Hugging Face's logo
Jun Young Baek's picture
1 2 76

Jun Young Baek

jupiterbjy
21world's profile picture Mi6paulino's profile picture swords's profile picture
·

AI & ML interests

None yet

Recent Activity

liked a model 13 days ago
PowerInfer/SmallThinker-3B-Preview
replied to bartowski's post 21 days ago
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
replied to bartowski's post 22 days ago
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
View all activity

Organizations

None yet

jupiterbjy's activity

liked a model 13 days ago
replied to bartowski's post 21 days ago
view reply

Ah I see, hope so too - Catching both size and speed means alot in mobile! Gotta preemptively thanks your future IQ4_NL quants ;)
replied to bartowski's post 22 days ago
view reply

Interesting, in this case will description "Legacy format, generally not worth using over similarly sized formats" of Q4_0 change to something like "ARM recommended (Do not use in Apple Silicons)" - or will IQ4_NL added in list and recommend that over Q4_0?
reacted to bartowski's post with 👍 22 days ago
view post
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
·
liked 6 models about 1 month ago
liked a Space about 1 month ago
liked 8 models about 1 month ago
liked a model about 2 months ago