HSDP + TP integration bug fixes #119819
Merged
HSDP + TP integration bug fixes #119819
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
Currently, HSDP validates that all intra/inter node PGs are the same. This makes sense if you are only using HSDP with no other forms of parallelism and is a nice but not necessary sanity check. However, if you want to mix HSDP with other forms, say tensor parallelism on the FFN of a transformer block, the intra/inter node PGs will be different for that layer. This check raises errors in this scenario, so we need to remove this assumption. Pull Request resolved: pytorch#112435 Approved by: https://github.com/wz337, https://github.com/Skylion007
Removes raising error if a device_mesh has a parent. The comment says that HSDP + TP is not supported, but I'm able to do 2D parallelism + HSDP fine. The only issues are: - this check - pytorch#118618 - a series of PRs related to checkpointing with 3D meshes that I will open We currently monkeypatch for the above which I am slowly upstreaming. I imagine torch will have a better, native integration eventually, but this check seems too aggressive in the meantime given DTensor now lets users do some things themselves (which is amazing 🎉)! Pull Request resolved: pytorch#118620 Approved by: https://github.com/Skylion007
If the user passes `device_mesh`, then we should not forward the process groups to the children during auto wrap and instead just rely on the `device_mesh` argument. This should fix pytorch#118906. Pull Request resolved: pytorch#119064 Approved by: https://github.com/wz337
Update all_gather to support HSDP + TP. Currently, the `_all_gather_dtensor` function for dtensors only replaces the first dimension with replicate (the FSDP dimension) and does not touch the second dimension (which is assumed to be the TP dimension). With HSDP, we have two dimensions ahead of the TP dimension as opposed to 1. This PR updates to replace all other dimensions with replicate to run the all-gather. Pull Request resolved: pytorch#118638 Approved by: https://github.com/fegin, https://github.com/awgu, https://github.com/wz337
Fixes pytorch#118639. Adds support to replicate across HSDP dimensions instead of sharding for shard placement Pull Request resolved: pytorch#119481 Approved by: https://github.com/Skylion007, https://github.com/wz337
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/119819
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New Failure, 3 Unrelated FailuresAs of commit 20a5c32 with merge base a8bd593 ( NEW FAILURE - The following job has failed:
FLAKY - The following jobs failed but were likely due to flakiness present on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
atalman
approved these changes
Feb 14, 2024
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.
Contains bug fixes for the HSDP + TP integration with device_mesh, which was announced in torch 2.2