How to run YuE music generation on a cloud GPU

Rent a GPU on Deploy, then open a terminal in JupyterLab or over SSH. The official repository’s default branch is YuE2. Its quick start installs the package and runs examples/generate.py, and the weights download on first use. The original inference script that takes genre.txt and lyrics.txt is on the YuE-v1 branch that the same README names. Copy the output directory off the instance before you destroy it. Destroying the instance deletes its disk.

Other music and speech installs: How to run ACE-Step 1.5 on a cloud GPU, How to run Stable Audio Open on a cloud GPU, How to run Fish Speech on a cloud GPU, and How to run GPT-SoVITS on a cloud GPU.

The default-branch LICENSE file is the Apache License, Version 2.0, January 2004 (read 2026-11-12). The README says that license covers the code, the agent skill, and the documentation. MODEL_LICENSE is titled “YuE2 model-weight license”. It says the YuE2-3B, YuE2-Vae, and YuE2-Vae-legacy checkpoint weights are licensed under CC BY-NC 4.0, and that an additional permission lets personal users, content creators, and musicians acting in an individual capacity monetize outputs. It says that permission does not authorize commercial use of the model weights by companies, and it points companies to the repository contact section. The README’s license table uses the words “Free to use” and “no fees or royalties payable to us”. Those sentences are the README’s license table. They are not a CloudGPU price and not a storage allowance.

The YuE-v1 README’s license section says the YuE model, including its weights, is released under the Apache License, Version 2.0. That sentence is about the YuE-v1 branch. This page does not treat it as the YuE2 MODEL_LICENSE.

The commands below are copied from the upstream files named in each step. They were not run on a CloudGPU instance for this page. Anything this page adds around them is marked.

What the README and the hardware doc say about memory

The YuE2 quick start says “Linux · Python 3.12 · NVIDIA GPU with BF16 support and 24 GB VRAM” (read 2026-11-12). This page does not check BF16 support for a catalog name.

The hardware doc prints a low-memory command and says stable length limits on 6/8/12 GB devices remain to be measured. It also reports a completed run on a Windows RTX 5060 Laptop with 8 GB. That card is not in the catalog below. This page does not treat that report as a catalog row.

The YuE-v1 README says that for GPUs with 24GB memory or less you should run up to 2 sessions to avoid out-of-memory errors, and that full song generation (many sessions, for example 4 or more) should use GPUs with at least 80GB memory. Its examples for that 80GB sentence are H800, A100, or multiple RTX 4090s with tensor parallel. It says the default is 2 sessions. Its execution-time sentences are not copied here.

GET /v1/public/compute-capabilities does not list GPUs. It returns feature flags. The memory column is vram_gb from gpu-prices.json, updated_at 2026-11-12T21:04:40.726Z. GET /v1/public/gpu-catalog the same day (generated_at 2026-11-12T21:04:46.496Z) listed the same names and the same vram_gb values, plus Ascend 910B at 64, which is not in gpu-prices.json and is not listed here. The catalog response also includes an availability field. This page does not copy it. Every gpu-prices.json row has gpu_count 1, so the v1 README’s “multiple RTX 4090s” example is not a row in that file. The catalog had no YuE template. Which listing you can start is on Deploy.

Rows below 24 do not meet the YuE2 quick start’s 24 GB sentence as a number: RTX 2080 (8), RTX 3080 (10), RTX 2080 Ti (11), RTX 3080 Ti (12), and V100 16G (16). Rows of 80 or higher meet the v1 README’s “at least 80GB” sentence as a number: A100 80G, A800 80G, H800, RTX 6000D (83), H20 96G, RTX PRO 6000 96G, RTX PRO 6000 Max-Q, and H20-3e 141G. The v1 README names H800 and A100 in that sentence. It does not name the other rows. Rows from 32 through 72 are neither “24GB memory or less” nor “at least 80GB”. This page does not add a third figure for them.

Catalog name vram_gb
RTX 2080 8
RTX 3080 10
RTX 2080 Ti 11
RTX 3080 Ti 12
V100 16G 16
A30 24
RTX 3090 24
RTX 4090 24
RTX 4090D 24
RTX 4080 SUPER 32G 32
RTX 5090 32
RTX 5090D 32
V100 32G 32
A100 40G 40
A800 40G 40
A40 48
RTX 4090 48G 48
RTX 4090D 48G 48
RTX A6000 48
NVIDIA RTX PRO 5000 72G 72
RTX PRO 5000 72G 72
A100 80G 80
A800 80G 80
H800 80
RTX 6000D 83
H20 96G 96
RTX PRO 6000 96G 96
RTX PRO 6000 Max-Q 96
H20-3e 141G 141

YuE2 on the default branch

On Deploy, choose Raw GPU. A clone without a branch name checks out the default branch. The quick start, copied:

git clone https://github.com/multimodal-art-projection/YuE.git
cd YuE
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install .
python examples/generate.py --output outputs/first-song

The quick start says to open outputs/first-song/audio.flac. It says the output directory also retains the score, semantic tokens, acoustic latents, generation settings, and model identities. It says model files download from Hugging Face on first use. The Python example in that section calls YuE2Pipeline.from_pretrained("m-a-p/YuE2-3B", device="cuda"). That repository id is the one the README prints.

The hardware doc says PyTorch is pinned to 2.10.0 elsewhere than macOS, and that you install the wheel before pip install .. Its NVIDIA line, which the quick start block above does not print:

python -m pip install torch==2.10.0 --index-url https://download.pytorch.org/whl/cu128

The same doc then says to run python -m pip install . and yue2 doctor. Its low-memory command, separate from the quick start:

yue2 generate --request examples/song.json --device cuda --low-vram --budget 8

The doc says that command’s stable length limits on 6/8/12 GB devices remain to be measured. This page does not treat --budget 8 as a catalog VRAM figure.

YuE-v1, with genre and lyrics files

The default-branch README says the original YuE code, documentation, and license are preserved on the YuE-v1 branch. The commits API for that branch listed 6d4f0b1f8ce6a55fb2392e959394c46e07ee334d, committed 2026-09-10T06:11:46Z. The Linux quick start there, copied. It recommends Python >=3.8:

conda create -n yue python=3.8
conda activate yue
conda install pytorch torchvision torchaudio cudatoolkit=11.8 -c pytorch -c nvidia
pip install -r <(curl -sSL https://raw.githubusercontent.com/multimodal-art-projection/YuE/YuE-v1/requirements.txt)

sudo apt update
sudo apt install git-lfs
git lfs install
git clone --branch YuE-v1 https://github.com/multimodal-art-projection/YuE.git
cd YuE/inference/
git clone https://huggingface.co/m-a-p/xcodec_mini_infer

The CoT inference command, copied. The snippet’s working directory is YuE/inference/, and --output_dir is ../output. The README says you may edit genre.txt and lyrics.txt, and that the language-model checkpoints download from Hugging Face:

cd YuE/inference/
python infer.py \
    --cuda_idx 0 \
    --stage1_model m-a-p/YuE-s1-7B-anneal-en-cot \
    --stage2_model m-a-p/YuE-s2-1B-general \
    --genre_txt ../prompt_egs/genre.txt \
    --lyrics_txt ../prompt_egs/lyrics.txt \
    --run_n_segments 2 \
    --stage2_batch_size 4 \
    --output_dir ../output \
    --max_new_tokens 3000 \
    --repetition_penalty 1.1

The same README prints dual-track and single-track in-context commands. This page does not repeat those blocks. It says the in-context checkpoint example is m-a-p/YuE-s1-7B-anneal-en-icl. Copy the rest of those commands from the README.

ComfyUI route, and no Gradio command in the quick starts

The YuE2 README says YuE2 has native nodes and links an official text-to-music workflow: https://github.com/Comfy-Org/workflow_templates/blob/main/templates/audio_yue2_text2music.json. A GET of that file on 2026-11-12 returned HTTP 200. This page does not paste the JSON. Installing ComfyUI is How to install ComfyUI and ComfyUI-Manager on a cloud GPU. The ComfyUI examples top-level directory list on 2026-11-12 had no YuE directory.

The YuE2 quick start does not print a Gradio launch. The YuE-v1 Linux quick start does not print one either. That README links third-party interfaces, including Gradio projects maintained outside the repository. This page does not install those. There is no port to forward for infer.py or examples/generate.py. If you open the ComfyUI workflow, the tunnel is the one on the ComfyUI install page, port 8188. How to use SSH port forwarding on a cloud GPU is the tunnel page.

When it fails

CUDA out of memory. The YuE-v1 README says to keep to 2 sessions on GPUs with 24GB memory or less, and to be careful raising --stage2_batch_size. The YuE2 hardware doc says the 6/8/12 GB length limits remain to be measured. If the process raises that error, use how to fix CUDA out of memory on a cloud GPU.

“CUDA error: no kernel image is available for execution on the device” on an RTX 5090. The YuE2 hardware doc’s NVIDIA line installs torch==2.10.0 from the cu128 index. The YuE-v1 conda line installs cudatoolkit=11.8. The RTX 5090 guide says that card is compute capability 12.0 (sm_120) and needs a PyTorch build for CUDA 12.8 or newer. This page did not run either install on an RTX 5090. Catalog RTX 5090 and RTX 5090D are 32. 32 is at or above the YuE2 quick start’s 24 GB sentence as a number, and it is neither of the two YuE-v1 memory sentences (24GB or less, or at least 80GB). That comparison is the memory number only.

Save the audio, then destroy

For the YuE2 quick start, copy outputs/first-song before you destroy the instance. For the YuE-v1 command above, --output_dir ../output is relative to YuE/inference/, so the directory is YuE/output. This page did not run either command. Save your work before you destroy a cloud GPU is the checklist. The terms say data on local instance storage is permanently deleted when the instance is destroyed.

The pricing page states the billing rule in these words:

Per minute, rounded up, with a 1-minute minimum. One hour is pre-authorised while an instance runs, and unused minutes go back to your balance when you destroy the instance. There is no subscription and no minimum term.

Prices are not copied here.

Frequently asked questions

How much VRAM do the YuE READMEs publish?

The YuE2 quick start, read 2026-11-12, says an NVIDIA GPU with BF16 support and 24 GB VRAM. The YuE-v1 README says GPUs with 24GB memory or less should run up to 2 sessions, and that full song generation with many sessions should use GPUs with at least 80GB memory. Its examples are H800, A100, or multiple RTX 4090s. Every gpu-prices.json row has gpu_count 1. Rows under 24 do not meet the YuE2 24 GB sentence as a number. Rows of 80 or higher meet the 80GB sentence as a number. Rows from 32 through 72 match neither YuE-v1 sentence. The hardware doc says 6/8/12 GB length limits remain to be measured. Which listing you can start is on https://cloudgpu.app/deploy.

Which license covers YuE2 and which covers YuE-v1?

The default-branch LICENSE file is the Apache License, Version 2.0. The README says that covers the code, the agent skill, and the documentation. MODEL_LICENSE says the YuE2-3B, YuE2-Vae, and YuE2-Vae-legacy weights are CC BY-NC 4.0, with an additional permission for individual creators to monetize outputs, and that commercial use of the weights by companies is not part of that permission. The README license table uses the words Free to use and no fees or royalties payable to us. Those sentences are that license table. They are not a CloudGPU price and not a storage allowance. The YuE-v1 README says the model, including its weights, is Apache License, Version 2.0. That is a different file from MODEL_LICENSE.

Where are genre.txt and lyrics.txt used?

On the YuE-v1 branch. The Linux quick start runs python infer.py with --genre_txt set to the example genre file and --lyrics_txt set to the example lyrics file, --run_n_segments 2, and --output_dir set to the output folder beside the inference directory. A clone without --branch YuE-v1 checks out the default branch, which is YuE2. The YuE2 quick start runs python examples/generate.py --output outputs/first-song and does not pass those two files.

Do the songs stay on the instance after I destroy it?

No. The YuE2 quick start writes outputs/first-song/audio.flac and keeps the score and tokens in that directory. The YuE-v1 command sets the output directory to the output folder beside the inference directory. Copy the directory that the command you ran wrote, before you destroy the instance. The terms say data on local instance storage is permanently deleted when the instance is destroyed.

Sources (read 2026-11-12, Asia/Shanghai)

Cards you can start are on the Deploy page. Prices are on https://cloudgpu.app/pricing and are not repeated here.

Last updated: 2026-11-12 (Asia/Shanghai).