Fix a bug that causes a crash when decoding for language models because initially targets is empty and only "infer_targets" is populated.

PiperOrigin-RevId: 317003338

Author: royaurko

tensor2tensor/utils/t2t_model.py

@@ -1718,7 +1718,7 @@ def estimator_spec_predict(self, features, use_tpu=False):
 
     inputs = features.get("inputs")
     if inputs is None:
-      inputs = features["targets"]
+      inputs = features.get("targets")
 
     predictions = {
         "outputs": outputs,