Skip to content

Rename FSDP Flat placement to RowAtomic - #7539

Merged
wujingyue merged 2 commits into
NVIDIA:mainfrom
wujingyue:rename-fsdp-row-atomic
Sep 23, 2026
Merged

wujingyue merged 2 commits into
NVIDIA:mainfrom
wujingyue:rename-fsdp-row-atomic

Conversation

@wujingyue

@wujingyue wujingyue commented Sep 21, 2026 •

Copy link
Copy Markdown
Contributor
  • I, the PR author, have personally reviewed every line of this PR.

What does this PR do?

Rename the experimental FSDP Flat placement to RowAtomic to describe its row-preserving sharding semantics.

By default, mFSDP cuts weights at row boundaries, keeping each row intact within a shard. The name Flat is confusing because it suggests treating a weight as a 1D flat tensor and potentially cutting in the middle of a row. RowAtomic makes the intended semantics explicit: each row is an indivisible unit of sharding.

Update imports, exports, tests, error messages, and design documentation. Sharding behavior is unchanged.

Validation

  • Ran the affected test_dbuffer.py, test_memory.py, test_owner_planning.py, and test_quantized_dbuffer.py files with two GPU ranks in the development container: 53 passed, 7 skipped per rank.
  • Ran isort on changed imports, Black and Ruff on changed Python files, and git diff --check.
  • The commit contains the same tested patch, moved to a separate worktree.

@copy-pr-bot

copy-pr-bot Bot commented Sep 21, 2026

Copy link
Copy Markdown

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

@wujingyue
wujingyue marked this pull request as ready for review September 21, 2026 06:19
@wujingyue
wujingyue requested a review from a team as a code owner September 21, 2026 06:19
@wujingyue wujingyue changed the title Rename experimental FSDP Flat placement to RowAtomic Rename FSDP Flat placement to RowAtomic Sep 23, 2026
Signed-off-by: Jingyue Wu <jingyuew@nvidia.com>
Signed-off-by: Jingyue Wu <jingyuew@nvidia.com>
@wujingyue
wujingyue enabled auto-merge September 23, 2026 07:12
@wujingyue
wujingyue force-pushed the rename-fsdp-row-atomic branch from 9b7ffc1 to bd4053a Compare September 23, 2026 07:13
@svcnvidia-nemo-ci svcnvidia-nemo-ci added the Approved All necessary approvals have been made label Sep 23, 2026
@wujingyue
wujingyue added this pull request to the merge queue Sep 23, 2026
@nemo-automation-bot

Copy link
Copy Markdown

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/35861658427

@github-merge-queue
github-merge-queue Bot removed this pull request from the merge queue due to failed status checks Sep 23, 2026
@wujingyue
wujingyue added this pull request to the merge queue Sep 23, 2026
@nemo-automation-bot

Copy link
Copy Markdown

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/35875329759

Merged via the queue into NVIDIA:main with commit ffb3c4f Sep 23, 2026
103 checks passed
@wujingyue
wujingyue deleted the rename-fsdp-row-atomic branch September 23, 2026 15:59

This branch was successfully deployed

2 active deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Approved All necessary approvals have been made complexity: low

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants