diff options
| author | Kawrakow <iwankawrakow@gmail.com> | 2024-10-13 13:34:30 +0300 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2024-10-13 13:34:30 +0300 |
| commit | 910a13409463f7aedb0a92be013a1b9bb50f4859 (patch) | |
| tree | 16e13e1fd3010549877408a0a62706b2bc5d5f0c /ggml/src/ggml-quants.h | |
| parent | c15de3654e0002537c8052fd6d52d879e778e88c (diff) | |
IQ2_KS: 2.1875 bpw non-linear quantization (#85)
* Experimenting
* iq2k: Try make_qx_quants for the scale
Slightly better for LLaMA-3.1, Gemma-2, slightly worse for
Qwen2.5
* iq2k with make_qx_quants: adjust scale
* iq2ks: basics
* iq2_ks: CUDA works
* iq2_ks: WIP
* iq2_ks: WIP
* iq2_ks: Zen4
* iq2_ks: AVX2
* iq2_ks: scalar dot product
* iq2_ks: ARM_NEON
* iq2_ks: Metal
* iq2_ks: faster Metal
LLaMA-3.1-8B:
PP-512 = 475.22 ± 0.37 t/s
TG-128 = 45.32 ± 0.03 t/s
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Diffstat (limited to 'ggml/src/ggml-quants.h')
| -rw-r--r-- | ggml/src/ggml-quants.h | 4 |
1 files changed, 4 insertions, 0 deletions
diff --git a/ggml/src/ggml-quants.h b/ggml/src/ggml-quants.h index bad7e9d9..a40a6d37 100644 --- a/ggml/src/ggml-quants.h +++ b/ggml/src/ggml-quants.h @@ -35,6 +35,8 @@ void quantize_row_q6_K_ref(const float * GGML_RESTRICT x, block_q6_K * GGML_REST void quantize_row_q8_K_ref(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k); void quantize_row_q8_K64_ref(const float * GGML_RESTRICT x, block_q8_K64 * GGML_RESTRICT y, int64_t k); +void quantize_row_iq2_xxs_ref(const float * GGML_RESTRICT x, block_iq2_xxs * GGML_RESTRICT y, int64_t k); +void quantize_row_iq2_xs_ref (const float * GGML_RESTRICT x, block_iq2_xs * GGML_RESTRICT y, int64_t k); void quantize_row_iq3_xxs_ref(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k); void quantize_row_iq4_nl_ref (const float * GGML_RESTRICT x, block_iq4_nl * GGML_RESTRICT y, int64_t k); void quantize_row_iq4_xs_ref (const float * GGML_RESTRICT x, block_iq4_xs * GGML_RESTRICT y, int64_t k); @@ -59,6 +61,8 @@ void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); void quantize_row_q8_K64(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_row_iq2_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_row_iq2_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); |