New SOTA quantization: 4.25 bpw IQ4_KS #83
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
added 13 commits
October 8, 2024 10:52
At the same 4.25 bpw PPL is always better, for some models significantly better. I'll rename to iq4_ks and keep it.
We get TG-128 = 126 t/s for LLaMA-3.1-8B, compared to 123 t/s for q4_0.
Also fix the breakage I caused with the dedicated work buffer quantization portion when the multiplication is not done via iqk_mul_mat.
I noticed that iq4_xs is wrong on Zen4 (and possibly AVX2). Again the same mistake of packing int32_t back to int16_t, which overflows occasionally (just occasionally, that's why the result doesn't look completely wrong, so I didn't notice).
After all, tt is a smaller variant of iq4_k.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
It is similar to
IQ4_Kwith the following differencefloatscale instead of per block instead of per super-blockggml_halfIt ends up being 4.25 bpw, so the same as
IQ4_XS. Why add it then? Because it has a lower quantization error thanIQ4_XS. For some models the difference is quite significant. The following table gives some examples. Quantization errorQerris defined asPPL(Q)/PPL(f16)-1Performance is similar to
IQ4_XSor even slightly better, except for TG on the M2-Max GPU, where it is ~2% slower (Apple Silicon does not like non-sequential memory access, but having the row scale stored at the beginning of the row causes an additional memory jump in the dot product kernel).The PR also adds a new quantization mix -
IQ3_KL(Lfor "large"). It fills the gap betweenIQ4_KandIQ4_K(and nowIQ4_KS). The following graph illustrates where this new mix sits for LLaMA-3.1-8B-Instruct.