Fix OOB decode crash in buffered RNNT/TDT pipeline

[naymaraq]

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An outdated branch is never blocking the merge of a PR.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

Problem

In the buffered transducer/RNNT pipeline, a request's final chunk left enc_lens_dec
unclamped. When right_padding_size is large relative to chunk_size, this length can
exceed the encoder output's time dimension (after left padding is sliced off), feeding an
out-of-bounds time index to the TDT decoder → CUDA index-out-of-bounds assert (crash).

Collection: [ASR]

Changelog

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

Results

Setup: nvidia/parakeet-tdt-0.6b-v2 (TDT, greedy_batch), buffered pipeline
(buffered_rnnt.yaml), full LibriSpeech test-other (2939 utts, 5.34 h), left_padding=5.6s,
EoU disabled, batch_size=64. old = pre-fix (HEAD~1), new = fix (HEAD).

case	chunk (s)	right (s)	old code	new code (fix)	WER Δ
old crashes (OOB)	0.08	0.64	💥 CUDA OOB crash	✅ 43.06%	crash → runs
old crashes (OOB)	0.16	0.64	💥 CUDA OOB crash	✅ 38.79%	crash → runs
old crashes (OOB)	0.08	1.12	💥 CUDA OOB crash	✅ 12.28%	crash → runs
old crashes (OOB)	0.16	1.12	💥 CUDA OOB crash	✅ 11.82%	crash → runs
old runs (in-bounds)	0.16	1.60	10.24%	10.24%	0.00 pp
old runs (in-bounds)	0.32	1.60	10.40%	10.40%	0.00 pp
old runs (in-bounds)	0.64	1.60	9.88%	9.88%	0.00 pp
old runs (in-bounds)	4.80	1.60	7.21%	7.21%	0.00 pp
old runs (in-bounds)	0.08	0.56	38.54%	38.52%	−0.01 pp
old runs (in-bounds)	0.16	0.56	32.90%	32.90%	−0.01 pp

• Crash resolved: OOB configs that aborted on old code now run.
• No regression: on already-working configs, WER changes by ≤ 0.02 pp (bit-identical
  on the right=1.60 rows) — the clamp is a no-op when nothing overflows.

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The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.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

[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.

[naymaraq]

/ok to test 7a8e9a0