Soften the token embeddings bump for experts >= 4

Author: Nexesenex

src/llama.cpp

@@ -17624,20 +17624,16 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             new_type = qs.params->token_embedding_type;
         } else {
             if (qs.model.hparams.n_expert >= 4) {
-                if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K   || ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S  || ftype == LLAMA_FTYPE_MOSTLY_Q2_K_L ||
-                         ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M  || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L ||
-                         ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XL) {
-                new_type = GGML_TYPE_Q4_K;
-                }
-                else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_S  || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS ||
-                    ftype == LLAMA_FTYPE_MOSTLY_IQ1_XL || ftype == LLAMA_FTYPE_MOSTLY_IQ1_XS) {
-                new_type = GGML_TYPE_IQ3_S;
-                }
-                else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS  || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S  || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ||
-                         ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S ||
-                         ftype == LLAMA_FTYPE_MOSTLY_IQ3_M   || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XL || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XL) {
-                new_type = GGML_TYPE_IQ4_XS;
-                }
+                if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S       || ftype == LLAMA_FTYPE_MOSTLY_Q2_K)    new_type = GGML_TYPE_Q3_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_L  || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S  || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M ||
+                         ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L  || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XL) new_type = GGML_TYPE_Q4_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_XS  || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S   || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M  ||
+                         ftype == LLAMA_FTYPE_MOSTLY_IQ1_XL) new_type = GGML_TYPE_IQ3_XXS;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS  || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S  ||
+                         ftype == LLAMA_FTYPE_MOSTLY_IQ2_M)  new_type = GGML_TYPE_IQ3_S;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XL  || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS ||
+                         ftype == LLAMA_FTYPE_MOSTLY_IQ3_S   || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M   || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XL ||
+                         ftype == LLAMA_FTYPE_MOSTLY_IQ2_XL) new_type = GGML_TYPE_IQ4_XS;
             }
             else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_L) {
                 if (qs.model.hparams.n_vocab >= 127999 || qs.model.hparams.n_head() <= 20) new_type = GGML_TYPE_Q2_K;