GNU social JP
  • FAQ
  • Login
GNU social JPは日本のGNU socialサーバーです。
Usage/ToS/admin/test/Pleroma FE

  • Public

    • Public
    • Network
    • Groups
    • Featured
    • Popular
    • People

Notices by Wanderer über dem Nebelmeer (wandereruber@poa.st), page 2

  1. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Saturday, 13-Jun-2026 21:51:02 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    in reply to
    • SuperDicq
    • kaia
    🤓
    uhm, actually 7B is the parameter count, not the quantization.
    A quantization to fit Fable into 7 GB ram would still have as many parameters. It would be a Q0.04, i.e. 0.04 bits per parameter
    In conversation about 22 days ago from poa.st permalink
  2. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Saturday, 13-Jun-2026 19:37:02 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    in reply to
    • SuperDicq
    • kaia
    that is how that works
    In conversation about 22 days ago from poa.st permalink
  3. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Friday, 12-Jun-2026 11:23:45 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    in reply to
    • Big Booty Monster Girls
    • Waldbrand
    >likes the Idea of keeping his flaccid one in the Vag for a long time after the act
    legend
    In conversation about 23 days ago from poa.st permalink
  4. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Friday, 12-Jun-2026 10:38:07 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    Thank you, Pewdiepie. You are far too kind.
    In conversation about 23 days ago from poa.st permalink

    Attachments


    1. https://i.poastcdn.org/fd/d5/89/fdd5891be40868057b1802718df1ac23a25258f389010792165ea76005f7930f.JPG
  5. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Friday, 12-Jun-2026 10:38:06 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    in reply to
    • <- ?? Striderpod ?? ->
    • verita84
    He doesn't need to because, unlike you, he didn't make the decision to roll the backend into this tool. Quite frankly, your decision to do this still seems absurd to me.

    I have it running on my GPU right now.

    >agentic hard with quants
    could you quantify this?
    In conversation about 23 days ago from poa.st permalink
  6. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Friday, 12-Jun-2026 10:38:06 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    in reply to
    • <- ?? Striderpod ?? ->
    • verita84
    best-in-class. Not aware of anything better than prometheus.
    I will also try @verita84_64 's kit.
    a bit buggy. I had to fix some annoyances. The biggest was that it would truncate messages and sometimes get context-size wrong. Deepseek V4 flash + opencode did it in minutes. If you need details later, just ask. I can make a patch file.

    I'm running Qwen3.6-35B-A3B-Q6 on pic rel
    In conversation about 23 days ago from poa.st permalink

    Attachments


    1. https://i.poastcdn.org/d0/bb/0c/d0bb0c3ba94b6da9b40b5201dc6ff9f09d8aaabcda872f211d073000ad891822.jpg
  7. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Friday, 12-Jun-2026 10:38:05 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    in reply to
    • <- ?? Striderpod ?? ->
    • verita84
    please don't vaguepost at me like this.

    >It all started with getting the Intel Arc to work properly on Gentoo without Docker and with newer support for images.
    I seriously doubt that you get significantly higher inference speeds than llamacpp or the forks that are out there, by rolling your own backend.

    >You have to do some guidance on the backend to get models to code properly. Think of it as holding their hand. The lower the model is, say 9b, the more help it needs.
    I don't appreciate being talked to this way. We're not on reddit and I'm not five years old. Besides that, Gemma4 E4B is decent at it. The benchmarks are out there, this is not some secret sauce.

    >Also, since I made the backend, I can handle large source code files on small GPU's. Nobody does that
    I can run 200k context with Q3.6-35B on beellama and use it in opencode/pi no problem. Again, please quantify what you're saying.
    In conversation about 23 days ago from poa.st permalink
  8. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Friday, 12-Jun-2026 10:38:04 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    in reply to
    • <- ?? Striderpod ?? ->
    • verita84
    You're really doing your software a disservice by talking nonsense like this, I gotta be real here.

    2. Please explain the technical details of what we talked about instead of copy-pasting LLM snippets at me. this is very disrespectful and you're wasting my time.

    1. If you are using llamacpp anyways, why the fuck does your docker container ship with it. That's literally my only point dude. Pewdiepie does not have to care about getting GPU performance code right because he doesn't ship any. So I could only guess you were implying you're doing it differently. Now you're telling me you also don't even modify it. So I was right! it *IS* an absurd decision, making the user install llamacpp in the docker container even if he has inference up and running elsewhere. Nobody does it like this. Not anythingllm, not openwebui, not sillytavern, not prometheus, not even fucking oobabooga.

    3. once again, how exactly did you "design it", if the inference is done by an external codebase anyways? Please quantify this.
    In conversation about 23 days ago from poa.st permalink
  9. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Friday, 12-Jun-2026 10:38:03 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    in reply to
    • <- ?? Striderpod ?? ->
    • verita84
    >I am quoting Claude

    >I never gave you LLM snippets.


    I think I just about had enough of this nonsense convo.


    >Go use PewDiePie's software lol.
    I am using it right now. You know why? because (halfway down the page, mind you) your github readme tells me to install gigabytes of shit in docker and it's a bit of a hassle to do make space for this in my vm for unidentified benefit.
    In conversation about 23 days ago from poa.st permalink
  10. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Friday, 12-Jun-2026 08:42:42 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    in reply to
    • Waldbrand
    @Waldbrand The imouto-firster is a groveling plebeian, bickers with childhood friend pleb, etc.

    Mommychads tower above all, the aristocratic few. It's that simple.
    In conversation about 23 days ago from poa.st permalink

    Attachments


    1. https://i.poastcdn.org/a6/1f/a0/a61fa085222120016b04569790d70f859fbab597c5f3f5a939e6bb7ab3a3f639.png
  11. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Thursday, 11-Jun-2026 22:22:58 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    found something funny on reddit
    atheist = racist
    In conversation about 24 days ago from poa.st permalink

    Attachments


    1. https://i.poastcdn.org/10/57/8b/10578b4de5d500c32a28e32a09f3a35a99166830ddb4fc7ed587106f1d3db432.JPG

    2. https://i.poastcdn.org/46/35/db/4635dbacc82da2cf8112cbb9931c5b7541f0367850ae8fc0eb0e22da25369cea.JPG
  12. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Thursday, 11-Jun-2026 11:42:55 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    in reply to
    • lainy
    • bronze
    • binkle
    • Wanderer über dem Nebelmeer
    • Blurry Moon
    update:
    Running beellama (cuda) on the same config is faster than llama-cpp vulkan, which is already faster than vanilla llama-cpp cuda.

    I can't use TurboQuant because it's slower. It needs cpu-moe = true and apparently my cpu is NOT moe. Costs me ~15% t/s

    I have not had ANY success with the dflash drafting. Logs show a lot of rejections. Maybe that's it. It's slow.


    The absolute kicker why I will keep using it:
    A 3X increase in prompt processing speed, that's on top of the inference speed increase. I have no idea what they did to achieve this.
    In conversation about a month ago from gnusocial.jp permalink
  13. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Thursday, 11-Jun-2026 11:42:55 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    in reply to
    • lainy
    • bronze
    • binkle
    • Blurry Moon
    saw your post the other day, coincidentally testing that right now.

    Cuda runs ~1t/s slower than Vulkan on vanilla cpp for my old Nvidia card, so I'll have to see.

    >large context
    I don't know why people have such trouble with this. I run
    swa-full = true
    context-shift = true
    ctx-size = 64000
    on Gemma

    and since a few updates ago that throws an error for Q3.6 so I just turned it off and
    context-shift = false
    swa-full = false
    ctx-size = 131072

    I think I could probably bump the Gemma numbers but I don't like 26b for coding so I had no need.
    In conversation about a month ago from poa.st permalink
  14. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Thursday, 11-Jun-2026 00:42:12 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    in reply to
    • lainy
    very disingenuous but I respect the grind
    In conversation about a month ago from poa.st permalink
  15. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Wednesday, 10-Jun-2026 22:11:21 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    in reply to
    artificialanalysis.ai for benchmarks

    from huggingface
    Qwen3.6 35B A3B (It's good enough for a quick script)
    Gemma4 26B A4B
    (maybe + heretic)

    Ollama does not show you default quantization+params. I found it too opaque in general.
    I personally use llama-cpp
    with --models-preset /path/to/a-config.ini
    In conversation about a month ago from gnusocial.jp permalink
  16. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Wednesday, 10-Jun-2026 21:47:35 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    >The Crazy physics of launching at 80 km/h from a truck already moving 80km/h
    the physics are not that crazy, but this is very cool
    In conversation about a month ago from poa.st permalink

    Attachments


  17. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Wednesday, 10-Jun-2026 21:23:05 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    in reply to
    • lainy
    I don't think they did.
    It's about ethics in video game journalism.
    In conversation about a month ago from poa.st permalink

    Attachments


    1. https://i.poastcdn.org/d6/dc/1c/d6dc1c36cc2d74f4f15d6ea112a903cf7bde1aab325732ff3a696ffc58a13593.JPG
  18. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Wednesday, 10-Jun-2026 02:39:53 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    in reply to
    • lainy
    "claude, write both" not an option?
    In conversation about a month ago from poa.st permalink
  19. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Wednesday, 10-Jun-2026 02:32:49 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    in reply to
    • lainy
    [something about noble bloodlines and golden hair and elf ears and zelda]
    In conversation about a month ago from poa.st permalink
  20. Embed this notice
    Wanderer über dem Nebelmeer (wandereruber@poa.st)'s status on Wednesday, 10-Jun-2026 02:30:48 JST Wanderer über dem Nebelmeer Wanderer über dem Nebelmeer
    in reply to
    • lainy
    giwtwm
    In conversation about a month ago from poa.st permalink
  • After
  • Before

User actions

    Wanderer über dem Nebelmeer

    Wanderer über dem Nebelmeer

    I hate niggers

    Tags
    • (None)

    Following 0

      Followers 0

        Groups 0

          Statistics

          User ID
          202805
          Member since
          20 Oct 2023
          Notices
          1741
          Daily average
          2

          Feeds

          • Atom
          • Help
          • About
          • FAQ
          • TOS
          • Privacy
          • Source
          • Version
          • Contact

          GNU social JP is a social network, courtesy of GNU social JP管理人. It runs on GNU social, version 2.0.2-dev, available under the GNU Affero General Public License.

          Creative Commons Attribution 3.0 All GNU social JP content and data are available under the Creative Commons Attribution 3.0 license.