load chat_template in from_pretrained (vLLM apply_chat_template fix)

tokenizer.py

@@ -580,22 +580,29 @@ class HybridDNATokenizer(PreTrainedTokenizer):
         return (save_directory,)
 
     @classmethod
-    def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
-        k = 6
-        auto_dna_tags = False
-
-        dna_config_path = os.path.join(pretrained_model_name_or_path, "dna_config.json")
-        tok_config_path = os.path.join(pretrained_model_name_or_path, "tokenizer_config.json")
-
-        if os.path.exists(dna_config_path):
-            with open(dna_config_path, "r") as f:
-                dna_config = json.load(f)
-            k = dna_config.get("k", 6)
-            auto_dna_tags = dna_config.get("auto_dna_tags", False)
-        elif os.path.exists(tok_config_path):
-            with open(tok_config_path, "r") as f:
-                tok_config = json.load(f)
-            k = tok_config.get("k", 6)
-            auto_dna_tags = tok_config.get("auto_dna_tags", False)
-
-        return cls(base_tokenizer_path=pretrained_model_name_or_path, k=k, auto_dna_tags=auto_dna_tags, **kwargs)
+    def from_pretrained(cls, pretrained_model_name_or_path, *args, **kwargs):
+        # Custom loader: PreTrainedTokenizer.from_pretrained is incompatible with this
+        # tokenizer's __init__ (it re-passes bos_token etc.), so construct directly from
+        # bio_config/dna_config. That bypasses standard metadata loading, so also read
+        # chat_template from tokenizer_config.json (vLLM / apply_chat_template need it).
+        import os as _os, json as _json
+        cfg = {}
+        for _name in ("bio_config.json", "dna_config.json"):
+            _p = _os.path.join(pretrained_model_name_or_path, _name)
+            if _os.path.exists(_p):
+                with open(_p, "r", encoding="utf-8") as _f:
+                    cfg = _json.load(_f)
+                break
+        _init = {"base_tokenizer_path": pretrained_model_name_or_path}
+        for _key in ("k", "tail", "auto_dna_tags"):
+            if _key in cfg:
+                _init[_key] = cfg[_key]
+        tok = cls(**_init, **kwargs)
+        if getattr(tok, "chat_template", None) is None:
+            _tc = _os.path.join(pretrained_model_name_or_path, "tokenizer_config.json")
+            if _os.path.exists(_tc):
+                with open(_tc, "r", encoding="utf-8") as _f:
+                    _c = _json.load(_f)
+                if _c.get("chat_template"):
+                    tok.chat_template = _c["chat_template"]
+        return tok