Adding Conformer encoder I/O-styled Transformer encoder

[tango4j]

What does this PR do?

Follow up work after the initial TF encoder PR (#15661). Many NeMo Speech AI maintainers are asking for the new Transformer Encoder implementations to have pre-encode, positional encoding features in conformer encoder.

Aligns the ASR TransformerEncoder module with the offline ConformerEncoder module surface while preserving Transformer-specific attention parameters and behavior.

Streaming encoder and adapter implementations are not included in this PR. These features will be added later on.

Tested LibriSpeech training with Transformer + CTC (BPE). Added transformer_ctc_bpe.yaml with the default configurations.

Collection: ASR

Changelog

• Updated TransformerEncoder to inherit NeMo module/export/access mixins and expose Conformer-style input/output type metadata.
• Added Conformer-style offline encoder utilities, including input_example, forward_for_export, forward_internal, bypass_pre_encode, feat_out, positional encoding, pad mask toggling, stochastic depth, and inter-CTC tensor capture.
• Added Conformer-style pre-encoder options while preserving the Transformer-native FeatureStacking path as subsampling="feature_stacking".
• Moved FeatureStacking into the shared ASR subsampling module so it can be imported from nemo.collections.asr.parts.submodules.subsampling.
• Added self_attention_model mirroring Conformer's positional-encoding switch: "rel_pos" (default), "abs_pos", and "no_pos" (None is accepted as a YAML alias for "no_pos").
• Implemented Transformer-XL relative PE on FlexAttention — (b)+(d) bias via a score_mod closure, (c) bias folded as Q + pos_bias_u; rel-shift is shared with ConformerEncoder via RelPositionMultiHeadAttention.rel_shift.
• Added Transformer encoder tests mirroring relevant Conformer encoder test procedures for stochastic depth and bypass pre-encode behavior.
• Added self_attention_model tests, including a T != n_heads regression for pos_bias_{u,v} broadcasting.
• Updated Transformer encoder tests to use typed-module keyword arguments and validate output lengths.
• Wrapped CPU forward tests in torch.no_grad() so FlexAttention's CPU path doesn't raise under model.train().

Usage

from nemo.collections.asr.modules.transformer_encoder import TransformerEncoder

encoder = TransformerEncoder(
    feat_in=128,
    d_model=512,
    n_heads=8,
    n_layers=17,
    subsampling="feature_stacking",
    subsampling_factor=4,
    self_attention_model="rel_pos",  # one of "rel_pos" | "abs_pos" | "no_pos" (or None)
)

encoded, encoded_len = encoder(audio_signal=features, length=feature_lengths)

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

Follow up work after the initial TF encoder PR (#15661)

[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.

[stevehuang52]

Shall we set default xscaling to False, since we already know that the layernorm will zero-out the effect of xscaling?

[tango4j]

Thanks for pointing this out. Setting this with default to False.

[tango4j]

/ok to test 5398604

[github-actions]

[🤖]: Hi @tango4j 👋,

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.

[nithinraok]

Thanks Taejin!

Have you had a chance to run training with this? Does it converges similarly with positional embedding enabled and how the results compare to previous runs.

[tango4j]

Thanks Taejin!

Have you had a chance to run training with this? Does it converges similarly with positional embedding enabled and how the results compare to previous runs.

@KunalDhawan is working on using this PR part to train his MoE transformer experiments.
If time allows, I will also try Librispeech-only training with several setups to do sanity check (to test the points you mentioned)

Recently, after doing some survey, It appeared to me that convnet frontend and positional encoding can affect the performance a lot. So I think we need to test these two configurations separately (ablations).

[tango4j]

@nithinraok
I have tested relative positional encoding with training job. Compared with FastConformer CTC vs Transformer CTC.
Having rel_pos in the Transformer encoder affects the performance a lot (10~20% more error). Need to set this as default.

Also figured that Filterbank Stacking feature is equally good as dw_striding (3 level convnet frontend). Better switching to Filterbank stacking to make this model low precision friendly.

@ipmedenn @KunalDhawan @stevehuang52
Now, transformer is fully equipped with relative positional encoding.

[tango4j]

/ok to test 6725930

[tango4j]

/ok to test 56423ba

[tango4j]

/ok to test 839fd85

[tango4j]

/ok to test 3b3ec63

[github-actions]

[🤖]: Hi @tango4j 👋,

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.