[tune] Allow to set buffer_length via tune.run #15810
Conversation
There was a problem hiding this comment.
Thanks for your contribution!
One thing I'm wondering is if we should really include this in the top level tune.run() API, or if this should be an option for the RayTrialExecutor only (which can be passed to tune.run()). The reason is that we try to prevent the tune.run() parameter list from growing too much, and technical parameters such as the result buffer length (especially when there's another way to set them, e.g. via environment variables) are prime candidates to exclude from the high level API.
Would you be happy with that as well?
Also cc @richardliaw would like to get your opinion here
python/ray/tune/tune.py
Outdated
| @@ -175,6 +176,8 @@ def run( | |||
| ``ray.tune.Stopper``, which allows users to implement | |||
| custom experiment-wide stopping (i.e., stopping an entire Tune | |||
| run based on some time constraint). | |||
| result_buffer_length (int): Maximum number of results to buffer. | |||
| If set, env var 'TUNE_RESULT_BUFFER_LENGTH' will be ignored. | |||
There was a problem hiding this comment.
| If set, env var 'TUNE_RESULT_BUFFER_LENGTH' will be ignored. | |
| If set, env var ``TUNE_RESULT_BUFFER_LENGTH`` will be ignored. |
|
I think it's okay to only include this in RayTrialExecutor. In my case, I just want to be able to infer and set the value of buffer_length from other parameters such as validation period. |
|
That sounds great. Would you be able to update the PR? I'm happy to approve afterwards and merge once CI passes |
llan-ml
force-pushed
the
tune_run_buffer_length
branch
from
a750003
to
de58dbe
Compare
|
updated and rebased |
See discussion here.
Checks
scripts/format.shto lint the changes in this PR.