Adjustments on attn_k

Author: Nexesenex

src/llama.cpp

@@ -15531,7 +15531,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             new_type = (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) ? GGML_TYPE_IQ4_XS : GGML_TYPE_IQ3_XXS;
         }
         else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_XL || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) {
-            new_type = (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) ? GGML_TYPE_Q4_K : GGML_TYPE_IQ3_XXS;
+            new_type = (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) ? GGML_TYPE_IQ4_XS : GGML_TYPE_IQ3_XXS;
         }
         else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XL) {
             new_type = (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) ? GGML_TYPE_Q4_K : GGML_TYPE_IQ3_S;
@@ -15573,23 +15573,21 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             }
             else new_type = GGML_TYPE_Q8_0;
         }
-        else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ1_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) &&
-                (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
-            new_type = GGML_TYPE_IQ1_M;
-        }
-        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_M && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
-            new_type = GGML_TYPE_IQ2_XXS;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
+            if (qs.model.hparams.n_gqa() >= 4 || qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_IQ2_XS;
+            else new_type = GGML_TYPE_IQ2_XXS;
         }
         else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_XL || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS) {
-            if (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_IQ2_XS;
-            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_XL) new_type = GGML_TYPE_IQ2_XXS;
+            if (qs.model.hparams.n_gqa() >= 4 || qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_IQ2_S;
+            else if (qs.model.hparams.n_gqa() >= 2) new_type = GGML_TYPE_IQ2_XS;
+            else new_type = GGML_TYPE_IQ2_XXS;
         }
-        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
-            new_type = GGML_TYPE_IQ2_S;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) {
+            if (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_IQ2_S;
         }
-        else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XL) &&
-                (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
-            new_type = GGML_TYPE_IQ3_XXS;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XL) {
+            if (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_IQ3_XXS;
+            else new_type = GGML_TYPE_IQ2_S;
         }
         else if ((ftype == LLAMA_FTYPE_MOSTLY_Q2_K || ftype == LLAMA_FTYPE_MOSTLY_Q2_K_L) &&
                 (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {