Q4_K_R4 (#129)

* Something is still wrong

* Simply don't see what is wrong

* q4_k_r4: finally works on Zen4

I had forgotten to prevent token_embd.weight being quantized
with q4_k_r4!

* q4_k_r4: AVX2

We get PP-512(LLaMA-3.1-8B) = 267 t/s on a Ryzen-5975WX.
This is ~30% better than Q4_K_S.

* q4_k_r4: NEON

We get PP-512(LLaMA-3.1-8B) = 110 t/s.
Not quite as good as q4_0_r4, but still a massive
improvement compared to he 69 t/s for q4_K.

* q4_k_r4: slightly better AVX2

PP-512 goes from 267 t/s to 282 t/s on Ryzen-5975WX

* Minor

* Minor

---------

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

1 files changed, 24 insertions, 2 deletions

diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 974e42b2..fadda3e3 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -888,6 +888,19 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .nrows                    = 1,
         .row_meta_size            = 0,
     },
+    [GGML_TYPE_Q4_K_R4] = {
+        .type_name                = "q4_k_r4",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_q4_K),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_q4_k_r4,
+        .from_float               = quantize_row_q4_k_r4,
+        .from_float_ref           = (ggml_from_float_t) quantize_row_q4_k_r4_ref,
+        .vec_dot                  = vec_dot_q4_k_r4_q8_k,
+        .vec_dot_type             = GGML_TYPE_Q8_K32,
+        .nrows                    = 1,
+        .row_meta_size            = 0,
+    },
     [GGML_TYPE_Q5_K] = {
         .type_name                = "q5_K",
         .blck_size                = QK_K,
@@ -4020,8 +4033,9 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_Q2_K:          wtype = GGML_TYPE_Q2_K;  break;
         case GGML_FTYPE_MOSTLY_Q3_K:          wtype = GGML_TYPE_Q3_K;  break;
         case GGML_FTYPE_MOSTLY_Q4_K:          wtype = GGML_TYPE_Q4_K;  break;
-        case GGML_FTYPE_MOSTLY_Q5_K:          wtype = GGML_TYPE_Q5_K;  break;
-        case GGML_FTYPE_MOSTLY_Q6_K:          wtype = GGML_TYPE_Q6_K;  break;
+        case GGML_FTYPE_MOSTLY_Q4_K_R4:       wtype = GGML_TYPE_Q4_K_R4;  break;
+        case GGML_FTYPE_MOSTLY_Q5_K:          wtype = GGML_TYPE_Q5_K;     break;
+        case GGML_FTYPE_MOSTLY_Q6_K:          wtype = GGML_TYPE_Q6_K;     break;
         case GGML_FTYPE_MOSTLY_IQ2_XXS:       wtype = GGML_TYPE_IQ2_XXS;  break;
         case GGML_FTYPE_MOSTLY_IQ2_XS:        wtype = GGML_TYPE_IQ2_XS;   break;
         case GGML_FTYPE_MOSTLY_IQ3_XXS:       wtype = GGML_TYPE_IQ3_XXS;  break;
@@ -10550,6 +10564,7 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q4_K_R4:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_IQ2_XXS:
@@ -10999,6 +11014,7 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q4_K_R4:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_IQ2_XXS:
@@ -11145,6 +11161,7 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q4_K_R4:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_IQ2_XXS:
@@ -14337,6 +14354,7 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q4_K_R4:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_IQ2_XXS:
@@ -14723,6 +14741,7 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q4_K_R4:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_IQ2_XXS:
@@ -15003,6 +15022,7 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q4_K_R4:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_IQ2_XXS:
@@ -15610,6 +15630,7 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q4_K_R4:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_IQ2_XXS:
@@ -22445,6 +22466,7 @@ size_t ggml_quantize_chunk(
         case GGML_TYPE_Q2_K:    result = quantize_q2_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_Q3_K:    result = quantize_q3_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_Q4_K:    result = quantize_q4_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q4_K_R4: result = quantize_q4_k_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_Q5_K:    result = quantize_q5_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_Q6_K:    result = quantize_q6_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ2_XXS: result = quantize_iq2_xxs(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;