Skip to content

Move MarkAliasAnalysisPreparePass before propagateShardingsPass #4274

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Priya2698 merged 3 commits into from
Apr 23, 2025
Merged

Move MarkAliasAnalysisPreparePass before propagateShardingsPass #4274

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
226a7db
move alias analysis before propagateShardings
Priya2698 Apr 18, 2025
cfb15cd
fix breaking test
Priya2698 Apr 18, 2025
afacb99
Merge branch 'main' into pm/alias_analysis
Priya2698 Apr 22, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 0 additions & 4 deletions csrc/alias_analysis.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -239,10 +239,6 @@ void AliasFinder::handle(const ViewOp* view) {
}

void AliasFinder::handle(const LoadStoreOp* set) {
if (isResharding(set)) {
return;
}

TensorView* in = dynamic_cast<TensorView*>(set->in());
if (in == nullptr) {
return;
Expand Down
16 changes: 10 additions & 6 deletions csrc/preseg_passes/pre_segmenter.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -39,12 +39,6 @@ namespace nvfuser::preseg_passes {
debug() << "========================================" << std::endl;
}

// For resharding across GPUs.
OptimizationPass<PropagateShardingsPass>::runPass(fusion);
OptimizationPass<InsertReshardingsPass>::runPass(fusion);
OptimizationPass<ReorderShardedAxisPass>::runPass(fusion);
OptimizationPass<MakeReshardingContiguousPass>::runPass(fusion);

// Replace TensorViews with zero extent. Outputs and inputs may still be empty
OptimizationPass<RemoveEmptyPass>::runPass(fusion);
// This pass should be placed before ConsecutiveCastPass as more
Expand Down Expand Up @@ -81,6 +75,16 @@ namespace nvfuser::preseg_passes {
OptimizationPass<MarkAliasesPreparePass>::runPass(fusion);
OptimizationPass<ExactMappedExtentSubstitutionPass>::runPass(fusion);
OptimizationPass<AllocationDomainPass>::runPass(fusion);

// All the multidevice passes are moved after allocation related passes:
// MarkAliasesPreparePass, and AllocationDomainPass Multidevice passes will
// try to set the allocation domain for tvs with device mesh which will
// conflict with these passes.
OptimizationPass<PropagateShardingsPass>::runPass(fusion);
OptimizationPass<InsertReshardingsPass>::runPass(fusion);
OptimizationPass<ReorderShardedAxisPass>::runPass(fusion);
OptimizationPass<MakeReshardingContiguousPass>::runPass(fusion);

OptimizationPass<RemoveBcastSqueeze>::runPass(fusion);
OptimizationPass<SegmentInplaceUpdatePass>::runPass(fusion);
OptimizationPass<TranslateNoReductionMatmulToMulSqueeze>::runPass(fusion);
Expand Down
6 changes: 4 additions & 2 deletions tests/cpp/test_alias_analysis.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -270,7 +270,9 @@ TEST_F(AliasAnalysisTest, BroadcastExpandDimensions) {
EXPECT_EQ(analysis.getRoot(expanded_tv), in);
}

TEST_F(AliasAnalysisTest, NoAliasForReshardingExprs) {
// See PR: https://github.com/NVIDIA/Fuser/pull/4274
// for alias analysis for resharding exprs
TEST_F(AliasAnalysisTest, AliasForReshardingExprs) {
Fusion fusion;
FusionGuard fg(&fusion);

Expand All @@ -288,7 +290,7 @@ TEST_F(AliasAnalysisTest, NoAliasForReshardingExprs) {
fusion.addOutput(out);

AliasAnalysisResult analysis = findAliases(&fusion);
EXPECT_TRUE(analysis.getRoot(out) == nullptr);
EXPECT_TRUE(analysis.getRoot(out) == in);
}

} // namespace nvfuser
2 changes: 1 addition & 1 deletion tests/cpp/test_multidevice_matmul.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -238,7 +238,7 @@ TEST_F(DistributedMatmulTest, Matmul_LayoutTN_Allgather) {
executor_cache.getMostRecentKernelRuntime();
EXPECT_THAT(
kernel_runtime->fusionSegments()->groups(),
Contains(HeuristicIs(SchedulerType::ExprEval)).Times(2));
Contains(HeuristicIs(SchedulerType::ExprEval)).Times(3));
}

TEST_F(DistributedMatmulTest, Matmul_LayoutNT_AllReduce) {
Expand Down