revert sharding grouping logic for vbe #2216
Closed
Conversation
This file contains 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
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D58989195 |
Summary: Pull Request resolved: pytorch#2215 Previously to handle multiple VBE TBE output which is 1d tensor ordered by rank, we grouped sharding info such that there would only be one TBE created per sharding module. This avoided the issue of concatting multiple 1d tensors that are ordered by rank (not a problem in on VBE bc of 2d output which we can concat on dim 1). This grouping which would be done only applies to specific UVM caching setups that used prefetch pipeline, as each sharding type could require multiple TBE to handle both HBM and UVM caching setups. In most cases the TBE could be fused for each sharding type, so we grouped by such. Each sharding module handles individual input dist, lookup, output dist, and by creating a sharding module per each TBE in EMO setups would cause regression, as there would be an increase in comms to handle the increased input dists and output dists. This diff removes the need for the grouping logic to circumvent the VBE TBE output concatenation by implementing output concatenation, which removes the necessity for specialized sharding grouping logic for specific EMO cases. Differential Revision: D58894728 Reviewed By: dstaay-fb, levythu
Summary: Pull Request resolved: pytorch#2216 reverting sharding grouping logic in EBC/VLE modules that supported specific UVM caching + prefetch pipeline uses cases to circumvent VBE TBE output concatenation. As concatenation is implemented in the preceding diff, this diff cleans up the logic left behind from grouping sharding by UVM caching kernel conditions to avoid VBE TBE output concatenation. Reviewed By: dstaay-fb, levythu Differential Revision: D58989195
|
This pull request was exported from Phabricator. Differential Revision: D58989195 |
joshuadeng
force-pushed
the
export-D58989195
branch
from
a8ba210 to
c5144fd
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.
Summary:
reverting sharding grouping logic in EBC/VLE modules that supported specific UVM caching + prefetch pipeline uses cases to circumvent VBE TBE output concatenation.
As concatenation is implemented in the preceding diff, this diff cleans up the logic left behind from grouping sharding by UVM caching kernel conditions to avoid VBE TBE output concatenation.
Differential Revision: D58989195