accumulate_grad warning when train convnext_tiny with pytorch_lightning and accumulate

[ray-lee-94]

[W accumulate_grad.h:170] Warning: grad and param do not obey the gradient layout contract. This is not an error, but may impair performance.
grad.sizes() = [96, 1, 7, 7], strides() = [49, 49, 7, 1]
param.sizes() = [96, 1, 7, 7], strides() = [49, 1, 7, 1] (function operator())

[rwightman]

@VCBE123 you'll have to provide a lot more detail of your environment, hardware, model etc. I do have an idea what could cause it, but I can't reproduce on PyTorch 1.10+ on 3090 and there were some incorrect triggers for the warning in older versions of PT. It is also as it says, a warning and may not be impacting anything.

[ray-lee-94]

I test the ConvNeXt family, they give the same warning. It sees not impact the training.

The environment:
torch=1.7.1
pytorch-lightning=1.5.6
timm=0.5.4
python=3.6.9
CUDA Version=11.2
GPU=GeForce RTX 2080Ti
Ubuntu=18.0.4

[rwightman]

@VCBE123 try newer PyTorch, 1.10.1, cuda 11.3