Update app.py

app.py

@@ -5,18 +5,18 @@ import numpy as np
 app = FastAPI()
 
 # Load ONNX model only
-session = InferenceSession("model.onnx")
+# session = InferenceSession("model.onnx")
 
 @app.post("/predict")
 async def predict(inputs: dict):
     # Expect pre-tokenized input from client
-    input_ids = np.array(inputs["input_ids"], dtype=np.int64)
-    attention_mask = np.array(inputs["attention_mask"], dtype=np.int64)
+    ##input_ids = np.array(inputs["input_ids"], dtype=np.int64)
+    #attention_mask = np.array(inputs["attention_mask"], dtype=np.int64)
     
     # Run model
-    outputs = session.run(None, {
-        "input_ids": input_ids,
-        "attention_mask": attention_mask
-    })
-    
-    return {"embedding": outputs[0].tolist()}
+    #outputs = session.run(None, {
+    #    "input_ids": input_ids,
+    #    "attention_mask": attention_mask
+    #})
+    return "Status ok"
+    #return {"embedding": outputs[0].tolist()}