Urle Sistiana

### Discussed in https://github.com/IrineSistiana/mosdns/discussions/547 Originally posted by **kkkgo** November 10, 2022 因为clash使用的格式是mmdb，为了节省空间，希望能复用clash的mmdb文件。

fix: lumina 2 timesteps handling

3

fix #2201 mainly replaced `get_noisy_model_input_and_timesteps` with the one from flux_train_utils.py. This also added `flux_shift` timestep sampling method and `--ip_noise_gamma` support.

feat: torch compile within checkpointing and activation memory budget for Lumina 2

7

This PR has two features to speed up Lumina 2 training: **checkpointing + torch.compile sub modules** Currently, when using gradient checkpointing, torch.compile will skip all frames (modules) inside the checkpointed...

During my training, I accidently found if I use the default `shift` timestep sampling method, even after thousands of steps, the model still learned nothing, the sample images are identical....