Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Expected Behavior

The command 'python main.py' should run successfully on AMD RADEON RX 7900 XTX with ROCm 7.2 without requiring NVIDIA-specific dependencies. This used to be the case.

Actual Behavior

Running:
python main.py
results in a ModuleNotFoundError: 'No module named comfy_aimdo'.
The dependency comfy_aimdo (introduced in commit #f8acd9c40) only supports NVIDIA GPUs.

Attempts to remove the dependency from requirements.txt and import statements lead to many additional NameError exceptions due to missing comfy_aimdo function references.

Steps to Reproduce

• Tested using an AMD RADEON RX 7900 XTX with ROCm 7.2 (although ROCm is irrelevant - this can be even tested/verified on NVIDIA simply by not installing comfy_aimdo requirement.

• Check out commit #f8acd9c40.

• Run python main.py
  -Observe the ModuleNotFoundError for comfy_aimdo.

• Next, check out the earlier working commit #873de5f

• Verify python main.py no longer produces comfy_aimdo errors.

Debug Logs

Line 34 main.py
ModuleNotFoundError: 'No module named comfy_aimdo'

(Sorry, I cannot post full logs for security reasons.)

Other

I haven't seen this reported already as an issue or in documentation. My apologies if it has.
My GPU is: AMD RADEON RX 7900 XTX with ROCm 7.2

The earlier working commit #873de5f does raise a secondary 'ModuleNotFoundError: 'no module named torchsde' in comfy/k_diffusion/sampling.py. (I don't think torchsde is available/needed for ROCm 7.2). Hopefully that dependency issue has already been fixed since.
For now I have mitigated it by commenting out the import on line 7 and the function call to BatchedBrownianTree() on line 142 of the file. Hopefully that call is not too important :) In any case, if it hasn't already been fixedit should be a separate but similar issue._