[Async TP] Fuse matmul-reduce-scatters when reduce scatters have multiple users, and save fused node for backward instead of reduce_scatter node #149875
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
…matmul reduce scatter
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/149875
Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. |
5509e6f to
8c1a679
Compare
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Outdated suggestions cannot be applied.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.
Fixes #149876
Stack
TL;DR
This PR implements support in async TP for saving the reduce-scatter result for backward, which previously would break the torchtitan AC policies: no AC, per op SAC, and per layer SAC.
Context
In torchtitan's LLama3 per op SAC policy, we want to save the output of
reduce_scatterops for backward, which is useful for TP. The reduce_scatter op is also saved for No AC (since all activations are saved) and per layer SAC (since we save the activations for N full layers, which do contain reduce-scatters for TP.However, doing this causes incompatibility with Async TP for the AC policies above, for 2 reasons:
The graph pattern matching specifically only matches on reduce scatter nodes with 1 user, but reduce_scatter nodes saved for backwards will have 2 users (the 2nd one being the return/output node, which saves it for backward).
The subgraph replacement logic which replaces the users of the
wait_tensorafter the reduce-scatter with the new fused node has no mechanism to save the fused_node for backward instead of the reduce-scatter node. This means we cannot directly replace the subgraph, since we can't delete nodes which still have users (in this case, the output node is still using the reduce-scatter node).To fix this, we do 2 things:
Add additional pattern matching logic to also match reduce-scatter nodes with 2 users, so we also perform fusion when reduce-scatter is saved for backward.
When replacing the subgraph with the fused node, detect if the reduce-scatter was saved for backward, and if so, save the result of the fused node for backward instead. This enables us to properly erase the subgraph and prevent the memory leak which occurred in [Async TP] Activations not cleared after backward when reduce_scatter_tensor saved for backward by per op SAC #149876
Other changes
Test plan
cc @H-Huang @awgu @kwen2501 @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov