Update handler.py

handler.py

@@ -38,13 +38,15 @@ class EndpointHandler:
         # Determine device
         self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
         
-        # Eagle 2.5 uses Qwen2VLProcessor - import and load directly
-        from transformers import Qwen2VLProcessor, Qwen2VLForConditionalGeneration
+        # Eagle 2.5 uses Qwen2VL architecture - use AutoProcessor with use_fast=False
+        # to avoid the broken Eagle2_5_VLVideoProcessorFast class
+        from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
         
-        self.processor = Qwen2VLProcessor.from_pretrained(
+        self.processor = AutoProcessor.from_pretrained(
             model_id,
             trust_remote_code=True,
             token=hf_token,
+            use_fast=False,  # Avoid the broken video processor
         )
         
         # Set padding side for batch processing