`sct_deepseg -task seg_spinal_rootlets_t2w` raises `unexpected keyword argument 'perform_everything_on_device'` error about spinalcordtoolbox

[joshuacwnewton]

On first glance, this looks like an upstream API change with nnunetv2.

• In requirements.txt, we specify:
  

  spinalcordtoolbox/requirements.txt

  Lines 22 to 23 in 9ce98c2

  	# 2.4.0/2.4.1 fail during inference due to https://github.com/spinalcordtoolbox/spinalcordtoolbox/issues/4444
  	nnunetv2!=2.4.0,!=2.4.1

• The latest version is 2.4.2, released on April 25th, 2024. Given the above specification, we should have been using 2.4.2 for about a month now. Given that there are no recent test failures, I assume that this issue isn't caused by any recent upstream changes.
• Checking nnUNetPredictor on nnunetv2's master branch shows that perform_everything_on_device is present.
• However, 6 months ago this argument used to be called perform_everything_on_gpu. This was changed in v2.3.1.

So, my best guess is that you have an older version of nnunetv2 in your local SCT copy (nnunetv2==2.3.0 or below), but SCT's code is meant to work with the newest versions of nnunetv2.

We can definitely amend the requirements.txt file to make this incompatibility clearer, but I would recommend updating your local venv too. :)

[joshuacwnewton]

Aha! Searching our issues for perform_everything_on_gpu returns:

• #4380
• MIC-DKFZ/nnUNet#1972

[valosekj]

Great, thank you! Indeed, I had an older version of nnunetv2 in my local SCT copy.

I pulled the updated requirements.txt, activated the sct_venv and ran pip install -r requirements.txt. And the model is now working!