Restore LoRA step scheduling functionality

custom_linear.py

@@ -264,6 +264,11 @@ class CustomLinear(nn.Linear):
         del weight, input, bias
         return out
 
+def update_lora_step(module, step):
+    for name, submodule in module.named_modules():
+        if isinstance(submodule, CustomLinear) and hasattr(submodule, "_step"):
+            submodule._step.fill_(step)
+
 def remove_lora_from_module(module):
     for name, submodule in module.named_modules():
         if hasattr(submodule, "lora_diffs"):

wanvideo/modules/model.py

@@ -21,6 +21,7 @@ from ...utils import log, get_module_memory_mb
 from ...cache_methods.cache_methods import TeaCacheState, MagCacheState, EasyCacheState, relative_l1_distance
 from ...multitalk.multitalk import get_attn_map_with_target
 from ...echoshot.echoshot import rope_apply_z, rope_apply_c, rope_apply_echoshot
+from ...custom_linear import update_lora_step
 
 from ...MTV.mtv import apply_rotary_emb
 from comfy.ldm.flux.math import apply_rope1 as apply_rope_comfy1
@@ -2251,10 +2252,7 @@ class WanModel(torch.nn.Module):
                 ip_scale = fantasy_portrait_input.get("strength", 1.0)
 
         if self.lora_scheduling_enabled:
-            for name, submodule in self.named_modules():
+            update_lora_step(self, current_step)
-                if isinstance(submodule, nn.Linear):
-                    if hasattr(submodule, 'step'):
-                        submodule.step = current_step
 
         # lynx
         lynx_x_ip = lynx_ref_feature = lynx_ref_buffer = lynx_ref_feature_extractor = None