fixing wrong stride retrieved from fusion

[jjsjann123]

We forgot to consider output with aliases when constructing kernel arguments.
This ended up giving us wrong arguments passed for kernel, producing wrong results.

[jjsjann123]

!test

[github-actions]

Description

• Correctly handle stride retrieval for aliased tensors

• Ensure kernel arguments are correctly constructed

---

Changes walkthrough 📝

	Relevant files
Bug fix	allocations.cpp Correct stride retrieval for aliased tensors                          csrc/runtime/allocations.cpp Adjusted condition to handle non-New allocation types Added logic to retrieve aliased tensor for ReuseBuffer type Ensured tensor evaluation and stride retrieval are correct +6/-1

---

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 No relevant tests

⚡ Recommended focus areas for review Possible Issue The condition if (alias_info.type != AllocationType::New) seems redundant and might lead to incorrect behavior if alias_info.type is neither New nor ReuseBuffer. if (alias_info.type != AllocationType::New) { // For reuse buffer alias, we need to get the aliased_io's size/stride if (alias_info.type == AllocationType::ReuseBuffer) {

[jjsjann123]

Without this patch, we'll ended up with passing the wrong kernel arguments when we have inplace updates on inputs with a non-trivial allocation domain.

The earlier refactor PR broke this test:

Fuser/tests/python/test_python_frontend.py

Lines 4949 to 4950 in 5111d3b

	fd.add_output(T6, T0)
	fd.add_output(T6)

Unfortunately the two PRs ( the refactor #3952 and the test #3970 ) were merged in parallel and never combined in the CI.

[csarofeen]

Approving, but we should understand this only works if we guarantee that input and outputs are exactly the same. i.e. they have exact matching of logical domains where every ID is exact matched in the same order and no other extraneous IDs (at least those that are not reduction). Even a broadcast would invalidate this approach. J, please follow up with a PR to make sure this is the case.