Add format versioning to ML-selector trainer - #942
Open
kevinjzhang wants to merge 4 commits into
Open
Conversation
Summary: Adds a shared training utility that filters candidate graphs by compressing caller-provided MultiInput samples at a target format version. Callers are responsible for supplying inputs that exercise the graph paths whose compatibility must be tested. Differential Revision: D114119859
Summary: Threads the resolved target format version through the clustering trainer: - clusterSuccessors() now filters the clustering (concat) codecs by ZL_Compressor_Node_getMinVersion against the target version and throws FormatVersionUnsupportedError if none survive (so the orchestrator can fall clustering back to zstd), and builds its round-trip CCtx at the target version. - train_cluster() resolves the version from TrainParams and passes it through Trainer::getTrainedClusteringConfig into CompressionUtils, whose per-sample benchmark CCtx now targets that version instead of ZL_MAX_FORMAT_VERSION. FormatVersionUnsupportedError moved to train_params.h (train_common) so it is visible to every trainer and the orchestrator without a BUCK dependency cycle. Differential Revision: D113974921
Summary: Adds format versioning to ACE training by filtering the list of graphs per format version. - Since it is useful to store both the format version and the per format version/ input type list of compatible graphs, refactor the stateless API to a ACECompressorBuilder class. This contains a cache and versions. - Additionally fix a minor bug with graph replacement Differential Revision: D113974924
Summary: Add targeting per format version to ML selector. The successor are obtained from the ML Selector graph, so when the format version is inadequate to support them, the graph must be rewritten to exclude the unsupported graphs. Implements this functionality. Reviewed By: terrelln Differential Revision: D113974920
|
@kevinjzhang has exported this pull request. If you are a Meta employee, you can view the originating Diff in D113974920. |
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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:
Add targeting per format version to ML selector.
The successor are obtained from the ML Selector graph, so when the format version is inadequate to support them, the graph must be rewritten to exclude the unsupported graphs. Implements this functionality.
Reviewed By: terrelln
Differential Revision: D113974920