how use nightly version??

[dltptkd000]

NEW installed.
how use blackwellgpu??
cant generation image

cpu 9950x3d
gpu 5080

---

19:06:31-269762 INFO Command line args: ['--upgrade', '--use-nightly'] upgrade=True use_nightly=True args=[]
19:06:36-372496 INFO Torch: torch==2.6.0+cu126 torchvision==0.21.0+cu126
19:06:36-374536 INFO Packages: diffusers==0.33.0.dev0 transformers==4.48.3 accelerate==1.3.0 gradio==3.43.2
pydantic==1.10.21
19:06:36-624454 INFO Device detect: memory=15.9 default=balanced
19:06:36-775178 INFO Engine: backend=Backend.DIFFUSERS compute=cuda device=cuda attention="Scaled-Dot-Product"
mode=no_grad
19:06:36-792037 WARNING Torch FP16 test fail: CUDA error: no kernel image is available for execution on the device
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace
below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.

19:06:36-795101 WARNING Torch BF16 test fail: CUDA error: no kernel image is available for execution on the device
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace
below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.

[Disty0]

Add --reinstall too. Torch is only installed / changed on the first install.

[brknsoul]

Don't forget to remove --reinstall after it's up and running, otherwise you'll be reinstalling every time you start SD.Next.

[itzolkmar]

Probably, this is not really a news to you. But probably it helps others.

2 days ago I got a new system with a Geforce 5090 and no AI tool worked, due to Blackwell incompatibilities of the installed pytorch versions. I found a good hint on the web and did the following:

I use Python3-11 and therefore loaded the following packages from https://download.pytorch.org/whl:
torch-2.7.0+cu128-cp311-cp311-win_amd64.whl
torchaudio-2.7.0+cu128-cp311-cp311-win_amd64.whl
torchvision-0.22.0+cu128-cp311-cp311-win_amd64.whl

Installation worked easily:
.\pip install ‘\torch 2.7\torch-2.7.0+cu128-cp311-cp311-win_amd64.whl’
.\pip install ‘\torch 2.7\torchaudio-2.7.0+cu128-cp311-cp311-win_amd64.whl’
.\pip install ‘\torch 2.7\torchvision-0.22.0+cu128-cp311-cp311-win_amd64.whl’

It seems torchaudio has already been installed with torch, cause it was yet found.

SDNext produced a nice apple picture for a test with Flux.1 Dev. In nearly no time. I have not updated my other install yet, so I can't say whether the speed benefit results from Vlad's last update or the 5090.
I will try this with Auto1111, FOOOCUS, ComfyUI, etc in the next days.

[brknsoul]

Torch 2.7.0 now supports cu128 and blackwell GPUs in their release version.
The correct command would be;

venv\scripts\python -m pip install -U torch==2.7.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

(In linux, change venv\scripts\ to venv/bin/)

I'm not sure if SD.Next (or any extensions) use torchaudio, but it doesn't hurt to have it in there.

[vladmandic]

torchaudio is not needed

otherwise, this is now default on dev branch and will be merged to master soon.

[brknsoul]

It's in my notes, because ComfyUI has various audio-related custom nodes. I wasn't sure if SD.Next (or extensions) could do audio stuff as well. It's only 11MB ;-)