Crack down fallback GGML_types

Author: Nexesenex

src/llama.cpp

@@ -19667,10 +19667,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             case GGML_TYPE_IQ1_M:
             case GGML_TYPE_Q2_K:
             case GGML_TYPE_Q3_K:
-            case GGML_TYPE_IQ4_XS: new_type = GGML_TYPE_IQ4_NL; break;
-            case GGML_TYPE_Q4_K:   new_type = GGML_TYPE_Q5_0;   break;
-            case GGML_TYPE_Q5_K:   new_type = GGML_TYPE_Q5_1;   break;
-            case GGML_TYPE_Q6_K:   new_type = GGML_TYPE_Q8_0;   break;
+            case GGML_TYPE_IQ4_XS:
+            case GGML_TYPE_Q4_K:   new_type = GGML_TYPE_IQ4_NL; break;
+            case GGML_TYPE_Q5_K:   new_type = GGML_TYPE_Q5_0;   break;
+            case GGML_TYPE_Q6_K:   new_type = GGML_TYPE_Q5_1;   break;
             default: throw std::runtime_error("\nUnsupported tensor size encountered\n");
         }
         if (tensor->ne[0] % ggml_blck_size(new_type) != 0) {