HTDemucs (LibTorch / CoreML Ready)

This repository contains TorchScript (.pt) exports of the Hybrid Transformer Demucs (htdemucs) model by Meta Research.

These models are optimized for C++ Inference (using LibTorch) on Apple Silicon (Metal/MPS) and CPU.

⚠️ License & Attribution

• Original Model Code: MIT License (Copyright Meta Platforms, Inc.)
• Pre-Trained Weights: CC-BY-NC 4.0 (Research constraints derived from training data).

Attribution:

Original work by Meta Research. Based on the paper "Hybrid Transformers for Music Source Separation" by Alexandre Défossez et al. Source Repository: facebookresearch/demucs

Disclaimer:

This is a format conversion only. No fine-tuning was performed. The weights are numerically identical to the original release, but packaged for C++ execution without Python dependencies.

Model Variants

Filename	Description	Chunk Size	Target Device
htdemucs_ft.pt	Fast Trace (Recommended)	8.0s	Metal (MPS) / GPU
htdemucs_6s.pt	6-Stem (Guitar/Piano)	8.0s	Metal (MPS) / GPU
htdemucs_cpu.pt	CPU Fallback	8.0s	CPU

Usage (C++)

These models are designed to be loaded directly in C++ using torch::jit::load():

#include <torch/script.h>

auto module = torch::jit::load("htdemucs_ft.pt");
module.to(torch::kMPS); // Or kCPU
module.eval();

// Input: [1, 2, Samples]
auto output = module.forward({input_tensor}).toTensor();

Integrity

SHA256 checksums are provided in SHA256SUMS.txt to verify file integrity.