Blackwell compatibility (Windows only)

[w-e-w]

Description

Switch to isntall early access pytorch wheels if the user is has a Blackwell GPU

wheels are upload to huggingface with permission from Nvidia
https://huggingface.co/w-e-w/torch-2.6.0-cu128.nv

from my testing all seems to work fine
apart from the optional --xformers

the current xformers pre-built only suppors pytorch 2.5.1

test was done with python 3.10 and 3.11 on 3090 and 1650 cards

---

if someone wish to test compatibility on non-blackwell cards
override the condiction at

stable-diffusion-webui/modules/launch_utils.py

Lines 333 to 339 in 0005760

	if all([
	os.environ.get('TORCH_INDEX_URL') is None,
	sys.version_info.major == 3,
	sys.version_info.minor in (10, 11, 12),
	platform.system() == "Windows",
	get_cuda_comp_cap() >= 10, # Blackwell
	]):

---

note blackwell support should be considered experimental / theoretical
feedback with actual hardware test is more the welcome

Checklist:

• I have read contributing wiki page
• I have performed a self-review of my own code
• My code follows the style guidelines
• My code passes tests

[catboxanon]

I have two GPUs installed, so the output returned by nvidia-smi can't be parsed.

>>> float(subprocess.check_output(['nvidia-smi', '--query-gpu=compute_cap', '--format=noheader,csv'], text=True))
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ValueError: could not convert string to float: '8.6\n7.5\n'

[w-e-w]

thanks I see what's happening
I didn't know that could happen
it turns out I actually did know that this could happen I just forgot, I've encountered this exact issue in the past
https://github.com/w-e-w/stable-diffusion-webui-GPU-temperature-protection/blob/f388d7081945576ca2361c365d16f5edb3250412/temperature_sensor_modules/nvidia_smi.py#L8

[w-e-w]

@catboxanon shoud be fixed