[scan] Reduce memory usage #8562
Merged
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
qihqi
approved these changes
Jan 15, 2025
This change optimizes the memory usage of scan by turning some copies into aliases. Specifically, it optimizes intermediate activations which are aliases of inputs. This is a common occurrence. In principle, we could have `forward` return all the intermediate activations, including those that are aliases to an input tensor. However, those inputs will then be duplicated as part of the output of a `scan` call, because we want to save all activations during the forward pass of a `scan`. The XLA compiler can't optimize away this duplication likely because they're behind a DynamicSlice + DynamicUpdateSlice, so we end up doubling the memory usage from those inputs. To reduce memory usage, we can have `forward` return the activations that don't alias to inputs, called `partial_activations`. The autograd implementation of `scan` will call `alias_input` to add back activations that are aliases of input tensors outside of a scan, turning the partial activations back to full activations.
tengyifei
force-pushed
the
scan/activation-alias-input
branch
from
fc707b8 to
385797d
Compare
qihqi
pushed a commit
that referenced
this pull request
Jan 16, 2025
tengyifei
added a commit
that referenced
this pull request
Jan 22, 2025
This was referenced Jan 22, 2025
Merged
tengyifei
added a commit
that referenced
this pull request
Jan 22, 2025
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Outdated suggestions cannot be applied.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.
This change optimizes the memory usage of scan by turning some copies into aliases. Specifically, it optimizes intermediate activations which are aliases of inputs. This is a common occurrence.
In principle, we could have
forwardreturn all the intermediate activations, including those that are aliases to an input tensor. However, those inputs will then be duplicated as part of the output of ascancall, because we want to save all activations during the forward pass of ascan. The XLA compiler can't optimize away this duplication likely because they're behind a DynamicSlice + DynamicUpdateSlice, so we end up doubling the memory usage from those inputs.To reduce memory usage, we can have
forwardreturn the activations that don't alias to inputs, calledpartial_activations. The autograd implementation ofscanwill callalias_inputto add back activations that are aliases of input tensors outside of a scan, turning the partial activations back to full activations.