feat(speechlm2): add Multi-Token Prediction (MTP) training support to SALMAutomodel

[Slyne]

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An outdated branch is never blocking the merge of a PR.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

Add Multi-Token Prediction (MTP) auxiliary training support to SALMAutomodel, enabling the model to predict multiple future tokens per step using an auxiliary head attached to a NemotronV3 base checkpoint that ships without one.

Collection: speechlm2

Changelog

• Add _mtp_enabled property to SALMAutomodel for a safe, pre-configure_model-safe check of whether the MTP head is attached
• Add _build_and_attach_mtp_head() to construct the NemotronV3 MTP head via Automodel factories and attach it as self.llm.mtp after the LLM is loaded (the base checkpoint has no MTP head and the normal from_pretrained injection path does not work for this architecture)
• FSDP2-shard the MTP head over the DP mesh in configure_model, since it is attached after the LLM's own FSDP wrapping
• Compute MTP auxiliary loss in the training step using calculate_mtp_loss from nemo_automodel, normalized by the same - - global token count as the main loss and scaled by dp_size to cancel FSDP gradient averaging
• Guard MTP against the packed-sequences (THD) path, which is not yet supported
• MTP is opt-in via mtp.enabled: true in the model config; relevant fields: loss_scaling_factor, num_nextn_predict_layers, hybrid_override_pattern, use_repeated_layer

Usage

Add the below to the config file to enable MTP heads

# In your SALM Hydra config (e.g. conf/salm_automodel.yaml):
model:
  mtp:
    enabled: true
    num_nextn_predict_layers: 1
    loss_scaling_factor: 0.1
    hybrid_override_pattern: "*"
    use_repeated_layer: false

The mtp_loss scalar is logged on every training step (prog_bar=True) so you can monitor it in your experiment tracker alongside the main loss.

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The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.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

This PR depends on the changes in the below PR:

• Related to feat(mtp): enable MTP to accept pre-fused input embeddings for multimodal models Automodel#2510

[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.

[pzelasko]

we should guard this with an extra condition in case we finetune a model that does have MTP already in base checkpoint (e.g. Nemotron Super)