fix(rnnt): guard the export decoder_type path on the actual signature, not hasattr

[Anai-Guo]

What does this PR do ?

EncDecRNNTModel.set_export_config() gates the decoder_type export option on
hasattr(self, 'change_decoding_strategy'), but EncDecRNNTModel defines
change_decoding_strategy() itself, so the check is always True. The else
branch is therefore unreachable, and a non-hybrid RNNT model exported with
decoder_type raises a confusing TypeError from deep inside the call instead of
the intended message.

https://github.com/NVIDIA-NeMo/Speech/blob/main/nemo/collections/asr/models/rnnt_models.py#L1103-L1111

Only the hybrid models actually accept the keyword:

class	change_decoding_strategy signature	accepts decoder_type?
EncDecRNNTModel (rnnt_models.py:425)	(self, decoding_cfg: DictConfig, verbose=True)	❌
EncDecRNNTBPEModel (rnnt_bpe_models.py:387)	(self, decoding_cfg: DictConfig, verbose: bool = True)	❌
EncDecHybridRNNTCTCModel (hybrid_rnnt_ctc_models.py:324)	(self, decoding_cfg=None, decoder_type=None, verbose=True)	✅
EncDecHybridRNNTCTCBPEModel (hybrid_rnnt_ctc_bpe_models.py:424)	(self, decoding_cfg=None, decoder_type=None, verbose=True)	✅

Collection: [ASR]

Changelog

• Guard the decoder_type branch on the parameter the branch actually needs
  ('decoder_type' in inspect.signature(...).parameters) instead of on
  hasattr, which cannot discriminate.
• Hybrid models keep their existing behaviour; non-hybrid models now get the
  intended Model does not have decoder type option error.

Usage

Before this change, on any non-hybrid RNNT model:

model.set_export_config({'decoder_type': 'ctc'})
# TypeError: change_decoding_strategy() got an unexpected keyword argument 'decoder_type'

After:

model.set_export_config({'decoder_type': 'ctc'})
# Exception: Model does not have decoder type option

Verification

I don't have a GPU here, so rather than importing NeMo I lifted each
change_decoding_strategy signature verbatim from the sources with ast, rebuilt
them, and replayed the exact call the guarded branch makes — under both the current
and the patched guard:

model	current (main)	with this PR
EncDecRNNTModel	TypeError: ... unexpected keyword argument 'decoder_type'	Exception: Model does not have decoder type option
EncDecRNNTBPEModel	TypeError: ... unexpected keyword argument 'decoder_type'	Exception: Model does not have decoder type option
EncDecHybridRNNTCTCModel	returns 'applied'	returns 'applied' (unchanged)
EncDecHybridRNNTCTCBPEModel	returns 'applied'	returns 'applied' (unchanged)

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? (no user-facing docs change; the documented behaviour is what now actually happens)
• Does the PR affect components that are optional to install? (No)

PR Type:

• Bugfix
• New Feature
• Documentation
• Test

🤖 Generated with Claude Code

[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.