IQ1_M_R4: better 1.75 bpw quants #187
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 5 commits
February 5, 2025 16:59
With the deltas being per group of 8, we cannot make use of the q8 sums stored in q8_1, so we get a tiny gain by using q8_0_x4.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.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.
Following in the foot steps of #185, this PR adds
IQ1_M_R4, a 4-row interleaved version ofIQ1_M.f16super-block scale (replaced with af16per row scale) and have changed the 3-bitIQ1_Mblock scales with 4 bit. Hence, we end up using the same 1.75 bpw asIQ1_M.IQ1_M_R4with a block size of 32. I wanted to have this because DeepSeek-Lite, the model I'm testing with, has a lot of tensors with row sizes not divisible by 256, so a significant fraction of tensors gets quantized toIQ4_NLwhen usingIQ1_Mllama.cpparrives at a context-512 perplexity (PPL(512)in what follows) of 20.75 for DeepSeek-Lite using 2.74 bpw withIQ1_M. TheIQ1_M_R4quantization in this PR getsPPL-512 = 8.85with 1.966 bpw for the repeating layers.IQ1_M_R4is much faster on the CPU compared toIQ1_M(see tables below). I never implemented iqk-style GEMM forIQ1_S/IQ1_M, so these quantization types run at the snail speed of mainlinellama.cpp.The following table compares prompt processing (pp512) and token generation (tg128) speed for LLaMA-3.1-8B on
AVX2(Ryzen-5975WX),Zen4(Ryzen-7950X) andARM_NEON(M2-Max CPU). I didn't use DeepSeek-Lite for this comparison to avoid the difference in quantization types one ends up with due to not all tensors having row sizes that are multiple of 256.