Upload configuration_llava_qwen.py with huggingface_hub

configuration_llava_qwen.py

@@ -0,0 +1,42 @@
+# Configuration for LLaVA Qwen3 model
+# Upload this file to your HuggingFace repo for trust_remote_code=True
+
+from transformers import Qwen2Config
+
+try:
+    from transformers import Qwen3Config
+    QWEN3_AVAILABLE = True
+except ImportError:
+    QWEN3_AVAILABLE = False
+    Qwen3Config = Qwen2Config  # Fallback
+
+
+class LlavaQwen3Config(Qwen3Config if QWEN3_AVAILABLE else Qwen2Config):
+    """Configuration class for LLaVA Qwen3 model."""
+
+    model_type = "llava_qwen3"
+
+    def __init__(
+        self,
+        mm_vision_tower=None,
+        mm_hidden_size=None,
+        mm_projector_type="mlp2x_gelu",
+        mm_vision_select_layer=-2,
+        mm_vision_select_feature="patch",
+        mm_patch_merge_type="flat",
+        mm_use_im_start_end=False,
+        mm_use_im_patch_token=False,
+        image_aspect_ratio="pad",
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+
+        self.mm_vision_tower = mm_vision_tower
+        self.mm_hidden_size = mm_hidden_size
+        self.mm_projector_type = mm_projector_type
+        self.mm_vision_select_layer = mm_vision_select_layer
+        self.mm_vision_select_feature = mm_vision_select_feature
+        self.mm_patch_merge_type = mm_patch_merge_type
+        self.mm_use_im_start_end = mm_use_im_start_end
+        self.mm_use_im_patch_token = mm_use_im_patch_token
+        self.image_aspect_ratio = image_aspect_ratio