MiniMax-H3: why set shift=2.22

[serend1p1ty]

13c2a16

[longxiang-ai]

it has been modified from 12.0, 3.0 -> 2.22, 2.22, while inference and the official manual remain 12.0, 3.0, maybe it is a bug? @mi804

[Artiprocher]

@serend1p1ty @longxiang-ai

Typically, a smaller shift is used during training to ensure the model spends enough time at each denoising stage, which helps guarantee good training performance.

During inference, a larger shift is applied so that the model stays longer in stages with high noise, leading to better inference quality.

The model's open-source release did not provide the shift used for training, so we adopted the empirical value of 2.22. Its sigma curve almost perfectly matches the ones we used when training FLUX.2 and Qwen-Image, both of which achieved excellent training results.