Revert to proper version that produced good BF16

ikawrakow#668 (comment)

Author: Thireus

convert_hf_to_gguf.py

@@ -3988,20 +3988,7 @@ def set_vocab(self):
         special_vocab._set_special_token(
             "bos", tokenizer.get_added_vocab()["<|endoftext|>"]
         )
-        special_vocab._set_special_token("eom", tokenizer.get_added_vocab()["<|observation|>"])  # 151338
-
-        # Fix chat template syntax error in GLM-4.5 models
-        if special_vocab.chat_template and isinstance(special_vocab.chat_template, str):
-            # Fix multiple syntax issues in GLM-4.5 chat template
-            template = special_vocab.chat_template
-            # Fix nested double quotes issue
-            template = template.replace('endswith("/nothink")', "endswith('/nothink')")
-            # Fix any other potential parentheses/tuple issues
-            template = template.replace(
-                "not visible_text(m.content).endswith('/nothink'))",
-                "not visible_text(m.content).endswith('/nothink')"
-            )
-            special_vocab.chat_template = template
+
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def set_gguf_parameters(self):
@@ -4048,7 +4035,7 @@ def modify_tensors(
             name = name.replace("language_model.", "")  # for multimodal variants
 
         # Handle main token embedding (but not layer-specific NextN embeddings)
-        if name == "model.embed_tokens.weight" and ".layers." not in name:
+        if name == "model.embed_tokens.weight":
             return [(self.map_tensor_name("token_embd.weight"), data_torch)]
 
         # Handle routed experts