[DeviceMesh] Implement a device mesh concatenate api for submesh and SPMD use case #163358
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
[ghstack-poisoned]
This was referenced Sep 19, 2025
Closed
pytorch-bot
bot
added
the
oncall: distributed
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/163358
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (1 Unrelated Failure)As of commit cc835fc with merge base 000f495 ( UNSTABLE - The following job is marked as unstable, possibly due to flakiness on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
…or submesh" Today FSDP needs to slicing out spmd mesh from root mesh here: https://github.com/pytorch/pytorch/blob/main/torch/distributed/fsdp/_fully_shard/_fsdp_param.py#L301. But essentially, users want is a concatenate of some submesh into a big mesh and used as a spmd mesh. This PR is tentatively trying to implement this API for users. One thing to note is that, all sub-mesh needs to slicing/flatten or unflatten from same root mesh otherwise the indices make no sense when it comes to mesh indexing and device allocation. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta ezyang msaroufim dcci [ghstack-poisoned]
…or submesh" Today FSDP needs to slicing out spmd mesh from root mesh here: https://github.com/pytorch/pytorch/blob/main/torch/distributed/fsdp/_fully_shard/_fsdp_param.py#L301. But essentially, users want is a concatenate of some submesh into a big mesh and used as a spmd mesh. This PR is tentatively trying to implement this API for users. One thing to note is that, all sub-mesh needs to slicing/flatten or unflatten from same root mesh otherwise the indices make no sense when it comes to mesh indexing and device allocation. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta ezyang msaroufim dcci [ghstack-poisoned]
ezyang
reviewed
Sep 21, 2025
ezyang
reviewed
Sep 21, 2025
ezyang
reviewed
Sep 21, 2025
torch/distributed/device_mesh.py
Outdated
| get_world_size(), | ||
| ) | ||
|
|
||
| for mesh_nd in pg_ranks_by_dim: |
Contributor
There was a problem hiding this comment.
?!?! Why do you need to do it for every mesh_nd? Is this because you're triggering comms to initialize PGs?
Contributor
Author
There was a problem hiding this comment.
so long story short, we need all ranks to call new_group which is hidden very deep in the stack to initialize PGs. Otherwise the code will hang.
…or submesh" Today FSDP needs to slicing out spmd mesh from root mesh here: https://github.com/pytorch/pytorch/blob/main/torch/distributed/fsdp/_fully_shard/_fsdp_param.py#L301. But essentially, users want is a concatenate of some submesh into a big mesh and used as a spmd mesh. This PR is tentatively trying to implement this API for users. One thing to note is that, all sub-mesh needs to slicing/flatten or unflatten from same root mesh otherwise the indices make no sense when it comes to mesh indexing and device allocation. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta ezyang msaroufim dcci [ghstack-poisoned]
…or submesh" Today FSDP needs to slicing out spmd mesh from root mesh here: https://github.com/pytorch/pytorch/blob/main/torch/distributed/fsdp/_fully_shard/_fsdp_param.py#L301. But essentially, users want is a concatenate of some submesh into a big mesh and used as a spmd mesh. This PR is tentatively trying to implement this API for users. One thing to note is that, all sub-mesh needs to slicing/flatten or unflatten from same root mesh otherwise the indices make no sense when it comes to mesh indexing and device allocation. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta ezyang msaroufim dcci [ghstack-poisoned]
…or submesh" Today FSDP needs to slicing out spmd mesh from root mesh here: https://github.com/pytorch/pytorch/blob/main/torch/distributed/fsdp/_fully_shard/_fsdp_param.py#L301. But essentially, users want is a concatenate of some submesh into a big mesh and used as a spmd mesh. This PR is tentatively trying to implement this API for users. One thing to note is that, all sub-mesh needs to slicing/flatten or unflatten from same root mesh otherwise the indices make no sense when it comes to mesh indexing and device allocation. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta ezyang msaroufim dcci [ghstack-poisoned]
…or submesh" Today FSDP needs to slicing out spmd mesh from root mesh here: https://github.com/pytorch/pytorch/blob/main/torch/distributed/fsdp/_fully_shard/_fsdp_param.py#L301. But essentially, users want is a concatenate of some submesh into a big mesh and used as a spmd mesh. This PR is tentatively trying to implement this API for users. One thing to note is that, all sub-mesh needs to slicing/flatten or unflatten from same root mesh otherwise the indices make no sense when it comes to mesh indexing and device allocation. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta ezyang msaroufim dcci [ghstack-poisoned]
…or submesh" Today FSDP needs to slicing out spmd mesh from root mesh here: https://github.com/pytorch/pytorch/blob/main/torch/distributed/fsdp/_fully_shard/_fsdp_param.py#L301. But essentially, users want is a concatenate of some submesh into a big mesh and used as a spmd mesh. This PR is tentatively trying to implement this API for users. One thing to note is that, all sub-mesh needs to slicing/flatten or unflatten from same root mesh otherwise the indices make no sense when it comes to mesh indexing and device allocation. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta ezyang msaroufim dcci [ghstack-poisoned]
…ubmesh and SPMD use case" Today FSDP needs to slicing out spmd mesh from root mesh here: https://github.com/pytorch/pytorch/blob/main/torch/distributed/fsdp/_fully_shard/_fsdp_param.py#L301. But essentially, users want is a concatenate of some submesh into a big mesh and used as a spmd mesh. This PR is tentatively trying to implement this API for users. One thing to note is that, all sub-mesh needs to slicing/flatten or unflatten from same root mesh otherwise the indices make no sense when it comes to mesh indexing and device allocation. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta ezyang msaroufim dcci [ghstack-poisoned]
Contributor
Author
|
@pytorchbot merge -i |
Collaborator
Merge startedYour change will be merged while ignoring the following 2 checks: trunk / linux-jammy-py3-clang12-executorch / test (executorch, 1, 1, lf.linux.2xlarge, unstable), trunk / linux-jammy-rocm-py3.10 / test (default, 2, 2, linux.rocm.gpu.gfx942.1) Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Contributor
|
@pytorchbot revert -m "probably need to revert this one too, its stacked with #166003 (comment)" -c ghfirst |
Collaborator
|
@pytorchbot successfully started a revert job. Check the current status here. |
pytorchmergebot
added a commit
that referenced
this pull request
Oct 24, 2025
…esh and SPMD use case (#163358)" This reverts commit 5a4997d. Reverted #163358 on behalf of https://github.com/clee2000 due to probably need to revert this one too, its stacked with #166003 (comment) ([comment](#163358 (comment)))
Collaborator
|
@fduwjj your PR has been successfully reverted. |
…ubmesh and SPMD use case" Today FSDP needs to slicing out spmd mesh from root mesh here: https://github.com/pytorch/pytorch/blob/main/torch/distributed/fsdp/_fully_shard/_fsdp_param.py#L301. But essentially, users want is a concatenate of some submesh into a big mesh and used as a spmd mesh. This PR is tentatively trying to implement this API for users. One thing to note is that, all sub-mesh needs to slicing/flatten or unflatten from same root mesh otherwise the indices make no sense when it comes to mesh indexing and device allocation. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta ezyang msaroufim dcci [ghstack-poisoned]
Contributor
Author
|
@pytorchbot merge |
Collaborator
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Contributor
Author
|
@fduwjj has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
1 similar comment
Contributor
Author
|
@fduwjj has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
pytorchmergebot
pushed a commit
that referenced
this pull request
Oct 27, 2025
…from root mesh (#165492) With concatenate API, we can directly combine two meshes together rather than getting the spmd mesh from root. Differential Revision: [D85409698](https://our.internmc.facebook.com/intern/diff/D85409698) Pull Request resolved: #165492 Approved by: https://github.com/fegin ghstack dependencies: #163358
tianrengao
pushed a commit
that referenced
this pull request
Oct 30, 2025
…SPMD use case (#163358) Today FSDP needs to slicing out spmd mesh from root mesh here: https://github.com/pytorch/pytorch/blob/main/torch/distributed/fsdp/_fully_shard/_fsdp_param.py#L301. But essentially, users want is a concatenate of some submesh into a big mesh and used as a spmd mesh. This PR is tentatively trying to implement this API for users. One thing to note is that, all sub-mesh needs to slicing/flatten or unflatten from same root mesh otherwise the indices make no sense when it comes to mesh indexing and device allocation. Pull Request resolved: #163358 Approved by: https://github.com/fegin
tianrengao
pushed a commit
that referenced
this pull request
Oct 30, 2025
…from root mesh (#165492) With concatenate API, we can directly combine two meshes together rather than getting the spmd mesh from root. Differential Revision: [D85409698](https://our.internmc.facebook.com/intern/diff/D85409698) Pull Request resolved: #165492 Approved by: https://github.com/fegin ghstack dependencies: #163358
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Stack from ghstack (oldest at bottom):
Today FSDP needs to slicing out spmd mesh from root mesh here: https://github.com/pytorch/pytorch/blob/main/torch/distributed/fsdp/_fully_shard/_fsdp_param.py#L301. But essentially, users want is a concatenate of some submesh into a big mesh and used as a spmd mesh. This PR is tentatively trying to implement this API for users.
One thing to note is that, all sub-mesh needs to slicing/flatten or unflatten from same root mesh otherwise the indices make no sense when it comes to mesh indexing and device allocation.
cc @H-Huang @awgu @wanchaol @fegin @wz337 @wconstab @d4l3k @pragupta @ezyang @msaroufim @dcci
Differential Revision: D85409698