Adding Q6_0 (#77)

* Adding q6_0 - basics + AVX2/Zen4 working

* Adding q6_0: CUDA dequantize works, but not mmvq

* Adding q6_0: CUDA mmvq works

* Adding q6_0: CUDA cpy, so Q6_0 can be used for KV-cache

* Add q6_0 to CPU flash attention

Disappointing result: for LlaMA-3.2-1B, q6_0 K- and V-cache
gives about the same PPL as q8_0 K-cache and q4_0 V-cache,
while needing the exact same RAM.
I.e., what was the point?

* q6_0: slightly better kv-cache result

Better than q8_0+q4_0, but not as good as q8_0+iq4_nl

* q6_0: works on ARM_NEON

* q6_0: dequantize works on Metal, but not vector dot product

* q6_0: it now works on Metal

Outperforms q5_0 by a significant margin. E.g.
| model                          |       size |     params | backend    | ngl | threads |          test |              t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | ------------: | ---------------: |
| llama 8B Q6_0                  |   6.08 GiB |     8.03 B | Metal      | 100 |       4 |         tg128 |     44.02 ± 0.08 |
| llama 8B Q5_0                  |   5.21 GiB |     8.03 B | Metal      | 100 |       4 |         tg128 |     40.13 ± 0.12 |
| llama 8B Q6_0                  |   6.08 GiB |     8.03 B | Metal      | 100 |       4 |         pp512 |    500.55 ± 0.32 |
| llama 8B Q5_0                  |   5.21 GiB |     8.03 B | Metal      | 100 |       4 |         pp512 |    448.02 ± 0.27 |

* q6_0: can now be used for kv-cache on Metal

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>

1 files changed, 5 insertions, 0 deletions

diff --git a/ggml/src/ggml-quants.h b/ggml/src/ggml-quants.h
index 775aa875..bad7e9d9 100644
--- a/ggml/src/ggml-quants.h
+++ b/ggml/src/ggml-quants.h
@@ -25,6 +25,7 @@ void quantize_row_q5_0_ref(const float * GGML_RESTRICT x, block_q5_0 * GGML_REST
 void quantize_row_q5_1_ref(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
 void quantize_row_q8_0_ref(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
 void quantize_row_q8_1_ref(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_0_ref(const float * GGML_RESTRICT x, block_q6_0 * GGML_RESTRICT y, int64_t k);
 
 void quantize_row_q2_K_ref(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
 void quantize_row_q3_K_ref(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
@@ -48,6 +49,7 @@ void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in
 void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 
 void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
@@ -72,6 +74,7 @@ void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRI
 void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 //void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q6_0(const block_q6_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 
 void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
@@ -98,6 +101,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 void ggml_vec_dot_q5_0_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 ggml_vec_dot_q5_1_q8_1(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 ggml_vec_dot_q8_0_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 ggml_vec_dot_q6_0_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 ggml_vec_dot_q2_K_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);
 void ggml_vec_dot_q3_K_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);
@@ -140,6 +144,7 @@ size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst,
 size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q6_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 
 void iq2xs_init_impl(enum ggml_type type);
 void iq2xs_free_impl(enum ggml_type type);