Revamp a bit output weight

for more granularity in low quants.

Author: Nexesenex

src/llama.cpp

@@ -16426,23 +16426,35 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
                 new_type = GGML_TYPE_Q8_0;
             }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q2_K   || ftype == LLAMA_FTYPE_MOSTLY_Q2_K_L) {
+                if (qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q6_K;
+                else new_type = GGML_TYPE_Q4_K;
+            }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
+                if (qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q6_K;
+                else if (qs.model.hparams.n_vocab >= 127999) new_type = GGML_TYPE_Q5_K;
+                else new_type = GGML_TYPE_Q6_K;
+            }
             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_expert >= 4) new_type = GGML_TYPE_Q6_K;
                 else if (qs.model.hparams.n_vocab >= 127999) new_type = GGML_TYPE_IQ4_XS;
+				else if (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_Q4_K;
+                else new_type = GGML_TYPE_IQ4_XS;
+            }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_XL || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) {
+                if (qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q6_K;
                 else new_type = GGML_TYPE_Q4_K;
             }
-            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_XL || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS ||
-                     ftype == LLAMA_FTYPE_MOSTLY_IQ2_S  || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XL ||
-                     ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q2_K   || ftype == LLAMA_FTYPE_MOSTLY_Q2_K_L) {
+            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_expert >= 4) new_type = GGML_TYPE_Q6_K;
                 else if (qs.model.hparams.n_vocab >= 127999) new_type = GGML_TYPE_Q4_K;
                 else new_type = GGML_TYPE_Q5_K;
             }
-            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) {
                 if (qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q6_K;
                 else new_type = GGML_TYPE_Q5_K;
             }
-            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S  || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) {
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_S  || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) {
                 if (qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q6_K;
                 else if (qs.model.hparams.n_vocab >= 127999) new_type = GGML_TYPE_Q5_K;
                 else new_type = GGML_TYPE_Q6_K;