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-quants.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-quants.h')
| -rw-r--r-- | ggml/src/ggml-quants.h | 5 |
1 files changed, 5 insertions, 0 deletions
diff --git a/ggml/src/ggml-quants.h b/ggml/src/ggml-quants.h index b6d69011..4753f342 100644 --- a/ggml/src/ggml-quants.h +++ b/ggml/src/ggml-quants.h @@ -42,6 +42,7 @@ void quantize_row_iq4_xs_ref (const float * GGML_RESTRICT x, block_iq4_xs * GGM void quantize_row_iq3_s_ref (const float * GGML_RESTRICT x, block_iq3_s * GGML_RESTRICT y, int64_t k); void quantize_row_iq2_s_ref (const float * GGML_RESTRICT x, block_iq2_s * GGML_RESTRICT y, int64_t k); void quantize_row_iq1_bn_ref (const float * GGML_RESTRICT x, block_iq1_bn * GGML_RESTRICT y, int64_t k); +void quantize_row_iq1_s_ref (const float * GGML_RESTRICT x, block_iq1_s * GGML_RESTRICT y, int64_t k); void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); @@ -66,6 +67,7 @@ void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, void quantize_row_iq3_s (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); void quantize_row_iq2_s (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); void quantize_row_iq1_bn (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_row_iq1_s (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); // Dequantization void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); @@ -148,6 +150,9 @@ void iq2xs_free_impl(enum ggml_type type); void iq3xs_init_impl(int grid_size); void iq3xs_free_impl(int grid_size); +void iq1s_process_1block(int block_size, const float * xb, const float * weight, int8_t * L, + float * the_scale, uint16_t * the_index, int * the_shift, float * pairs, float * sumx, float * sumw); + #if defined(__ARM_FEATURE_SVE) extern int ggml_sve_cnt_b; #endif |