Remove an unnecessary corner case

[wujingyue]

No description provided.

[wujingyue]

!test

[github-actions]

Review updated until commit 8dec3f9

Description

• Removed unnecessary 0-dim tensor filtering logic in contiguous inner dimensions mapper

• Simplified function by eliminating corner case check that was deemed unnecessary

• Function now proceeds directly to dimension matching logic after contiguity computation

Changes walkthrough

Relevant files

Enhancement

vectorize_helper.cpp Remove 0-dim tensor filtering corner case                                csrc/scheduler/vectorize_helper.cpp Removed block that filtered out 0-dimensional tensors (lines with alloc_no_reductions_size < 1 check) Eliminated early return path for empty allocation sizes Streamlined function logic by removing unnecessary corner case handling Function now proceeds directly to dimension matching validation +0/-5

PR Reviewer Guide

Here are some key observations to aid the review process:

🧪 PR contains tests

⚡ Recommended focus areas for review

Corner Case Removal Validation The removal of the 0-dimensional tensor check (alloc_no_reductions_size < 1) needs validation. Verify that: 1) This corner case cannot occur in practice, 2) Any existing code paths that might produce 0-dim tensors are properly handled elsewhere, 3) The behavior when alloc_no_reductions_size < 0 is well-defined and won't cause crashes or incorrect results. auto alloc_no_reductions_size = alloc_no_reductions.size(); NVF_ERROR_EQ(alloc_no_reductions_size, contiguity.size()); // Order is important, need to make sure dimensions match up correctly with

Test failures

• (Medium, 3) Shape mismatch in Thunder nvFuser higher-order inplace alias update tests

  Test Name	A100	GB200	H100	Source
  thunder.tests.test_update_aliases.test_higher_order_inplace_alias_update_nvfuser_cuda_thunder.dtypes.float32	❌	❌	❌

• (Medium, 1) Scalar numerical mismatch in thunder nanogpt NVFuser CUDA test_networks

  Test Name	H100	Source
  thunder.tests.test_networks.test_nanogpt_complete_autograd_nvfuser_cuda_thunder.dtypes.float32	❌

[greptile-apps]

Greptile Overview

Greptile Summary

• Removes an early-return branch in ContiguousInnerDimensionsMapper::getContigMergeOfInnerSize for the (commented) “0-dim tensor” case.
• The function still computes the contiguous inner-extent product by walking allocation dims from inner to outer and stopping when contiguity/mapping/device-split constraints break.
• No behavioral changes elsewhere; change is localized to csrc/scheduler/vectorize_helper.cpp and relies on existing identity product initialization when there are no eligible dims.

Confidence Score: 5/5

• This PR is safe to merge with minimal risk.
• Change is a small deletion of an early-return that was redundant with the existing identity initialization (oneVal()) and loop structure; core invariants are still enforced by NVF_ERROR_EQ and contiguity handling, and no other code paths are modified.
• No files require special attention

Important Files Changed

Filename	Overview
csrc/scheduler/vectorize_helper.cpp	Removes an early-return branch in ContiguousInnerDimensionsMapper::getContigMergeOfInnerSize; remaining code already handles 0-dim tensors via size checks and returns the identity product when no alloc dims exist.

Sequence Diagram

sequenceDiagram
  participant Caller as Scheduler/Heuristic
  participant Mapper as ContiguousInnerDimensionsMapper
  participant TV as TensorView
  participant TD as TensorDomain
  participant Dep as DependencyCheck
  participant IR as SimplifyingIrBuilder

  Caller->>Mapper: getContigMergeOfInnerSize(tv)
  Mapper->>TV: getMaybeAllocationDomain()
  Mapper->>Mapper: mappedLogicalIds(tv)
  Mapper->>TD: noReductions(alloc)
  Mapper->>TV: domain()->contiguity()
  alt contiguity includes reductions
    Mapper->>Mapper: filter contiguity to alloc_no_reductions
  end
  Mapper->>Mapper: assert alloc_no_reductions.size == contiguity.size
  loop iterate inner alloc dims (reverse)
    Mapper->>Mapper: skip if extent==1 or broadcast
    Mapper->>Dep: getAllExprsBetween(logicalDomain, alloc_iid)
    Mapper->>Mapper: validate only device splits
    Mapper->>IR: divExpr(getProjectedExtent, num_devices)
    Mapper->>IR: mulExpr(product, sharded_extent)
  end
  Mapper-->>Caller: simplifyExpr(product_of_inner_extents)

Loading

[greptile-apps]

_{1 file reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Additional Comments (1)

csrc/scheduler/vectorize_helper.cpp
0-dim contiguity invariant

Removing the of_tv_alloc_no_reductions_size < 1 early-return means 0-dim tensors now hit NVF_ERROR(of_tv_alloc_no_reductions_size == contiguity.size()). If domain()->contiguity() can be empty for 0-dim while getMaybeAllocationDomain() still has entries (or vice versa), this will now fail even though product_of_inner_extents is already 1. Please confirm the invariant (both sizes are 0 for 0-dim) or adjust logic so 0-dim tensors still return 1 without requiring contiguity alignment.

[wujingyue]

!test

[greptile-apps]

_{1 file reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}