About missing "ctx" argument in autograd function "ApplyRotaryEmbeddingInplace"

[sairights]

Thanks for opensourcing your great work HDiT!
In

k-diffusion/k_diffusion/models/image_transformer_v2.py

Line 204 in 21d12c9

def forward(x, theta, conj):

, it seems that the ctx is missing. But strangely, the function could run smoothly without the ctx argument on PyTorch 2.1. As I switch to PyTorch-1.12, the code is not runnable. I am wondering about the reason...