Handle missing validation_ds in RNNT BPE transcribe

[basnijholt]

Important

The Update branch button must only be pressed in very rare occassions.
An outdated branch is never blocking the merge of a PR.
Please reach out to the automation team before pressing that button.

What does this PR do ?

Handle RNNT BPE checkpoints that restore with validation_ds: null and then fail during transcribe() setup.

Collection: ASR

Changelog

• Treat missing or null validation_ds as an empty config in EncDecRNNTBPEModel._setup_transcribe_dataloader.
• Keep the current transcribe default of use_start_end_token=False when there is no validation dataset config.
• Add a regression test for this path.

Usage

This showed up in agent-cli while adding a Whisper-compatible NeMo backend for NVIDIA Parakeet models: basnijholt/agent-cli#436

Our daemon use case is:

ag daemon install whisper -- --backend nemo --model parakeet-unified-en-0.6b

The nvidia/parakeet-unified-en-0.6b checkpoint restores as an RNNT BPE model with validation_ds: null. Before this change, transcribe() crashes before decoding audio:

AttributeError: 'NoneType' object has no attribute 'get'

agent-cli has a downstream workaround for this in basnijholt/agent-cli#576, but NeMo can handle the checkpoint config directly.

GitHub Actions CI

The GitHub Actions CI will run automatically when the "Run CICD" label is added to the PR.
To re-run CI remove and add the label again.
To run CI on an untrusted fork, a NeMo user with write access must first click "Approve and run".

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

• New Feature
• Bugfix
• Documentation

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

• Model that exposed this: https://huggingface.co/nvidia/parakeet-unified-en-0.6b
• Downstream workaround in agent-cli: Fix NeMo daemon runtime and unified model load basnijholt/agent-cli#576
• Local check: uv run --frozen --no-sync pytest tests/collections/asr/test_asr_rnnt_encoder_model_bpe.py -q (2 passed, 8 skipped)

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.