Skip to content

Transducer Decoding: avoid alignments to store non-blank confidence#15315

Open
artbataev wants to merge 25 commits intomainfrom
vbataev/rnnt_decoding_refactor
Open

Transducer Decoding: avoid alignments to store non-blank confidence#15315
artbataev wants to merge 25 commits intomainfrom
vbataev/rnnt_decoding_refactor

Conversation

@artbataev
Copy link
Collaborator

@artbataev artbataev commented Jan 26, 2026

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An outdated branch is never blocking the merge of a PR.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 ?

Store non-blank confidence in BatchedHyps to avoid costly alignments structure for confidence when possible (if blank label confidence is ignored):

  • confidence is used in different applications, but current implementation with alignments makes decoding significantly slower
  • by default blank label confidence is ignored
  • WIND RNN-T decoding (making RNN-T as fast as TDT) is incompatible with saving results at each step. This PR unblocks Faster RNN-T Label-Looping Decoding with WIND #14451

Additionally, this PR unifies and simplifies RNN-T and TDT decoding.

Collection: [ASR]

Changelog

  • Add specific line by line info of high level changes in this PR.

Usage

  • You can potentially add a usage example below
# Add a code snippet demonstrating how to use this 

GitHub Actions CI

The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.

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

  • Related to # (issue)

Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
…es in `torch_impl`.

Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
@github-actions github-actions bot added the ASR label Jan 26, 2026
@artbataev artbataev changed the title RNN-T and TDT decoding: avoid alignments to store non-blank confidence ASR Transducer Decoding: avoid alignments to store non-blank confidence Jan 26, 2026
@artbataev artbataev changed the title ASR Transducer Decoding: avoid alignments to store non-blank confidence ransducer Decoding: avoid alignments to store non-blank confidence Jan 26, 2026
@artbataev artbataev changed the title ransducer Decoding: avoid alignments to store non-blank confidence Transducer Decoding: avoid alignments to store non-blank confidence Jan 26, 2026
artbataev and others added 11 commits January 26, 2026 20:26
# Conflicts:
#	nemo/collections/asr/parts/utils/rnnt_utils.py
Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
Signed-off-by: artbataev <artbataev@users.noreply.github.com>
Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
@artbataev artbataev marked this pull request as ready for review January 26, 2026 19:37
@github-actions github-actions bot removed the Run CICD label Jan 27, 2026
self.reset_cuda_graphs_state()
return True

# fusion models-related methods
Copy link
Collaborator

@lilithgrigoryan lilithgrigoryan Jan 27, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Are these changes related to the PR? Seems like this is more related to per-stream biasing PR

# Conflicts:
#	nemo/collections/asr/parts/submodules/transducer_decoding/rnnt_label_looping.py
#	nemo/collections/asr/parts/submodules/transducer_decoding/tdt_label_looping.py
Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
Signed-off-by: Vladimir Bataev <vbataev@nvidia.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants