Adding bf16 Sortformer train and inference

[tango4j]

What does this PR do ?

This PR adds bf16 precision training and inference of Sortformer diarizer models.
Hardware: Starting from Ampere (e.g. A100) architecture, native bf16 operation is supported.

Collection: [Note which collection this PR will affect]

ASR/speaker_task

Changelog

NeMo/nemo/collections/asr/losses/bce_loss.py
NeMo/examples/speaker_tasks/diarization/neural_diarizer/e2e_diarize_speech.py

Usage

Although model weights are FP32, e2e_diarize_speech.py script automatically converts the precision to bf16 then perform inference based on bf16.

python $BASEPATH/neural_diarizer/e2e_diarize_speech.py \
    precision="bf16" \ 
    model_path=/path/to/diar_sortformer_4spk_v1.nemo \
    batch_size=1 \
    dataset_manifest=/path/to/diarization_manifest.json

for training, specify the following configuration:

trainer.precision="bf16"

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)

[KunalDhawan]

LGTM, thanks Taejin!

[KunalDhawan]

Let's also add the bf16-mixed option and maybe add a small comment about possible gains expected with bf16 training/inference?

[ipmedenn]

LGTM!