Relax stride check in unit dimensions.

[mzient]

Category:

Bug fix (non-breaking change which fixes an issue)

Description:

This PR relaxes the stride check in unit dimensions. Some older versions of torch always give stride 1 for unit dimensions in DLPack capsules, causing an unnecessary error on contiguous data.
Example:
shape = (1, 3, 15, 20)
dlpack strides: (1, 300, 20, 1)

Additional information:

Affected modules and functionalities:

Key points relevant for the review:

Tests:

• Existing tests apply
• New tests added
  • Python tests
  • GTests
  • Benchmark
  • Other
• N/A

Checklist

Documentation

• Existing documentation applies
• Documentation updated
  • Docstring
  • Doxygen
  • RST
  • Jupyter
  • Other
• N/A

DALI team only

Requirements

• Implements new requirements
• Affects existing requirements
• N/A

REQ IDs: N/A

JIRA TASK: N/A

[dali-automaton]

CI MESSAGE: [48103266]: BUILD STARTED

[greptile-apps]

Greptile Summary

This PR fixes a DLPack interoperability bug where older PyTorch versions report stride 1 for unit-extent dimensions instead of the expected contiguous stride, causing CheckContiguousTensor to reject valid contiguous tensors unnecessarily. The fix replaces the previous stride_from_shape_collapsed heuristic with a simpler, more correct rule: skip the stride check entirely for any dimension whose extent is 1 (since the stride is never dereferenced for a unit dimension). A companion early-return for zero-volume tensors is also a sound cleanup.

Confidence Score: 5/5

Safe to merge — the fix is minimal, correct, and well-targeted at the reported interoperability issue.

No P0 or P1 issues found. The new logic is simpler and strictly more correct than what it replaces: skipping the stride check for unit extents is sound because a stride is only meaningful when the index can be non-zero. The stride_from_shape accumulation is unaffected (multiplying by 1 is a no-op). The zero-volume early-return is a clean orthogonal improvement. All prior review concerns have been addressed or resolved in the thread.

No files require special attention.

Vulnerabilities

No security concerns identified.

Important Files Changed

Filename	Overview
dali/python/backend_impl.cc	Simplifies CheckContiguousTensor by skipping the stride check for unit-extent dimensions and adding an early return for zero-volume tensors — both changes are logically correct and the stride_from_shape accumulation continues to work properly since multiplying by 1 is a no-op.
dali/test/python/experimental_mode/test_interop.py	Adds test_from_torch parameterized over CPU/GPU, exercising the leading-unit-dimension case (shape (1,2,2)) that previously failed with DLPack strides from older PyTorch builds; the version guard is justified as discussed in the review thread.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[CheckContiguousTensor called] --> B{num_strides == num_extents?}
    B -->|No| C[DALI_ENFORCE fails]
    B -->|Yes| D[Scan dims for shape == 0]
    D -->|Any dim is 0| E[Return early - zero-volume tensor]
    D -->|No zero dims| F[stride_from_shape = element_size]
    F --> G{i >= 0?}
    G -->|No| H[All strides valid]
    G -->|Yes| I{shape at i == 1?}
    I -->|Yes - unit extent| J[Skip stride check]
    I -->|No| K{strides at i == stride_from_shape?}
    K -->|No| L[DALI_ENFORCE fails]
    K -->|Yes| M[stride_from_shape times shape at i, decrement i]
    J --> M
    M --> G

Loading

_{Reviews (2): Last reviewed commit: "Remove dead variable." | Re-trigger Greptile}

[greptile-apps]

Dead variable left over from refactor

last_non_one_dim is initialised but never read after the simplification — it was part of the old stride_from_shape_collapsed logic that was removed. Most compilers will warn about it.

Suggested change

	int64_t last_non_one_dim = 1;
	int64_t stride_from_shape = element_size;

[greptile-apps]

Test guards against the version that triggers the bug

The PR description states the problem occurs in older PyTorch versions (strides of 1 for unit dimensions in DLPack). This test skips on < 2.6.0, meaning it never runs on the versions that actually exposed the bug. Consider adding a separate test path (or removing the version guard) to confirm the fix works on the affected older versions, or at minimum document why this gate is needed.

[mzient]

@greptileai: The tests fails with PyTorch 2.7.1, so the check against 2.6.0 doesn't defeat it.

[greptile-apps]

You're right—if the stride issue still occurs in PyTorch 2.7.1, then the >= 2.6.0 guard makes sense. The test will run on 2.7.1+ and verify the fix handles those cases. I misread the situation as the bug only affecting older versions.

Tip: You can give me more context with .greptile/rules so I get this right next time.

[dali-automaton]

CI MESSAGE: [48104327]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [48104565]: BUILD STARTED

[dali-automaton]

CI MESSAGE: [48104565]: BUILD PASSED