bartowski/phi-4-GGUF
Text Generation
β’
Updated
β’
20.9k
β’
35
rtuuuuuuuur
urtuuuu
AI & ML interests
None yet
Recent Activity
liked
a model
4 days ago
bartowski/phi-4-GGUF
liked
a model
4 days ago
microsoft/phi-4
new activity
16 days ago
deepseek-ai/DeepSeek-V3:How can we thank you enough, whale bros?
Organizations
None yet
urtuuuu's activity
liked
2
models
4 days ago
How can we thank you enough, whale bros?
10
#1 opened 18 days ago
by
KrishnaKaasyap
liked
a
model
17 days ago
GGUFs eventually ?
6
#1 opened 21 days ago
by
HMasaki
A bit annoying, isn't it? Some time ago I asked you for arm version of gemma-2-9b-it-abliterated. So now it won't work again. I guess there is no Q4_0 ?
reacted to
bartowski's
post with π
20 days ago
Post
15876
Looks like Q4_0_N_M file types are going away
Before you panic, there's a new "preferred" method which is online (I prefer the term on-the-fly) repacking, so if you download Q4_0 and your setup can benefit from repacking the weights into interleaved rows (what Q4_0_4_4 was doing), it will do that automatically and give you similar performance (minor losses I think due to using intrinsics instead of assembly, but intrinsics are more maintainable)
You can see the reference PR here:
https://github.com/ggerganov/llama.cpp/pull/10446
So if you update your llama.cpp past that point, you won't be able to run Q4_0_4_4 (unless they add backwards compatibility back), but Q4_0 should be the same speeds (though it may currently be bugged on some platforms)
As such, I'll stop making those newer model formats soon, probably end of this week unless something changes, but you should be safe to download and Q4_0 quants and use those !
Also IQ4_NL supports repacking though not in as many shapes yet, but should get a respectable speed up on ARM chips, PR for that can be found here: https://github.com/ggerganov/llama.cpp/pull/10541
Remember, these are not meant for Apple silicon since those use the GPU and don't benefit from the repacking of weights
Before you panic, there's a new "preferred" method which is online (I prefer the term on-the-fly) repacking, so if you download Q4_0 and your setup can benefit from repacking the weights into interleaved rows (what Q4_0_4_4 was doing), it will do that automatically and give you similar performance (minor losses I think due to using intrinsics instead of assembly, but intrinsics are more maintainable)
You can see the reference PR here:
https://github.com/ggerganov/llama.cpp/pull/10446
So if you update your llama.cpp past that point, you won't be able to run Q4_0_4_4 (unless they add backwards compatibility back), but Q4_0 should be the same speeds (though it may currently be bugged on some platforms)
As such, I'll stop making those newer model formats soon, probably end of this week unless something changes, but you should be safe to download and Q4_0 quants and use those !
Also IQ4_NL supports repacking though not in as many shapes yet, but should get a respectable speed up on ARM chips, PR for that can be found here: https://github.com/ggerganov/llama.cpp/pull/10541
Remember, these are not meant for Apple silicon since those use the GPU and don't benefit from the repacking of weights
-
- 15 replies
Notably better than Phi3.5 in many ways, but something is wrong.
8
#5 opened 27 days ago
by
phil111
llama.cpp...
1
#1 opened about 1 month ago
by
urtuuuu
liked
a
model
about 1 month ago
Help with llama
3
#1 opened about 1 month ago
by
urtuuuu
Misleading advertisement as reasoning model
8
#5 opened about 2 months ago
by
esozbek
First impression: Impressed
1
#2 opened 3 months ago
by
urtuuuu
liked
a
model
3 months ago
liked
2
models
3 months ago
Observation: 4-bit quantization can't answer the Strawberry prompt
12
#2 opened 3 months ago
by
ThePabli
there are 3 "r"s in the playful "strawrberry"?
5
#6 opened 3 months ago
by
JieYingZhang
ARM version?
1
#1 opened 3 months ago
by
urtuuuu