Improve compute cost for all2all #159
Merged
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
The all2all implementation performs additional input/output copies depending on the in_shard / out_shard dims, see https://github.com/pytorch/pytorch/blob/afdd4247a2251b3f4c2f4b402cb625f46d6784ba/torch/csrc/distributed/c10d/Functional.cpp#L597-L617 for more details
Need to figure out a way of deciding if the input is contiguous or not
meta-cla
bot
added
the
CLA Signed
fmassa
commented
Sep 26, 2025
| # suppose 70% efficiency for the operator | ||
| compute_efficiency = 0.70 | ||
| compute_time = flops / gpu_flops * 1e6 # us | ||
| compute_time = max(compute_time / compute_efficiency, kernel_launch_overhead) |
Contributor
Author
There was a problem hiding this comment.
I removed this, but this is functionally the same as before because we already perform a max(..., kernel_launch_overhead) for the compute_read_write_time.
…sa/improve_a2a_cost
…sa/improve_a2a_cost
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Outdated suggestions cannot be applied.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.
All2all performs a memory copy in the following cases:
For now, I assume that the input is not contiguous. I should improve this in the future.
Also refactors the read-write cost into a helper function to centralize things.