Added TF specific documentation to DistributedEmbedding.
#94
Conversation
This file contains hidden or 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
cantonios
requested changes
May 21, 2025
hertschuh
force-pushed
the
embedding_tf
branch
3 times, most recently
from
4939b13 to
2bd6f5a
Compare
cantonios
approved these changes
May 21, 2025
| In addition to `tf.Tensor`, `DistributedEmbedding` accepts `tf.RaggedTensor` | ||
| and `tf.SparseTensor` as inputs for the embedding lookups. Ragged tensors | ||
| must be ragged in dimension 1. Note that if weights are passed, each weight |
There was a problem hiding this comment.
Is dimension 1 the dimension with index 0? Does TF actually support any other kind of raggedness?
There was a problem hiding this comment.
I meant dimension with index 1. TF supports any ragged dimension and in fact multiple ragged dimensions, a tensor can be twice ragged.
https://www.tensorflow.org/api_docs/python/tf/RaggedTensor#multiple_ragged_dimensions
https://www.tensorflow.org/api_docs/python/tf/RaggedTensor#attributes (see ragged_rank)
| In addition to `tf.Tensor`, `DistributedEmbedding` accepts `tf.RaggedTensor` | ||
| and `tf.SparseTensor` as inputs for the embedding lookups. Ragged tensors | ||
| must be ragged in dimension 1. Note that if weights are passed, each weight | ||
| tensor must be of the same type as the inputs for that particular feature |
There was a problem hiding this comment.
"same type" might be a bit confusing because we expect the datatype for indices to be integers, but weights to be floats.
Same class?
hertschuh
added a commit
that referenced
this pull request
Jun 3, 2025
* Ignore shard_map attr error in mypy. (#97) * Added TF specific documentation to `DistributedEmbedding`. (#94) * Fix symbolic calls for `EmbedReduce`. (#98) `EmbedReduce` was inheriting the behavior from `Embedding` and not correctly applying the reduction. * Move `DistributedEmbedding` declaration to its own file. (#99) Having it in `__init__.py` doesn't play nice with pytype. * Remove dependency on `tree` and use `keras.tree`. (#100) Keras can already depend on either `dmtree` or `optree` and use whichever is best or available on the current platform. * Only enable JAX on linux_x86_64. (#101) * Add out_sharding argument WrappedKerasInitializer. (#102) This is for forward-compatibility. Latest versions of JAX introduce the `out_sharding` argument. * Use Python 3.10 style type annotations. (#104) Now that we require Python 3.10, we can use the shorter annotation style, which should improve the readability of the documentation. * Do not bundle test utils in wheel. (#105) * Update version number to 0.2.1 (#106) As 0.2.0 was just released. * Fix invalid escape sequence in unit test. (#108) * Replace leftover `unflatten_as` to `pack_sequence_as`. (#109) This instance was missed as it is only run on TPU. * Make the declaration of `Nested` compatible with pytype. (#110) Which doesn't support `|` between forward declarations using a string. * Add ragged support for default_device placement on JAX. (#107) Requires calling `preprocess`. Internally, we currently convert ragged inputs to dense before passing to the embedding call(...) function. * Add documentation for using DistributedEmbedding with JAX. (#111) * `api_gen` now excludes backend specific code. (#103) This: - Allows development (`api_gen` / git presubmit hooks) without all backends and backend specific dependencies installed and working. For instance, jax_tpu_embedding currently doesn't import on MacOS Sequoia, this allows running `api_gen` regardless. - Makes sure we don't accidentally create and honor exports that are backend specific. * Enable preprocess calls with symbolic input tensors. (#113) This allows us to more-easily create functional models via: ```python preprocessed_inputs = distributed_embedding.preprocess(symbolic_inputs, symbolic_weights) outputs = distributed_embedding(preprocessed_inputs) model = keras.Model(inputs=preprocessed_inputs, outputs=outputs) ``` * Check for jax_tpu_embedding on JAX backend. (#114) This is to allow users to potentially run Keras RS _without_ the dependency. If a user doesn't have `jax-tpu-embedding` installed, but are on `linux_x86_64` and has a sparsecore-capable TPU available, and if they try to use `auto` or `sparsecore` placement with distributed embedding, will raise an error informing them to install the dependency. --------- Co-authored-by: C. Antonio Sánchez <[email protected]> Co-authored-by: hertschuh <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
No description provided.