• 0 Posts
  • 143 Comments
Joined 2 years ago
cake
Cake day: July 14th, 2023

help-circle
  • Look up “LLM quantization.” The idea is that each parameter is a number; by default they use 16 bits of precision, but if you scale them into smaller sizes, you use less space and have less precision, but you still have the same parameters. There’s not much quality loss going from 16 bits to 8, but it gets more noticeable as you get lower and lower. (That said, there’s are ternary bit models being trained from scratch that use 1.58 bits per parameter and are allegedly just as good as fp16 models of the same parameter count.)

    If you’re using a 4-bit quantization, then you need about half that number in VRAM. Q4_K_M is better than Q4, but also a bit larger. Ollama generally defaults to Q4_K_M. If you can handle a higher quantization, Q6_K is generally best. If you can’t quite fit it, Q5_K_M is generally better than any other option, followed by Q5_K_S.

    For example, Llama3.3 70B, which has 70.6 billion parameters, has the following sizes for some of its quantizations:

    • q4_K_M (the default): 43 GB
    • fp16: 141 GB
    • q8: 75 GB
    • q6_K: 58 GB
    • q5_k_m: 50 GB
    • q4: 40 GB
    • q3_K_M: 34 GB
    • q2_K: 26 GB

    This is why I run a lot of Q4_K_M 70B models on two 3090s.

    Generally speaking, there’s not a perceptible quality drop going to Q6_K from 8 bit quantization (though I have heard this is less true with MoE models). Below Q6, there’s a bit of a drop between it and 5 and then 4, but the model’s still decent. Below 4-bit quantizations you can generally get better results from a smaller parameter model at a higher quantization.

    TheBloke on Huggingface has a lot of GGUF quantization repos, and most, if not all of them, have a blurb about the different quantization types and which are recommended. When Ollama.com doesn’t have a model I want, I’m generally able to find one there.


  • I recommend a used 3090, as that has 24 GB of VRAM and generally can be found for $800ish or less (at least when I last checked, in February). It’s much cheaper than a 4090 and while admittedly more expensive than the inexpensive 24GB Nvidia Tesla card (the P40?) it also has much better performance and CUDA support.

    I have dual 3090s so my performance won’t translate directly to what a single GPU would get, but it’s pretty easy to find stats on 3090 performance.


  • To be clear, I’m measuring the relative humidity of the air in the drybox at room temp (72 degrees Fahrenheit / 22 degrees Celsius), not of the filament directly. You can use a hygrometer to do this. I mostly use the hygrometer that comes bundled with my dryboxes (I use the PolyDryer and have several extra PolyDryer Boxes, but there are much cheaper options available) but you can buy a hygrometer for a few bucks or get a bluetooth / wifi / connected one for $15-$20 or so.

    If you put filament into a sealed box, it’ll generally - depending on the material - end up in equilibrium with the air. So the measurement you get right away will just show the humidity of the room, but if the filament and desiccant are both dry, it’ll drop; if the desiccant is dry and the filament is wet, it’ll still drop, but not as low.

    Note also that what counts as “wet” varies by material. For example, from what I’ve read, PLA can absorb up to 1% or so of its mass as moisture, PETG up to 0.2%, Nylon up to 7-8%… silica gel desiccant beads up to 40%. So when I say they’ll be in equilibrium, I’m referring to the percentage of what that material is capable of absorbing. It isn’t a linear relationship as far as I know, but if it were, that would mean that: if the humidity of the air is 10% and the max moisture the material could retain is 1%, then the material is currently retaining 0.1% moisture by mass. If my room’s humidity is kept at 40%, it’ll absorb moisture until it’s at 0.4% moisture by mass.

    That said, this doesn’t measure it perfectly, since while most filament materials absorb moisture from the air when the humidity is higher, they don’t release it as easily. Heating it both allows the air to hold more moisture and allows the filament (and desiccant) to release more moisture.



  • What have you done to clean the bed? From the link to the textured sheet, you should be cleaning it between every print - after it cools - with 90% IPA, and if you still have adhesion issues, you should clean it with warm water and a couple drops of dish soap.

    Has the TPU been dried? I don’t normally print with TPU but my understanding is that it needs to be lower humidity than PLA; I use dryboxes for PLA and target a humidity of 15% or lower and don’t use them if they raise above 20%. The recommendation I saw for TPU was to dry it for 7 hours at 70 degrees Celsius, to target 10% humidity (or at least under 20%) and to print directly from a drybox. Note that compared to other filaments, TPU can’t recover as well from having absorbed moisture - if the filament has gotten too wet, it’ll become too brittle if you dry it out as much as is needed. At that point you would need to start with a fresh roll, which would ideally go into a dryer and then drybox immediately.

    You should be able to set different settings for the initial layer to avoid stringing, i.e., slower speeds and longer retraction distance. It’s a bit more complicated but you can also configure the speed for a specific range of layers to be slower - i.e., setting it to slow down again once you get to the top of the print. For an example of that, see https://forum.prusa3d.com/forum/prusaslicer/bed-flinger-slower-y-movement-as-function-of-z/

    What’s the max speed you’re printing at? My understanding is that everything other than travel should all be the same speed at a given layer, and no higher than 25 mm/s. And with a bed slinger I wouldn’t recommend a much higher travel, either.

    In addition to a brim, have you tried adding supports?








  • From the feature comparison at https://github.com/meichthys/foss_note_apps only two FOSS apps support handwriting: Joplin (with a plugin) which gets a subjective 6/10 score, and TriliumNext, which gets a subjective 2/10 score. I personally dislike Joplin but many people love it, so I recommend giving it a shot. EDIT: I installed Joplin using the APK from the site and both the handwriting and Excalidraw plugins were “not available on mobile,” so I have to rescind my recommendation. On my iOS device, the plugins didn’t even show up in the search.

    I think TriliumNext is great, but the mobile experience is still lacking (though they are tracking several issues to improve here). There’s no dedicated mobile app but they at least have a PWA. It also needs to be self-hosted, but doing so is straightforward if you’re already using Docker. The handwriting is done via a built-in Excalidraw integration.

    Here are some options not captured in that list:

    Obsidian is not open source, but also has an Excalidraw plugin. I’ve not used it yet but I’ve seen multiple discussions saying that it’s very well done and has additional functionality on top of base Excalidraw. There’s also an open source (MIT) plugin for Obsidian that adds support for handwritten notes. I only use Obsidian on my work computer and haven’t used it either, though I plan to install the Excalidraw plugin Monday.

    StylusLabs Write is FOSS (AGPL 3.0), multiplatform, and has a free Android apk available. Note that the Google Play version has had updates suspended. I just learned about it and don’t know how it otherwise measures up. I’m planning to check it out, though.

    You can use any note app that has Excalidraw support, so long as you don’t need your handwritten text to be OCRed. That means that the following are all options:




  • They put their repo first on the list.

    Right. And are we talking about the list for OBS or of repos in general? I doubt Fedora sets the priority on a package level. And if they don’t, and if there are some other packages in Flathub that are problematic, then it makes sense to prioritize their own repo over them.

    That said, if those problematic packages come from other repositories, or if not but there’s another alternative to putting their repo first that would have prevented unofficial builds from showing up first, but wouldn’t have deprioritized official, verified ones like OBS, then it’s a different story. I haven’t maintained a package on Flathub like the original commenter you replied to but I don’t get the impression that that’s the case.



  • I primarily use Standard Notes. It’s a fantastic tool and I can use it anywhere, online or offline. It’s not great for collaboration, though, and it doesn’t have a canvas option. But I use it for scratch pads, for todo lists, for project tracking, for ideas, plans, plotting for my tabletop (Monster of the Week) game, software design and architecture, for drafting comments, etc…

    Standard Notes also has a ton of options for automated backups. I get a daily email with a backup of my notes; I can host my notes on my home server and the corporate one; I can also set up automated backups on any desktop.

    I don’t use it for saving links. I’m still using Raindrop.io for that, even though I’m self-hosting both Linkding and Linkwarden.

    For sharing and collaboration, I either publish to Listed with Standard Notes or use Hedgedoc, which is great for collaboration and does a great job presenting nodes, too.

    For canvas notes, I use GoodNotes on a tablet or the Onyx Boox’s default Notes app. I’d love a better FOSS, self-hosted option, especially for the Boox, but my experiences thus far have been negative (especially on the Boox).

    I’ve been trying out SilverBullet lately, since I want to try out cross-note querying and all that, but I’m too stuck in my habits and keep going back to Standard Notes. I think I’ll have better luck if I choose one app and go with it.

    I also have a collection of Mnemosyne notebooks that I use with fountain pens (mostly the Lamy 2000, but also quite commonly a Platinum 3776 or a Twsbi). Side note: the Lamy 2000 was my first fountain pen and after getting it I went deep into fountain pens. I explored a ton of different options, found a lot of nice pens across a number of brands… and yet how I still haven’t found something that I consistently like more. The Pilot VP is great but deceptive; a fancy clicky pen that only holds 30 minutes of ink (in a converter, at least) is decidedly inconvenient.

    I’ve also been checking out Obsidian on my work computer. So far I haven’t seen anything that makes me prefer it over my existing set of tools.


  • Hedgedoc is fantastic. If you’re okay with your notes app being web-only (without an app or even a PWA) and you don’t need canvas notes or multi-note queries, you should check it out.

    First, every note is Markdown, but it supports a ton of things natively. It has native Vim, Emacs, and Sublime (the default) editors and it’s built to be great for collaboration (if you want).

    It also has

    • syntax highlighting for a ton of languages
    • Mermaid.js support
    • LaTeX support
    • easy drag and drop image uploads
    • a solid mobile interface (for a webapp in your browser, at least)
    • built in revision history
    • support for other diagram tools, like graphviz, flowchart.js
    • a bunch of other little Markdown enhancements that make using it feel oddly intuitive

    And best of all, they have a Hedgehog for the icon! (I may be biased.)