Implementation:Facebookresearch Audiocraft GeneralUtils

Knowledge Sources	Facebookresearch_Audiocraft
Domains	Utilities, Training
Last Updated	2026-02-14 01:00 GMT

Overview

Concrete tool for widely-used helper functions including tensor operations, sampling strategies, distributed training utilities, and hashing provided by the AudioCraft library.

Description

The utils module provides general-purpose utility functions used across AudioCraft. Key functions include sample_top_k and sample_top_p for nucleus/top-k sampling during generation, get_pool_executor for parallel processing, length_to_mask for creating attention masks from sequence lengths, hash_trick for feature hashing, and various tensor manipulation helpers. It also includes flashy_audiocraft_export for exporting training artifacts.

Usage

Import specific functions from this module when you need sampling utilities for token generation, mask creation for variable-length sequences, or parallel execution helpers.

Code Reference

Source Location

• Repository: Facebookresearch_Audiocraft
• File: audiocraft/utils/utils.py
• Lines: 1-326

Signature

def sample_top_k(probs: torch.Tensor, k: int) -> torch.Tensor:
    """Sample from the top-k most probable tokens."""

def sample_top_p(probs: torch.Tensor, p: float) -> torch.Tensor:
    """Sample using nucleus (top-p) sampling."""

def length_to_mask(lengths: torch.Tensor, max_len: tp.Optional[int] = None) -> torch.Tensor:
    """Create a boolean mask from sequence lengths."""

def hash_trick(module_name: str, seed: int, max_val: int) -> int:
    """Feature hashing for weight sharing across conditioners."""

def get_pool_executor(num_workers: int, mp_context=None):
    """Get a ProcessPoolExecutor or DummyPoolExecutor."""

Import

from audiocraft.utils.utils import sample_top_k, sample_top_p, length_to_mask, hash_trick

I/O Contract

Inputs

Name	Type	Required	Description
probs	torch.Tensor	Yes	Probability distribution [B, card] (for sampling functions)
k	int	Yes	Top-k parameter
p	float	Yes	Top-p (nucleus) threshold
lengths	torch.Tensor	Yes	Sequence lengths [B] (for mask creation)

Outputs

Name	Type	Description
sampled_token	torch.Tensor	Sampled token indices [B, 1]
mask	torch.Tensor	Boolean mask [B, T]

Usage Examples

from audiocraft.utils.utils import sample_top_k, sample_top_p, length_to_mask

import torch

# Top-k sampling
probs = torch.softmax(logits, dim=-1)  # [B, vocab_size]
tokens = sample_top_k(probs, k=250)

# Top-p sampling
tokens = sample_top_p(probs, p=0.9)

# Create attention mask from lengths
lengths = torch.tensor([100, 80, 120])
mask = length_to_mask(lengths, max_len=120)  # [3, 120]

Related Pages

• Principle:Facebookresearch_Audiocraft_Token_Sampling_Strategies