Less than 24 GB of VRAM — quick reference
The requirement stays: an NVIDIA GPU with 24 GB (YuE2's validated setup) or an Apple Silicon
Mac with 32 GB+. This page is for people running on smaller cards anyway. Linux numbers are
measured (LINUX_CUDA.md); Windows numbers are estimates (§4).
1. Your card
| GPU |
Works? |
What you get |
What to set |
| 24 GB+ (RTX 3090, 4090, 5090, …) |
Yes |
Everything |
Nothing |
| 16 GB (RTX 4060 Ti 16G, 4070 Ti Super, 4080, 5060 Ti 16G, 5070 Ti, 5080, …) |
Yes |
Everything, including the longest song the app can make (6 min) at any CFG |
Nothing |
| 12 GB (RTX 3060 12G, 4070, 4070 Super, 4070 Ti, 5070, …) — Linux or Docker |
With limits |
Songs of ~5 min at CFG 1.0 · songs ≤ 3 min at CFG 1.5 · longer at CFG 1.5: §3 |
MEMORY BUDGET 12, CFG SCALE empty |
| 12 GB — Windows |
Borderline |
Songs ≤ 3 min at CFG 1.0 to start · CFG 1.5: FP8 or Docker (§3) |
MEMORY BUDGET 12, CFG SCALE empty, max_tokens 4 500 |
| 8–11 GB (RTX 4060, 3070, 3080 10G, all RTX 20 series, …) |
No |
— |
Rent a cloud GPU (RunPod, …) with the Docker image; 16 GB there is enough |
2. Where to set it
- Top right → STUDIO (the SONG view has no settings).
- Settings rail (right) → MEMORY BUDGET →
12. Repeat after every launch — it resets to 24.
The next GENERATE applies it (or click LOAD / APPLY).
- 01 // GENERATE → CFG SCALE → leave empty (= 1.0).
- Song length → ADVANCED // SAMPLING → semantic phase → max_tokens; the Estimated audio
length line updates. Shorten the lyrics to match, or the song is cut off at the limit.
| max_tokens |
2 200 (Preview preset) |
4 500 |
6 000 |
7 200 |
9 000 (default) |
| Song length |
≈ 1:30 |
3:00 |
4:00 |
4:48 |
6:00 |
3. 12 GB: what you want → what to set
Start at the top; go down one row only if that row is not what you want.
| You want |
Set |
Result |
| ~5-min song, CFG 1.0 |
MEMORY BUDGET 12, nothing else |
Works (measured, ~5 min per song on an NVIDIA L4) |
| CFG 1.5, song ≤ 3 min |
+ max_tokens 4 500, shorter lyrics |
Works (measured with a 2-min song) |
| CFG 1.5, song ≤ 4:45 |
+ max_tokens 7 200 |
Works with almost no margin; long lyrics or a long score can push it over |
| CFG 1.5, full length |
QUANTIZATION → fp8 (RTX 40 series or newer; not RTX 30) |
Works, ~4× slower (23 min instead of 5½ on an L4) |
| Windows: the Linux rows above |
Docker image under Docker Desktop + WSL2 |
Linux numbers and speed |
Every one of these changes (CFG, length, FP8, Windows vs Linux) turns the same seed into a
different song. That is normal.
4. Windows users, read this
Windows is not a supported platform. Install, launch and Cover were verified once (Windows 11,
RTX 2070 8 GB); the maintainer has not generated a song on Windows, so the Windows numbers on
this page are the Linux measurements plus the known Windows overhead. Reports are welcome.
| What you see |
Why |
What to do |
Slower than the Linux figures (~1.5×); STATUS shows backend=torch-eager (… cannot run FlashAttention …) |
Windows builds of PyTorch have no FlashAttention, so the app uses the slower decoder — it also needs ~0.7 GB more VRAM, which is why 12 GB is "borderline" |
Expected. For Linux speed and memory, run the Docker image under Docker Desktop + WSL2 |
Crash at the first generate step: USE_FLASH_ATTENTION was not enabled for build |
App older than 0.9.0 |
Update (Pinokio: Update). 0.9.0+ switches decoder by itself |
Crash at start (fcntl), or Cover fails with non-English folder names or lyrics |
App older than 0.9.0 |
Update |
A song takes hours; STATUS shows device=cpu |
PyPI's Windows torch is CPU-only |
Pinokio installs the CUDA build for you. Manual install, in the app's venv: pip install torch==2.10.0 --index-url https://download.pytorch.org/whl/cu128 --force-reinstall --no-deps |
| Out of memory on 12 GB with the §3 settings |
The desktop, browser and other apps share the card |
Close other GPU apps. If the CPU has built-in graphics, plug the monitor into the motherboard |
5. Out of memory?
The app keeps running. Change one thing, press GENERATE again: on 12 GB, go one row down in §3;
on Windows, see the last row of §4.
6. Leave these alone
| Setting |
Why |
| OFFLOAD AR WEIGHTS |
Does not lower the peak (measured) |
| ODE STEPS |
Changes render time and sound, not memory |
| DTYPE, BACKEND, VAE CORE FRAMES |
auto / torch / auto; the app picks the right path |
| MEMORY BUDGET below your card's size |
Saves nothing; runs fail sooner |
Less than 24 GB of VRAM — quick reference
The requirement stays: an NVIDIA GPU with 24 GB (YuE2's validated setup) or an Apple Silicon
Mac with 32 GB+. This page is for people running on smaller cards anyway. Linux numbers are
measured (LINUX_CUDA.md); Windows numbers are estimates (§4).
1. Your card
2. Where to set it
12. Repeat after every launch — it resets to 24.The next GENERATE applies it (or click LOAD / APPLY).
length line updates. Shorten the lyrics to match, or the song is cut off at the limit.
3. 12 GB: what you want → what to set
Start at the top; go down one row only if that row is not what you want.
Every one of these changes (CFG, length, FP8, Windows vs Linux) turns the same seed into a
different song. That is normal.
4. Windows users, read this
Windows is not a supported platform. Install, launch and Cover were verified once (Windows 11,
RTX 2070 8 GB); the maintainer has not generated a song on Windows, so the Windows numbers on
this page are the Linux measurements plus the known Windows overhead. Reports are welcome.
backend=torch-eager (… cannot run FlashAttention …)USE_FLASH_ATTENTION was not enabled for buildfcntl), or Cover fails with non-English folder names or lyricsdevice=cpupip install torch==2.10.0 --index-url https://download.pytorch.org/whl/cu128 --force-reinstall --no-deps5. Out of memory?
The app keeps running. Change one thing, press GENERATE again: on 12 GB, go one row down in §3;
on Windows, see the last row of §4.
6. Leave these alone
auto/torch/auto; the app picks the right path