[Question] How is np_random set for AsyncVectorEnvs and why does the generator change?

[Daraan]

Question

Hello :)
In my setting I change the np_random generator of my envs

seed_sequence = np.random.SeedSequence(
            env_seed,
            spawn_key=(worker_index, env_context.vector_index, env_runner.config.in_evaluation),
        )
rng = np.random.default_rng(seed_sequence)
rngs = rng.spawn(env.num_envs if _is_async(env) else 1)
# env: SyncVectorEnv | AsyncVectorEnv
env.set_attr("np_random", rngs)
env.np_random  # <-- if async new generator, based on rngs but not the same as rngs

For sync environments the generators are like I expected identical.

For an async ones the final env.np_random that I get back from the pipe is not the same. What I wonder is, why does entropy and seed_sequence of the bit generator differ?
What I can tell is that my original rngs are used to create new random generators. I would like to now how and where in the code this happens.
In the end I would like to test if the framework that I use does not overwrite my own generators, with their own seed logic, which currently a bit difficult.
Thank you very much if you can provide me some insights.

[pseudo-rnd-thoughts]

Could you provide more information about the setup, but in short, you shouldn't be manually setting the np_randm. Rather, this should be done in reset(seed=seed), where this should assign the np_random generator.

[Daraan]

Could you provide more information about the setup, but in short, you shouldn't be manually setting the np_randm. Rather, this should be done in reset(seed=seed), where this should assign the np_random generator.

Sure, thanks.
As a framework I use ray within reset is not under my control. I make it so that reset(seed=None) is called and therefore np_random is not changed, i.e. the one I set manually beforehand is not discarded.

[pseudo-rnd-thoughts]

Ahh ok, could you reset the environment before you pass it to Rllib?