Merge pull request #1025 from krahets/patch-1

Fix sinusoidal_embedding calculation for bf16 precision.

Author: Artiprocher

diffsynth/models/wan_video_dit.py

@@ -362,7 +362,7 @@ def forward(self,
                 **kwargs,
                 ):
         t = self.time_embedding(
-            sinusoidal_embedding_1d(self.freq_dim, timestep))
+            sinusoidal_embedding_1d(self.freq_dim, timestep).to(x.dtype))
         t_mod = self.time_projection(t).unflatten(1, (6, self.dim))
         context = self.text_embedding(context)