Q6_K_R4 (#130)

* Adding q6_k_r4

* q6_k_r4: 1st functional AVX2 version

* q6_k_r4: AVX2 and simple Zen4

"Simple" as in processing 4 instead of 8 rows at once.
On Zen4 we get PP-512(LLaMA-3.1-8B) = 238.3 t/s vs
195.2 t/s for Q6_K. TG-128 @ 1 thread is 7.94 t/s
vs 5.38 t/s for Q6_K.

* q6_k_r4: 1st NEON version

PP-512(LLaMA-3.1-8B) = 78 t/s vs 57.6 t/s for q6_K.
TG-128 is slightly lower rthan q6_K for low number of threads,
becomes very slightly better at 8 threads.

* q6_k_r4: slightly faster NEON

PP-512(LLaMA-3.1-8B) = 83.25 t/s

* q6_k_r4: slightly faster Zen4

238.3 t/s -> 243.2 t/s

---------

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

1 files changed, 13 insertions, 2 deletions

diff --git a/src/llama.cpp b/src/llama.cpp
index 18c6e111..3b617b06 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -3840,6 +3840,7 @@ struct llama_model_loader {
                 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;
                 case GGML_TYPE_Q6_K:    ftype = LLAMA_FTYPE_MOSTLY_Q6_K;    break;
+                case GGML_TYPE_Q6_K_R4: ftype = LLAMA_FTYPE_MOSTLY_Q6_K_R4; break;
                 case GGML_TYPE_IQ2_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS; break;
                 case GGML_TYPE_IQ2_XS:  ftype = LLAMA_FTYPE_MOSTLY_IQ2_XS;  break;
                 case GGML_TYPE_IQ2_KS:  ftype = LLAMA_FTYPE_MOSTLY_IQ2_KS;  break;
@@ -4552,6 +4553,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q5_K_S:   return "Q5_K - Small";
         case LLAMA_FTYPE_MOSTLY_Q5_K_M:   return "Q5_K - Medium";
         case LLAMA_FTYPE_MOSTLY_Q6_K:     return "Q6_K";
+        case LLAMA_FTYPE_MOSTLY_Q6_K_R4:  return "Q6_K_R4";
         case LLAMA_FTYPE_MOSTLY_IQ2_XXS:  return "IQ2_XXS - 2.0625 bpw";
         case LLAMA_FTYPE_MOSTLY_IQ2_XS:   return "IQ2_XS - 2.3125 bpw";
         case LLAMA_FTYPE_MOSTLY_IQ2_KS:   return "IQ2_KS - 2.1875 bpw";
@@ -15757,7 +15759,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
                       ftype == LLAMA_FTYPE_MOSTLY_IQ4_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_KSS) && !qs.has_output) {
                 new_type = GGML_TYPE_IQ5_K;
             }
-            else if (new_type != GGML_TYPE_Q8_0 && new_type != GGML_TYPE_Q8_0_R4 && new_type != GGML_TYPE_IQ6_K) {
+            else if (new_type != GGML_TYPE_Q8_0 && new_type != GGML_TYPE_Q8_0_R4 && new_type != GGML_TYPE_IQ6_K && new_type != GGML_TYPE_Q6_K_R4) {
                 new_type = GGML_TYPE_Q6_K;
             }
         }
@@ -15791,6 +15793,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             else if (new_type == GGML_TYPE_Q4_K_R4) {
                 new_type = GGML_TYPE_Q4_K;
             }
+            else if (new_type == GGML_TYPE_Q6_K_R4) {
+                new_type = GGML_TYPE_Q6_K;
+            }
             else if (new_type == GGML_TYPE_Q4_0_R4) {
                 new_type = GGML_TYPE_Q4_0;
             }
@@ -16062,7 +16067,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
         new_type == GGML_TYPE_IQ1_M   || new_type == GGML_TYPE_IQ4_K   || new_type == GGML_TYPE_IQ2_K  ||
         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_IQ2_KS  || new_type == GGML_TYPE_IQ4_KSS || new_type == GGML_TYPE_Q6_K_R4) {
         int nx = tensor->ne[0];
         int ny = tensor->ne[1];
         if (nx % QK_K != 0) {
@@ -16102,6 +16107,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             case GGML_TYPE_IQ5_K:
             case GGML_TYPE_Q5_K:   new_type = GGML_TYPE_Q6_0;   break;
             case GGML_TYPE_IQ6_K:
+            case GGML_TYPE_Q6_K_R4:
             case GGML_TYPE_Q6_K:   new_type = GGML_TYPE_Q8_0;   break;
             default: throw std::runtime_error("\nUnsupported tensor size encountered\n");
         }
@@ -16194,6 +16200,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_Q5_K_S:
         case LLAMA_FTYPE_MOSTLY_Q5_K_M:  default_type = GGML_TYPE_Q5_K;    break;
         case LLAMA_FTYPE_MOSTLY_Q6_K:    default_type = GGML_TYPE_Q6_K;    break;
+        case LLAMA_FTYPE_MOSTLY_Q6_K_R4: default_type = GGML_TYPE_Q6_K_R4; break;
         case LLAMA_FTYPE_MOSTLY_IQ2_XXS: default_type = GGML_TYPE_IQ2_XXS; break;
         case LLAMA_FTYPE_MOSTLY_IQ2_XS:  default_type = GGML_TYPE_IQ2_XS;  break;
         case LLAMA_FTYPE_MOSTLY_IQ2_KS:  default_type = GGML_TYPE_IQ2_KS;  break;
@@ -16597,6 +16604,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_K;
                 else chunk_size_multiplier = 4;
             }
+            else if (new_type == GGML_TYPE_Q6_K_R4) {
+                if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q6_K;
+                else chunk_size_multiplier = 4;
+            }
             else if (new_type == GGML_TYPE_IQ2_BN_R4) {
                 if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ2_BN;
                 else chunk_size_multiplier = 4;