JackKnifeAI/htdemucs-executorch

htdemucs for ExecuTorch (Vulkan)

Hybrid Transformer Demucs (htdemucs) by Alexandre Défossez, Simon Rouard and Francisco Massa at Meta AI (https://github.com/facebookresearch/demucs), exported by JackKnifeAI to ExecuTorch with the Vulkan backend for the Studio in Killa YuE (https://github.com/JackKnifeAI/killa-yue). Weights and code are MIT licensed; the original copyright notice is kept below.

• htdemucs_vulkan.pte: the model, fixed 7.8 s windows (343980 samples, 44.1 kHz stereo); the app does the STFT mask and inverse STFT. 302594560 bytes, SHA-256 7c8db21cb65c11b2befeb8e2e6538b56c6ad9d275e82d0151e1d3d08cc22f296
• libexecutorch_ffi.so: the ExecuTorch runtime for Android arm64, built from https://github.com/pytorch/executorch (BSD licence). 111533272 bytes, SHA-256 cb8391689f77d86b9f6d7866985695485d5bc7651243e6ca699764e51c175d71

On a Galaxy S25 (Adreno 830): 24.4 s per window, 3.95x real time, 1.7 GB peak memory.

MIT License, Copyright (c) Meta Platforms, Inc. and affiliates.