Implementation:Facebookresearch Audiocraft MagnetLMModel

Knowledge Sources	Facebookresearch_Audiocraft MAGNeT
Domains	Audio_Generation, Transformer, Masked_Generation
Last Updated	2026-02-14 01:00 GMT

Overview

Concrete tool for non-autoregressive masked audio token generation using iterative parallel decoding provided by the AudioCraft library.

Description

MagnetLMModel extends the base LMModel transformer to support MAGNeT-style masked generation. Instead of autoregressive left-to-right token prediction, it uses iterative parallel decoding: tokens are initially fully masked, then progressively revealed over multiple decoding steps using a cosine masking schedule. Each codebook level is decoded independently with its own set of decoding steps, and classifier-free guidance is annealed from a maximum to minimum coefficient across steps.

Usage

Import this class when building or loading MAGNeT models for text-to-music or text-to-sound generation. It replaces the autoregressive generation of standard LMModel with faster parallel masked decoding.

Code Reference

Source Location

• Repository: Facebookresearch_Audiocraft
• File: audiocraft/models/lm_magnet.py
• Lines: 1-500

Signature

class MagnetLMModel(LMModel):
    def __init__(self, subcodes_context: int = 5, compression_model_framerate: int = 50,
                 segment_duration: int = 10, span_len: int = 3, **kwargs):
        """
        Args:
            subcodes_context: Number of cross-attention steps for sub-codebooks.
            compression_model_framerate: Frame rate of the compression model.
            segment_duration: Duration of audio segments in seconds.
            span_len: Length of atomic masking spans.
        """

Import

from audiocraft.models.lm_magnet import MagnetLMModel

I/O Contract

Inputs

Name	Type	Required	Description
sequence	torch.Tensor	Yes	Input token sequences [B, K, T]
conditions	list	Yes	List of ConditioningAttributes
condition_tensors	dict	No	Pre-computed condition tensors
stage	int	No	Current codebook stage for training

Outputs

Name	Type	Description
logits	torch.Tensor	Predicted logits [B, K, T, card]
mask	torch.Tensor	Valid token mask [B, K, T]

Usage Examples

Loading via MAGNeT Model

from audiocraft.models import MAGNeT

# MagnetLMModel is loaded internally by MAGNeT.get_pretrained
model = MAGNeT.get_pretrained('facebook/magnet-small-10secs')
model.set_generation_params(
    use_sampling=True,
    top_p=0.9,
    temperature=3.0,
    decoding_steps=[20, 10, 10, 10],
)
wav = model.generate(['dog barking in a park'])

Related Pages

• Principle:Facebookresearch_Audiocraft_Masked_Parallel_Token_Generation