Remove lazy loading and load model on instantiation

Load the model immediately when the CrossEncoder class is instantiated
instead of waiting for the first predict() call.

modeling_zeranker.py

@@ -108,6 +108,16 @@ def load_model(
     return tokenizer, model
 
 
+def __init__(self, *args: Any, **kwargs: Any) -> None:
+    # Load the model immediately on instantiation
+    self.inner_tokenizer, self.inner_model = load_model(global_device)
+    self.inner_model.eval()
+    self.inner_model.gradient_checkpointing_disable()
+    self.inner_yes_token_id = self.inner_tokenizer.encode(
+        "Yes", add_special_tokens=False
+    )[0]
+
+
 def predict(
     self,
     query_documents: list[tuple[str, str]] | None = None,
@@ -125,14 +135,6 @@ def predict(
             raise ValueError("query_documents or sentences must be provided")
         query_documents = [[sentence[0], sentence[1]] for sentence in sentences]
 
-    if not hasattr(self, "inner_model"):
-        self.inner_tokenizer, self.inner_model = load_model(global_device)
-        self.inner_model.eval()
-        self.inner_model.gradient_checkpointing_disable()
-        self.inner_yes_token_id = self.inner_tokenizer.encode(
-            "Yes", add_special_tokens=False
-        )[0]
-
     model = self.inner_model
     tokenizer = self.inner_tokenizer
 
@@ -208,10 +210,10 @@ def to_device(self: _CE, new_device: torch.device) -> None:
     global_device = new_device
 
 
+_CE.__init__ = __init__
 _CE.predict = predict
+_CE.to = to_device
 
 from transformers import Qwen3Config
 
 ZEConfig = Qwen3Config
-
-_CE.to = to_device