Don't duplicate frozen parameters during predict() #20851
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
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## master #20851 +/- ##
==========================================
+ Coverage 82.04% 82.25% +0.20%
==========================================
Files 559 559
Lines 52367 52374 +7
Branches 8096 8096
==========================================
+ Hits 42964 43078 +114
+ Misses 7427 7305 -122
- Partials 1976 1991 +15
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Sentry. |
On the Jax backend we were not using donate_argnums during predict. This works when a model is mostly trainable, but when a model is mostly or all frozen, this will result in 2x the memory jump (which is why we use donate_argnums for fit and evaluate). This change adds donate_argnums to the predict function to avoid the memory spike. But because this means all incoming state (including the trainable variables) will be deleted by jax, this means we need to sync the trainable variables state much like in fit and evaluate. An alternative would be to change the predict_step signature (so we could only donate non-trainable variables), but this would be a breaking change and confusing.
mattdangerw
force-pushed
the
predict-memory-use-fix
branch
from
e4eefb4 to
84a4897
Compare
|
Notably this will come up for lora + transformers. gemma = keras_hub.CausalLM.from_preset("gemma...")
gemma.backbone.enable_lora(4)
gemma.fit(...)
gemma.predict(example) # double optimal memory usage do to doubled frozen params |
fchollet
approved these changes
Feb 4, 2025
google-ml-butler
bot
added
kokoro:force-run
ready to pull
google-ml-butler
bot
removed
awaiting review
ready to pull
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.
On the Jax backend we were not using donate_argnums during
predict(). This works when a model is mostly trainable, but when a model is mostly or all frozen, this will result in 2x the memory jump (which is why we use donate_argnums for fit and evaluate).This change adds
donate_argnumsto the predict function to avoid the memory spike. But because this means all incoming state (including the trainable variables) will be deleted by jax, this means we need to sync the trainable variables state much like in fit and evaluate. An alternative would be to change the predict_step signature (so we could only donate non-trainable variables), but this would be a breaking change and confusing.