sm_120 RTX 5080 CUDA kernel crash with hunyuan.py - "no kernel image is available"

[PranThow]

Summary

When running FramePack on an RTX 5080 (Blackwell architecture, CUDA capability sm_120), the pipeline crashes with:

RuntimeError: CUDA error: no kernel image is available for execution on the device

This happens inside hunyuan.py, specifically on this line:

llama_attention_length = int(llama_attention_mask.sum())

Additionally, PyTorch emits the following warning:

UserWarning: NVIDIA GeForce RTX 5080 with CUDA capability sm_120 is not compatible with the current PyTorch installation.

Environment

• GPU: NVIDIA GeForce RTX 5080
• Driver: Studio 577.00 (latest)
• CUDA Version: 12.6+
• Torch Version: 2.6.0+cu126
• Operating System: Windows 11
• FramePack Version: framepack_cu126_torch26 one-click package

Output of torch.cuda.get_device_capability(0)

(12, 0)

[shanevcantwell]

You need CUDA 12.8 for RTX 50x0 cards.

Install: https://developer.nvidia.com/cuda-12-8-0-download-archive

Then paste the following to a command line:
pip install -U torch torchvision torchaudio xformers --index-url https://download.pytorch.org/whl/cu128

[SLoneTiger]

You need CUDA 12.8 for RTX 50x0 cards.

Install: https://developer.nvidia.com/cuda-12-8-0-download-archive

Then paste the following to a command line: pip install -U torch torchvision torchaudio xformers --index-url https://download.pytorch.org/whl/cu128

Hi shanevcantwell.
Could you please help with this issue? After installing "CUDA 12.8 for RTX 50x0 cards" from the link above, go to my framepack folder, open a CMD line and copy/paste the pip command above, I get:

D:\AI_Framepack_cu126_torch26>pip install -U torch torchvision torchaudio xformers --index-url https://download.pytorch.org/whl/cu128
Fatal error in launcher: Unable to create process using '"D:\webui-forge\system\python\python.exe"  "D:\AI_Framepack_cu126_torch26\system\python\Scripts\pip.exe" install -U torch torchvision torchaudio xformers --index-url https://download.pytorch.org/whl/cu128': The system cannot find the file specified.

Could you tell me what am I doing wrong?
Thanks in advance.

EDIT: After a reboot, the pip command worked without errors, and it updated, hopefully this should fix framepack.

[shanevcantwell]

You need CUDA 12.8 for RTX 50x0 cards.
Install: https://developer.nvidia.com/cuda-12-8-0-download-archive
Then paste the following to a command line: pip install -U torch torchvision torchaudio xformers --index-url https://download.pytorch.org/whl/cu128

Hi shanevcantwell. Could you please help with this issue? After installing "CUDA 12.8 for RTX 50x0 cards" from the link above, go to my framepack folder, open a CMD line and copy/paste the pip command above, I get:

D:\AI_Framepack_cu126_torch26>pip install -U torch torchvision torchaudio xformers --index-url https://download.pytorch.org/whl/cu128
Fatal error in launcher: Unable to create process using '"D:\webui-forge\system\python\python.exe"  "D:\AI_Framepack_cu126_torch26\system\python\Scripts\pip.exe" install -U torch torchvision torchaudio xformers --index-url https://download.pytorch.org/whl/cu128': The system cannot find the file specified.

EDIT: After a reboot, the pip command worked without errors, and it updated, hopefully this should fix framepack.

If I had to guess, I'd say that instead of the real pip command, some script is catching the command line and trying to expand your attempt to run pip install <etc> to instead call python pip install <etc>. However, it's expanding the module 'pip' as the full path to your executable D:\AI_Framepack_cu126_torch26\system\python\Scripts\pip.exe for python.exe to call.

...also that the attempt to craft the command line is fundamentally wrong because it should be calling python -m pip install <etc>.

As for why a reboot fixed it, were you maybe in the active venv one run vs not active venv the other? If you ever want to try to track it down, which pip will be your friend to see what's actually being called. (Sometimes under Windows you want where pip instead.)

[SLoneTiger]

Thanks for the reply, no, I realized I was running pip command from the wrong location, once I ran the command in the python.exe directory it worked, I was able to update "torch torchvision torchaudio" for CUDA 129, (RTX 5090 laptop CUDA 12.9), but after updating framepack still refuses to work returning:

File "D:\AI_Framepack_cu126_torch26\system\python\lib\site-packages\sageattention\core.py", line 666, in sageattn_qk_int8_pv_fp8_cuda
    assert SM89_ENABLED, "SM89 kernel is not available. Make sure you GPUs with compute capability 8.9."
AssertionError: SM89 kernel is not available. Make sure you GPUs with compute capability 8.9.

Reading more I tried to downgrade "torch torchvision torchaudio" from 2.8.0+cu129 to 2.7.0+cu129 but pip was unable to downgrade since it appears there are missing files in the cu129 location. The only option to downgrade was 2.7.0+cu128. And using this version stil get error:

File "D:\AI_Framepack_cu126_torch26\system\python\lib\site-packages\sageattention\quant.py", line 279, in per_channel_fp8
    v_transposed_permutted = torch.empty((b, head_dim, h_kv, padded_len), dtype=v.dtype, device=v.device)
RuntimeError: CUDA error: no kernel image is available for execution on the device

[shanevcantwell]

That binary that @lllyasviel has up here is no longer relevant.
Get rid of d:\AI_Framepack_cu126_torch26\ altogether.
Do not touch framepack_cu126_torch26.7z at all at all.

Start in the directory you want Framepack to live under. Enter the following by copying this whole block and right-clicking into your cmd.exe window:

git clone https://github.com/lllyasviel/Framepack
cd Framepack
python -m venv venv
venv\Scripts\activate
pip install torch torchvision torchaudio xformers --index-url https://download.pytorch.org/whl/cu128
pip install -r requirements.txt

(ignore this part at the end. Idk what's up with Windows pip but this doesn't ever work for me, and doesn't matter for this repo alone:)
[notice] A new release of pip is available: 25.0.1 -> 25.2
[notice] To update, run: python.exe -m pip install --upgrade pip

Then you can:
venv\Scripts\activate (each time you run)
python demo_gradio.py (or python demo_gradio_f1.py)

deactivate

OR

Just get this and run install.bat 128 then run_goan.bat. Learning pip was one of the most irritating roadblocks I faced early on with machine learning, and so I make simplifying installation a primary principle in my projects.
https://github.com/shanevcantwell/goan

[SLoneTiger]

Thanks.
I deleted the entire framepack install and started from scratch following your steps, all went well until I tried to generate something, keeping the defaults, no changes, after clicking "Start":

* Running on local URL:  http://0.0.0.0:7860

To create a public link, set `share=True` in `launch()`.
Unloaded DynamicSwap_LlamaModel as complete.
Unloaded CLIPTextModel as complete.
Unloaded SiglipVisionModel as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Unloaded DynamicSwap_HunyuanVideoTransformer3DModelPacked as complete.
Loaded CLIPTextModel to cuda:0 as complete.
Unloaded CLIPTextModel as complete.
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Loaded SiglipVisionModel to cuda:0 as complete.
latent_padding_size = 27, is_last_section = False
Unloaded SiglipVisionModel as complete.
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
  0%|                                                                                                                             | 0/25 [00:00<?, ?it/s]CUDA error (C:/a/xformers/xformers/third_party/flash-attention/hopper\flash_fwd_launch_template.h:188): invalid argument

(venv) D:\AI_Framepack\Framepack>
(venv) D:\AI_Framepack\Framepack>