Fix exporting streaming zipformer models as non-streaming models

[csukuangfj]

See errors reported at
#2086 (comment)

cc @nshmyrev

Summary by CodeRabbit

• Bug Fixes
  • Improved convolution handling for short input sequences so results are now computed correctly when the sequence length is smaller than the kernel size.
  • Preserved the existing behavior for longer sequences, with no change to standard processing.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: d3dd37e6-462d-4ebd-b902-1687d1afa41e

📥 Commits

Reviewing files that changed from the base of the PR and between 3bf7070 and bf4992a.

📒 Files selected for processing (1)

• egs/librispeech/ASR/zipformer/scaling_converter.py

---

📝 Walkthrough

Walkthrough

In NonStreamingChunkCausalDepthwiseConv1d.forward, the left_edge/right_edge scaling operands now branch on whether seq_len < kernel_size: short sequences slice those tensors to seq_len, while the existing zero-padding and concatenation path runs for seq_len >= kernel_size.

Changes

Short-sequence edge scaling fix

Layer / File(s)	Summary
Conditional left/right edge scaling egs/librispeech/ASR/zipformer/scaling_converter.py	Adds a seq_len < kernel_size branch that slices left_edge and right_edge to seq_len; the original zero-padding concatenation path is retained for seq_len >= kernel_size.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~5 minutes

Poem

A kernel so wide, the sequence so short,
The rabbit sliced edges of the proper sort.
No zeros appended where none should go,
Just trim to the length and let tensors flow.
🐇✂️

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately summarizes the change to fix exporting streaming Zipformer models as non-streaming models.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[gemini-code-assist]

Code Review

This pull request modifies the forward method in scaling_converter.py to handle sequence lengths shorter than the kernel size by introducing an if/else conditional block. The reviewer suggests refactoring this to a unified padding and slicing approach using max to eliminate the conditional block, which ensures the code remains ONNX-export friendly and avoids potential issues with ONNX runtimes.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

The class NonStreamingChunkCausalDepthwiseConv1d is specifically designed to be ONNX-export friendly by avoiding conditionals where possible (as noted in its docstring).\n\nBy using a unified padding and slicing approach with max(0, seq_len - self.kernel_size), we can completely eliminate the if/else control flow block. This avoids generating ONNX If nodes, which can be problematic or inefficient for certain ONNX runtimes (e.g., TensorRT or mobile runtimes), while keeping the code much more concise.

        pad_len = max(0, seq_len - self.kernel_size)\n        pad = torch.zeros(\n            left_edge.shape[0], pad_len, device=left_edge.device, dtype=left_edge.dtype\n        )\n        left_edge = torch.cat((left_edge, pad), dim=-1)[:, :seq_len]\n        right_edge = torch.cat((pad, right_edge), dim=-1)[:, -seq_len:]

[nshmyrev]

@csukuangfj wonderful, many thanks!