Anyone know where I can order really short 16awg UL listed PSU power cables (not extension)?
Anyone else getting hit with quota limits on rclone to Google Drive suddenly this week?
Training on Real-ESRGAN paid significant dividends. In my case the results of fine tune training of about 1000 image pairs of VHS vs HD showed significantly better results than Topazās Dione Dehalo 2x FPS (no other Topaz models yielded sane results in my case).
This came at some expense. The training took about 8 days running on an nvidia A4000 and required very nearly all of the A4000ās 16GB of vram. The primary card in my system is a 3070 and it was not capable of performing the training. That said, I am not proficient in the nuances of pytorch, but ootb, the training definitely required the whole A4000.
What resolution did you train at? If itās at or under 960*960 you can do FP16 and have reduced VRAM consumption
The stills after cropping out letterbox and edge falloff from the VHS were 706x421.
Do you know where to set the precision? I havenāt found any documentation on the options file.
For training no, for txt to img at least you omit --no-half and --precision full and put in --disable-nan-check
This is Real-ESRGAN, not stable diffusion though.
Oh, I use automatic 1111 so it bundles them together, idk then
Ha, RTFM I guessā¦ although itās not mentioned in the training instructions. I should have another set to train on coming up. Iāll try it out then.
Playing with this again. It is now definitely a ādonāt recommendā. After running a while, the entire chassis is too hot to touch. There are only 2 sensors. The package sits around 60C idle and some other sensor (unclear what it is) always stays in a reasonable range but the 32GB optane I put in there IDLES AT 80C! After a couple hours, it just detaches and the system crashes. Curiously, the 2.5" SSD never overheats. I bet the RAM is getting cooked though but no way to know for sureā¦
It took me a while to find this out because SMART isnāt compatible with NVME on OpenBSD (of course). I grabbed the temps from a live ubuntu image after OpenBSD crashed.
I am going to keep the system and just strap a 140mm fan to it, but definitely should be avoided for anyone who isnāt running their system in an igloo.
I repasted the CPU with kryonaut extreme but it made no real difference. The chassis/heatsink is simply hitting its thermal capacity.
I wonder if Protectli products suffer from the same issue. Many look close to identical to the unit you ordered.
I think the heat sinks on those are larger so idk. I suspect that hardware that is sold/marketed in āwestern marketsā will have some basic QA that is clearly missing in this case. This unit doesnāt even have a serial number. Nowhere to get BIOS updates. Missing basic stuff like that but itās like half the price.
Could also be just some discarded batch or trial run from some big Chinese factory that was bought out by somebody. Next thing you know itās being sold out in Aliexpress.
You check the thermal compound?
Also might be worth not having the side cover on where the ram / ssd is
You check the thermal compound?
I repasted the CPU with kryonaut extreme but it made no real difference. The chassis/heatsink is simply hitting its thermal capacity.
Also might be worth not having the side cover on where the ram / ssd is
Probably would help, especially if I mounted it so that the opening was facing up but I just strapped a Noctua A14 to it and called it a day. Now just gonna see if its stable after getting fried.
I am considering just using the 2.5" and giving up on the octane, but itās kind of a perfect application for those little 32GB ones.
I wish there were more QT browser options. I always end up gravitating towards Gnome applications just because I have to have GTK installed anyway unless I want to use Konqueror or something. Try not to have both GTK and QT installed if I can help it.
Related, I acquired a ~5 year old X1 Carbon which means I have a Linux ādesktopā, and a Linux laptop for the first time actually if you donāt count the netbook that just ran stock debian.
Unrelated I finally got around to setting up redundant routers via CARP/pfsync at home (OpenBSD of course). It was pretty easy!
The one caveat that I didnāt anticipate was that the backup hostās dedicated IP is inaccessible from outside of its subnet.
For example, letās say you have a shared carp address of 192.168.1.1. That works fine. The master has a dedicated IP of 192.168.1.2 and the backup has an IP of 192.168.1.3. This is normal. Dedicated IPs are a requirement because the carp participants need IP connectivity before they can use in the carp protocol to establish the shared address.
BUT, in this configuration, you cannot ping or otherwise connect to 192.168.1.3 unless you are within the 192.168.1.0 subnet. So if you have a management vlan that you are trying to access externally, you can easily find yourself in a situation where you are unable to connect to the backup host at all. You would need a bastion on the mgmt subnet or be on that subnet yourself to administer the backup carp host.
This is also the case for the WAN address, except you can connect to the backup host from anywhere externally. But if you tried to connect to it from within the LAN, it wonāt work. So if I have 192.0.0.0/29 for my WAN, and carp is 192.0.0.1, master host is 192.0.0.2 and backup host is 192.0.0.3, I can connect to any of those over WAN (remotely), but I cannot connect to 192.0.0.3 from within the LAN.
So what to do about this? Iām not completely sure, but I think Iāll try setting up /30 mgmt subnets for each carp host. Kind of awkward but should work?
I acquired a ~5 year old X1 Carbon which means I have a Linux ādesktopā, and a Linux laptop for the first time
Hyprland is very cool. Asymmetrical multi-monitor configurations are awkward but otherwise no complaints so far. Also trying to use eww
which has a steeper learning curve but should ultimately pay off. Hypothetically, it can be used for bar, launcher and notifications. Not sure Iāll dedicate the time to make all of those happen, but it would be pretty slick. For now Iām just enamored with the silky Wayland animations. Itās the first time the polish on a Linux desktop UX made me feel embarrassed for macOSā¦ of course, Iām always embarrassed for Windows.
Also havenāt gotten sushi or the nemo equivalent working yet in hyprland. Coming from macOS, I use this functionality a lot so hopefully thereās a way to make it work.
I have to have GTK installed anyway unless I want to use Konqueror or something
I used Falcon in the past, the browser that KDE officially accepted in their programs suite and now technically supersedes Konqueror. Falcon used to be named QupZilla.
You would need a bastion on the mgmt subnet
*mulbling about how everyone should have at least 8 ports on their servers*
So what to do about this?
Just SSH to the master via the 1.1 IP, then do an ifconfig or hostname to check on which one youāre connected, then use that to ssh into the other, because theyāre on the same subnet.
Just SSH to the master via the 1.1 IP, then do an ifconfig or hostname to check on which one youāre connected, then use that to ssh into the other, because theyāre on the same subnet.
In general I was trying to avoid SSH forwarding while relying exclusively on my yubikeys but using the routers as bastion hosts would be very convenient since they inherently have access to all the subnets.