Artur Niederfahrenhorst

[RLlib] Remove unneeded args from offline learning examples

1

## Why are these changes needed? We largely got rid of replay buffers in the context of offline RL in RLlib, but the examples have not beed changed. #26665 ##...

## Why are these changes needed? This PR aims to make much of our core algorithm code more readable. Many algorithms access three functions: `self.workers.local_worker()` and `self.workers.local_worker().set_global_vars()` and `self.workers.remote_workers()` (see...

[RLlib] Move learning_starts logic into execution plans

2

## Why are these changes needed? * `learning_starts` should be renamed to something more descriptive: `num_steps_sampled_before_learning_starts` * Should be moved out of replay buffer config according to our philosophy: Algorithm...

Signed-off-by: Artur Niederfahrenhorst ## Why are these changes needed? Make episodic buffer conform replay buffer API. ## Checks - [x] I've signed off every commit(by using the -s flag, i.e.,...

Signed-off-by: Artur Niederfahrenhorst ## Why are these changes needed? SAC's config has a paramter "optimization_config" that will not be used during execution because "optimization" is accessed. We have this discrepancy...

### What happened + What you expected to happen Running any library script with tf metal 0.6.0 gives me the following error: ``` Fatal Python error: Bus error Thread 0x000000032e3a7000...

Epsilon greedy exploration ignores custom action distributions (duplicate)

2

### This is a temporary duplicate of #26604 , since our team is unable to push to the PR in question. ## Checks - [ ] I've signed off every...

### This is a temporary duplicate of #27527, since our team is unable to push to the PR in question ## Checks - [ ] I've signed off every commit(by...

[RLlib] Add filters to connector pipeline

3

## Why are these changes needed? In our efforts to include our diverse experience processing steps into the connector pipeline, this PR includes filters. As long as connectors can be...

[RLlib] Bump torch versions and CUDA toolkit

2

## Why are these changes needed? After conducting a row of tests, we have detected a throughput regression in PyTorch, primarily in our APPO release test. This has been linked...