ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 19:31:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

Adding IQ1_KT - 1.75 bpw SOTA quants (#616)

Browse Source 
* iq1_kt: basics

* iq1_kt: CUDA dequantize

Testing with LlaMA-3.1-8B-Instruct, we get almost the same PPL
as iq2_xxs, so about 0.2 bpw fewer bits for the same quality.

* iq1_kt: CUDA MMQ

* iq1_kt: CUDA MMVQ

* iq1_kt: AVX2 GEMM/GEMV

* iq1_kt: convert/repack to q8_0_r8 (AVX2)

* iq1_kt: slightly faster GEMV

18.6 t/s -> 19.4 t/s

* iq1_kt: NEON GEMM/GEMV

Pathetic as usual

* iq1_kt: slightly faster NEON - still pathetic

* iq1_kt: tiny bit better GEMV on NEON

* iq1_kt: convert/repack to q8_0_r8 (NEON)

* iq1_kt: very slightly faster convert/repack to q8_0_r8 on NEON

* Adding frgotten file

* iq1_kt: add to constants.py

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-07-20 10:05:23 +02:00
committed by GitHub
parent 07673c6c33
commit f989fb03bd
21 changed files with 930 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
gguf-py/gguf/constants.py
View File

@@ -1322,6 +1322,7 @@ class GGMLQuantizationType(IntEnum):
IQ4_KT = 155
IQ3_KS = 156
IQ2_KL = 157
IQ1_KT = 158
Q4_0_R8 = 202
Q5_0_R4 = 206
Q8_0_R8 = 208
@@ -1539,6 +1540,7 @@ GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = {
GGMLQuantizationType.IQ4_KT : ( 256, 128),
GGMLQuantizationType.IQ3_KS : ( 256, 102),
GGMLQuantizationType.IQ2_KL : ( 256, 86),
GGMLQuantizationType.IQ1_KT : ( 256, 56),
GGMLQuantizationType.Q4_0_R8 : ( 32, 18),
GGMLQuantizationType.Q5_0_R4 : ( 32, 22),
GGMLQuantizationType.Q8_0_R8 : ( 32, 34),