5005 cores the stuff that felt realtime or almost realtime is now uncomfortably slow. Charts are slow. Connecting is slow.

Le sigh.

[#]HomeLab

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-09 at 21:01

Update: I found a bad stick of memory, which was probably the primary culprit. Haven't tried hooking up to the disk array without that stick yet, but probably going to wait until I get the new power supply (going from 260W to 380W should maybe help, hopefully?)

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-09 at 13:14

I will (hopefully) take a number of lessons to heart from this:

1. Production means other people rely on it. Experiments and FAFOing need to happen on separate hardware.

2. Critical stuff needs to be on separate and ideally redundant hardware. I'll be moving #HomeAssistant and friends as well as network controllers to a different machine today.

[#]homelab #selfhosted

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-09 at 11:46

Upgraded my server with a GPU yesterday morning. Worked great most of the day but then everything fell over at like 10pm last night.

Every disk in my drive array started throwing errors and ZFS decided to lock things down. I suspect the HBA went bad when I was moving things around for the GPU.

I have backups of the important stuff and extra hardware to limp along until a new HBA shows up, but the media collection will remain inaccessible for the duration. Family is gonna be so bummed.

[#]homelab #selfhosted

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-06 at 22:42

Is a 15k rpm fan necessary for cooling an Nvidia P40?

No.

Is it sufficient?

Boy howdy.

[#]homelab

=> More informations about this toot | View the thread

Written by Pete Keen on 2025-01-03 at 22:57

My new GPU arrived and while I wait on parts for its permanent home I decided to try installing it in my Dell Wyse 5070 Extended.

When I try to do stuff that approaches max VRAM the driver crashes with an error like "GPU has fallen off the bus" which apparently points a finger at power supply issues. That tracks because I have it sitting in a PCIe 2.0 x4 slot.

Hopefully it works more consistently in my server which has a GPU-specific x16 slot (currently occupied by an HBA). I need to free up that slot, which means I need to switch to some other 10Gbit card, which means I need to relocate an m.2 drive.

[#]HomeLab

=> View attached media

=> More informations about this toot | View the thread

=> This profile with reblog | Go to zrail@hachyderm.io account