Skip to content

Add Nemotron 3.5 ASR Streaming#46565

Draft
eustlb wants to merge 117 commits into
mainfrom
add-nemotron-3.5-asr
Draft

Add Nemotron 3.5 ASR Streaming#46565
eustlb wants to merge 117 commits into
mainfrom
add-nemotron-3.5-asr

Conversation

@eustlb

@eustlb eustlb commented Jun 11, 2026

Copy link
Copy Markdown
Collaborator

What does this PR do?

@eustlb eustlb changed the base branch from main to nemotron-asr June 18, 2026 14:51
@github-actions

Copy link
Copy Markdown
Contributor

Thank you for your contribution 🤗!

CI Security Gate — automatic approval blocked

This PR was not automatically approved for CI because the security gate failed.

Possible reasons:

  • The PR touches 50 or more files — only PRs with fewer than 50 changed files are automatically approved
  • A changed file is outside the allowed directories (src/, tests/, docs/, utils/), has a disallowed extension (only .py, .txt, .md permitted outside tests/ and docs/), or is not .md/.yml inside docs/
  • A new high-severity security issue was detected in the changed Python files (Bandit check)

See the workflow run for the exact violations.

A maintainer can review and manually approve CI if a finding is a false positive.

eustlb added 16 commits June 18, 2026 18:18
…n mixin

The base model was renamed nemotron_asr -> nemotron_asr_streaming (NemotronAsr* ->
NemotronAsrStreaming*) and evolved (Parakeet-based RNN-T generation, refactored encoder
subsampling, stateful processor num_lookahead API). Repoint Nemotron3_5Asr onto it:

- generation: Nemotron3_5AsrGenerationMixin now subclasses NemotronAsrStreamingGenerationMixin
  and only overrides generate() to stash prompt_ids (the offline encode and every streaming
  chunk read it via get_audio_features). ~370 lines removed.
- config/modeling/processing/feature-extraction: inherit the NemotronAsrStreaming* classes;
  forward returns the streaming cache fields (encoder_past_key_values, padding_cache).
- processor __call__ re-based on the new API (default_num_lookahead_tokens via
  set_num_lookahead_tokens; dropped streaming_latency_ms).
- conversion: updated encoder subsampling weight mapping (conv_in + depthwise/pointwise).
- re-added auto registrations (wiped by the rename); encoder reuses NemotronAsrStreamingEncoder.
Base automatically changed from nemotron-asr to main June 19, 2026 16:27
@github-actions

Copy link
Copy Markdown
Contributor

[For maintainers] Suggested jobs to run (before merge)

run-slow: auto, nemotron3_5_asr

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant