3·
19 hours agoAs a memory-poor user (hence the 8gb vram card), I consider Q8+ to be is higher precision, Q4-Q5 is mid-low precision (what i typically use), and below that is low precision
Xylight
I’m the developer of the Photon client. Try it out
- 6 Posts
- 26 Comments
Joined 2 years ago
Cake day: July 22nd, 2023
You are not logged in. If you use a Fediverse account that is able to follow users, you can follow this user.
It’s a webp animation. Maybe your client doesn’t display it right, i’ll replace it with a gif
Regarding your other question, I tend to see better results with higher params + lower precision, versus low params + higher precision. That’s just based on “vibes” though, I haven’t done any real testing. Based on what I’ve seen, Q4 is the lowest safe quantization, and beyond that, the performance really starts to drop off. unfortunately even at 1 bit quantization I can’t run GLM 4.6 on my system
1·1 day agoi never thought about that. i assumed the first page was just a joke website like “days since last JavaScript framework” always being zero
this made me mad so i made a single, ultra minimal html page in 5 minutes that you can just paste in your url box
data:text/html;base64,PCFkb2N0eXBlaHRtbD48Ym9keSBzdHlsZT10ZXh0LWFsaWduOmNlbnRlcjtmb250LWZhbWlseTpzYW5zLXNlcmlmO2JhY2tncm91bmQ6IzAwMDtjb2xvcjojMmYyPjxoMT5JcyBpdCBETlM/PC9oMT48cCBzdHlsZT1mb250LXNpemU6MTJyZW0+WWVzsource code:
<!doctypehtml><body style=text-align:center;font-family:sans-serif;background:#000;color:#2f2><h1>Is it DNS?</h1><p style=font-size:12rem>Yes
71·10 days agoA lot of LLMs now use intentionally synthesized, or AI generated training data. It doesn’t seem to affect them too adversely.
You can run Qwen3 4b thinking at q4 quantization at 2.5GB, which is probably a better model too
there’s also a “small” and “micro” variant, which are 32b a6b MoE and 3b dense models respectively
rsync for backups? I guess it depends on what kind of backup
for redundant backups of my data and configs that I still have a live copy of, I use restic, it compresses extremely well
I have used rsync to permanently move something to another drive though
I tried this again, with a gguf quantized to Q4_K_M. It works quite well and can generate in ~7 minutes! Thanks!
My guy stop following me around
I don’t trust an external third party to manage the coordination server.
Headscale has an issue open for wireguard only exit nodes though, I guess ill wait for that.
I tried self-hosting tailscale with headscale, but you cannot have a wireguard only exit node with headscale–and so I can’t have mullvad as my exit node.
If it turns on with mobile data automatically, that turns off my Mullvad VPN.
I’d like to try this but 22B is just outside of the range that i can get an acceptable t/s :(
11·1 month agoThe thing with this is that many people want to switch their setting to be non default, and it’s impossible to have both user configurable and nice css at the same time. One has to go.
Edit: Oops, just realized I replied to a 2 month old comment. My bad
GNOME: Designers trying to Develop a desktop. KDE: Developers trying to Design a desktop.
My residential IP (an inconspicuous Comcast network) gets blocked by this garbage on so many websites. I don’t know why VPNs are less suspicious to sites than my random california ip
Irrelevant but its funny how in English “x aren’t shit anymore” can mean it got better or it got worse simultaneously
I encounter VPN blocks everywhere frequently. I usually just reroll my selected server until the block goes away