Adding IQ1_KT - 1.75 bpw SOTA quants (#616)

* 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>

1 files changed, 7 insertions, 0 deletions

diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h
index 6dc439b8..1dc1ff6e 100644
--- a/ggml/src/ggml-common.h
+++ b/ggml/src/ggml-common.h
@@ -630,6 +630,13 @@ typedef struct {
 static_assert(sizeof(block_iq2_ks) == sizeof(uint16_t) + QK_K/64 + QK_K/4, "wrong iq2_ks block size/padding");
 
 typedef struct {
+    uint8_t  sh[QK_K/32]; // 4-bit scales + 13th bits for groups of 8
+    uint8_t  ql[QK_K/8];  //  low 8 bits for groups of 8
+    uint8_t  qh[QK_K/16]; // high 4 bits for groups of 8
+} block_iq1_kt;
+static_assert(sizeof(block_iq1_kt) == QK_K/8 + QK_K/16 + QK_K/32, "wrong iq1_kt block size/padding");
+
+typedef struct {
     uint8_t  scales[QK_K/64];
     uint8_t  ql[QK_K/4];
 } block_iq2_kt;