Update MagpieTTS' Inference Parameter Configuration

[blisc]

What does this PR do ?

This PR separates model specific and run specific inference parameters. Model specific parameters are not stored inside a ModelInferenceParameters dataclass within the model itself. This parameters should be stored in the model yaml, currently this needs to be done manually. In particular, the latest checkpoint: https://huggingface.co/nvidia/magpie_tts_multilingual_357m has been updated to include a inference_parameters field in the top level yaml.
Run specific inference parameters, such as applying local transformer or using CFG, remain as arguments to the inference functions.
Updates inference defaults, and merges the nemo/collections/tts/modules/magpietts_inference/inference.py configuration with this new setup.

Collection: tts

Changelog

• Update MagpieTTS Inference Configurations and enables LT and prior in do_tts()

PR Type:

• New Feature
• Bugfix
• Documentation

[XuesongYang]

LGTM

[Copilot]

Pull request overview

This PR introduces an InferenceParameters dataclass to centralize inference configuration for MagpieTTS models and enables attention prior by default in the do_tts() method. The main purpose is to make attention prior active by default for improved inference quality while simplifying parameter management across the codebase.

Key changes:

• Added InferenceParameters dataclass with apply_attention_prior=True by default (previously False)
• Refactored infer_batch() to read parameters from self.inference_parameters instead of method arguments
• Simplified do_tts() API by removing explicit temperature, topk, max_decoder_steps, and cfg_scale parameters

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 6 comments.

File	Description
nemo/collections/tts/models/magpietts.py	Introduced InferenceParameters dataclass with attention prior enabled by default; refactored infer_batch() to use instance parameters; simplified do_tts() signature
nemo/collections/tts/models/magpietts_preference_optimization.py	Updated test_step() and generate_and_reward() to set inference_parameters attributes instead of passing them as arguments
nemo/collections/tts/modules/magpietts_inference/inference.py	Modified standard inference to set model's inference_parameters before calling infer_batch()

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[rfejgin]

Should use_local_transformer_for_inference be in InferenceParameters?

[blisc]

No, the intension of InferenceParameters is to save model specific hyperparameters so we can save it as part of the yaml config and subsequently, part of the nemo file. Since some parameters like estimate_alignment_from_layers are model-specific, I want to keep this new class to contain only those parameters. Options like use_cfg or use_LT are run-specific and should lie elsewhere.

[subhankar-ghosh]

Left one minor comment, otherwise LGTM given standard and long-form inference has been tested and nothing breaks.

[rfejgin]

Looks good to me (some internal discussions ongoing but not needed for this PR)

[paarthneekhara]

LGTM.

[github-actions]

[🤖]: Hi @blisc 👋,

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