Big NDArray generate Operand Tensor meet protobuf exceeded maximum protobuf size of 2GB ?

Hi :
from spark DataFrame generate org.tensorflow.ndarray.DoubleNdArray , after I want to generate Operand[TFloat64] tensor , meet error

scala> val featureVector = SparkConverter.sparkDataframeFeatureVectorConvertTfTensor(finalInputDf,"final_features" )
featureVector: org.tensorflow.ndarray.DoubleNdArray = org.tensorflow.ndarray.impl.dense.DoubleDenseNdArray@e3f6a6a0
scala> val ft  = tf.constant(featureVector)
[libprotobuf ERROR external/com_google_protobuf/src/google/protobuf/message_lite.cc:451] tensorflow.AttrValue exceeded maximum protobuf size of 2GB: 6279090916
org.tensorflow.exceptions.TFInvalidArgumentException: AttrValue missing value with expected type 'tensor'
         for attr 'value'
        ; NodeDef: {{node Const}}; Op<name=Const; signature= -> output:dtype; attr=value:tensor; attr=dtype:type>
  at org.tensorflow.internal.c_api.AbstractTF_Status.throwExceptionIfNotOK(AbstractTF_Status.java:87)
  at org.tensorflow.EagerOperationBuilder.execute(EagerOperationBuilder.java:314)
  at org.tensorflow.EagerOperationBuilder.build(EagerOperationBuilder.java:77)
  at org.tensorflow.EagerOperationBuilder.build(EagerOperationBuilder.java:64)
  at org.tensorflow.op.core.Constant.create(Constant.java:1350)
  at org.tensorflow.op.core.Constant.tensorOf(Constant.java:521)
  at org.tensorflow.op.Ops.constant(Ops.java:1669)
  ... 59 elided

but if I filter some small part Dataframe is ok

scala> val featureVector = SparkConverter.sparkDataframeFeatureVectorConvertTfTensor(finalInputDf.filter(col("pay_status").equalTo(1)),"final_features" )
featureVector: org.tensorflow.ndarray.DoubleNdArray = org.tensorflow.ndarray.impl.dense.DoubleDenseNdArray@627077a

scala> val ft_small  = tf.constant(featureVector)
ft_small: org.tensorflow.op.core.Constant[org.tensorflow.types.TFloat64] = <Const 'Const_2'>

scala> ft_small.asTensor().numBytes()
res43: Long = 1058424696

Commencez par le chemin d’appel signalé de Ops.constant et Constant.tensorOf, en utilisant comme entrées contrastées le DoubleNdArray généré par SparkConverter et le DataFrame filtré plus petit. Reproduisez l’erreur de taille de protobuf et déterminez le comportement attendu pour les tenseurs de plus de 2GB ; l’issue ne définit aucune modification demandée ni aucun test clair de fin.