diff --git a/nodes.py b/nodes.py
index 0e62cf6..5cbf133 100644
--- a/nodes.py
+++ b/nodes.py
@@ -4079,6 +4079,16 @@ class WanVideoSampler:
                     gen_video_samples = gen_video_samples[:, :, :-1*miss_lengths[0]]
                 
                 del noise, latent
+                if force_offload:
+                    if model["manual_offloading"]:
+                        transformer.to(offload_device)
+                        mm.soft_empty_cache()
+                        gc.collect()
+                try:
+                    print_memory(device)
+                    torch.cuda.reset_peak_memory_stats(device)
+                except:
+                    pass
                 return {"video": gen_video_samples[0].permute(1, 2, 3, 0).cpu()},
             
             #region normal inference
diff --git a/wanvideo/modules/model.py b/wanvideo/modules/model.py
index 166d8a6..bef99ab 100644
--- a/wanvideo/modules/model.py
+++ b/wanvideo/modules/model.py
@@ -1455,6 +1455,7 @@ class WanModel(ModelMixin, ConfigMixin):
 
         # MultiTalk
         if multitalk_audio is not None:
+            self.audio_proj.to(self.main_device)
             audio_cond = multitalk_audio.to(device=x.device, dtype=x.dtype)
             first_frame_audio_emb_s = audio_cond[:, :1, ...] 
             latter_frame_audio_emb = audio_cond[:, 1:, ...] 
@@ -1470,6 +1471,7 @@ class WanModel(ModelMixin, ConfigMixin):
             multitalk_audio_embedding = self.audio_proj(first_frame_audio_emb_s, latter_frame_audio_emb_s) 
             human_num = len(multitalk_audio_embedding)
             multitalk_audio_embedding = torch.concat(multitalk_audio_embedding.split(1), dim=2).to(x.dtype)
+            self.audio_proj.to(self.offload_device)
 
         # convert ref_target_masks to token_ref_target_masks
         token_ref_target_masks = None