Repository navigation
fix: keep zigzag CP indexes on the activation device - #214
Merged
Merged
Conversation
split_cp_inputs built the zigzag rank index on CPU and then called .cuda(). On Ascend that index lands on CUDA while the activation is on NPU, and index_select fails before any context-parallel slice is taken. The same call fails on CPU. Build the two-element index on the activation device instead. Rank 0 of cp_size 2 still owns chunks 0 and 3, so the token order is unchanged.
hjh0119
approved these changes
Oct 9, 2026
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
split_cp_inputsbuilt the zigzag rank index on CPU and moved it with.cuda(). On Ascend the activation is on NPU, soindex_selectfails before the context-parallel slice is taken. The same.cuda()call fails when the activation is on CPU.The index is now created on the activation's device. Rank 0 with
cp_size2 still owns chunks 0 and 3, so the token order is unchanged.Checked with
python3 -m pytest tests/test_split_cp_inputs.py. The test stubs Megatron parallel state and checks that an 8-token CPU sequence splits to[0, 1, 6, 7]. Also ranflake8on the changed files.