How to run FramePack on a cloud GPU

Rent a GPU on Deploy in the RTX 30, 40 or 50 series the FramePack README names, then open a terminal in JupyterLab or over SSH. Install from the official repository, let the program download its weights, and run python demo_gradio.py. Copy ./outputs 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, except the clone line, which the Linux section does not print. They were not run on a CloudGPU instance for this page.

Memory and which catalog GPUs match

The FramePack README (read 2026-11-12) says:

  • “Nvidia GPU in RTX 30XX, 40XX, 50XX series that supports fp16 and bf16. The GTX 10XX/20XX are not tested.”
  • “At least 6GB GPU memory.”
  • “To generate 1-minute video (60 seconds) at 30fps (1800 frames) using 13B model, the minimal required GPU memory is 6GB.”

The same section gives the author’s desktop speed: “on my RTX 4090 desktop it generates at a speed of 2.5 seconds/frame (unoptimized) or 1.5 seconds/frame (teacache).” That is the README author’s measurement, not a measurement on CloudGPU. This page does not extend it to other cards.

GET /v1/public/compute-capabilities does not list GPUs. The memory column is vram_gb from gpu-prices.json, read 2026-11-12. These catalog names are RTX 30, 40 or 50 series, and each vram_gb is above 6:

Catalog name vram_gb
RTX 3080 10
RTX 3080 Ti 12
RTX 3090 24
RTX 4080 SUPER 32G 32
RTX 4090 24
RTX 4090D 24
RTX 4090 48G 48
RTX 4090D 48G 48
RTX 5090 32
RTX 5090D 32

RTX 2080 (8) and RTX 2080 Ti (11) are 20-series names. The README says the 10 and 20 series are not tested, so this page does not say they run FramePack. The other names in that catalog file (V100, A100, A800, A30, A40, RTX A6000, RTX 6000D, the RTX PRO listings, H20, H800) are not in the README’s RTX 30/40/50 sentence. This page does not claim they run FramePack. Which of the ten names above you can start is on Deploy. This page does not say which listings are available.

Install on Linux

On Deploy, choose Raw GPU. The README’s Linux section does not print a git clone line. Clone the repository that README belongs to, then run the two pip lines and the GUI command from that directory. The clone is the extra line:

git clone https://github.com/lllyasviel/FramePack
cd FramePack

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
pip install -r requirements.txt

python demo_gradio.py

The README says the program supports PyTorch attention, xformers, flash-attn and sage-attention, and that it uses PyTorch attention unless you install another kernel. Its sage-attention example is pip install sageattention==1.0.6. The next sentence says you are “highly recommended to first try without sage-attention since it will influence results.” This page did not install it.

The README says “the models will be downloaded automatically” and “You will download more than 30GB from HuggingFace.” It does not name Hub ids. demo_gradio.py in that repository (read 2026-11-12) loads:

  • hunyuanvideo-community/HunyuanVideo for the text encoders, tokenizers and VAE
  • lllyasviel/flux_redux_bfl for the image encoder
  • lllyasviel/FramePackI2V_HY for the transformer

Those ids are from that file, not from the README. The download happens on first launch. Before that launch, turn on download acceleration in the same shell with cloudgpu-net on, then turn it off afterwards. That command is described in the download speed guide. It is not in the FramePack README. How to download a Hugging Face model covers hf download if you want to fetch a Hub repo yourself. Doing that ahead of the GUI was not checked.

demo_gradio.py accepts --share, --server and --port. In that file --server defaults to 0.0.0.0 and --port has no default. This page does not state a port number. Use the port the process prints with SSH port forwarding.

The README’s Windows one-click package is not the path on this Linux instance.

No ComfyUI route in the official README

The FramePack README documents python demo_gradio.py and does not document a ComfyUI install. This page does not add one. If you are choosing a card for the ComfyUI template itself, see how to choose a rental GPU for ComfyUI. For a video model whose own docs do include ComfyUI, see how to run Wan 2.2 and which GPU for Wan.

When it fails

CUDA out of memory. The GUI in demo_gradio.py has a slider labeled “GPU Inference Preserved Memory (GB) (larger means slower)” with minimum 6. Its info text says: “Set this number to a larger value if you encounter OOM. Larger value causes slower speed.” That is the upstream control. How to fix CUDA out of memory on a cloud GPU is the general PyTorch page. This page did not measure either one.

“CUDA error: no kernel image is available” on an RTX 5090. The README includes the 50 series, and the Linux install pins the cu126 PyTorch index. Catalog RTX 5090 and RTX 5090D are 32 GB, which is above the 6GB minimum. The RTX 5090 kernel-image guide says that card needs a PyTorch build for CUDA 12.8 or newer. This page did not run the cu126 line on an RTX 5090. If that error appears, follow that guide.

Save the videos, then destroy

demo_gradio.py sets outputs_folder = './outputs/'. Copy that folder off the instance before you destroy it (rsync or scp). The GUI also shows finished frames. 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 FramePack repository LICENSE file is Apache 2.0. The weights the script loads are separate Hub repos. Read the licence on each of those model pages before commercial use. This page does not claim those weights are Apache 2.0.

Frequently asked questions

How much VRAM does FramePack need?

The FramePack README says at least 6GB of GPU memory, including for a 1-minute video (60 seconds) at 30fps (1800 frames) with the 13B model (read 2026-11-12). It asks for an Nvidia GPU in the RTX 30, 40 or 50 series that supports fp16 and bf16, and it says the GTX 10 and 20 series are not tested.

Which rented GPUs match that?

From https://cloudgpu.app/gpu-prices.json, read 2026-11-12, the RTX 30, 40 and 50 series names are RTX 3080 (10), RTX 3080 Ti (12), RTX 3090 (24), RTX 4080 SUPER 32G (32), RTX 4090 (24), RTX 4090D (24), RTX 4090 48G (48), RTX 4090D 48G (48), RTX 5090 (32) and RTX 5090D (32). Each of those vram_gb values is above 6. RTX 2080 (8) and RTX 2080 Ti (11) are 20-series cards, which the README says are not tested. This page does not claim the other catalog names run FramePack. Which listing you can start is on https://cloudgpu.app/deploy. This page does not say which listings are available.

Does FramePack have an official ComfyUI install?

The FramePack README documents its own Gradio program, python demo_gradio.py, and does not document a ComfyUI install. This page does not add one.

Where are the videos saved?

demo_gradio.py in the official repository sets the output folder to ./outputs/ (read 2026-11-12). The README says the models download automatically and that you will download more than 30GB from Hugging Face. Copy ./outputs off the instance before you destroy it. Destroying the instance deletes its disk.

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).