Reduce unnecessary GreedyPerfPartitioner calls from MemoryBalancedPartitioner #2914
+20
−9
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
…titioner Summary: MemoryBalancedPartitioner works by adjusting the max memory on devices and calling GreedyPerfPartitioner repeatedly. The max memory is adjusted with a binary search procedure to identify a more memory efficient plan than what GreedyPerfPartitioner gives by default. The search boundaries for the binary search procedure were inefficient which this diff addresses. * **Upper bound** * **Before:** Max device HBM (e.g. 80 GB) * **After:** Max HBM usage of the default plan since there is no point in searching for plans that use more max memory than what the default plan uses. * **Lower bound:** * **Before:** [Avg. HBM per Device] = [Total HBM Needed Across All Shards] / [World Size] * **After:** max([Avg. HBM per Device], [Max HBM Needed Across All Shards]). A feasible solution requires at least the max HBM that the biggest shard needs so there is no point in searching for options below that. Making these changes can have impact in two ways: 1. Search procedure is more efficient leading to plans with lower memory 2. We can reduce `search_count` to get comparable plans as before while calling `GreedyPerfPartitioner` less number of times from `MemoryBalancedPartitioner`. The default impact without further changes from pytorch#1 should lead to a marginal max memory improvement. Differential Revision: D73598477
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D73598477 |
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Outdated suggestions cannot be applied.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:
MemoryBalancedPartitioner works by adjusting the max memory on devices and calling GreedyPerfPartitioner repeatedly. The max memory is adjusted with a binary search procedure to identify a more memory efficient plan than what GreedyPerfPartitioner gives by default.
The search boundaries for the binary search procedure were inefficient which this diff addresses.
Making these changes can have impact in two ways:
search_countto get comparable plans as before while callingGreedyPerfPartitionerless number of times fromMemoryBalancedPartitioner.The default impact without further changes from #1 should lead to a marginal max memory improvement.
Differential Revision: D73598477