diff options
| author | Kawrakow <iwankawrakow@gmail.com> | 2025-02-05 13:49:39 +0200 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2025-02-05 13:49:39 +0200 |
| commit | 8b7536bda8b65107794c4df710f14ddfde430160 (patch) | |
| tree | 97a9dea70458bddcef51c734e22026ac51b51ed7 /ggml/src/ggml-common.h | |
| parent | ecf111a11ca56ff0731308f94bd6c5e96658b6ef (diff) | |
IQ1_S_R4: better 1.5 bpw quants (#185)
* iq1_s_r4: basics - quantize/dequantize
* iq1_s_r4: gemm/gemv works on AVX2/Zen4
* Don't forget to make sure we have a multiple of 4 rows per thread
* iq1_s_r4: this is better
* iq1_s_r4: fix Zen4 after AVX2 changes
* iq1_s_r4: NEON gemm/gemv
* iq1_s_r4: more bits for shared experts
With this mix we arrive at PPL(512) = 9.4140
for Deepseek-Lite using 1.766 bpw for the repeating layers.
On the Ryzen-7950X we get PP-512 = 494 t/s and
TG-128 = 52 t/s @ 16 threads.
* Forgotten counter increment
* iq1_s_r4: slightly faster AVX2/Zen4 gemm/gemv
* Compiler warnings
---------
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 | 6 |
1 files changed, 6 insertions, 0 deletions
diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h index 023b0b63..14813161 100644 --- a/ggml/src/ggml-common.h +++ b/ggml/src/ggml-common.h @@ -485,6 +485,12 @@ typedef struct { } block_iq1_s; static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding"); +typedef struct { + uint8_t qs[16]; + uint16_t qh[4]; +} block_iq1_s_r4; +static_assert(sizeof(block_iq1_s_r4) == 24, "wrong iq1_s_r4 block size/padding"); + // 1.75 bpw typedef struct { uint8_t qs[QK_K/8]; // grid index, low 8 bits |