diff options
| author | Kawrakow <iwankawrakow@gmail.com> | 2024-12-20 09:12:48 +0100 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2024-12-20 09:12:48 +0100 |
| commit | f4170f78bdfe69050c60bf3b51370d6c97593489 (patch) | |
| tree | 99f68c0b9eeeeba657d80b6c5247d7efee8c840f /ggml/src/ggml-common.h | |
| parent | dfa12b7f91b9d71bc11f5042151bde949da04d61 (diff) | |
IQ3_XXS_R4 (#153)
* iq3_xxs_r4: 1st shot on Zen4
PP-512: 107 t/s -> 137 t/s
TG-128(1 thread): 2.64 t/s -> 3.44 t/s
* iq4_xxs_r4: WIP
* iq4_xxs_r4: 1st shot at AVX2
Note: there is a bug in the AVX2 implementation for nrc_y = 1
for IQ quants with blocks of 32. I have fixed it for now by
using the nrc_y > 1 implementation (which works) also for nrc_y = 1.
* iq3_xxs_r4: NEON
---------
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 | 7 |
1 files changed, 7 insertions, 0 deletions
diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h index 6e87dbaa..b9da9771 100644 --- a/ggml/src/ggml-common.h +++ b/ggml/src/ggml-common.h @@ -423,6 +423,13 @@ typedef struct { } block_iq3_xxs; static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_half) + 3*(QK_K/8), "wrong iq3_xxs block size/padding"); +typedef struct { + ggml_half d[4]; + uint8_t sas[QK_K/2]; + uint8_t qs[QK_K]; +} block_iq3_xxs_r4; +static_assert(sizeof(block_iq3_xxs_r4) == 4*sizeof(block_iq3_xxs), "wrong iq3_xxs_r4 block size/padding"); + // 3.4375 bpw #define IQ3S_N_SCALE QK_K/64 typedef struct { |