Introduce BaseAttentionBias.has_value().
#920
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
`BaseAttentionBias.value()` is also used to check whether a bias value exists. However, calling `value()` creates the actual bias on the CPU, which is quadratic O(T^2) — especially when debugging long sequence lengths (e.g., 32k). Hence, we introduce a lightweight `has_value()` method. Normally, unused tensors are pruned from the graph during XLA compilation, and using `BaseAttentionBias.value()` to check for bias presence relies on XLA pruning. But `BaseAttentionBias.value()` is still expensive on the CPU for unittests and debugging, so `has_value()` saves Python runtime. The new `has_value()` method checks whether `value()` actually exists by calling `jax.eval_shape`. Since `jax.eval_shape` invokes `value()` through a tracer, it doesn’t materialize the actual value (proposed by John Peebles).
markblee
approved these changes
Jan 13, 2025
| Raises: | ||
| ValueError: If the bias has no value. | ||
| """ | ||
| if not self.has_value(): |
There was a problem hiding this comment.
I suppose multiple calls to jax.eval_shape are automatically cached?
There was a problem hiding this comment.
I couldn't find caching mechanism in jax impl. Would you mind asking how it's related to this PR? Can I merge it?
@api_boundary
def eval_shape(fun: Callable, *args, **kwargs):
"""Compute the shape/dtype of ``fun`` without any FLOPs.
This utility function is useful for performing shape inference. Its
input/output behavior is defined by::
There was a problem hiding this comment.
Oh, in eval_shape, we make two calls to jax.eval_shape(self.value) (once in has_value and once in return) -- so ideally we only trace once. I checked and it seems to be the case, so please feel free to ignore.
There was a problem hiding this comment.
Ah, I see. Good point. Thank you for check!
github-merge-queue bot
pushed a commit
that referenced
this pull request
Jan 13, 2025
`BaseAttentionBias.value()` is also used to check whether a bias value exists. However, calling `value()` creates the actual bias on the CPU, which is quadratic O(T^2) — especially when debugging long sequence lengths (e.g., 32k). Hence, we introduce a lightweight `has_value()` method. Normally, unused tensors are pruned from the graph during XLA compilation, and using `BaseAttentionBias.value()` to check for bias presence relies on XLA pruning. But `BaseAttentionBias.value()` is still expensive on the CPU for unittests and debugging, so `has_value()` saves Python runtime. The new `has_value()` method checks whether `value()` actually exists by calling `jax.eval_shape`. Since `jax.eval_shape` invokes `value()` through a tracer, it doesn’t materialize the actual value (proposed by John Peebles).
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.
BaseAttentionBias.value()is also used to check whether a bias value exists. However, callingvalue()creates the actual bias on the CPU, which is quadratic O(T^2) — especially when debugging long sequence lengths (e.g., 32k). Hence, we introduce a lightweighthas_value()method.Normally, unused tensors are pruned from the graph during XLA compilation, and using
BaseAttentionBias.value()to check for bias presence relies on XLA pruning. ButBaseAttentionBias.value()is still expensive on the CPU for unittests and debugging, sohas_value()saves Python runtime.The new
has_value()method checks whethervalue()actually exists by callingjax.eval_shape. Sincejax.eval_shapeinvokesvalue()through a tracer, it doesn’t materialize the actual value (proposed by John Peebles).