[Data] OperatorFusionRule: capability-based fusion check#64244
Open
ShockYoungCHN wants to merge 1 commit into
Open
[Data] OperatorFusionRule: capability-based fusion check#64244ShockYoungCHN wants to merge 1 commit into
ShockYoungCHN wants to merge 1 commit into
Conversation
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request generalizes operator fusion in Ray Data by introducing a mechanism for physical operators to declare if they can absorb an upstream TaskPoolMapOperator's MapTransformer (via absorbs_upstream_map_transformer) and how to perform the fusion (via fuse_with_upstream_map_transformer). The AllToAllOperator is updated to implement this capability, and the operator fusion rule is refactored to support any such 'absorber' operator rather than being hardcoded to AllToAllOperator. Feedback suggests making the logical operator rebuilding in _build_fused_logical_op more defensive by using .get() and getattr to avoid potential KeyError or AttributeError issues.
Important
The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.
ShockYoungCHN
force-pushed
the
capbased-fusion
branch
from
c927ee5 to
eb8a825
Compare
ray-gardener
Bot
added
data
ShockYoungCHN
force-pushed
the
capbased-fusion
branch
from
eb8a825 to
8f9daa2
Compare
…ber side
Refactors the "down absorbs up" fusion path so the rule no longer
isinstance-checks specific operator classes on the absorber side. The
absorber owns the construction of the fused replacement; the rule owns
DAG-level bookkeeping (logical-op replacement in _op_map).
Before:
- _fuse_all_to_all_operators_in_dag isinstance-checked AllToAllOperator
- _get_fused_all_to_all_operator built the fused AllToAllOperator with
the upstream MapTransformer stashed on TaskContext, plus rebuilt the
corresponding RandomShuffle/Repartition logical op.
- Any new sink op (future shuffle backends, write sinks, etc.) wanting
the same fusion would have required modifying the rule.
After:
- PhysicalOperator gains a pair of capability methods:
- absorbs_upstream_map_transformer() -> bool (default False)
- fuse_with_upstream_map_transformer(transformer) -> PhysicalOperator
(default raises NotImplementedError)
- AllToAllOperator overrides both. Its fuse_with_upstream_map_transformer
carries the prior _get_fused_all_to_all_operator transform-fn logic
verbatim: the new bulk_fn wraps the original transform_fn and stashes
the upstream MapTransformer (plus its ray_remote_args) on TaskContext,
so the existing shuffle scheduler picks them up and wraps each map task.
Behavior-preserving.
- OperatorFusionRule:
- _is_absorber(op) reduces to op.absorbs_upstream_map_transformer().
- _get_fused_absorber_operator(down, up) reduces to
down.fuse_with_upstream_map_transformer(up.get_map_transformer())
plus shared bookkeeping (_build_fused_logical_op handles the
RandomShuffle/Repartition logical-op rebuild).
- _can_fuse adds a third branch for any capability-implementing
downstream; it skips the rule's MapOperator-oriented logical-op
compat checks (absorbers handle compat in their fuse method).
The legacy _get_fused_all_to_all_operator helper is removed; the
fuse_with_upstream_map_transformer method on AllToAllOperator is the
only place that knows the AllToAllOperator construction details.
No behavior change for current users (AllToAllOperator is the only
absorber in master). The interface is the seam future sink ops can use
to opt into fusion without modifying this rule.
Signed-off-by: Yuanzhuo Yang <yuanzhuoyang@gmail.com>
ShockYoungCHN
force-pushed
the
capbased-fusion
branch
from
8f9daa2 to
b848141
Compare
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.
Refactor "down absorbs up" fusion to capability-based dispatch
Move fusion construction from the rule into the operator itself. The rule no longer do isinstance-check specific classes on the absorber side, it asks the op via two new PhysicalOperator hooks:
AllToAllOperator implements both, carrying the same transform-fn wrapping logic that was previously in _get_fused_all_to_all_operator.
FuseOperators changes:
No behavior change, AllToAllOperator is the only absorber. It opens the seam for future sink ops to opt into fusion without modifying the rule.