Add Customization Capabilities to Cache-Aware Models

[artbataev]

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An outdated branch is never blocking the merge of a PR.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

Refactors cache-aware streaming script:

• unifies the interface with speech_to_text_eval.py and speech_to_text_streaming_infer_rnnt.py (using Hydra/OmegaConf instead of Argparse)
• exposes decoding parameters for customization purposes - to allow passing LM and/or phrase boosting params

Known issues: currently, cache-aware models do not support torch.bfloat16/float16 directly, only via AMP. This should be fixed in future.

Collection: [ASR]

Changelog

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

Usage

python ${NEMO_DIR}/examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py \
  model_path=${ASR_MODEL_PATH} \
  dataset_manifest=${DATASET} \
  compute_dtype=float32 \
  matmul_precision=high \
  allow_mps=true \
  batch_size=${BATCH_SIZE} \
  output_path="${RESULTS_DIR} \
  rnnt_decoding.strategy=greedy_batch \
  rnnt_decoding.greedy.boosting_tree.key_phrases_file=${KEY_WORDS_LIST} \
  rnnt_decoding.greedy.boosting_tree.context_score=1.0 \
  rnnt_decoding.greedy.boosting_tree.depth_scaling=2.0 \
  rnnt_decoding.greedy.boosting_tree_alpha=${BT_ALPHA}

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The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.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

[Copilot]

Pull Request Overview

This PR refactors the cache-aware streaming script to unify its interface with other ASR scripts by replacing argparse with Hydra/OmegaConf configuration management and exposing decoding parameters for customization.

• Migrates from argparse to Hydra/OmegaConf configuration system
• Extracts common device and dtype selection utilities to shared module
• Adds support for customizable decoding parameters including language model and phrase boosting

Reviewed Changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
nemo/collections/asr/parts/utils/transcribe_utils.py	Adds common device and dtype selection utilities
nemo/collections/asr/parts/submodules/tdt_beam_decoding.py	Adds NotImplementedError for unsupported partial hypotheses
nemo/collections/asr/parts/submodules/rnnt_beam_decoding.py	Adds NotImplementedError for unsupported partial hypotheses
examples/asr/transcribe_speech.py	Updates to use shared dtype selection utility
examples/asr/asr_chunked_inference/rnnt/speech_to_text_streaming_infer_rnnt.py	Updates to use shared device and dtype utilities
examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py	Major refactor from argparse to Hydra configuration with dataclass config

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[Copilot]

There is a typographical error with '§' symbol before '[70,1]'. This should be removed.

Suggested change

	For models which support multiple lookaheads, the default is the first one in the list of model.encoder.att_context_size. To change it, you may use att_context_size, for example att_context_size=§[70,1].
	For models which support multiple lookaheads, the default is the first one in the list of model.encoder.att_context_size. To change it, you may use att_context_size, for example att_context_size=[70,1].

[Copilot]

[nitpick] The processed_signal assignment on line 206 could be chained with line 205 to reduce redundant variable assignment: processed_signal = streaming_buffer.get_all_audios()[0].to(compute_dtype)

[Copilot]

[nitpick] Similar to the offline case, this dtype conversion could be optimized by moving it inline or consolidating the conversion logic into a helper function if this pattern appears frequently.

[nithinraok]

LGTM. Neat work. Thanks

[nithinraok]

Shouldn;t any CI-CD tests be updated that relies on this script? @artbataev

[KunalDhawan]

This is great work, thank you @artbataev! I tested it locally on my end and everything works well. Added minor comments and a few nitpicks, other than that LGTM

[KunalDhawan]

need to change this to speech_to_text_cache_aware_streaming_infer.py

[artbataev]

Fixed

[KunalDhawan]

Same as above, change to speech_to_text_cache_aware_streaming_infer.py

[artbataev]

Fixed

[KunalDhawan]

Could we add an example here to show how to do word boosting and LM rescoring?

[artbataev]

I added an example and also a link to the documentation. Documentation about word boosting would be updated shortly in #14800

[KunalDhawan]

should we set compute_dtype to be float32 by default? This is because currently by default compute_dtype is None and amp is false, if a user calls the inference script as follows:

 python examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py model_path=model.nemo dataset_manifest=manifest.json batch_size=16 att_context_size=[70,13]

They'll run into the NotImplementedError on line 310: Compute dtype None is not yet supported for cache-aware models, use float32 instead

[artbataev]

Thanks, good catch! Fixed (and added comments).

[artbataev]

Shouldn;t any CI-CD tests be updated that relies on this script? @artbataev

@nithinraok I do not see tests for streaming inference scripts in tests/functional_tests, looks like only transcribe_speech.py is tested. Do you suggest adding a bunch of tests there? I can add simple tests for cache-aware + (new) streaming here or in future PR.

[github-actions]

[🤖]: Hi @artbataev 👋,

We wanted to let you know that a CICD pipeline for this PR just finished successfully.

So it might be time to merge this PR or get some approvals.

//cc @chtruong814 @ko3n1g @pablo-garay @thomasdhc

[KunalDhawan]

LGTM, thank you!

[nithinraok]

Shouldn;t any CI-CD tests be updated that relies on this script? @artbataev

@nithinraok I do not see tests for streaming inference scripts in tests/functional_tests, looks like only transcribe_speech.py is tested. Do you suggest adding a bunch of tests there? I can add simple tests for cache-aware + (new) streaming here or in future PR.

Yes, pls add one for cache-aware as we are prioritzing cache-aware models

[artbataev]

Shouldn;t any CI-CD tests be updated that relies on this script? @artbataev

@nithinraok I do not see tests for streaming inference scripts in tests/functional_tests, looks like only transcribe_speech.py is tested. Do you suggest adding a bunch of tests there? I can add simple tests for cache-aware + (new) streaming here or in future PR.

Yes, pls add one for cache-aware as we are prioritzing cache-aware models

I made a separate PR with tests, will finalize shortly #14823