glm4 unclamp for all except vulkan

Author: LostRuins

gpttype_adapter.cpp

@@ -1907,16 +1907,18 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
         printf("Warning: Only GGUF models can use max context above 16k. Max context lowered to 16k.\n");
         clamped_max_context_length = 16384;
     }
-    if (isGguf && file_format_meta.model_architecture == GGUFArch::ARCH_GLM4 && kcpp_data->n_batch > 16) {
+
+    #if defined(GGML_USE_VULKAN)
+    if (isGguf && file_format_meta.model_architecture == GGUFArch::ARCH_GLM4 && kcpp_data->n_ubatch > 16) {
         if(debugmode==1)
         {
-            printf("GLM-4 is broken on larger batch sizes. Clamp ignored in debug.\n");
+            printf("GLM-4 is broken on larger batch sizes in Vulkan. Clamp ignored in debug.\n");
         } else {
-            printf("GLM-4 is broken on larger batch sizes. Clamping batch size to 16.\n");
-            kcpp_data->n_batch = kcpp_data->n_ubatch = 16;
+            printf("GLM-4 is broken on larger batch sizes in Vulkan. Clamping ubatch size to 16.\n");
+            kcpp_data->n_ubatch = 16;
         }
-
     }
+    #endif
 
     kcpp_data->n_ctx = clamped_max_context_length;
     max_context_limit_at_load = clamped_max_context_length;

klite.embd

@@ -3440,7 +3440,7 @@ Current version indicated by LITEVER below.
 		"name":"GLM-4",
 		"user":"<|user|>\\n",
 		"user_end":"",
-		"assistant":"<|assistant|>",
+		"assistant":"<|assistant|>\\n",
 		"assistant_end":"",
 		"system":"<|system|>\\n",
 		"system_end":"",