Do not merge, please review first. Tests still required to ensure logging and training is the same
Also pending some clarifications from the metaworld/garage work group regarding the training flow in general and whether or not updating tasks every epoch is required in v2.
One other issue, possibly very relevant: runtimes are higher. The change is likely in the underlying libraries, metaworld and/or garage, since this PR installs from master instead. It is impossible to tell whether these changes that led to higher runtime actually fixed something or not, so it is a risk (or it is the other way, that the branch we were pulling from prior to this PR is actually the latest version that fixes a bug, but hasn't been merged yet). On a p3.2x is taking 31 sec to update gradient vs 16 sec in the prior version.
Do not merge, please review first. Tests still required to ensure logging and training is the same
Also pending some clarifications from the metaworld/garage work group regarding the training flow in general and whether or not updating tasks every epoch is required in v2.
One other issue, possibly very relevant: runtimes are higher. The change is likely in the underlying libraries, metaworld and/or garage, since this PR installs from master instead. It is impossible to tell whether these changes that led to higher runtime actually fixed something or not, so it is a risk (or it is the other way, that the branch we were pulling from prior to this PR is actually the latest version that fixes a bug, but hasn't been merged yet). On a p3.2x is taking 31 sec to update gradient vs 16 sec in the prior version.