# YuE2 3B isolated quality test

Prepared, non-starting evaluation of `m-a-p/YuE2-3B` with the standard
`m-a-p/YuE2-Vae` listening decoder. The source is pinned to the official
`yue2-v0.1.6` commit `9c6c4b349be978b06a9d0d958471a07a6cdeff4d`.

Preparation on Athena is complete. The model and VAE files were checked
against their published `weights_manifest.json` SHA-256 values. The Docker
image is built, but no YuE2 container has been created or started.

## Safety and isolation

- This experiment is not part of the profile controller or dashboard.
- The Compose service uses the `manual` profile, has no restart policy and
  cannot start through an ordinary `docker compose up`.
- Only the RTX 5080 is exposed to the container.
- Building and downloading do not load the model or use a GPU.
- Do not start it while another Athena GPU job is active.

## Persistent files

```text
/data/models/yue2/
├── YuE2-3B/
└── YuE2-Vae/

/data/music/yue2/
```

The initial control request is a true empty-lyrics instrumental request. No
invented `[Instrumental]` lyrics marker is used.

## Community WebUI

The `community-ui` profile runs Ladypoly's YuE2 WebUI at the pinned commit
`8fc05609bde5dcd345d7c6d57fff3da0839164c5`. Only the WebUI layer is copied
from that repository. Athena continues to use the verified official YuE2
`0.1.6` runtime and the existing model cache; the WebUI fork's bundled model
code and Windows installers are not used.

The UI is bound only to Athena's localhost on port 8014 and stores complete
takes below `/data/music/yue2`. Optional Windows-only installers for llama.cpp
and stable-diffusion.cpp are not part of the Athena setup. Manual composition,
generation, result playback, score editing and the take library work without
those optional components.

### Uploaded-audio remix (SheetSage2)

The **Remix a take** drawer also accepts WAV, FLAC, MP3, M4A, OGG, Opus and
AAC uploads. SheetSage2 transcribes the recording into an editable ABC melody
and chord plan, then YuE2 renders that structure in a newly selected style.
It does not preserve the original samples, singer or production verbatim.

SheetSage2 is kept in `/opt/yue2/.venv-sheetsage2`, while its persistent model
files live below `/data/models/yue2`. The environment intentionally reuses the
image's PyTorch 2.10/CUDA 12.8 runtime: the upstream cu126 recipe is not
Blackwell-capable. Required persistent directories are:

```text
/data/models/yue2/
├── SheetSage2/
└── MERT-v2-FullSong/
```

The local `SheetSage2/config.json` must point `base_model_name_or_path` to
`/opt/yue2/models/MERT-v2-FullSong`, allowing the complete transcription path
to run offline. Analysis and generation share the RTX 5080 and therefore run
sequentially; the UI parks YuE2 before starting SheetSage2.

The small integration patch under `community-webui-patches/` fixes the
community release's missing `refreshArt()` function on Linux and permits the
native YuE2 empty-lyrics request for true instrumentals. It deliberately does
not modify the model runtime.

```sh
docker compose --profile community-ui up -d yue2-ui
```

The original small German playground remains available as a stopped fallback
on localhost port 8016 through the `playground-fallback` profile. Do not run
both frontends concurrently because both can submit work to the same GPU.

## Manual test (only after GPU availability was checked)

From `/opt/mike-ai/yue2-3b` on Athena:

```sh
docker compose --profile manual run --rm yue2-test generate \
  --offline \
  --device cuda:0 \
  --budget 16 \
  --request /workspace/requests/instrumental-synthwave.json \
  --output /workspace/runs
```

Start with the official unquantized BF16 path. If and only if this fails from
VRAM pressure, repeat with `--quantization fp8 --offload-ar`; keep the outputs
separate because that is a different inference configuration.

YuE2 is newly released and officially specifies a 24-GB BF16 GPU. Readiness of
this image and the downloaded weights is not evidence that the 16-GB RTX 5080
run will fit or that its audio quality is acceptable.