model: support youtu-vl model (#18479)

* Support Youtu-VL Model

* merge code

* fix bug

* revert qwen2 code & support rsplit in minja.hpp

* update warm info

* fix annotation

* u

* revert minja.hpp

* fix

* Do not write routed_scaling_factor to gguf when routed_scaling_factor is None

* fix expert_weights_scale

* LGTM after whitespace fixes

* fix

* fix

* fix

* layers to layer_index

* enum fix

---------

Co-authored-by: Xuan-Son Nguyen <son@huggingface.co>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

gguf-py/gguf/gguf_writer.py

@@ -1129,11 +1129,40 @@ class GGUFWriter:
         self.add_uint32(Keys.ClipVision.Projector.SCALE_FACTOR, value)
 
     def add_vision_n_wa_pattern(self, value: int) -> None:
+        """Add window attention pattern interval for vision models.
+
+        This defines the pattern interval for window attention vs full attention layers.
+        For example, if n_wa_pattern=4, then layers 3, 7, 11, ... use full attention,
+        while other layers use window attention.
+
+        Used by models like Qwen2.5-VL where full attention layers follow a regular pattern.
+        """
         self.add_uint32(Keys.ClipVision.N_WA_PATTERN, value)
 
+    def add_vision_wa_layer_indexes(self, layers: Sequence[int]) -> None:
+        """Add explicit layer indexes that use full attention in vision models.
+
+        This specifies the exact layer indices (0-based) that should use full attention
+        instead of window attention. All other layers will use window attention.
+
+        Args:
+            layers: List of layer indices that use full attention (e.g., [3, 7, 11, 15])
+
+        Used by models like YoutuVL where full attention layers are explicitly specified
+        rather than following a regular pattern.
+
+        Difference from add_vision_n_wa_pattern:
+        - n_wa_pattern: Defines a regular interval pattern (every Nth layer uses full attention)
+        - wa_layer_indexes: Explicitly lists which layers use full attention (irregular pattern)
+        """
+        self.add_array(Keys.ClipVision.WA_LAYER_INDEXES, layers)
+
     def add_vision_is_deepstack_layers(self, layers: Sequence[bool]) -> None:
         self.add_array(Keys.ClipVision.IS_DEEPSTACK_LAYERS, layers)
 
+    def add_vision_window_size(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.WINDOW_SIZE, value)
+
     # audio models
 
     def add_audio_projection_dim(self, value: int) -> None: