Supported pytorch added for CUDA 11.3 and CUDA 11.7, dynamically. (Windows only)

[kavyamali]

Also called the early_access_blackwell_wheels under def prepare_environment():

Description

This change adds dynamic detection using nvidia-smi to determine Compute Capability for Nvidia GPUs with CC<7 on Windows. It then installs supported, latest Pytorch for the same. In all other cases, it falls back to default logic or that mentioned in early_access_blackwell_wheels. This has been tested on a GTX 970. No issues.

I also apologize for the confusion in my previous PR, as the changes were clearly bogus there.

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

[w-e-w]

sorry no offense your actions are very suspicious

I also apologize for the confusion in my previous PR, as the changes were clearly bogus there.

assuming that you are responding to #17158 this meant that you deleted your old account (whitch itself it's a new account that has no other activity)
that is not normal behavior

Also called the early_access_blackwell_wheels under def prepare_environment():

calling early_access_blackwell_wheels makes zero sense

the PR description feels like LLM generated
I can't tell if the code is generated and the clame of it works on GTX970 is just a hallucination of an LLM

---

I think you are a bot, or at least someone that may not understand what there are doing
I can't trust your code and can't trust your clame of it works on older GPU
so sorry I'm not going to merge this PR

[kavyamali]

sorry no offense your actions are very suspicious

I also apologize for the confusion in my previous PR, as the changes were clearly bogus there.

assuming that you are responding to #17158 this meant that you deleted your old account (whitch itself it's a new account that has no other activity) that is not normal behavior

Also called the early_access_blackwell_wheels under def prepare_environment():

calling early_access_blackwell_wheels makes zero sense

the PR description feels like LLM generated I can't tell if the code is generated and the clame of it works on GTX970 is just a hallucination of an LLM

I think you are a bot, or at least someone that may not understand what there are doing I can't trust your code and can't trust your clame of it works on older GPU so sorry I'm not going to merge this PR

I am sorry, I understand why you feel my actions might be suspicious. The reason I called early_access_blackwell was to make it work under torch command, as that was the logic of my code. I wanted to make sure both work. I can provide you screenshots or proof for it working on a 970. The reason I deleted my old account was because the entire branch there was a mess.

If you still don't want to merge the PR, I understand. Thanks for your time. And I am still a beginner, so I'd appreciate guidance.

[kavyamali]

sorry no offense your actions are very suspicious

I also apologize for the confusion in my previous PR, as the changes were clearly bogus there.

assuming that you are responding to #17158 this meant that you deleted your old account (whitch itself it's a new account that has no other activity) that is not normal behavior

Also called the early_access_blackwell_wheels under def prepare_environment():

calling early_access_blackwell_wheels makes zero sense

the PR description feels like LLM generated I can't tell if the code is generated and the clame of it works on GTX970 is just a hallucination of an LLM

I think you are a bot, or at least someone that may not understand what there are doing I can't trust your code and can't trust your clame of it works on older GPU so sorry I'm not going to merge this PR

Here are the screenshots.

[kavyamali]

sorry no offense your actions are very suspicious

I also apologize for the confusion in my previous PR, as the changes were clearly bogus there.

assuming that you are responding to #17158 this meant that you deleted your old account (whitch itself it's a new account that has no other activity) that is not normal behavior

Also called the early_access_blackwell_wheels under def prepare_environment():

calling early_access_blackwell_wheels makes zero sense

the PR description feels like LLM generated I can't tell if the code is generated and the clame of it works on GTX970 is just a hallucination of an LLM

I think you are a bot, or at least someone that may not understand what there are doing I can't trust your code and can't trust your clame of it works on older GPU so sorry I'm not going to merge this PR

Lastly, for the clarification of my logic. The torch_command variable didn't call for the defined variables, which was responsible for smi checking and installing separate pytorch versions. So I created a condition, calling those variables, and if the state was None, it'd fallback to original.

[w-e-w]

have you read early_access_blackwell

stable-diffusion-webui/modules/launch_utils.py

Lines 331 to 334 in 6685e53

	def early_access_blackwell_wheels():
	"""For Blackwell GPUs, use Early Access PyTorch Wheels provided by Nvidia"""
	print('deprecated early_access_blackwell_wheels')
	if all([

when something is "deprecated" it means it is something that was used before but is not to be used anymore
basically you by using that function means that you either don't know what the deprecated term mean or you never read the function

[kavyamali]

have you read early_access_blackwell

stable-diffusion-webui/modules/launch_utils.py

Lines 331 to 334 in 6685e53

	def early_access_blackwell_wheels():
	"""For Blackwell GPUs, use Early Access PyTorch Wheels provided by Nvidia"""
	print('deprecated early_access_blackwell_wheels')
	if all([

when something is "deprecated" it means it is something that was used before but is not to be used anymore basically you by using that function means that you either don't know what the deprecated term mean or you never read the function

Oh, I see. I didn't know that. Thanks for correcting me.

[kavyamali]

have you read early_access_blackwell

stable-diffusion-webui/modules/launch_utils.py

Lines 331 to 334 in 6685e53

	def early_access_blackwell_wheels():
	"""For Blackwell GPUs, use Early Access PyTorch Wheels provided by Nvidia"""
	print('deprecated early_access_blackwell_wheels')
	if all([

when something is "deprecated" it means it is something that was used before but is not to be used anymore basically you by using that function means that you either don't know what the deprecated term mean or you never read the function

The reason I called my variable was because it didn't read the logic and proceeded with usual installation process. I thought calling the earlier variable might be a good idea too. This is my first PR, so I'm sorry for the mistake.

[w-e-w]

and why did you delete your previous account and create new account and make a new PR thing
that feels like something like a LLM code PR farm would do

[kavyamali]

and why did you delete your previous account and create new account and make a new PR thing that feels like something like a LLM code PR farm would do

I made that account when I was 13, and it has nothing valuable but mess of code. So I made a new one. I'm not aware of LLM farming.

[w-e-w]

I made that account when I was 13, and it has nothing valuable but mess of code. So I made a new one. I'm not aware of LLM farming.

if my memory serves me I'm pretty sure that previous account was created less than two weeks ago

[kavyamali]

have you read early_access_blackwell

stable-diffusion-webui/modules/launch_utils.py

Lines 331 to 334 in 6685e53

	def early_access_blackwell_wheels():
	"""For Blackwell GPUs, use Early Access PyTorch Wheels provided by Nvidia"""
	print('deprecated early_access_blackwell_wheels')
	if all([

when something is "deprecated" it means it is something that was used before but is not to be used anymore basically you by using that function means that you either don't know what the deprecated term mean or you never read the function

So, will creating a new PR without the early_access_blackwell variable called be worthy?

[kavyamali]

I made that account when I was 13, and it has nothing valuable but mess of code. So I made a new one. I'm not aware of LLM farming.

if my memory serves me I'm pretty sure that previous account was created less than two weeks ago

N
It was made around 2022, had three repositories.

[kavyamali]

I made that account when I was 13, and it has nothing valuable but mess of code. So I made a new one. I'm not aware of LLM farming.

if my memory serves me I'm pretty sure that previous account was created less than two weeks ago

Here is a mail for that account from 2022.

[w-e-w]

from what I can find Maxwell cards shoud work with cu126

try something for me install with cu126 and see it it works

use the dev branch (without your changes from this PR)
either set TORCH_INDEX_URL in webui-user.bat
like so

set TORCH_INDEX_URL=https://download.pytorch.org/whl/cu126

or directly modify the code (temporarily)

stable-diffusion-webui/modules/launch_utils.py

Line 351 in 6685e53

torch_index_url = os.environ.get('TORCH_INDEX_URL', "https://download.pytorch.org/whl/cu128")

- 'TORCH_INDEX_URL', "https://download.pytorch.org/whl/cu128"
+ 'TORCH_INDEX_URL', "https://download.pytorch.org/whl/cu126"

then do a clean install see if it works

[w-e-w]

and if it doesn't work with pytorch 2.7 then it seems to work with 2.6
so might not be a need to drop down to 1.12.1+cu113

[kavyamali]

and if it doesn't work with pytorch 2.7 then it seems to work with 2.6 so might not be a need to drop down to 1.12.1+cu113

I've tried cu12x versions, even cu118 doesn't work. Pytorch dropped full library support for CC5.2 after cu113, 1.12.1. But if you say so, I'll try again with cu126.

[w-e-w]

your are on 970 right?
base on the data hear it seems to work with
2.1.2+cu121 (this is the pytorch version that is used on the master branch)
https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html

[kavyamali]

your are on 970 right? base on the data hear it seems to work with 2.1.2+cu121 (this is the pytorch version that is used on the master branch) https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html

Wierd then. When I try installing either of those, it says 'MemoryError'. I can say cu118 works on oobabooga, with full GPU acceleration. But it doens't on automatic1111. I tried compiling pytorch from source, that failed too. But I can confirm cu126 or latest one mentioned doens't work. And besides, there is little to no improvement from 11.x to 12.x performance for such old GPUs. cu113 is the safest, I believe.

[kavyamali]

your are on 970 right? base on the data hear it seems to work with 2.1.2+cu121 (this is the pytorch version that is used on the master branch) https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html

Update: --no-cache-dir does makes all pytorch versions upto cu118 work. cu121 is targeted for pascal, still doesn't install.

Edit: It might still work with memory optimisations, but it's too much work. Waiting for your call.

[kavyamali]

your are on 970 right? base on the data hear it seems to work with 2.1.2+cu121 (this is the pytorch version that is used on the master branch) https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html

Hey, I know it has been a while, but I have some important updates. I recently freshly installed latest Nvidia drivers, and tried reinstalling CUDA 12.8(cu128) pyotrch with --no-cache-dir command. Now, every pytorch version installs successfully. But cu128 pytorch is the only one that doesn't officially support maxwell. This is the install log:
log.txt

And all Cuda versions up to cu126 work fine now. Only cu128 gives an error at the start:

log1.txt