model : add text-only support for Kimi-VL (and find special tokens in text_config) (ggml-org#15051)

* basic kimi-vl textmodel conversion

* check config["text_config"] for special tokens

Author: gabriellarson

convert_hf_to_gguf.py

@@ -6389,6 +6389,7 @@ def prepare_tensors(self):
 
 @ModelBase.register("DeepseekV2ForCausalLM")
 @ModelBase.register("DeepseekV3ForCausalLM")
+@ModelBase.register("KimiVLForConditionalGeneration")
 class DeepseekV2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
 
@@ -6491,6 +6492,13 @@ def set_gguf_parameters(self):
     _experts: list[dict[str, Tensor]] | None = None
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # skip vision tensors and remove "language_model." for Kimi-VL
+        if "vision_tower" in name or "multi_modal_projector" in name:
+            return []
+
+        if name.startswith("language_model."):
+            name = name.replace("language_model.", "")
+
         # rename e_score_correction_bias tensors
         if name.endswith("e_score_correction_bias"):
             name = name.replace("e_score_correction_bias", "e_score_correction.bias")

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