Making TF BART-like models XLA and AMP compliant #10191
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
|
I succeed to fix Marian and Pegasus, and my first guess was the good one. I basically reworked a bit how the embedding was created, and now it works in XLA_GPU. Of course, all the corresponding slow tests are passing, and the weights are properly loaded. |
jplu
force-pushed
the
tf-bart-xla-amp
branch
from
0940def to
dfb3d5e
Compare
LysandreJik
reviewed
Feb 16, 2021
| class TFMarianSinusoidalPositionalEmbedding(tf.keras.layers.Embedding): | ||
| class TFMarianSinusoidalPositionalEmbedding(tf.keras.layers.Layer): |
There was a problem hiding this comment.
This was causing trouble to XLA?
There was a problem hiding this comment.
Yes, more precisely, XLA on GPU. This is because the weights with the tf.keras.layers.Embeddings are initialized on CPU and the model is run on GPU, in XLA you cannot access from one to the other.
This is because the embeddings are created in the __init__ of the classes instead of being created in the build().
LysandreJik
reviewed
Feb 16, 2021
Comment on lines
+280
to
+283
| def test_saved_model_creation(self): | ||
| # This test is too long (>30sec) and makes fail the CI | ||
| pass | ||
|
|
sgugger
approved these changes
Feb 16, 2021
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Outdated suggestions cannot be applied.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.
What does this PR do?
This PR makes the TF BART-like models compliant with AMP and XLA. The main issue for XLA was all the asserts, XLA is not compliant with them (see the TF doc), so I had to disable them if the model is run with another mode than eager.
TF Marian and Pegasus have still their XLA test locked because they are not working for XLA_GPU. I need to investigate more in order to better understand why. My first guess is because of the
TFXSinusoidalPositionalEmbeddingclass.