Use ATen ops on meta tensor to compute output shapes and strides for ExprEval segments

[zasdfgbnm]

No description provided.

[zasdfgbnm]

!test

[zasdfgbnm]

Fixes #4888

[github-actions]

Review updated until commit 5f89380

Description

• Use ATen ops for ExprEval segment output shapes and strides

• Compute and update contiguity for non-broadcast and non-reduction dims

• Fix contiguity validation by passing TensorView to validation function

• Add helper to infer contiguity from sizes and strides

---

Changes walkthrough 📝

	Relevant files
Bug fix	nodes.cpp Improve contiguity validation logic                                            csrc/ir/nodes.cpp Added check for broadcast or reduction domains when setting contiguity Improved error message clarity for contiguity validation +2/-1      tensor_metadata.cpp Fix contiguity validation via TensorView                                  csrc/tensor_metadata.cpp Updated validateAllocationSizesAndStrides to accept TensorView instead of contiguity vector Fetch contiguity directly from TensorView in validation Improved error message with TensorView name +6/-3
Enhancement	fusion_cache_utils.cpp Add contiguity computation and update logic                            csrc/runtime/fusion_cache_utils.cpp Added _computeContiguity helper to infer contiguity from sizes and strides Enhanced updateWithSegmentOutputs to conditionally update contiguity Integrated ATen tensor metadata for ExprEval segment outputs Added debug prints for tensor properties during contiguity update +75/-1    fusion_kernel_runtime.cpp Use ATen ops for ExprEval output inference                              csrc/runtime/fusion_kernel_runtime.cpp Use ExpressionEvaluator to compute output sizes for ExprEval segments Route ExprEval outputs through ATen-style stride computation Pass update_contiguity flag when updating segment outputs +22/-5    fusion_cache_utils.h Extend updateWithSegmentOutputs signature                                csrc/runtime/fusion_cache_utils.h Added optional update_contiguity parameter to updateWithSegmentOutputs Defaulted new parameter to false for backward compatibility +2/-1

---

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 No relevant tests

⚡ Recommended focus areas for review Debug Logging The code includes uncommented std::cout statements for debugging, which should not be present in production code. std::cout << tv->toString() << std::endl; tv->printTransforms(); const at::Tensor& tensor = group_runtime_outputs[group_out_i].as<at::Tensor>(); const std::vector<int64_t> sizes = tensor.sizes().vec(); const std::vector<int64_t> strides = tensor.strides().vec(); // const auto [sizes, strides] = inferAndValidateAllocationSizesAndStrides(tensor, tv, ExpressionEvaluator()); std::vector<std::optional<bool>> contiguity = _computeContiguity(sizes, strides); std::cout << "sizes: " << sizes << std::endl; std::cout << "strides: " << strides << std::endl; // std::cout << "contiguity: " << contiguity << std::endl; Possible Issue The logic for handling reduction dimensions in contiguity computation may incorrectly assume the index increment; the commented-out line suggests uncertainty in the implementation. int64_t index_with_reduction = 0; for (const auto id : tv->domain()->maybeAllocation()) { if (id->isReduction()) { contiguity_with_reduction.push_back(std::nullopt); } else { contiguity_with_reduction.push_back(contiguity[index_with_reduction++]); // (void)index_with_reduction; // contiguity_with_reduction.push_back(false); } } Function Signature Change The function validateAllocationSizesAndStrides now takes a TensorView* instead of a contiguity vector, which may affect how contiguity is validated and could introduce inconsistencies if not handled properly across call sites. void validateAllocationSizesAndStrides( const std::vector<IterDomain*>& alloc_dom, TensorView* tv, c10::IntArrayRef sizes, c10::IntArrayRef strides) { const std::vector<std::optional<bool>>& contiguity = tv->getContiguity(); NVF_ERROR(alloc_dom.size() == contiguity.size());

[zasdfgbnm]

!test