Yuxuan Zhang

Still the 5B model, currently I am rushing to complete the preparation work for the open-source of the I2V model

You can check solutions likes SVD (not the same), where the input channels are doubled for processing image embeddings

还在做，没成功，相关PR可以在diffuser中查看

已经完成，可以在sat的readme中看到怎么合并

More than 100,000 GPU hours of H100 as told in X before

These are limited by the company's requirements, and we are not at liberty to disclose them directly. However, I believe you can estimate based on the data I mentioned; the...

安装torch请使用mac的安装方式，mps，另外需要将模型启动换成mps

改为MPS为什么会有cuda错误，或许你应该clone我们最新的代码，并将运行设备启动到mps上 即 .to('mps')

你需要吧 pipe.enable_model_cpu_offload() pipe.enable_sequential_cpu_offload() 这两行注释掉，然后确保你的mac有64G的内存（表测的是显存而不是内存，如果你使用的是内存推理，实际占用远远大于显存） pipe.to("mps")

这个问题我好像也暂时解决不了，目前没有mac设备能测试，如果有了，我们测一下反馈