In the pretraining process, directly loading the safetensors of diffusion model from https://huggingface.co/Lightricks/LTX-Video/blob/main/ltx-video-2b-v0.9.safetensors cause lots of state dict mismatch.