check config["text_config"] for special tokens

Author: gabriellarson

convert_hf_to_gguf.py

@@ -6113,19 +6113,6 @@ def set_vocab(self):
             self.gguf_writer.add_token_merges(merges)
 
             special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
-
-            # fix for Kimi-VL: Use <|im_end|> as EOS token instead of [EOS]
-            # This ensures text generation stops correctly at sentence boundaries, rather than at commas (which would happen with wrong EOS token)
-            if self.hf_arch == "KimiVLForConditionalGeneration":
-                im_end_id = None
-                for i, token in enumerate(tokens):
-                    if token == "<|im_end|>":
-                        im_end_id = i
-                        break
-                if im_end_id is not None:
-                    logger.info(f"Kimi-VL: Overriding EOS token from {special_vocab.special_token_ids.get('eos', 'N/A')} to <|im_end|> (ID: {im_end_id})")
-                    special_vocab.special_token_ids["eos"] = im_end_id
-
             special_vocab.add_to_gguf(self.gguf_writer)
         else:
             raise NotImplementedError(f"Deepseek pre-tokenizer {tokpre!r} is not supported yet!")

gguf-py/gguf/vocab.py

@@ -312,7 +312,11 @@ def _try_load_from_config_json(self, path: Path) -> bool:
         with open(config_file, encoding = 'utf-8') as f:
             config = json.load(f)
         for typ in self.special_token_types:
-            self._set_special_token(typ, config.get(f'{typ}_token_id'))
+            token_id = config.get(f'{typ}_token_id')
+            # If not found at root, check in text_config (for multimodal models like Kimi-VL)
+            if token_id is None and 'text_config' in config:
+                token_id = config['text_config'].get(f'{typ}_token_id')
+            self._set_special_token(typ, token_id)
         return True