[Core|Tune]: Static/Class variables with `tune.Trainable`

[Rohan138]

Description

Is it possible to use static/class variables or otherwise save attributes from an object without using Trainable.save? If not, is there some way to explicitly warn users against this using metaclasses?

Here's an example script that might fall under ray gotchas:

from ray import tune
from ray.rllib.utils.annotations import override
from typing import Dict

class Inner(tune.Trainable):
    b = 0
    @override(tune.Trainable)
    def setup(self, config: Dict):
        self.a = 0

    @override(tune.Trainable)
    def step(self):
        self.a += 1
        Inner.b += 1
        return {"inner": self.a, "inner_cls": Inner.b}

class Outer(tune.Trainable):
    @override(tune.Trainable)
    def setup(self, config: Dict):
        self.ctr = 0

    @override(tune.Trainable)
    def step(self):
        tune.run(Inner, stop={"training_iteration": 5})
        self.ctr += 1
        return {"outer": self.ctr}

def run(cls, stop):
    obj = cls()
    obj.setup(config={})
    for _ in range(stop):
        print(obj.step())

class PlainInner:
    b = 0
    def setup(self, config: Dict):
        self.a = 0

    def step(self):
        self.a += 1
        PlainInner.b += 1
        return {"inner": self.a, "inner_cls": PlainInner.b}

class PlainOuter:
    def setup(self, config: Dict):
        self.ctr = 0

    def step(self):
        run(PlainInner, 5)
        self.ctr += 1
        return {"outer": self.ctr}

if __name__=="__main__":
    tune.run(Outer, stop={"training_iteration": 5})
    run(PlainOuter, 5)

Outer.b is 5, but PlainOuter.b is 25.

Use case

No response

[matthewdeng]

Not sure if I understand the original question, but does this explain the behavior you're seeing?