IQ4_XS_R4 (#123)

* Adding iq4_xs_r4

This is a 1st working version on Zen4.
We get PP-512(LLaMA-3.1-8B) = 226 t/s, so 16% slower
than iq4_nl_x4.

* iq4_xs_r4: WIP

* iq4_xs_r4: Use AVX2 version for matrix x vector on Zen4

* iq4_xs_r4: NEON

We get PP-512(LLaMA-3.1-8B) = 115.6 t/s on M2-Max,
up from 68.2 t/s for iq4_xs!

* DRY

---------

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

1 files changed, 17 insertions, 3 deletions

diff --git a/src/llama.cpp b/src/llama.cpp
index f307fd89..e2abc235 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -3850,6 +3850,7 @@ struct llama_model_loader {
                 case GGML_TYPE_IQ2_BN:  ftype = LLAMA_FTYPE_MOSTLY_IQ2_BN;  break;
                 case GGML_TYPE_IQ4_NL:  ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL;  break;
                 case GGML_TYPE_IQ4_NL_X4:ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL_X4;break;
+                case GGML_TYPE_IQ4_XS_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS_R4;break;
                 case GGML_TYPE_Q4_0_R4: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_R4; break;
                 case GGML_TYPE_Q5_0_R4: ftype = LLAMA_FTYPE_MOSTLY_Q5_0_R4; break;
                 case GGML_TYPE_Q6_0_R4: ftype = LLAMA_FTYPE_MOSTLY_Q6_0_R4; break;
@@ -4559,6 +4560,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_IQ1_M:    return "IQ1_M - 1.75 bpw";
         case LLAMA_FTYPE_MOSTLY_IQ4_NL:   return "IQ4_NL - 4.5 bpw";
         case LLAMA_FTYPE_MOSTLY_IQ4_NL_X4:return "IQ4_NL_X4 - 4.5 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ4_XS_R4:return "IQ4_XS_R4 - 4.25 bpw";
         case LLAMA_FTYPE_MOSTLY_Q4_0_R4:  return "Q4_0_R4 - 4.5 bpw";
         case LLAMA_FTYPE_MOSTLY_Q5_0_R4:  return "Q5_0_R4 - 5.5 bpw";
         case LLAMA_FTYPE_MOSTLY_Q6_0_R4:  return "Q6_0_R4 - 6.5 bpw";
@@ -15779,6 +15781,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             else if (new_type == GGML_TYPE_IQ4_NL_X4) {
                 new_type = GGML_TYPE_IQ4_NL;
             }
+            else if (new_type == GGML_TYPE_IQ4_XS_R4) {
+                new_type = GGML_TYPE_IQ4_XS;
+            }
             else if (new_type == GGML_TYPE_Q4_0_R4) {
                 new_type = GGML_TYPE_Q4_0;
             }
@@ -15852,7 +15857,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             new_type = qs.i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
         }
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
-        else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL_X4 ||
+        else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS ||
+                  ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL_X4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS_R4 ||
                   ftype == LLAMA_FTYPE_MOSTLY_IQ4_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_KSS) && qs.model.hparams.n_gqa() >= 2) {
             new_type = GGML_TYPE_IQ5_K;
         }
@@ -15883,6 +15889,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             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_X4) new_type = GGML_TYPE_Q5_K;
+            else if (new_type == GGML_TYPE_IQ4_XS_R4) new_type = GGML_TYPE_Q5_K;
             else if (new_type == GGML_TYPE_Q5_K) new_type = GGML_TYPE_Q6_K;
         }
         ++qs.i_attention_wv;
@@ -15947,7 +15954,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
         }
         else if (i_layer < n_layer/8 && !qs.has_imatrix &&
                 (ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS ||
-                 ftype == LLAMA_FTYPE_MOSTLY_IQ4_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_KSS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL_X4)) {
+                 ftype == LLAMA_FTYPE_MOSTLY_IQ4_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_KSS ||
+                 ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL_X4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS_R4)) {
             new_type = GGML_TYPE_Q5_K;
         }
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q6_K;
@@ -15973,7 +15981,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
                     ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL  ||
                     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_IQ4_NL_X4) {
+                    ftype == LLAMA_FTYPE_MOSTLY_IQ2_K  || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K  ||
+                    ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL_X4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS_R4) {
                     new_type = GGML_TYPE_Q5_K;
                 }
             } else {
@@ -16183,6 +16192,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_IQ2_BN:  default_type = GGML_TYPE_IQ2_BN;  break;
         case LLAMA_FTYPE_MOSTLY_IQ4_NL:  default_type = GGML_TYPE_IQ4_NL;  break;
         case LLAMA_FTYPE_MOSTLY_IQ4_NL_X4:default_type = GGML_TYPE_IQ4_NL_X4;break;
+        case LLAMA_FTYPE_MOSTLY_IQ4_XS_R4:default_type = GGML_TYPE_IQ4_XS_R4;break;
         case LLAMA_FTYPE_MOSTLY_Q4_0_R4: default_type = GGML_TYPE_Q4_0_R4; break;
         case LLAMA_FTYPE_MOSTLY_Q5_0_R4: default_type = GGML_TYPE_Q5_0_R4; break;
         case LLAMA_FTYPE_MOSTLY_Q6_0_R4: default_type = GGML_TYPE_Q6_0_R4; break;
@@ -16548,6 +16558,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ4_NL;
                 else chunk_size_multiplier = 4;
             }
+            else if (new_type == GGML_TYPE_IQ4_XS_R4) {
+                if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ4_XS;
+                else chunk_size_multiplier = 4;
+            }
             else if (new_type == GGML_TYPE_Q4_0_R4) {
                 if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_0;
                 else chunk_size_multiplier = 4;