Move getTvsWithDifferentSharding to cpp

[wujingyue]

The function no longer needs to be template and is large enough to justify putting in cpp.

[wujingyue]

!test

[github-actions]

Description

• Moved getTvsWithDifferentSharding from header to source file

• Removed template parameter from function signature

---

Changes walkthrough 📝

Relevant files

Enhancement

utils.cpp Added `getTvsWithDifferentSharding` implementation              csrc/multidevice/utils.cpp Added getTvsWithDifferentSharding function implementation +35/-0    utils.h Updated `getTvsWithDifferentSharding` signature                    csrc/multidevice/utils.h Removed getTvsWithDifferentSharding function implementation Updated function signature to use const std::vector& instead of template +1/-33

---

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 No relevant tests

⚡ Recommended focus areas for review Performance Impact Moving the function to C++ may have implications on performance. Ensure that the move does not introduce any performance regressions. std::unordered_set<TensorView*> getTvsWithDifferentSharding( TensorView* ref, const std::vector<TensorView*>& tvs) { std::unordered_set<TensorView*> ret; const auto& reference_dom = ref->getLoopDomain(); FusionGuard fg(ref->fusion()); auto ca_map = ComputeAtMap(FusionGuard::getCurFusion()); std::unordered_map<IterDomain*, IterDomain*> concrete_to_reference_map; for (auto id : reference_dom) { auto ca_id = ca_map.getConcreteMappedID(id, IdMappingMode::PERMISSIVE_RESIZE); concrete_to_reference_map[ca_id] = id; } for (TensorView* tv : tvs) { if (ref->getDeviceMesh().vector() != tv->getDeviceMesh().vector()) { ret.insert(tv); continue; } for (auto id : tv->getLoopDomain()) { auto ca_id = ca_map.getConcreteMappedID(id, IdMappingMode::PERMISSIVE_RESIZE); if (concrete_to_reference_map.count(ca_id) > 0) { auto ref_id = concrete_to_reference_map.at(ca_id); if ((ref_id->isDeviceDim() || id->isDeviceDim()) && ref_id->getParallelType() != id->getParallelType()) { ret.insert(tv); break; } } } } return ret; } Template Removal The function was previously templated. Verify that removing the template does not affect the flexibility or correctness of the function. std::unordered_set<TensorView*> getTvsWithDifferentSharding( TensorView* ref, const std::vector<TensorView*>& tvs);