Pytorch CUDA / GPU error

1

There is an error in the log upon starting easydiffusion:
CUDA initialization: The NVIDIA driver on your system is too old (found version 10020).

To Reproduce
Steps to reproduce the behavior:

  1. Run ‘Start Stable Diffusion UI.cmd’.
  2. Error will show in log window, cannot enable GPU
  3. NVIDIA-SMI 441.66 Driver Version: 441.66 CUDA Version: 10.2
  4. EasyDiffusion will not allow enabling GPU, stays in CPU mode only.

Expected behavior
Program should allow GPU function, GPU has 4GB VRAM.

  • OS: Windows 7 X64
  • Memory: 32GB RAM
  • GPU: Nvidia M2000M 4GB VRAM
  • Browser: Firefox
  • Version: 115.0.3

If possible, do I need to downgrade pytorch version?
I believe I am using latest Nvidia driver and Cuda version for this GPU.
Have tried with invokeai and automatic1111 as well. Can generate via CPU only.

Here is the actual error from pytorch:

git version 2.42.0.windows.1
conda 23.7.3

COMSPEC=C:\Windows\system32\cmd.exe

Name AdapterRAM DriverDate DriverVersion

Intel(R) HD Grap… 1073741824 20190529000000.0… 21.20.16.5076
NVIDIA Quadro M2… 4293918720 20191206000000.0… 26.21.14.4166

Python 3.8.5
Launching uvicorn

C:\ED\installer_files\env\lib\site-packages\torch\cuda_init_.py:107: UserWarn
ing: CUDA initialization: The NVIDIA driver on your system is too old (found ver
sion 10020). Please update your GPU driver by downloading and installing a new v
ersion from the URL: Download The Latest Official NVIDIA Drivers Alternatively, go
to: https://pytorch.org to install a PyTorch version that has been compiled wit
h your version of the CUDA driver. (Triggered internally at C:\actions-runner_w
ork\pytorch\pytorch\builder\windows\pytorch\c10\cuda\CUDAFunctions.cpp:109.)
return torch._C._cuda_getDeviceCount() > 0
22:26:52.572 INFO MainThread forcing full precision for device: cpu

2

Update your NVIDIA driver as the error suggests to use a newer PyTorch binary with CUDA 11 or 12 runtime dependencies.

3

I believe I am already on the latest nvidia driver version for the m2000m: 441.66

4

The latest driver still supports Maxwell (although deprecated) so you could still update the driver.

5

I believe the 441.66 is the latest form NVIDIA for windows 7 unless I am overlooking something.

Thanks for your help, I see you have helped MANY on here with various pytorch issues!

6

I see a recent driver for your M2000 GPU for Windows:

image
image1346×621 65 KB

I haven’t tried it, but you could still give it a go. Alternatively, you would be bound to older PyTorch releases with CUDA 10.x runtime dependencies.

7

Got it. Any way to “force” windows 7 to accept this graphics driver if I unpack it? My windows 7 is x64 as well.

Alternatively, how could I downgrade to an older pytorch version so it can ‘see’ this GPU?

Thanks again

8

I don’t know if you are seeing any issues while installing the driver and am also not deeply familiar with Windows itself.
If you want to downgrade PyTorch, search for the latest binary with CUDA 10.2 here (which seems to be 1.12.1).

9

I was able to upgrade the video driver to Version 442.74, but still get the same error. I may try to force a newer driver to see if it works…

I used the tool nvcleanstall

Update: the above release 570 seems to only work on Windows 10 / 11 unless I can modify it somehow. Looks like I may try to downgrade the pytorch version to 1.12.1 as you suggested…

Currently I am at Nvidia driver version 442.74 and CUDA 10.2 according to nvidia-smi

10

I have a feeling this is related - all of these pytorch versions are no longer available and I believe they may be the ones needed as they support CUDA 10.2:

python -m pip install torch==1.11.0+cu102 torchvision==0.12.0+cu102 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu102

python -m pip install torch==1.12.0+cu102 torchvision==0.13.0+cu102 torchaudio==0.12.0 --extra-index-url https://download.pytorch.org/whl/cu102

python -m pip install torch==1.12.1+cu102 torchvision==0.13.1+cu102 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu102

11

I have a feeling this is not going anywhere. Like all “open source” projects, they are plagued by errors and mismatching files. Even though they claim a card is “supported”, it does not mean it truly is. I have wasted almost a week on this and have nothing to show for it. Looks like I’ll scrap it and try to use an external GPU

12

Our current binaries still support Maxwell architectures, which were released in 2014, so over 10 years ago. However, to use the newer CUDA runtime dependencies you would need to update to any NVIDIA driver supporting CUDA 12 (or you could update to the PyTorch binaries with CUDA 11.8 [which we also still build] but would also need to update the NVIDIA driver to a version supporting CUDA 11).

I’ve provided the valid links to these drivers, but since you are still on Windows 7, which is at EOL since Jan 14 2020, you are unable to install these.

I don’t know how else to help but you might want to consider using a supported OS to be able to use any recent SW stack.