Fix CPU ReverseSequence returning zeros for zero-length sequences

[tairenpiao]

Description

ReverseSequenceImpl skips a batch with continue when seq_len == 0, but the output isn't pre-copied from the input, so that column is left as zeros instead of unchanged. Removing the early continue lets the existing [seq_len, max_seq_len) loop copy the whole column through.

Adds ReverseSequenceTest.ZeroSequenceLength.

Motivation and Context

Fixes #28490. Labelled [OpenVINO] but the bug is in the CPU EP.

[Copilot]

Pull request overview

Fixes an incorrect CPU ReverseSequence result when a batch’s sequence_lens[i] == 0, where the output column was previously left uninitialized (observed as zeros) instead of matching the input. The fix removes the early continue so the existing [seq_len, max_seq_len) copy loop runs and preserves the unchanged portion as required by the operator semantics.

Changes:

• Remove the seq_len == 0 early-continue in the CPU ReverseSequenceImpl so the “copy-through” loop executes for zero-length sequences.
• Add a regression unit test covering sequence_lens containing a zero (ReverseSequenceTest.ZeroSequenceLength).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
onnxruntime/core/providers/cpu/tensor/reverse_sequence.cc	Removes the seq_len == 0 early exit so zero-length batches are copied through instead of leaving output uninitialized.
onnxruntime/test/providers/cpu/tensor/reverse_sequence_test.cc	Adds a regression test validating that a zero sequence length leaves the corresponding batch column unchanged.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.