Optimize time & memory for Wan

Author: Fabrice-TIERCELIN

src/diffusers/models/transformers/transformer_wan.py

@@ -106,7 +106,8 @@ def apply_rotary_emb(
                 freqs_cos: torch.Tensor,
                 freqs_sin: torch.Tensor,
             ):
-                x1, x2 = hidden_states.unflatten(-1, (-1, 2)).unbind(-1)
+                x1 = hidden_states[..., 0::2]
+                x2 = hidden_states[..., 1::2]
                 cos = freqs_cos[..., 0::2]
                 sin = freqs_sin[..., 1::2]
                 out = torch.empty_like(hidden_states)