UD stands for "Unsloth-Dynamic" which upcasts important layers to higher bits. Non UD is just standard llama.cpp quants. Both still use our calibration dataset.
Please consider authoring a single, straightforward introductory-level page somewhere that explains what all the filename components mean, and who should use which variants.
The green/yellow/red indicators for different levels of hardware support are really helpful, but far from enough IMO.
Is there some indication on how the different bit quantization affect performance? IE I have a 5090 + 96GB so I want to get the best possible model but I don't care about getting 2% better perf if I only get 5 tok/s.
It takes download time + 1 minute to test speed yourself, you can try different quants, it's hard to write down a table because it depends on your system ie. ram clock etc. if you go out of gpu.
I guess it would make sense to have something like max context size/quants that fit fully on common configs with gpus, dual gpus, unified ram on mac etc.
My experience with IRC has always been the same, including present day libera.
Almost all the channels are dominated by a few terminally-online people with zero emotional intelligence and the biggest god complexes you've ever seen. Everything is black-and-white, and daring to suggest otherwise just gets you attacked.
Some will say "just use /ignore", but that's not very helpful when most conversations always involve said problematic people and their walls of text... you just constantly see one-sided conversations now.
> They have no backend systems access and just run through the AWS equivalent of "reboot it", "defrag your disk"
To be fair I would bet money that the overwhelmingly vast majority of support tickets are exactly those kind of issues, and ones that refer to actual bugs on their end are, comparatively, extremely rare, and should have to be escalated through normal procedures to weed out common problems.
I know I will get downvoted for this and people will just say "you're doing it wrong" or "the library wasn't designed for that" but I still think it would be really helpful for some people if the source of the data to swap out could be retrieved from parsing the response data as JSON and grabbing a certain named key.
Yes one can add a callback function to parse the data as JSON, do error handling etc. and then swap the text/HTML, but having that built in to a hx-foo attribute would be a lot easier.
Htmx has events you can listen to like htmx after response. You can think of it almost like a middleware. After the response comes in, your callback is triggered and you can make the callback look up some attribute given the calling parents attribute that you might call hx-json-key.
Yes you have to add this yourself, but you only need to add one js function once and be done with it.
I've used the callback pattern for custom error handling for all hx responses.
Re-using existing API calls that I can't change, where I want to replace the text of an html element with the text from a specific key in a json response.
Wireguard to where? Another ISP/VPN that can also sell/MITM your traffic just as well? Non-residential exit IPs are also very often blocked by many websites.
Residential ISPs are well setup for monitoring home traffic (and legally required to in most places). A VPS in a different jurisdiction is vastly less likely to be good at it.
You occasionally get blocked, but not that often if you can putup with a few more captchas. Can't remember it ever being more then a minor inconvenience and well worth this cost.
My home internet is 5G, and many, many websites are blocked or have infinite captcha loops... even well-known sites. Etsy is blocked. Reddit/Discord/Locals is blocked. Archive.is only captcha loops. Even libera IRC is blocked. Trying to buy products online often gets the order flagged or canceled as a potential bot or VPN. IPs are rotated often so I unfortunately have to share bad-reputation IPs with people who keep the addresses on global blacklists like DroneBL that are used by many sites. Even 4chan blocks most of the IPs I get because other people post CP from there.
Trying to use a VPS/cloud IP or well-known VPN provider, the experience for me is just as bad or worse.
For some, the issue is a lot worse than you think.
If you think the courts are going to keep letting it happen anyway without any consequences... have some faith. Justice moves slowly but hope is not lost.
Thanks for posting that link, but after reading it, I'm not nearly as hopeful as you are:
> Also on January 12, 2026, the Court ordered Respondents to respond to the Petition by January 15, 2026 at 11:00 a.m., certifying the true cause and proper duration of Petitioner’s confinement and showing cause as to why the writ should not be granted in this case. Respondents entered an appearance that day but have not filed any response to the Petition.
The government didn't even bother responding to the habeas petition because they knew what they were doing was unconstitutional, but it still successfully sows fear.
Worse, ICE detained the man again less than 24 hours after he was released on the judge's orders and called it a "mistake": https://youtu.be/jmoF63Msk0Y
I know who the domestic terrorists are, and it sure as hell isn't Renee Good or Alex Pretti.
I hope you are right, but the administration does not tend to follow court orders. They break the law and then use every legal mechanism at their disposal to slow walk cases through the system. And then, as I say, they do not follow rulings except under the most extreme pressure.
I have about as much faith in the courts ability to actually walk any of this back as I do in their ability to return the families they have kidnapped from my community.
I'll relax when it gets through SCOTUS (who has overturned nearly 90% of lower conservative court rulings to be in Trump's favor) and not a moment before.
Source:
reply