horseee
commented
4 months ago Hi @ledrose ,
We tested the issue you mentioned and found that the possible cause of this issue is a conflict between the cpu offloa and our implementation of deepcache (hook a new forward function). When the model's submodules are called, the forward function of the original model is loaded, while the one needed by DeepCache is removed.
You can try our old version of the code here(implemented differently, just rewrite the forward function). We tested it and it is compatible with enable_model_cpu_offload(). But you need to downgrade the version of diffusers to ~0.24.0 (Higher version may have some import error since the API in diffusers changed )
Found a weird problem with DeepCache. If you use cpu offload with DeepCache first generated batch will be normal (expected speedup and quality), but all next batches will be very distorted (and generation is speed up even more). It does not happen with pipe.enable_sequential_cpu_offload() and when using without offload. It can be fixed by helper.disable() after every batch and helper.enable() again before next batch generation but as it is not required in pipeline without cpu offload I decided not consider it as expected behavior. Code for recreating this issue and example of distortion here: https://www.kaggle.com/code/ledrose/deepcache-cpu-offload-bug