librekitty on nostr

Recent Notes

note164ute...

librekitty @librekitty 1775673185

i got confused thinking about silent hill: downpour

note1mrpv5...

librekitty @librekitty 1775671135

i love it!
the difference between 26b (moe) and 31b for writing is pretty big though
but i think i'll use 26b for coding and stuff, faster and more VRAM for context

i'm yet to try the smaller models on my phone, excited to try that too

eardiod · 17h

Did you use the compressed model from nvidia? I so far tried only 4bit og, but this new nvidia compression seems to be even better

librekitty @librekitty 1775670916

my card is too old to run NVFP4 💔
yeah i'm using the classic q4_k_m

eardiod · 9h

makes two of us, mine is also too old :)

Uno · 1d

I'm testing my new client! Specifically DMs, I'm gonna send you a message all you gotta do is reply if you get it.

librekitty @librekitty 1775621907

you got it!

librekitty

@librekitty 1775621892

i finally got gemma4 to fit in VRAM, the KV cache in a brutal size 🤖🙀

quantization is such a blessing 🙏

eardiod · 17h

Did you use the compressed model from nvidia? I so far tried only 4bit og, but this new nvidia compression seems to be even better

Uno · 1d

I need one volunteer to help me test something real quick, if you will be on Nostr for 5 minutes let me know.

librekitty @librekitty 1775621812

what's up?

1🚀1

Uno · 1d

I'm testing my new client! Specifically DMs, I'm gonna send you a message all you gotta do is reply if you get it.

🇵🇸 whoever loves Digit · 1d

The most giga based people have some misspellings on the record because they use nostr where you can't edit podts

librekitty @librekitty 1775603975

more clients need to support edits
or standardize them

🇵🇸 whoever loves Digit · 1d

But let us see pre edited versions of course

Stacking Functions · 1d

+1 for newpipe. Grayjay is also worth a look since you get access to other libraries besides youtube, like odysee, rumble, bitchute, podcasts, etc

librekitty @librekitty 1775602882

iirc Grayjay is proprietary software

GrapheneOS · 1d

nostr:nprofile1qy2hwumn8ghj7un9d3shjtnyd968gmewwp6kyqpqwp2lcskf5cu2k44dq4cpdgk7lt8f0d0wecc68fayvun6mprcn8jsdf6gu6 nostr:nprofile1qy2hwumn8ghj7un9d3shjtnyd968gmewwp6kyqpq0r8xl2njyepcw2zwv3a6dyufj4e4ajx86hz6v4ehu4gnpupxxp7spmxr3j nostr:nprofile1qy2hwumn8ghj7un9d3shjtnyd968gmewwp6kyqpqjfn4ghffz7uq7urll...

librekitty · 2d

anyone else still need to use tricks to remember how to spell? 🔡 to this day i still spell restaurant 🍽️ as "rest😴 + aura✨ + nt🧠"

librekitty @librekitty 1775536309

btw, on an unrelated spelling note, homeowner is secretly hoMEOWner 🏠🐈‍⬛

1❤️1💜1

blackcat · 1d

🐈‍⬛🖤

librekitty

@librekitty 1775536231

anyone else still need to use tricks to remember how to spell? 🔡

to this day i still spell restaurant 🍽️ as "rest😴 + aura✨ + nt🧠"

101❤️3💜2🤙1

librekitty · 2d

titus · 1d

Jim Carrey taught me how to spell beautiful when I saw Bruce Almighty as a youngster https://www.youtube.com/watch?v=6K3UpktQH9w&t=25

Kayne · 1d

I pronounce words in my mind how they are spelt, silent letters aren't silent, and anything that's spelt with re instead of er for instance, like centre or litre it's re instead of er when I hear the word in my mind.

skepticles · 1d

“The way I remember my name is-“ -Luanne Platter

🇵🇸 whoever loves Digit · 1d

The most giga based people have some misspellings on the record because they use nostr where you can't edit podts

librekitty · 2d

librekitty @librekitty 1775518139

you can also go the CPU route with tons of RAM, but inference speed will be terrible compared to GPU accellerated

nicodemus · 1d

This is true, GPUs are faster for inference. But you'll also be consuming 1500 watts, have to deal with those thermal issues, and still struggle to fit a model larger than 32B with decent quantization. Alternatively, the 395 chips and their NPU are doing pretty good. Combine 2 of them and you're lo...

Gigi · 2d

considering buying hardware to run everything locally. Would should I buy? #asknostr

librekitty @librekitty 1775518007

intel is seriously competitive for price-to-VRAM, but i don't know about compatibility

NVIDIA is usually the clear winner for performance, 5xxx series/blackwell has support for NVFP4 quantized models
https://developer.nvidia.com/blog/introducing-nvfp4-for-efficient-and-accurate-low-precision-inference/

but you could also do like, multiple 3090s or something

hope this helps

librekitty · 2d

you can also go the CPU route with tons of RAM, but inference speed will be terrible compared to GPU accellerated

zaytun · 1d

I think dual 3090s would be preferable to fx a dgx spark with regards to inference speed, no? vRAM speed is higher I believe. Downside is model size limit is obviously lower on 48 gb vRAM than 128gb unified of the dgx spark.