Add sharding propagation support in nnx.eval_shape (clone of #5111)#5247
Add sharding propagation support in nnx.eval_shape (clone of #5111)#5247samanklesaria wants to merge 2 commits intogoogle:mainfrom
Conversation
Summary of ChangesHello @samanklesaria, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed! This pull request enhances Highlights
🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console. Changelog
Activity
Using Gemini Code AssistThe full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips. Invoking Gemini You can request assistance from Gemini at any point by creating a comment using either
Customization To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a Limitations & Feedback Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here. You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension. Footnotes
|
There was a problem hiding this comment.
Code Review
The pull request introduces sharding propagation support in nnx.eval_shape by adding logic to handle jax.sharding.NamedSharding for variables. This ensures that the sharding information is correctly propagated during shape evaluation. Additionally, new test cases have been added to tests/nnx/spmd_test.py to validate this new functionality. The changes appear to be correct and well-tested.
| global_mesh = jax.sharding.get_mesh() | ||
| if global_mesh.axis_sizes == (): | ||
| global_mesh = None | ||
| mesh = var.get_metadata("mesh", None) or global_mesh | ||
| if mesh is not None: | ||
| pspec = get_var_pspec(var) | ||
| sharding = jax.sharding.NamedSharding(mesh=mesh, spec=pspec) | ||
| var.set_value(jax.ShapeDtypeStruct(shape=var.shape, dtype=var.dtype, sharding=sharding)) |
There was a problem hiding this comment.
The logic for applying sharding within _to_variable seems to correctly handle cases where a mesh is explicitly provided in metadata or a global mesh is available. This ensures that jax.ShapeDtypeStruct is created with the appropriate sharding information.
| def test_eval_shape_with_sharding0(self): | ||
| # based on https://github.com/google/flax/issues/5110 | ||
| mesh1 = jax.make_mesh((2, 2), ("a", "b"), (jax.sharding.AxisType.Auto, jax.sharding.AxisType.Auto)) | ||
| mesh2 = jax.make_mesh((1, 4), ("c", "d"), (jax.sharding.AxisType.Auto, jax.sharding.AxisType.Auto)) | ||
|
|
||
| class Model(nnx.Module): | ||
| def __init__(self): | ||
| self.p1 = nnx.Linear(16, 16, rngs=nnx.Rngs(0), kernel_metadata={"out_sharding": ("a", "b"), "mesh": mesh1}) | ||
| self.p2 = nnx.Linear(16, 16, rngs=nnx.Rngs(0), kernel_metadata={"out_sharding": ("c", "d"), "mesh": mesh2}) | ||
|
|
||
| abs_model = nnx.eval_shape(lambda: Model()) | ||
| assert isinstance(abs_model.p1.kernel.sharding, jax.sharding.NamedSharding) | ||
| assert abs_model.p1.kernel.sharding.mesh is mesh1 | ||
| assert abs_model.p1.kernel.sharding.spec == jax.P("a", "b") | ||
| assert isinstance(abs_model.p2.kernel.sharding, jax.sharding.NamedSharding) | ||
| assert abs_model.p2.kernel.sharding.mesh is mesh2 | ||
| assert abs_model.p2.kernel.sharding.spec == jax.P("c", "d") |
There was a problem hiding this comment.
The test_eval_shape_with_sharding0 test case effectively verifies the sharding propagation when meshes are explicitly defined in the kernel_metadata for different linear layers. This is a good test for ensuring that custom mesh configurations are respected.
| def test_eval_shape_with_sharding1(self): | ||
| class Model(nnx.Module): | ||
| def __init__(self): | ||
| self.linear = nnx.Linear(10, 10, rngs=nnx.Rngs(0), kernel_metadata={"out_sharding": ("a", "b")}) | ||
|
|
||
| mesh = jax.make_mesh((2, 2), ("a", "b"), (jax.sharding.AxisType.Auto, jax.sharding.AxisType.Auto)) | ||
| with jax.set_mesh(mesh): | ||
| abs_model = nnx.eval_shape(lambda: Model()) | ||
| assert isinstance(abs_model.linear.kernel.sharding, jax.sharding.NamedSharding) | ||
| assert abs_model.linear.kernel.sharding.mesh is mesh | ||
| assert abs_model.linear.kernel.sharding.spec == jax.P("a", "b") |
There was a problem hiding this comment.
The test_eval_shape_with_sharding1 test case correctly validates sharding propagation when a global mesh is set using jax.set_mesh. This covers a common use case where sharding is implicitly applied based on the global mesh context.
Fixes #5110, clone of #5111 but rebased against current main.