diff options
| author | Kawrakow <iwankawrakow@gmail.com> | 2024-12-04 15:20:07 +0100 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2024-12-04 15:20:07 +0100 |
| commit | f64de08203aaee95ca755336de3e1db85d990198 (patch) | |
| tree | 9af01056e0b304ee5df5792f25d82066931eb4d6 /ggml/src/iqk/iqk_quantize.h | |
| parent | f1f4eb988fe5ee969100cd0d3782fd7460d13949 (diff) | |
IQ4_XS_R4 (#123)
* Adding iq4_xs_r4
This is a 1st working version on Zen4.
We get PP-512(LLaMA-3.1-8B) = 226 t/s, so 16% slower
than iq4_nl_x4.
* iq4_xs_r4: WIP
* iq4_xs_r4: Use AVX2 version for matrix x vector on Zen4
* iq4_xs_r4: NEON
We get PP-512(LLaMA-3.1-8B) = 115.6 t/s on M2-Max,
up from 68.2 t/s for iq4_xs!
* DRY
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Diffstat (limited to 'ggml/src/iqk/iqk_quantize.h')
| -rw-r--r-- | ggml/src/iqk/iqk_quantize.h | 6 |
1 files changed, 6 insertions, 0 deletions
diff --git a/ggml/src/iqk/iqk_quantize.h b/ggml/src/iqk/iqk_quantize.h index 3349c675..ad2294c5 100644 --- a/ggml/src/iqk/iqk_quantize.h +++ b/ggml/src/iqk/iqk_quantize.h @@ -93,6 +93,12 @@ size_t quantize_q6_0_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT ds void dequantize_row_q6_0_r4(const block_q6_0_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); void vec_dot_q6_0_r4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void quantize_row_iq4_xs_r4_ref(const float * GGML_RESTRICT x, block_iq4_xs_r4 * GGML_RESTRICT y, int64_t k); +void quantize_row_iq4_xs_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +size_t quantize_iq4_xs_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +void dequantize_row_iq4_xs_r4(const block_iq4_xs_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void vec_dot_iq4_xs_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); + #ifdef __cplusplus } #endif |