Skip to content

Latest commit

 

History

History
42 lines (30 loc) · 1.9 KB

models.md

File metadata and controls

42 lines (30 loc) · 1.9 KB

Model artefacts

This is what the hosting repository's README said before the source moved into it. It describes the models-v1 tag, which is superseded: the shipping artefacts are safetensors under models-v2, and their digests are pinned in crates/stemd-server/src/models/preset.rs.

It is kept because the provenance and licence position below apply to every release, not only to the tag it was written for.


Hosting only. This repository carries the traced models that stemd downloads on first run. The source lives elsewhere.

models-v1

Two TorchScript artefacts traced for Apple MPS, each with its manifest:

file preset size sha256
htdemucs_mps.pt Quality 161 MB 9a5d56dac50cc58258df8576d8fa76f61d8943fbb117c43f55f06ce96c609d65
htdemucs_mps.json 290 B f26265bc876887eac6934cc347a9a90c8da0b6b89d96199410180fed8c0564cc
hdemucs_mmi_mps.pt Speed 320 MB dfed3230fb735772502bb4f9453810b84bf574a7d03ed6dc823b28b4f6327975
hdemucs_mmi_mps.json 296 B 00b131cc07b352f977c005485b991502d13f8c05a44813c3b3fe6169e2a18688

stemd pins these digests and verifies them before putting a download in place, so replacing an asset here will make existing installs fail loudly rather than quietly load something else. Publish a new tag instead of editing this one.

The artefacts are traced for MPS and will not load on CPU.

Provenance

Traced from facebookresearch/demucs pretrained models: htdemucs (v4) and hdemucs_mmi.

The demucs code is MIT. The pretrained weights are distributed separately from that repository and carry no explicit licence: the question was asked in demucs#327 in 2022, went unanswered, and the repository is now archived. These artefacts are derived works of those weights and are published on that understanding.