Update llama-quant.cpp

Update llama-quant.cpp

Author: Nexesenex

src/llama-quant.cpp

@@ -255,7 +255,7 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
             && qs.model.hparams.n_gqa() >= 2)
             new_type = GGML_TYPE_Q5_K;
         else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS) && qs.model.hparams.n_gqa() >= 2) {
-            new_type = GGML_TYPE_Q5_K;
+            new_type = GGML_TYPE_Q6_K;
         }
         // else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
                 // use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K;
@@ -360,13 +360,13 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
                 // if (use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q6_K;
             // }
         // }
-        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q5_K;
+        // else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q5_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K;
         else if (i_layer < n_layer/8 && (ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS)
             && !qs.has_imatrix && use_more_bits(i_layer, n_layer)) {
             new_type = GGML_TYPE_Q5_K;
         }
-        else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S && use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q6_K;
+        // else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S && use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q6_K;
 		else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && i_layer < n_layer/8) {
             new_type = GGML_TYPE_Q5_K;