-
Notifications
You must be signed in to change notification settings - Fork 5.8k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[RLlib] [Release tests] Make rollout fragment length auto for release tests #30079
[RLlib] [Release tests] Make rollout fragment length auto for release tests #30079
Conversation
Signed-off-by: Avnish <[email protected]>
@@ -10,7 +10,7 @@ ddpg-hopperbulletenv-v0: | |||
config: | |||
actor_hiddens: [256, 256] | |||
critic_hiddens: [256, 256] | |||
n_step: 3 | |||
n_step: 1 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Here, too, remove rollout_fragment_length (or set to "auto") further below entirely.
@@ -26,7 +26,7 @@ cql-halfcheetahbulletenv-v0: | |||
tau: 0.005 | |||
target_entropy: auto | |||
no_done_at_end: false | |||
n_step: 3 | |||
n_step: 1 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Wait, let's do this the other way around and keep the n_step setting.
So the new config logic is: rollout_fragment_length
is "auto" b/c it's a lot of times dependent on other settings, like train_batch_size, num_envs_per_worker, and num_rollout_workers (PPO) or just needs to be at least n_step
to make sense.
The RolloutWorkers now infer this setting from these other settings, if "auto", otherwise, take the user provided values, but this could lead to errors, like in this case.
So let's just remove the rollout_fragment_length setting entirely here (or set to its default: "auto").
@@ -19,7 +19,7 @@ sac-halfcheetahbulletenv-v0: | |||
tau: 0.005 | |||
target_entropy: auto | |||
no_done_at_end: false | |||
n_step: 3 | |||
n_step: 1 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Here, too, remove rollout_fragment_length (or set to "auto") further below entirely.
okie sg :) |
So I'm going to try this, but I suspect that we might need to increase the timestep limits on these tests in response, in order to meet the minimum reward threshold. |
Signed-off-by: Avnish <[email protected]>
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Perfect, thanks for the investigation and fixes @avnishn !
…te_release_test_n_step_frag_length
Signed-off-by: Avnish <[email protected]>
lets merge pending the release tests all getting kicked off successfully: |
Signed-off-by: Avnish <[email protected]>
Signed-off-by: Avnish <[email protected]>
… tests (ray-project#30079) Signed-off-by: Avnish <[email protected]> Signed-off-by: Weichen Xu <[email protected]>
SAC, DDPG, and CQL release tests were broken since we started enforcing that the
rollout_fragment_length
must be greater than or equal ton_step
when computing n step returns.This error was triggered I think by #29854 and eventually propagated to the release tests. Its likely that validation wasn't being run properly before #29854 .
Signed-off-by: Avnish [email protected]
Why are these changes needed?
Related issue number
Checks
git commit -s
) in this PR.scripts/format.sh
to lint the changes in this PR.