def GetPhyBlobNumpy(i, phy_blob_object):
name = "{}/{}".format(self.logical_blob_name, i)
blob_register.SetObject4BlobName(name, phy_blob_object)
return (
eager_blob_util.EagerPhysicalBlob(name).numpy_list()
if self.is_tensor_list
else eager_blob_util.EagerPhysicalBlob(name).numpy()
)
def FetchBlobNumpy(blob_object):
consistent_blob_name = None
def BoxingToSingleDevice(builder):
parallel_conf = placement_pb.ParallelConf()
parallel_conf.device_tag = blob_object.parallel_desc_symbol.device_tag
parallel_conf.device_name.append("{}:{}".format(0, 0))
tmp_parallel_desc_symbol = builder.GetParallelDescSymbol(parallel_conf)
tmp_op_arg_parallel_attr = op_arg_util.OpArgParallelAttribute(
tmp_parallel_desc_symbol,
blob_object.op_arg_parallel_attr.sbp_parallel,
blob_object.op_arg_parallel_attr.opt_mirrored_parallel,
)
with oneflow.scope.placement(
self.parallel_conf.device_tag, list(self.parallel_conf.device_name)
):
tmp_blob_object = boxing_util.BoxingTo(
builder, blob_object, tmp_op_arg_parallel_attr
)
nonlocal consistent_blob_name
consistent_blob_name = "{}-consistent".format(self.logical_blob_name)
if not blob_register.HasObject4BlobName(consistent_blob_name):
blob_register.SetObject4BlobName(
consistent_blob_name, tmp_blob_object
)
vm_util.LogicalRun(BoxingToSingleDevice)
return eager_blob_util.EagerPhysicalBlob(consistent_blob_name).numpy()