1、Implement cfg_trunc calculation directly using timesteps, without intermediate steps.

2、Deprecate and remove the guidance_scale parameter because it used in inference not train

3、Add inference command-line arguments --ct for cfg_trunc_ratio and --rc for renorm_cfg to control CFG truncation and renormalization during inference.

library/lumina_models.py

@@ -1081,7 +1081,7 @@ class NextDiT(nn.Module):
         cap_feats: Tensor,
         cap_mask: Tensor,
         cfg_scale: float,
-        cfg_trunc: int = 100,
+        cfg_trunc: float = 0.25,
         renorm_cfg: float = 1.0,
     ):
         """