PR #20808: [GSPMD] Partitions collective permute instructions in manual sharding group. #84637
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
![copybara-service[bot]](https://avatars.githubusercontent.com/in/44061?s=80&v=4){kind=small}
copybara-service
bot
force-pushed
the
exported_pr_714284894
branch
from
d1419f9 to
de85df5
Compare
…al sharding group. Imported from GitHub PR openxla/xla#20808 This is a small fix in GSPMD partitioning for partitioning collective permutes instructions added in manual sharding group. In JAX, we can add `ppermute` instruction in shard_map. In cases where we have shard_map with auto axes specified, collective permuting an operand even with the same sharding will end up with an `all-gather` and then collective permute, which leads to inefficient collectives. The correct and efficient way is to partition the collective permute as an element-wise op. The unit test added provides a repro. Also, the JAX unit test in https://github.com/jax-ml/jax/blob/fa9c7edf736516052df6eab22947bc627d0deca3/tests/shard_map_test.py#L2167 gives a real-world JAX example. Copybara import of the project: -- 8ee6ecd51f6e4aae8e3d92a6a439a60f53ab02ae by Yunlong Liu <yunlongl@x.ai>: A hacky fix on partitioning collective permute. -- e50e87696defb290f7561a7808ee42ebbc11e144 by Yunlong Liu <yunlongl@x.ai>: Local change. -- 84eb38597c783a4488774823c2c464296a8c54c7 by Yunlong Liu <yunlongl@x.ai>: Simplifies sharding in tests. Merging this change closes #20808 PiperOrigin-RevId: 714851861
copybara-service
bot
force-pushed
the
exported_pr_714284894
branch
from
de85df5 to
b5d22c7
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.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.
PR #20808: [GSPMD] Partitions collective permute instructions in manual sharding group.
Imported from GitHub PR openxla/xla#20808
This is a small fix in GSPMD partitioning for partitioning collective permutes instructions added in manual sharding group.
In JAX, we can add
ppermuteinstruction in shard_map. In cases where we have shard_map with auto axes specified, collective permuting an operand even with the same sharding will end up with anall-gatherand then collective permute, which leads to inefficient collectives. The correct and efficient way is to partition the collective permute as an element-wise op.The unit test added provides a repro. Also, the JAX unit test in https://github.com/jax-ml/jax/blob/fa9c7edf736516052df6eab22947bc627d0deca3/tests/shard_map_test.py#L2167 gives a real-world JAX example.
Copybara import of the project:
--
8ee6ecd51f6e4aae8e3d92a6a439a60f53ab02ae by Yunlong Liu yunlongl@x.ai:
A hacky fix on partitioning collective permute.
--
e50e87696defb290f7561a7808ee42ebbc11e144 by Yunlong Liu yunlongl@x.ai:
Local change.
--
84eb38597c783a4488774823c2c464296a8c54c7 by Yunlong Liu yunlongl@x.ai:
Simplifies sharding in tests.
Merging this change closes #20808
FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#20808 from yliu120:cp_sharding_2 84eb38597c783a4488774823c2c464296a8c54c7