diff options
| author | Kawrakow <iwankawrakow@gmail.com> | 2025-02-06 14:08:52 +0200 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2025-02-06 14:08:52 +0200 |
| commit | 7f61b3068e18728e5e7e2b95546ff03dd2fd41ac (patch) | |
| tree | f175a942a6ebd2d2d8b08c46fa71d9f6fbad50e7 /ggml/src/ggml-common.h | |
| parent | a6f9f2ec9af92b5a13f035db054aac2fd2efaee7 (diff) | |
IQ1_M_R4: better 1.75 bpw quants (#187)
* iq1_m_r4: basics (quantize/dequantize)
* iq1_m_r4: Zen4 gemm
* iq1_m_r4: neon gemm
* iq1_m_r4: switch to q8_0_x4 also on AVX2/Zen4
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.
* iq1_m_r4: rename mul_mat_iq1_m_r4_q8_1 to mul_mat_iq1_m_r4_q8_0
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Diffstat (limited to 'ggml/src/ggml-common.h')
| -rw-r--r-- | ggml/src/ggml-common.h | 8 |
1 files changed, 8 insertions, 0 deletions
diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h index 14813161..679353be 100644 --- a/ggml/src/ggml-common.h +++ b/ggml/src/ggml-common.h @@ -499,6 +499,14 @@ typedef struct { } block_iq1_m; static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32, "wrong iq1_m block size/padding"); +// 1.75 bpw - blocks of 32 with 4 interleaved rows = 128 quants +typedef struct { + uint8_t qs[16]; // grid index, low 8 bits + uint8_t qh[ 8]; // grid index, high 3 bits + grid shift bits (for two groups of 8) + uint8_t scales[4]; // 4-bit block scales +} block_iq1_m_r4; +static_assert(sizeof(block_iq1_m_r4) == 28, "wrong iq1_m_r4 block size/padding"); + // // Bitnet and TriLM - implemented as 1.625 bpw // |