Training erroe with accelerate.backward(loss)

[juxingyiwan]

When I run the default training script for sd1.5, an error occurred on the line accelerate.backward(loss). The detailed error is as follows:

Traceback (most recent call last):
  File "/mnt/lustre/zhuguibo/dcy/BrushNet/examples/brushnet/train_brushnet.py", line 1403, in <module>
    main(args)
  File "/mnt/lustre/zhuguibo/dcy/BrushNet/examples/brushnet/train_brushnet.py", line 1303, in main
    accelerator.backward(loss)
  File "/mnt/lustre/zhuguibo/anaconda3/envs/diffusers/lib/python3.9/site-packages/accelerate/accelerator.py", line 2248, in backward
    loss.backward(**kwargs)
  File "/mnt/lustre/zhuguibo/anaconda3/envs/diffusers/lib/python3.9/site-packages/torch/_tensor.py", line 487, in backward
    torch.autograd.backward(
  File "/mnt/lustre/zhuguibo/anaconda3/envs/diffusers/lib/python3.9/site-packages/torch/autograd/__init__.py", line 200, in backward
    Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
RuntimeError: CUDA error: an illegal instruction was encountered
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

Any one met the same issue?

[VanJessie]

Hello friend, I have encountered a similar problem. Just like when using sd1.5, the output error showed that this weight was not quite compatible. But I don't know how to solve it. Could you please tell me if you have finally found a solution?