fix(paraformer): trim timestamp predictor padding per sample

[John6666cat]

Summary

Fix per-sample timestamp predictor padding handling in Paraformer dynamic batches.

When the predictor has tail_mask=True, timestamp inputs are now trimmed per sample using:

• encoder_out_lens[i]
• one additional tail frame when tail_threshold > 0
• the available widths of both timestamp tensors as an upper bound

Behavior is unchanged when tail_mask=False, when the predictor has no tail_mask attribute, or when the tail threshold is zero.

A CPU-only, network-free regression test is included.

Type of change

• Bug fix
• Documentation
• Example or demo
• Runtime or deployment
• Benchmark or evaluation
• Model/training change

Validation

Focused regression test:

python -m pytest -q tests/test_paraformer_timestamp_padding.py

On the unmodified baseline:

3 failed, 2 passed

With this change:

5 passed

The tests cover:

• per-sample trimming with tail_mask=True

• unchanged behavior with tail_mask=False

• unchanged behavior when tail_mask is absent

• zero tail_threshold

• clamping to both timestamp tensor widths

• python -m compileall funasr examples tests

• Docs or links checked

• Runtime/deployment command tested

Also checked:

git diff --check

User impact

In dynamic batches, shorter samples can otherwise pass padded predictor frames into timestamp post-processing.

This change limits timestamp processing to each sample's valid encoder extent when tail masking is enabled, preventing padded tail frames from affecting generated timestamps while preserving the existing behavior for other predictor configurations.

Notes for reviewers

The scope is intentionally narrow and limited to Paraformer timestamp padding.

This is separate from the pred_timestamp / speaker-output contract fix handled in PR A.

No EParaformer, English/BPE, or broader timestamp flag normalization changes are included.

The regression test is CPU-only and requires no network access or model download.

[LauraGPT]

I reproduced the focused result (5 passed), but the two timestamp PRs are not currently green together. Applying #3553 then #3554 and running python -m pytest -q tests/test_paraformer_timestamp_contract.py tests/test_paraformer_timestamp_padding.py tests/test_punc_model_none.py gives 20 passed, 1 failed: #3553 constructs a lightweight Paraformer without predictor, while the new #3554 branch reads self.predictor directly and raises AttributeError. Please preserve the intended production behavior while making the access compatible, e.g. bind predictor = getattr(self, "predictor", None) before checking getattr(predictor, "tail_mask", None). Please add a combined regression (or extend the fixture) so both PRs remain green together. DCO is also missing on 947d7116eaffea2dde962a49d5251e02eb55eebb; amend with --signoff before the force-with-lease push.

[LauraGPT]

Fresh current-main revalidation confirms the existing requested change without adding scope. Against main@4300a6b2a158e126ff5d97ed55816b43ae88522a:

• fix(auto-model): preserve explicit timestamp and speaker text #3553 alone: 16 passed
• fix(auto-model): preserve explicit timestamp and speaker text #3553 + fix(paraformer): trim timestamp predictor padding per sample #3554 as currently published: 20 passed, 1 failed with AttributeError: Paraformer has no attribute predictor
• after only binding predictor = getattr(self, "predictor", None) and using that value for the two getattr calls: 23 passed across both PR regressions, test_punc_model_none.py, and test_timestamp_tools.py

Environment: Python 3.12, NumPy 1.26.4, SciPy 1.12.0, Torch 2.11.0+cu130. The server-wide NumPy 2.x/SciPy ABI mismatch was isolated and is not attributed to this PR.

Please apply the minimal compatibility change on #3554 and amend both PR commits with your own Signed-off-by trailers. I will re-run the exact replacement heads; I will not add DCO on the contributor’s behalf.

[John6666cat]

Thanks. I applied the requested minimal compatibility change by binding predictor = getattr(self, "predictor", None) and using that value for the two checks. I also added coverage for the missing predictor attribute case, amended the commit with my DCO sign off, and force pushed the replacement head. New SHA: 6042ace.

[LauraGPT]

The requested compatibility fix and missing-predictor regression are correct. I applied exact heads #3553 449414665d36953f3bd83e514c31125a228c0fe4 and #3554 6042ace22d617a69c47735503cb67be04c7825fd together on current main@55b662ccf9ea77237ba9253b3bddd953d4184f84; all 24 focused timestamp, padding, punctuation-none, and timestamp-tool tests pass. py_compile and git diff --check also pass.

The only remaining blocker is the trailer syntax. The commit currently contains:

Signed-off-by: John6666 186692226+John6666cat@users.noreply.github.com

DCO requires angle brackets:

Signed-off-by: John6666 <186692226+John6666cat@users.noreply.github.com>

Please ensure user.name and user.email are configured, then run git commit --amend --no-edit --signoff and git push --force-with-lease. Please verify the resulting message with git log -1 --format=%B. No code changes are requested.

[LauraGPT]

Revalidated the exact cleaned replacement head b2fb22d1addcfa0284b3ca0c63bf2e02d02a6ad4 after #3553 on current main@917ca2f4a673b4ee8b0aea4b64fbe2d6c2034a67.

• the accidental C# obj/Debug artifacts are absent; only the two intended files remain
• the source tree is identical to the previously reviewed implementation
• the corrected Signed-off-by trailer satisfies DCO
• the combined #3553 + #3554 focused suite passes 24/24
• py_compile and git diff --check pass

Approved for merge after #3553.

[John6666cat]

Thanks for revalidating the cleaned replacement head and the combined suite. I appreciate the thorough review.