diff --git a/comfy/ldm/kandinsky5/model.py b/comfy/ldm/kandinsky5/model.py
index 064591aa3..6436dd304 100644
--- a/comfy/ldm/kandinsky5/model.py
+++ b/comfy/ldm/kandinsky5/model.py
@@ -30,15 +30,14 @@ class TimeEmbeddings(nn.Module):
         self.model_dim = model_dim
         self.max_period = max_period
         self.register_buffer("freqs", get_freqs(model_dim // 2, max_period), persistent=False)
-        self.dtype = operation_settings.get("dtype")
         operations = operation_settings.get("operations")
-        self.in_layer = operations.Linear(model_dim, time_dim, bias=True, device=operation_settings.get("device"), dtype=self.dtype)
+        self.in_layer = operations.Linear(model_dim, time_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
         self.activation = nn.SiLU()
-        self.out_layer = operations.Linear(time_dim, time_dim, bias=True, device=operation_settings.get("device"), dtype=self.dtype)
+        self.out_layer = operations.Linear(time_dim, time_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
 
     def forward(self, timestep):
         args = torch.outer(timestep, self.freqs.to(device=timestep.device))
-        time_embed = torch.cat([torch.cos(args), torch.sin(args)], dim=-1).to(self.dtype)
+        time_embed = torch.cat([torch.cos(args), torch.sin(args)], dim=-1).to(torch.bfloat16) #todo dtype
         time_embed = self.out_layer(self.activation(self.in_layer(time_embed)))
         return time_embed