Q3_K_R4 (#134)

* q3_k_r4: Zen4 works, but not as good as it should be

238 t/s, so sloghtly slower than q6_k_r4.

* q3_k_r4: NEON

We get PP-512(LLaMA-3.1-8B) = 106.9 t/s.
This is 1.93X faster than q3_K_S!

---------

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

1 files changed, 15 insertions, 3 deletions

diff --git a/src/llama.cpp b/src/llama.cpp
index dc6d307d..9f41724f 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -3836,6 +3836,7 @@ struct llama_model_loader {
                 case GGML_TYPE_Q8_0:    ftype = LLAMA_FTYPE_MOSTLY_Q8_0;    break;
                 case GGML_TYPE_Q2_K:    ftype = LLAMA_FTYPE_MOSTLY_Q2_K;    break;
                 case GGML_TYPE_Q3_K:    ftype = LLAMA_FTYPE_MOSTLY_Q3_K_M;  break;
+                case GGML_TYPE_Q3_K_R4: ftype = LLAMA_FTYPE_MOSTLY_Q3_K_R4; break;
                 case GGML_TYPE_Q4_K:    ftype = LLAMA_FTYPE_MOSTLY_Q4_K_M;  break;
                 case GGML_TYPE_Q4_K_R4: ftype = LLAMA_FTYPE_MOSTLY_Q4_K_R4; break;
                 case GGML_TYPE_Q5_K:    ftype = LLAMA_FTYPE_MOSTLY_Q5_K_M;  break;
@@ -4548,6 +4549,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q3_K_S:   return "Q3_K - Small";
         case LLAMA_FTYPE_MOSTLY_Q3_K_M:   return "Q3_K - Medium";
         case LLAMA_FTYPE_MOSTLY_Q3_K_L:   return "Q3_K - Large";
+        case LLAMA_FTYPE_MOSTLY_Q3_K_R4:  return "Q3_K_R4";
         case LLAMA_FTYPE_MOSTLY_Q4_K_S:   return "Q4_K - Small";
         case LLAMA_FTYPE_MOSTLY_Q4_K_R4:  return "Q4_K_R4";
         case LLAMA_FTYPE_MOSTLY_Q4_K_M:   return "Q4_K - Medium";
@@ -15792,6 +15794,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             else if (new_type == GGML_TYPE_IQ4_XS_R4) {
                 new_type = GGML_TYPE_IQ4_XS;
             }
+            else if (new_type == GGML_TYPE_Q3_K_R4) {
+                new_type = GGML_TYPE_Q3_K;
+            }
             else if (new_type == GGML_TYPE_Q4_K_R4) {
                 new_type = GGML_TYPE_Q4_K;
             }
@@ -15904,6 +15909,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
         else if (qs.model.hparams.n_gqa() >= 4) {
             if      (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_IQ3_XXS) new_type = GGML_TYPE_IQ3_S;
             else if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_IQ3_S  ) new_type = GGML_TYPE_Q4_K;
+            else if (new_type == GGML_TYPE_Q3_K_R4) new_type = GGML_TYPE_Q4_K_R4;
             else if (new_type == GGML_TYPE_Q4_K || new_type == GGML_TYPE_IQ4_XS) new_type = GGML_TYPE_Q5_K;
             else if (new_type == GGML_TYPE_IQ4_NL) new_type = GGML_TYPE_Q5_K;
             else if (new_type == GGML_TYPE_IQ4_NL_R4) new_type = GGML_TYPE_Q5_K;
@@ -15935,7 +15941,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) {
             if (qs.model.hparams.n_vocab >= 127999 && (qs.model.type == MODEL_8B || qs.model.type == MODEL_70B))
                 new_type = GGML_TYPE_Q4_K;
-        } 
+        }
     } else if (name.find("ffn_down") != std::string::npos) {
         auto info = layer_info(qs.i_ffn_down, qs.n_ffn_down, name.c_str());
         int i_layer = info.first, n_layer = info.second;
@@ -16003,7 +16009,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
                     ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S   ||
                     ftype == LLAMA_FTYPE_MOSTLY_IQ3_M  || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_K   ||
                     ftype == LLAMA_FTYPE_MOSTLY_IQ2_K  || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K  || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_R4 ||
-                    ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS_R4) {
+                    ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS_R4 || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_R4) {
                     new_type = GGML_TYPE_Q5_K;
                 }
             } else {
@@ -16073,7 +16079,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
         new_type == GGML_TYPE_IQ5_K   || new_type == GGML_TYPE_IQ3_K   || new_type == GGML_TYPE_Q4_K_R4 ||
         new_type == GGML_TYPE_IQ6_K   || new_type == GGML_TYPE_IQ4_KS  || new_type == GGML_TYPE_IQ4_XS_R4 ||
         new_type == GGML_TYPE_IQ2_KS  || new_type == GGML_TYPE_IQ4_KSS || new_type == GGML_TYPE_Q6_K_R4 ||
-        new_type == GGML_TYPE_Q5_K_R4) {
+        new_type == GGML_TYPE_Q5_K_R4 || new_type == GGML_TYPE_Q3_K_R4) {
         int nx = tensor->ne[0];
         int ny = tensor->ne[1];
         if (nx % QK_K != 0) {
@@ -16101,6 +16107,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             case GGML_TYPE_IQ1_M:
             case GGML_TYPE_Q2_K:
             case GGML_TYPE_Q3_K:
+            case GGML_TYPE_Q3_K_R4:
             case GGML_TYPE_IQ2_K:
             case GGML_TYPE_IQ3_K:
             case GGML_TYPE_IQ4_KSS:
@@ -16201,6 +16208,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_Q3_K_S:
         case LLAMA_FTYPE_MOSTLY_Q3_K_M:
         case LLAMA_FTYPE_MOSTLY_Q3_K_L:  default_type = GGML_TYPE_Q3_K;    break;
+        case LLAMA_FTYPE_MOSTLY_Q3_K_R4: default_type = GGML_TYPE_Q3_K_R4; break;
         case LLAMA_FTYPE_MOSTLY_Q4_K_S:
         case LLAMA_FTYPE_MOSTLY_Q4_K_M:  default_type = GGML_TYPE_Q4_K;    break;
         case LLAMA_FTYPE_MOSTLY_Q4_K_R4: default_type = GGML_TYPE_Q4_K_R4; break;
@@ -16608,6 +16616,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q8_0;
                 else chunk_size_multiplier = 4;
             }
+            else if (new_type == GGML_TYPE_Q3_K_R4) {
+                if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q3_K;
+                else chunk_size_multiplier = 4;
+            }
             else if (new_type == GGML_TYPE_Q4_K_R4) {
                 if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_K;
                 else chunk_size_multiplier = 4;