How to run HunyuanVideo on a cloud GPU

Rent a GPU on Deploy, then open a terminal in JupyterLab or over SSH, or open the ComfyUI template. Install HunyuanVideo from the official repository, download the weights that repository documents, and run its single-GPU command or the ComfyUI files it links. Copy the videos off the instance before you destroy it. Destroying the instance deletes its disk.

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.

Memory the README publishes

The HunyuanVideo README (read 2026-11-12) gives this table for batch size 1:

Setting (height/width/frame) GPU peak memory
720px1280px129f 60GB
544px960px129f 45GB

The same section says: “The minimum GPU memory required is 60GB for 720px1280px129f and 45G for 544px960px129f.” It also says: “We recommend using a GPU with 80GB of memory for better generation quality.” And: “The model is tested on a single 80G GPU.” The FP8 section says FP8 weights save “about 10GB of GPU memory.” This page does not subtract that 10GB from the minimums above.

GET /v1/public/compute-capabilities does not list GPUs. It returns feature flags. The memory column below is vram_gb from gpu-prices.json, read 2026-11-12. The same names are on the rental inventory. Which listing you can start is on Deploy. This page does not say which listings are available.

Catalog name vram_gb Against the README
A800 80G 80 meets the 80GB recommendation
A100 80G 80 meets the 80GB recommendation
H800 80 meets the 80GB recommendation
RTX 6000D 83 meets the 80GB recommendation
RTX PRO 6000 96G 96 meets the 80GB recommendation
RTX PRO 6000 Max-Q 96 meets the 80GB recommendation
H20 96G 96 meets the 80GB recommendation
NVIDIA RTX PRO 5000 72G 72 meets the 60GB line, below the 80GB recommendation
RTX PRO 5000 72G 72 meets the 60GB line, below the 80GB recommendation
RTX 4090 48G 48 meets the 45G setting, below the 60GB line
RTX 4090D 48G 48 meets the 45G setting, below the 60GB line
A40 48 meets the 45G setting, below the 60GB line
RTX A6000 48 meets the 45G setting, below the 60GB line

Every other name in that file has vram_gb below 45, so it does not meet either published minimum. That includes A100 40G and A800 40G (40), RTX 4080 SUPER 32G, V100 32G, RTX 5090 and RTX 5090D (32), RTX 4090, RTX 4090D, RTX 3090 and A30 (24), V100 16G (16), RTX 3080 Ti (12), RTX 2080 Ti (11), RTX 3080 (10) and RTX 2080 (8).

HunyuanVideo 1.5 is a different repository. Its README says “Minimum GPU Memory: 14 GB (with model offloading enabled)” (read 2026-11-12). That line is not the table above. The AI short film guide already cites it. The install on this page is for Tencent-Hunyuan/HunyuanVideo.

Install on Linux

On Deploy, choose Raw GPU for the repository commands, or the ComfyUI template for the ComfyUI files further down. The README’s Linux section starts with conda. If conda is not on PATH, use how to set up a conda environment on a cloud GPU. This page did not check whether the Raw GPU image already has conda.

The README says it recommends CUDA 12.6 or 11.8 for the manual installation. Copied from that section:

git clone https://github.com/Tencent-Hunyuan/HunyuanVideo
cd HunyuanVideo

conda create -n HunyuanVideo python==3.10.9
conda activate HunyuanVideo

# For CUDA 11.8
pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu118
# For CUDA 12.6
pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu126

python -m pip install -r requirements.txt
python -m pip install ninja
python -m pip install git+https://github.com/Dao-AILab/flash-attention.git@v2.8.3
python -m pip install xfuser==0.4.0

The README says flash attention v2 “requires CUDA 11.8 or above.” If that pip line fails, see how to install FlashAttention on a cloud GPU. That guide is not part of the HunyuanVideo README.

The same README has a “float point exception” block. It was not run here. Option 1 prints export LD_LIBRARY_PATH=/opt/conda/lib/python3.8/site-packages/nvidia/cublas/lib/ while step 1 creates Python 3.10.9. The path is copied as written:

pip install nvidia-cublas-cu12==12.4.5.8
export LD_LIBRARY_PATH=/opt/conda/lib/python3.8/site-packages/nvidia/cublas/lib/

Option 2 in the README forces the CUDA 11.8 PyTorch build, then reinstalls requirements.txt, ninja, flash-attention v2.8.3 and xfuser==0.4.0.

The README also documents Docker images hunyuanvideo/hunyuanvideo:cuda_12 and hunyuanvideo/hunyuanvideo:cuda_11. This page does not repeat those docker run flags. They were not verified on a rented instance.

Leave download acceleration off while pip runs. The download speed guide says cloudgpu-net off before pip, then cloudgpu-net on in the shell that downloads weights. cloudgpu-net is not in the HunyuanVideo README.

Download the weights

The install block leaves the shell inside HunyuanVideo. ckpts/README.md (read 2026-11-12) prints cd HunyuanVideo and cd HunyuanVideo/ckpts because each of its snippets starts in the parent of the clone. Pasting those cd lines after the install block fails: there is no HunyuanVideo directory inside HunyuanVideo, and the preprocess script then cannot see hyvideo/ or ckpts/. The block below keeps that file’s download and preprocess commands and changes only the directory steps so one shell can run install, download, and inference in order. Those cd lines are not what the file prints. None of this was run.

python -m pip install "huggingface_hub[cli]"
huggingface-cli download tencent/HunyuanVideo --local-dir ./ckpts

cd ckpts
huggingface-cli download xtuner/llava-llama-3-8b-v1_1-transformers --local-dir ./llava-llama-3-8b-v1_1-transformers

cd ..
python hyvideo/utils/preprocess_text_encoder_tokenizer_utils.py --input_dir ckpts/llava-llama-3-8b-v1_1-transformers --output_dir ckpts/text_encoder

The file then prints one more huggingface-cli download command whose --local-dir is ./text_encoder_2. Copy that line from ckpts/README.md and run it from ckpts/, then cd ... This page does not repeat that Hub id. ./text_encoder_2 is under ckpts/. The preprocess command above runs from the repository root, after cd ... Turn cloudgpu-net off again before the next pip install, if you still have it on.

That file also has a sentence about how long the download may take. This page does not repeat timing claims. It also shows an HF-Mirror example. This page does not include it. The Hugging Face download guide uses the current hf download command name. The Hunyuan file prints huggingface-cli download. Swapping the name was not checked against this repository.

Run one GPU

The README’s single-GPU command is copied below. It also prints cd HunyuanVideo in front of that command. After the download block, cd .. has already put the shell at the repository root, so that cd is not repeated here.

python3 sample_video.py \
    --video-size 720 1280 \
    --video-length 129 \
    --infer-steps 50 \
    --prompt "A cat walks on the grass, realistic style." \
    --flow-reverse \
    --use-cpu-offload \
    --save-path ./results

--use-cpu-offload is in that command. The README’s description of the flag is “Use CPU offload for the model load to save more memory, necessary for high-res video generation.” The 60GB and 45G lines above are still the minimums the README states. This page does not claim the flag lowers those minimums.

The FP8 command in the README sets DIT_CKPT_PATH to the FP8 weight file, then adds --dit-weight, --use-fp8 and --use-cpu-offload. The weight file it links is mp_rank_00_model_states_fp8.pt under tencent/HunyuanVideo.

The README also has python3 gradio_server.py --flow-reverse. A comment in that section shows SERVER_NAME=0.0.0.0 SERVER_PORT=8081 as a manual setting, not as a default this page measured. Reach the port with SSH port forwarding.

The README has a multi-GPU torchrun --nproc_per_node=8 example. This page does not treat an 8-GPU machine as something Deploy offers.

ComfyUI

The HunyuanVideo README links a native ComfyUI example, ComfyUI Hunyuan Video (read 2026-11-12). That page does not publish a VRAM number. It names these files from Comfy-Org/HunyuanVideo_repackaged:

File Folder under ComfyUI/models
clip_l.safetensors text_encoders
llava_llama3_fp8_scaled.safetensors text_encoders
hunyuan_video_vae_bf16.safetensors vae
hunyuan_video_t2v_720p_bf16.safetensors diffusion_models

For image-to-video it also names llava_llama3_vision.safetensors in clip_vision, and either hunyuan_video_image_to_video_720p_bf16.safetensors or hunyuan_video_v2_replace_image_to_video_720p_bf16.safetensors in diffusion_models. Those names are on the example page. The shell lines below are assembled from that page. They are not printed there, and they were not run:

pip install -U "huggingface_hub"
cloudgpu-net on
hf download Comfy-Org/HunyuanVideo_repackaged \
  split_files/text_encoders/clip_l.safetensors \
  split_files/text_encoders/llava_llama3_fp8_scaled.safetensors \
  split_files/vae/hunyuan_video_vae_bf16.safetensors \
  split_files/diffusion_models/hunyuan_video_t2v_720p_bf16.safetensors \
  --local-dir ~/hunyuan-dl
cloudgpu-net off

Move each file into the folder in the table. Whether the ComfyUI template already contains the workflow was not checked. How to choose a rental GPU for ComfyUI is the card page for that template. For a model this repository already documents end to end, see how to run Wan 2.2 and which GPU for Wan.

When it fails

CUDA out of memory. The README minimums are the first check: a listing whose vram_gb is below 45 does not meet either published line. Then use how to fix CUDA out of memory on a cloud GPU.

“CUDA error: no kernel image is available” on an RTX 5090. The Hunyuan README pins torch==2.7.1 for cu118 or cu126. The RTX 5090 kernel-image guide says that card needs a PyTorch build for CUDA 12.8 or newer. This page did not run either pin on an RTX 5090. Catalog RTX 5090 and RTX 5090D are 32 GB in gpu-prices.json, which is below the 45G minimum, so those listings do not meet the README’s memory line even if the kernel loads.

Save the videos, then destroy

The single-GPU command uses --save-path ./results. Copy that folder off the instance before you destroy it (rsync or scp). Save your work before you destroy a cloud GPU is the checklist.

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.

The weights are under the Tencent Hunyuan Community License Agreement in LICENSE.txt in the HunyuanVideo repository. It is not an Apache-style licence. Read that file before commercial use. The README also contains a comparison table against hosted video products. This page does not reproduce that table.

Frequently asked questions

How much GPU memory does the HunyuanVideo README require?

The HunyuanVideo README, read 2026-11-12, says the minimum GPU memory is 60GB for 720px1280px129f and 45G for 544px960px129f, and it recommends a GPU with 80GB of memory. It also says FP8 weights save about 10GB of GPU memory. This page does not turn that sentence into a lower minimum.

Which rented GPUs meet that requirement?

Compare those figures with vram_gb in https://cloudgpu.app/gpu-prices.json, read 2026-11-12. Listings at 80 or above are A800 80G, A100 80G, H800, RTX 6000D, RTX PRO 6000 96G, RTX PRO 6000 Max-Q and H20 96G. NVIDIA RTX PRO 5000 72G and RTX PRO 5000 72G are 72, which meets the 60GB line and is below the 80GB recommendation. RTX 4090 48G, RTX 4090D 48G, A40 and RTX A6000 are 48, which meets the 45G setting and is below the 60GB line. Listings below 45 do not meet either published minimum. Which listing you can start is on https://cloudgpu.app/deploy. This page does not say which listings are available.

Is this the same model as HunyuanVideo 1.5?

No. This page follows the HunyuanVideo repository. HunyuanVideo 1.5 is a separate repository, and its README says the minimum GPU memory is 14 GB with model offloading enabled (read 2026-11-12). That 14 GB line is not the 45G or 60GB table on this page.

Where are the videos saved?

The single-GPU command in the README passes --save-path ./results. Copy that folder off the instance before you destroy it. Destroying the instance deletes its disk. The pricing page states: 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.

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).