Comfy UI Support for linux 10 series Nvidia GPUs

[rdemon77]

Recently, comfyui linux updated to require pytorch 2.7. This means that it does not work on linux 10 series Nvidia GPUs. The last version of pytorch compatable with 10 series gpus is 2.6.

I get the following error when running Comfy 0.39.0 on pytorch 2.6 on linux.

...
  File "/home/roger/miniconda3/envs/comfyui_2b/lib/python3.12/site-packages/comfy_kitchen/backends/eager/conv3d.py", line 37, in
    @torch.library.custom_op("comfy_kitchen::fp16_conv3d_out", mutates_args=("out",))
     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/roger/miniconda3/envs/comfyui_2b/lib/python3.12/site-packages/torch/_library/custom_ops.py", line 121, in inner
    schema_str = torch.library.infer_schema(fn, mutates_args=mutates_args)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/roger/miniconda3/envs/comfyui_2b/lib/python3.12/site-packages/torch/_library/infer_schema.py", line 106, in infer_schema
    error_fn(
  File "/home/roger/miniconda3/envs/comfyui_2b/lib/python3.12/site-packages/torch/_library/infer_schema.py", line 58, in error_fn
    raise ValueError(
ValueError: infer_schema(func): Parameter stride has unsupported type list[int]. The valid types are: dict_keys([<class 'torch.Tensor'>, typing.Optional[torch.Tensor], typing.Sequence[torch.Tensor], typing.List[torch.Tensor], typing.Sequence[typing.Optional[torch.Tensor]], typing.List[typing.Optional[torch.Tensor]], <class 'int'>, typing.Optional[int], typing.Sequence[int], typing.List[int], typing.Optional[typing.Sequence[int]], typing.Optional[typing.List[int]], <class 'float'>, typing.Optional[float], typing.Sequence[float], typing.List[float], typing.Optional[typing.Sequence[float]], typing.Optional[typing.List[float]], <class 'bool'>, typing.Optional[bool], typi

This seems to be caused by 3D mesh generation support.

I was able to work around this by commenting out 3D mesh generation in the following files:

• /home/roger/miniconda3/envs/comfyui_2b/lib/python3.12/site-packages/comfy_kitchen/backends/eager/init.py
  /home/roger/miniconda3/envs/comfyui_2b/lib/python3.12/site-packages/comfy_kitchen/backends/eager/sol_attn.py
  /home/roger/miniconda3/envs/comfyui_2b/lib/python3.12/site-packages/comfy_kitchen/backends/eager/group_norm_pad3d.py

Going forward, I request the comfyui dev team provide a provide an option to disable functionality that requires pytorch 2.7. Us gpu peasants have to work with what we have.

[chaObserv]

Try installing pytorch with cuda 12.6.

[rdemon77]

Does pytorch 2.7 work with cuda 12.6 on 10 series GPUs?

[chaObserv]

I think so. There are some reported cases on the pytorch forum: https://discuss.pytorch.org/t/what-version-of-pytorch-is-compatible-with-nvidia-geforce-gtx-1080/222056.

[mrhy7176]

Yes, per PyTorch's own notes. The 2.8 release notes say Maxwell and Pascal were dropped only from the CUDA 12.8/12.9 builds and tell those cards to use CUDA 12.6, and a PyTorch notice from 4 Sept says 2.14 is the last release with cu126 wheels (from 2.15 it is build-from-source). So pip install torch==2.14.0 torchvision==0.29.0 --index-url https://download.pytorch.org/whl/cu126 should meet ComfyUI's 2.7 requirement on a 10-series card. I haven't tested it on one myself. I work on CloudGPU.