Fix chat template detection

Changes suggested by @ubergarm - ikawrakow#668 (comment)

Author: Thireus

src/llama.cpp

@@ -1740,8 +1740,8 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_DEEPSEEK_3,
     LLM_CHAT_TEMPLATE_COMMAND_R,
     LLM_CHAT_TEMPLATE_LLAMA_3,
-    LLM_CHAT_TEMPLATE_CHATGML_3,
-    LLM_CHAT_TEMPLATE_CHATGML_4,
+    LLM_CHAT_TEMPLATE_CHATGLM_3,
+    LLM_CHAT_TEMPLATE_CHATGLM_4,
     LLM_CHAT_TEMPLATE_MINICPM,
     LLM_CHAT_TEMPLATE_EXAONE_3,
     LLM_CHAT_TEMPLATE_RWKV_WORLD,
@@ -1781,8 +1781,8 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "deepseek3",         LLM_CHAT_TEMPLATE_DEEPSEEK_3        },
     { "command-r",         LLM_CHAT_TEMPLATE_COMMAND_R         },
     { "llama3",            LLM_CHAT_TEMPLATE_LLAMA_3           },
-    { "chatglm3",          LLM_CHAT_TEMPLATE_CHATGML_3         },
-    { "chatglm4",          LLM_CHAT_TEMPLATE_CHATGML_4         },
+    { "chatglm3",          LLM_CHAT_TEMPLATE_CHATGLM_3         },
+    { "chatglm4",          LLM_CHAT_TEMPLATE_CHATGLM_4         },
     { "minicpm",           LLM_CHAT_TEMPLATE_MINICPM           },
     { "exaone3",           LLM_CHAT_TEMPLATE_EXAONE_3          },
     { "rwkv-world",        LLM_CHAT_TEMPLATE_RWKV_WORLD        },
@@ -20712,7 +20712,17 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     //  - qs.n_attention_wv == 3 * model.hparams.n_layer for Encoder-Decoder models
     //  - model.arch == LLM_ARCH_DECI                    for Deci-Nemotron   models
     //
-    GGML_ASSERT((qs.n_attention_wv == 0 || qs.n_attention_wv == (int)model.hparams.n_layer || qs.n_attention_wv == 3 * (int)model.hparams.n_layer || model.arch == LLM_ARCH_DECI) && "n_attention_wv is unexpected");
+    //GGML_ASSERT((qs.n_attention_wv == 0 || qs.n_attention_wv == (int)model.hparams.n_layer || qs.n_attention_wv == 3 * (int)model.hparams.n_layer || model.arch == LLM_ARCH_DECI) && "n_attention_wv is unexpected");
+    // allow any count for GLM4-MoE, but still enforce for all others
+    if (model.arch != LLM_ARCH_GLM4_MOE) {
+        GGML_ASSERT(
+             qs.n_attention_wv == 0
+          || qs.n_attention_wv == (int)model.hparams.n_layer
+          || qs.n_attention_wv == 3 * (int)model.hparams.n_layer
+          || model.arch == LLM_ARCH_DECI
+          && "n_attention_wv is unexpected"
+        );
+    }
 
     size_t total_size_org = 0;
     size_t total_size_new = 0;
@@ -23841,6 +23851,11 @@ static llm_chat_template llama_chat_detect_template(const std::string & tmpl) {
                 return LLM_CHAT_TEMPLATE_LLAMA_2;
             }
         }
+    } else if (tmpl_contains("[gMASK]sop")) {
+        // chatglm3-6b
+        return LLM_CHAT_TEMPLATE_CHATGLM_3;
+    } else if (tmpl_contains("[gMASK]<sop>")) {
+        return LLM_CHAT_TEMPLATE_CHATGLM_4;
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>")) {
         return LLM_CHAT_TEMPLATE_PHI_3;
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
@@ -23873,11 +23888,6 @@ static llm_chat_template llama_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_COMMAND_R;
     } else if (tmpl_contains("<|start_header_id|>") && tmpl_contains("<|end_header_id|>")) {
         return LLM_CHAT_TEMPLATE_LLAMA_3;
-    } else if (tmpl_contains("[gMASK]sop")) {
-        // chatglm3-6b
-        return LLM_CHAT_TEMPLATE_CHATGML_3;
-    } else if (tmpl_contains("[gMASK]<sop>")) {
-        return LLM_CHAT_TEMPLATE_CHATGML_4;
     } else if (tmpl_contains(LU8("<用户>"))) {
         // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
         return LLM_CHAT_TEMPLATE_MINICPM;
@@ -24160,7 +24170,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|start_header_id|>assistant<|end_header_id|>\n\n";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_3) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_3) {
         // chatglm3-6b
         ss << "[gMASK]" << "sop";
         for (auto message : chat) {
@@ -24170,7 +24180,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|assistant|>";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_4) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_4) {
         ss << "[gMASK]" << "<sop>";
         for (auto message : chat) {
             std::string role(message->role);