[RLLib] Make movement of tensors to device only happen once #36091
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: avnishn <[email protected]>
Signed-off-by: Avnish <[email protected]>
…appen once Signed-off-by: avnishn <[email protected]>
avnishn
requested review from
sven1977,
gjoliver,
ArturNiederfahrenhorst,
smorad,
maxpumperla,
kouroshHakha and
krfricke
as code owners
avnishn
changed the title
Signed-off-by: Avnish <[email protected]>
…into fix_minibatching_gpu
Signed-off-by: Avnish <[email protected]>
sven1977
reviewed
Jun 6, 2023
sven1977
reviewed
Jun 6, 2023
rllib/core/learner/learner.py
Outdated
| for minibatch in batch_iter(batch, minibatch_size, num_iters): | ||
| # Convert minibatch into a tensor batch (NestedDict). | ||
| tensor_minibatch = self._convert_batch_type(minibatch) | ||
| # Convert minibatch into a tensor batch (NestedDict). |
There was a problem hiding this comment.
Can we add: "... on the correct device (e.g. GPU)"? This would clarify further.
Also note then that "we only perform copying to the correct device once so we do not have to move data in each minibatch iteration below". something like this.
sven1977
reviewed
Jun 6, 2023
sven1977
reviewed
Jun 6, 2023
sven1977
reviewed
Jun 6, 2023
rllib/utils/minibatch_utils.py
Outdated
| @@ -1,5 +1,6 @@ | |||
| import math | |||
|
|
|||
|
|
|||
There was a problem hiding this comment.
I can get rid of these, they're left over from previous commits
sven1977
reviewed
Jun 6, 2023
rllib/policy/sample_batch.py
Outdated
| @@ -1631,7 +1635,22 @@ def _concat_values(*values, time_major=None) -> TensorType: | |||
| time_major: Whether to concatenate along the first axis | |||
| (time_major=False) or the second axis (time_major=True). | |||
| """ | |||
| return np.concatenate(list(values), axis=1 if time_major else 0) | |||
| if torch and isinstance(values[0], torch.Tensor): | |||
There was a problem hiding this comment.
Can you explain why we need these changes (add comment)?
There was a problem hiding this comment.
Also, let's use torch.is_tensor().
sven1977
reviewed
Jun 6, 2023
rllib/policy/sample_batch.py
Outdated
| return torch.cat(values, dim=1 if time_major else 0) | ||
| elif isinstance(values[0], np.ndarray): | ||
| return np.concatenate(values, axis=1 if time_major else 0) | ||
| elif tf and isinstance(values[0], tf.Tensor): |
There was a problem hiding this comment.
Can we use tf.is_tensor() instead?
Signed-off-by: Avnish <[email protected]>
avnishn
changed the title
sven1977
reviewed
Jun 7, 2023
Co-authored-by: Sven Mika <[email protected]> Signed-off-by: Avnish Narayan <[email protected]>
sven1977
approved these changes
Jun 7, 2023
kouroshHakha
approved these changes
Jun 7, 2023
Co-authored-by: kourosh hakhamaneshi <[email protected]> Signed-off-by: Avnish Narayan <[email protected]>
Signed-off-by: Avnish <[email protected]>
Signed-off-by: Avnish <[email protected]>
arvind-chandra
pushed a commit
to lmco/ray
that referenced
this pull request
Aug 31, 2023
…ect#36091) Signed-off-by: Avnishn <[email protected]> Co-authored-by: Sven Mika <[email protected]> Co-authored-by: kourosh hakhamaneshi <[email protected]> Signed-off-by: e428265 <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Signed-off-by: avnishn [email protected]
Our current minibatching logic in the learner stack forces individual minibatches to be moved to the gpu after they have been sliced. This is wasteful since it creates unnecessary copies of batches, and adds unnecessary movements of the batch over to gpu. This pr addresses this by moving the whole batch to the gpu first, then doing any minibatching operations on it.
Why are these changes needed?
Related issue number
Checks
git commit -s) in this PR.scripts/format.shto lint the changes in this PR.method in Tune, I've added it in
doc/source/tune/api/under thecorresponding
.rstfile.