Streaming transducer inference: fix memory usage, improve WER

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An outdated branch is never blocking the merge of a PR.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

Fix streaming inference pipeline introduced in #9106

• fix memory consumption when using small chunks
  • fix memory used by BatchedHyps when merging (extend only if necessary)
  • use PYTORCH_CUDA_ALLOC_CONF with expandable_segments:True: this can improve memory usage (reserved memory) by >3x (in some cases, where chunk=1 frame (0.08s) – more than 10x)
• fix number of frames to decode in the last chunk: improves WER on datasets where end of the audio still contains speech (no silence)

Streaming results comparison

Configuration:

• model: nvidia/parakeet-rnnt-1.1b
• contexts: left=10, chunk=0.16, right=0.8
• batch size: 32
• compute dtype: bfloat16

	LS test-other	AMI
WER before	3.77	17.90
WER this PR	3.77	17.79
time before	2:57	3:45
time this PR	2:53	3:40
GPU memory allocated before	8.02 GB	8.02 GB
GPU memory allocated this PR	8.01 GB	8.01 GB
GPU memory reserved before	26.17 GB	26.17 GB
GPU memory reserved this PR	8.05 GB	8.05 GB

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

 python examples/asr/asr_chunked_inference/rnnt/speech_to_text_streaming_infer_rnnt.py \
    pretrained_name="nvidia/parakeet-rnnt-1.1b" \
    audio_dir="<dir_with_audio>" \
    output_filename="decoded.jsonl" \
    right_context_secs=1 \
    chunk_secs=0.16 \
    left_context_secs=10 \
    batch_size=32

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The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.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

[andrusenkoau]

Great work, thank you!

[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