Changing PYUObject.device.party is same as sf.to?

[Krout0n]

A minimum reproducible code

import secretflow as sf
import pandas as pd

sf.init(['alice', 'bob'], num_cpus=8, log_to_driver=False)

alice, bob = sf.PYU('alice'), sf.PYU('bob')
alice_val = alice(lambda: 1)()
bob_val = bob(lambda: 2)()
bob_owned_alice_val = alice_val.to(bob)

bob(lambda x, y: x + y)(bob_owned_alice_val, bob_val) # nothing wrong with this line
bob(lambda x, y: x + y)(alice_val, bob_val) # this line causes raising AssertionError, nothing wrong

alice_val.device.party = 'bob'
bob(lambda x, y: x + y)(alice_val, bob_val) # this line causes nothing, but is this same as sf.to ?

I know assigning PYUObject.device.party is not right way but I was just wondering. Is this consistent with the expected behavior?

[rivertalk]

Hi @Krout0n , it's really a good question.

I know assigning PYUObject.device.party is not right way but I was just wondering. Is this consistent with the expected behavior?

For the moment, I think the answer is positive, the behavior is the same as alice_val.to since the underline distributed engine will fetch the content of alice_val to bob automatically (lazy and silently).

But the it is not a right to do this, we should use val.to to handle this semantic, since the fetching implementation may be changed in the future, but the to semantic will be stable.