IQ2_XXS_R4 (#154)

* iq2_xxs_r4: Zen4

Disapointing gain: 134.7 t/s -> 151.1 t/s for PP-512
TG-128 is better: 3.45 -> 4.61 t/s @ 1 thread

* Minor

* iq2_xxs_r4: NEON

---------

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

1 files changed, 17 insertions, 4 deletions

diff --git a/src/llama.cpp b/src/llama.cpp
index 0171539d..df700c12 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -3850,6 +3850,7 @@ struct llama_model_loader {
                 case GGML_TYPE_Q6_K_R4: ftype = LLAMA_FTYPE_MOSTLY_Q6_K_R4; break;
                 case GGML_TYPE_Q8_K_R8: ftype = LLAMA_FTYPE_MOSTLY_Q8_K_R8; break;
                 case GGML_TYPE_IQ2_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS; break;
+                case GGML_TYPE_IQ2_XXS_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4; 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;
                 case GGML_TYPE_IQ2_S:   ftype = LLAMA_FTYPE_MOSTLY_IQ2_S;   break;
@@ -4578,6 +4579,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q6_K_R4:  return "Q6_K_R4";
         case LLAMA_FTYPE_MOSTLY_Q8_K_R8:  return "Q8_K_R8";
         case LLAMA_FTYPE_MOSTLY_IQ2_XXS:  return "IQ2_XXS - 2.0625 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4:return "IQ2_XXS_R4 - 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";
         case LLAMA_FTYPE_MOSTLY_IQ2_S:    return "IQ2_S - 2.5 bpw";
@@ -15798,6 +15800,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
                      ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4) {
                 new_type = !qs.has_output ? GGML_TYPE_IQ4_K : GGML_TYPE_Q5_K;
             }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4) {
+                new_type = !qs.has_output ? GGML_TYPE_IQ4_K_R4 : GGML_TYPE_Q5_K_R4;
+            }
             else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ3_S || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS ||
                       ftype == LLAMA_FTYPE_MOSTLY_IQ4_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_KSS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_KS_R4) && !qs.has_output) {
                 new_type = GGML_TYPE_IQ5_K;
@@ -15812,7 +15817,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             new_type = qs.params->token_embedding_type;
         } else {
             if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS ||
-                ftype == LLAMA_FTYPE_MOSTLY_IQ1_S   || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
+                ftype == LLAMA_FTYPE_MOSTLY_IQ1_S   || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M  ||
+                ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4) {
                 new_type = GGML_TYPE_Q2_K;
             }
             else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
@@ -15887,8 +15893,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             }
         }
     } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S ||
-               ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M    || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M ||
-               ftype == LLAMA_FTYPE_MOSTLY_IQ2_KS) {
+               ftype == LLAMA_FTYPE_MOSTLY_IQ2_S   || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M  || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M ||
+               ftype == LLAMA_FTYPE_MOSTLY_IQ2_KS  || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4) {
         if (name.find("attn_v.weight") != std::string::npos) {
             if      (qs.model.hparams.n_gqa() >= 4 || qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_IQ4_K;
             else if (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_IQ3_K;
@@ -16182,7 +16188,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
         new_type == GGML_TYPE_Q5_K_R4 || new_type == GGML_TYPE_Q3_K_R4 || new_type == GGML_TYPE_Q2_K_R4 ||
         new_type == GGML_TYPE_IQ4_K_R4|| new_type == GGML_TYPE_Q8_K_R8 || new_type == GGML_TYPE_IQ3_K_R4||
         new_type == GGML_TYPE_IQ2_K_R4|| new_type == GGML_TYPE_IQ5_K_R4|| new_type == GGML_TYPE_IQ4_KS_R4 ||
-        new_type == GGML_TYPE_IQ3_XXS_R4) {
+        new_type == GGML_TYPE_IQ3_XXS_R4 || new_type == GGML_TYPE_IQ2_XXS_R4) {
         int nx = tensor->ne[0];
         int ny = tensor->ne[1];
         if (nx % QK_K != 0) {
@@ -16201,6 +16207,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
     if (convert_incompatible_tensor) {
         switch (new_type) {
             case GGML_TYPE_IQ2_XXS:
+            case GGML_TYPE_IQ2_XXS_R4:
             case GGML_TYPE_IQ2_XS:
             case GGML_TYPE_IQ2_KS:
             case GGML_TYPE_IQ2_S:
@@ -16332,6 +16339,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_Q6_K_R4: default_type = GGML_TYPE_Q6_K_R4; break;
         case LLAMA_FTYPE_MOSTLY_Q8_K_R8: default_type = GGML_TYPE_Q8_K_R8; break;
         case LLAMA_FTYPE_MOSTLY_IQ2_XXS: default_type = GGML_TYPE_IQ2_XXS; break;
+        case LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4:default_type = GGML_TYPE_IQ2_XXS_R4; 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;
         case LLAMA_FTYPE_MOSTLY_IQ2_S:   default_type = GGML_TYPE_IQ2_XS;  break;
@@ -16685,6 +16693,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             }
             if (!params->ignore_imatrix_rules && !imatrix &&
                 (new_type == GGML_TYPE_IQ2_XXS ||
+                 new_type == GGML_TYPE_IQ2_XXS_R4 ||
                  new_type == GGML_TYPE_IQ2_XS  ||
                  new_type == GGML_TYPE_IQ2_S   ||
                  new_type == GGML_TYPE_IQ1_S   ||
@@ -16787,6 +16796,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ4_KS;
                 else chunk_size_multiplier = 4;
             }
+            else if (new_type == GGML_TYPE_IQ2_XXS_R4) {
+                if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ2_XXS;
+                else chunk_size_multiplier = 4;
+            }
             else if (new_type == GGML_TYPE_IQ3_XXS_R4) {
                 if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ3_XXS;
                 else chunk_size_multiplier = 4;