Installation Guide For nVidia 50 Series Card

[alloystorm]

Since this is not mentioned in the readme and 50 series users will run into issues if simply following the "One-CLick" package. I'm listing my installation steps below to help others.

This is assuming you have the required python version installed, the readme suggest 3.10 and my 3.11 works fine too. Use python --version to check your versions.

1. Create and activate a virtual environment

python -m venv framepack_venv
framepack_venv\Scripts\activate.bat

2. Clone FramePack

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

3. Install requirements

cd FramePack
pip install -r requirements.txt

4. Install nightly pytorch with cu128

pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

5. Run FramePack

python demo_gradio.py

[anderson-ads-dev]

Great guide! Worked like a charm on my 50-series GPU.

Just a small tip for anyone running into issues — if you're using Python 3.13, you might hit compatibility problems with some libraries like sentencepiece or torch.

Fix: Install Python 3.12 or 3.11 instead. You don’t need to remove 3.13 — Python supports side-by-side versions.

During installation:
Uncheck "Add to PATH" (if 3.13 is already there)

Then just create a virtual environment using the right version:

py -3.12 -m venv framepack_venv
framepack_venv\Scripts\activate

This uses the compatible Python version without affecting your system setup.

Hope that helps

[EmmJayAye]

Can you also provide a guide for 30 series GPUs please?

My installation crashes when "Loading checkpoint shards" reaches 67%
I have a 6GB 3060. Technically it should work on this card.

[alloystorm]

30 series should be exactly the same except that you don't need step 4.

You can try to reinstall the stable version of pytorch and see if that helps. Just go to pytorch webstie and select your OS and you can see the command below.

[hoge46773]

In my environment, even when using the method described in this thread or after upgrading PyTorch from the One-Click Package (CUDA 12.6 + PyTorch 2.6) to PyTorch 2.7 or 2.8, my PC always crashes with a blue screen when I try to launch it.

I've tried both versions of CUDA (12.8 and 12.8.1) and both PyTorch 2.7 and 2.8, but the result is always the same.

I've given up for now, but I wanted to report that some users with RTX 5000 series GPUs may experience similar issues.

Environment:

• Windows 10
• Ryzen 7 3700X
• 32GB RAM
• ASUS PRIME GeForce RTX 5070 Ti 16GB GDDR7 OC Edition
• GPU Driver 576.02

Tried versions:

• CUDA 12.8 & 12.8.1
• cuDNN v8.9.7
• PyTorch 2.7 & 2.8

[Compositaes]

If the Blue Screen mentioned about Memory Management Error, You may try to increase the pagefile manually to (over 80GB or even more). Hope it may help you. However, I am stucking on last stage after sampling with CUDA out of memory error(5060Ti 8Gb Memory). Perhaps, 50xx series is relatively too new in compatibility.

[hoge46773]

Previously I was getting a blue-screen at the “Downloading shards:” step, but by allocating at least 51,200 MB of virtual memory, it now starts up without crashing. If you don’t have enough free space on your C: drive, switch your dump type from “Complete memory dump” to “Small memory dump” or “Kernel memory dump,” then set up pagefiles on another drive so that the total virtual memory is at least 51,200 MB—this should allow it to launch.

Below is what I did:
To support xformers later on, I upgraded my system Python to Python 3.10.11.
Then I installed FramePack.

python -m venv framepack_venv
framepack_venv\Scripts\activate.bat

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

cd FramePack
pip install -r requirements.txt

python.exe -m pip install --upgrade pip

pip3 uninstall torch torchvision torchaudio -y

pip3 install torch==2.7.0+cu128 torchvision==0.22.0+cu128 torchaudio==2.7.0+cu128 --index-url https://download.pytorch.org/whl/cu128

pip3 install xformers==0.0.30

pip3 install flash-attn sageattention

python demo_gradio.py

As of 2025-05-03, this appears to be the correct combination. As long as you keep each file at the matching version, you should even be able to update from the One-Click Package.

I installed into D:\FramePack, so to simplify future launches I created a run.bat with the following contents.

cd /d D:\FramePack
python demo_gradio.py --server 127.0.0.1 --inbrowser

When I first tried generating, I only got color-noise videos.
It turned out the model data was corrupted, so I deleted the hf_download folder, restarted to re-download the model files, and then it began producing normal output.

以前は「Downloading shards:」の所でブルースクリーンになっていましたが、仮想メモリを51200以上確保すればブルースクリーンにならずに起動できました。
Cドライブの空き容量が少ない人は完全メモリダンプから小さいメモリダンプ、もしくはカーネルメモリダンプに変えた上で、他のドライブでも仮想メモリを設定し、合計で51200以上になるようにすれば起動すると思います。

行ったことの記録を残しておきます。
まず後々使うxformersのために、システムのpythonをpython3.10.11にアップグレード
それからframepackをインストール

略

上記が250503時点での正しい組み合わせのようです。
各ファイルのバージョンの組み合わせさえ守ればOne-Click Packageからでもアップデートできると思われます。

私の場合はD:\FramePackにインストールしたので、次回以降の起動を楽にするために以下の内容でrun.batを作りました。

略

とりあえず生成を試してみた所、カラーノイズ動画しか生成されませんでした。
原因はモデルデータの破損ということだったので、一旦hf_downloadフォルダを削除してから起動して、モデルデータをダウンロードし直した所、正常に出力されるようになりました。

[hoge46773]

I was planning to make the One-Click Package (CUDA 12.6 + PyTorch 2.6) work as well, but in the middle of the process, I realized I had made a mistake.
One-Click Package (CUDA 12.6 + Pytorch 2.6) の方も使えるようにしようと思って作業途中に自分のミスに気が付きました。

So I redid everything as follows for the One-Click Package (CUDA 12.6 + PyTorch 2.6).
One-Click Package (CUDA 12.6 + Pytorch 2.6) のために以下のようにやり直しました。

In this example, the directory is D:\framepack_cu126_torch26.
例としてD:\framepack_cu126_torch26の場合

"D:\framepack_cu126_torch26\system\python\python.exe" -m pip install ^
torch==2.7.0+cu128 ^
torchvision==0.22.0+cu128 ^
torchaudio==2.7.0+cu128 ^
--index-url https://download.pytorch.org/whl/cu128

"D:\framepack_cu126_torch26\system\python\python.exe" -m pip install ^
xformers==0.0.30 ^
--no-deps ^
--upgrade ^
--target "D:\framepack_cu126_torch26\system\python\Lib\site-packages"

"D:\framepack_cu126_torch26\system\python\python.exe" -m pip install ^
sageattention==1.0.6 ^
--no-deps ^
--upgrade ^
--target "D:\framepack_cu126_torch26\system\python\Lib\site-packages"

"D:\framepack_cu126_torch26\system\python\python.exe" -m pip install -U "triton-windows<3.4"

Flash Attention doesn’t have a wheel distribution yet, so you have to build it from source, which is such a hassle that I gave up.
Flash Attentionはホイール版がまだないのでソースからビルドする必要があり、面倒なので私は辞めました。

[Grey3016]

You only need one attention, install sage (after triton)

[RaptorJesus675]

this has worked for me! thanks! but how can i installed sage to this? I' haven't had any luck on that so far.

[NftGamer666]

With python 3.10.11 installed, this worked for me on my 5060ti 16GB:

Create and activate a virtual environment
python -m venv framepack_venv
framepack_venv\Scripts\activate.bat

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

Install requirements
cd FramePack
pip install -r requirements.txt
pip install --upgrade pip

Install pytorch with cu128, Transformers, Triton and Sage
pip3 install torch==2.7.0+cu128 torchvision==0.22.0+cu128 torchaudio==2.7.0+cu128 --index-url https://download.pytorch.org/whl/cu128
pip3 install xformers==0.0.30
pip3 install -U triton-windows
pip3 install sageattention

Run FramePack
python demo_gradio.py

[xedious]

Worked perfectly for my 5070ti. Thanks.

[AssassinWarlord]

when installing sageattention i get an error:

(framepack_venv) C:\KI\FramePack50\FramePack>pip3 install sageattention
Collecting sageattention
Using cached sageattention-2.2.0.tar.gz (67 kB)
Installing build dependencies ... done
Getting requirements to build wheel ... error
error: subprocess-exited-with-error

× Getting requirements to build wheel did not run successfully.
│ exit code: 1
╰─> [18 lines of output]
Traceback (most recent call last):
File "C:\KI\FramePack50\framepack_venv\Lib\site-packages\pip_vendor\pyproject_hooks_in_process_in_process.py", line 389, in
main()
File "C:\KI\FramePack50\framepack_venv\Lib\site-packages\pip_vendor\pyproject_hooks_in_process_in_process.py", line 373, in main
json_out["return_val"] = hook(**hook_input["kwargs"])
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\KI\FramePack50\framepack_venv\Lib\site-packages\pip_vendor\pyproject_hooks_in_process_in_process.py", line 143, in get_requires_for_build_wheel
return hook(config_settings)
^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Dave\AppData\Local\Temp\pip-build-env-faldc77c\overlay\Lib\site-packages\setuptools\build_meta.py", line 332, in get_requires_for_build_wheel
return self._get_build_requires(config_settings, requirements=[])
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Dave\AppData\Local\Temp\pip-build-env-faldc77c\overlay\Lib\site-packages\setuptools\build_meta.py", line 302, in _get_build_requires
self.run_setup()
File "C:\Users\Dave\AppData\Local\Temp\pip-build-env-faldc77c\overlay\Lib\site-packages\setuptools\build_meta.py", line 318, in run_setup
exec(code, locals())
File "", line 36, in
ModuleNotFoundError: No module named 'torch'
[end of output]

note: This error originates from a subprocess, and is likely not a problem with pip.
ERROR: Failed to build 'sageattention' when getting requirements to build wheel

[Chiyaoyaoi]

@AssassinWarlord
try
pip install sageattention --prefer-binary

[AssassinWarlord]

@AssassinWarlord try pip install sageattention --prefer-binary

Works fine, thanks :) is it also possible to install Flash Attention? dont know which version i must install here

[Miguelox081]

I got it to load, but when I try the sanity check or other image, I click on start generation, it goes to the "Start Sampling" stage, then it gives me an error of "connection erroed out"
Any Ideas?
Edit:
Actually I saw this error message in cmd:
OutOfMemoryError: CUDA out of memory. Tried to allocate 72.00 MiB. GPU 0 has a total capacity of 15.92 GiB of which 9.33 GiB is free. Of the allocated memory 5.15 GiB is allocated by PyTorch, and 34.55 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)

I have all drives with more than 60 gb free.

[NftGamer666]

Try by changing this setting "GPU Inference Preserved Memory (GB) (larger means slower)" to a larger value.

[Giribot]

Hello!

This installation is a bit of a rock 'n' roll affair.
Couldn't we create a script to run the first time that examines the PC and its components (RTX, CPU, etc.) and installs the correct configuration?
I did everything manually on mine, correcting my error messages one by one... (CU128, etc.)

Thanks!

[screwballl]

I've already been using Fooocus with CUDA 12.8 on Linux Mint 22.2 (Ryzen 9600X, 32GB RAM, 16GB RTX5060Ti), and Python 3.12. I have prior commands to setup/fix whenever driver updates or kernel updates break the Fooocus expected standard, which I applied here. I have had prior issues with cu121-cu126, which is why i found 128 works best in my experience.

Downloaded and unpacked to fast NVME drive:
cd /mnt/760680ED0680B023/AI/framepack_cu126_torch26
python3 -m venv fp_env
source fp_env/bin/activate
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
pip install --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
pip install -r webui/requirements.txt
python3 webui/demo_gradio.py --server 127.0.0.1 --inbrowser

[Reborn315]

Layman here. alloystorm, i copy your code into windows command prompt and none of them work except for creating/activating the venv. can you help me understand why? i dont have coding knowledge but i really want this to work.

[Compositaes]

It may due to your system didn't PATH the python directory. You may download the python from https://www.python.org/downloads/ . During installation , check the box which mentioned about 'PATH the directory'. After that , reboot the system and Try again. Hope It may help you.
*simply you may type 'PATH C:\your python directory' in Command prompt before you try those pip command.

[buddhaya7]

Tks your explain, but I met a trouble on your step 5(Run FramePack):

D:\test\FramePack>python demo_gradio.py
Traceback (most recent call last):
File "D:\test\FramePack\demo_gradio.py", line 7, in
import gradio as gr
ModuleNotFoundError: No module named 'gradio'

D:\test\FramePack>

[NathaliaLietuvaite]

Hi,

The Solution
Open your terminal (the same one you are using, at the D:\test\FramePack> prompt).

Run the following command to install gradio:

Bash
pip install gradio

After the installation successfully completes, try running your script again:

Bash
python demo_gradio.py
Pro-Tip (for future reference): If you ever run into a similar issue and the simple pip install command doesn't seem to work, a more robust command is:

Bash
python -m pip install gradio

This explicitly tells your system to use the pip module associated with the python executable you are running, which avoids confusion if you
have multiple versions of Python installed.

[buddhaya7]

tks your reply, however i feel sad for RTX 50 series. of course it not mean your solution is invalid.
actually, i try all steps again. in step 3: pip install -r requirements.txt
i got the error message as below, but my previous test lab has not appear it, so inconceivable...

D:\test\FramePack>pip install -r requirements.txt
Collecting accelerate==1.6.0 (from -r requirements.txt (line 1))
Using cached accelerate-1.6.0-py3-none-any.whl.metadata (19 kB)
Collecting diffusers==0.33.1 (from -r requirements.txt (line 2))
Using cached diffusers-0.33.1-py3-none-any.whl.metadata (19 kB)
Collecting transformers==4.46.2 (from -r requirements.txt (line 3))
Using cached transformers-4.46.2-py3-none-any.whl.metadata (44 kB)
Collecting gradio==5.23.0 (from -r requirements.txt (line 4))
Using cached gradio-5.23.0-py3-none-any.whl.metadata (16 kB)
Collecting sentencepiece==0.2.0 (from -r requirements.txt (line 5))
Using cached sentencepiece-0.2.0.tar.gz (2.6 MB)
Installing build dependencies ... done
Getting requirements to build wheel ... error
error: subprocess-exited-with-error

× Getting requirements to build wheel did not run successfully.
│ exit code: 1
╰─> [48 lines of output]
Traceback (most recent call last):
File "C:\Users\chin\AppData\Local\Programs\Python\Python313\Lib\site-packages\pip_vendor\pyproject_hooks_in_process_in_process.py", line 353, in
main()
~~~~^^
File "C:\Users\chin\AppData\Local\Programs\Python\Python313\Lib\site-packages\pip_vendor\pyproject_hooks_in_process_in_process.py", line 335, in main
json_out['return_val'] = hook(**hook_input['kwargs'])
~~~~^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\chin\AppData\Local\Programs\Python\Python313\Lib\site-packages\pip_vendor\pyproject_hooks_in_process_in_process.py", line 118, in get_requires_for_build_wheel
return hook(config_settings)
File "C:\Users\chin\AppData\Local\Temp\pip-build-env-o6rrinaa\overlay\Lib\site-packages\setuptools\build_meta.py", line 331, in get_requires_for_build_wheel
return self._get_build_requires(config_settings, requirements=[])
~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\chin\AppData\Local\Temp\pip-build-env-o6rrinaa\overlay\Lib\site-packages\setuptools\build_meta.py", line 301, in _get_build_requires
self.run_setup()
~~~~~~~~~~~~~~^^
File "C:\Users\chin\AppData\Local\Temp\pip-build-env-o6rrinaa\overlay\Lib\site-packages\setuptools\build_meta.py", line 512, in run_setup
super().run_setup(setup_script=setup_script)
~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\chin\AppData\Local\Temp\pip-build-env-o6rrinaa\overlay\Lib\site-packages\setuptools\build_meta.py", line 317, in run_setup
exec(code, locals())
~~~~^^^^^^^^^^^^^^^^
File "", line 128, in
File "C:\Users\chin\AppData\Local\Programs\Python\Python313\Lib\subprocess.py", line 414, in check_call
retcode = call(*popenargs, **kwargs)
File "C:\Users\chin\AppData\Local\Programs\Python\Python313\Lib\subprocess.py", line 395, in call
with Popen(*popenargs, **kwargs) as p:
~~~~~^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\chin\AppData\Local\Programs\Python\Python313\Lib\subprocess.py", line 1036, in init
self._execute_child(args, executable, preexec_fn, close_fds,
~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
pass_fds, cwd, env,
^^^^^^^^^^^^^^^^^^^
...<5 lines>...
gid, gids, uid, umask,
^^^^^^^^^^^^^^^^^^^^^^
start_new_session, process_group)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\chin\AppData\Local\Programs\Python\Python313\Lib\subprocess.py", line 1548, in _execute_child
hp, ht, pid, tid = _winapi.CreateProcess(executable, args,
~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^
# no special security
^^^^^^^^^^^^^^^^^^^^^
...<4 lines>...
cwd,
^^^^
startupinfo)
^^^^^^^^^^^^
FileNotFoundError: [WinError 2] 系統找不到指定的檔案。
[end of output]

note: This error originates from a subprocess, and is likely not a problem with pip.

[notice] A new release of pip is available: 24.2 -> 25.3
[notice] To update, run: python.exe -m pip install --upgrade pip
error: subprocess-exited-with-error

× Getting requirements to build wheel did not run successfully.
│ exit code: 1
╰─> See above for output.

note: This error originates from a subprocess, and is likely not a problem with pip.

D:\test\FramePack>

[NathaliaLietuvaite]

Hi,

Here is the step-by-step guide for a Solution.

This is the most stable solution. The error happens because the package sentencepiece==0.2.0 does not have a pre-built version ("wheel") for the brand-new Python 3.13. By using an older, more established version like Python 3.11, pip can download a ready-made package and won't need to compile it from source code.

---

Solution: Change Python Version (The Safest Way)

Step 1: Install Python 3.11

1. Go to the official Python website: https://www.python.org/downloads/
2. Find and download an installer for Python 3.11 (e.g., Python 3.11.x).
3. Run the installer.
4. IMPORTANT: On the first screen of the installer, make sure to check the box at the bottom that says "Add python.exe to PATH".
5. Complete the installation.

Step 2: Re-create the Virtual Environment

1. Open a new Command Prompt (CMD) or PowerShell window (to ensure it recognizes the new Python version).
2. Navigate to your project folder:

   cd D:\test\FramePack

3. (Optional, but recommended) If you already have a folder named venv in this directory, delete it.
4. Create a new virtual environment, explicitly telling it to use Python 3.11. The py launcher will find it:

   py -3.11 -m venv venv

   This command creates a new folder named venv which contains this specific Python version.

Step 3: Activate and Install

1. Activate the new virtual environment:

   .\venv\Scripts\activate

   Your command prompt should now change to show (venv) at the beginning, like this: (venv) D:\test\FramePack>

2. With the environment active, try installing the requirements again:

   pip install -r requirements.txt

This time, pip should be able to find and install the pre-built sentencepiece package for Python 3.11, and the installation should complete without the C++ build error.

Hope it helps!

[marcolucifora-crypto]

I have installed python 3.11 and all other modules.
When I launch i get the message: "Flash Attn is not installed"
Also after launching web interface after generating 2 seconds of video everything crashes cause it says "not enough memory" !
Consider that my system has 16gb of ram, rtx 5060, and a core i7 14650HX
Any thoughts how to solve this issue?

[Grey3016]

I have installed python 3.11 and all other modules.

When I launch i get the message: "Flash Attn is not installed"

Also after launching web interface after generating 2 seconds of video everything crashes cause it says "not enough memory" !

Consider that my system has 16gb of ram, rtx 5060, and a core i7 14650HX

Any thoughts how to solve this issue?

I would try increasing the paging file to 60-80gb