reduce transpose operations to speedup

src/infer_compiler_registry/register_diffusers/spatio_temporal_oflow.py

@@ -366,7 +366,6 @@ def forward(
         """
         # 1. Input
         batch_frames, _, height, width = hidden_states.shape
-        hidden_states_in = hidden_states
         num_frames = image_only_indicator.shape[-1]
         batch_size = batch_frames // num_frames
 
@@ -382,9 +381,11 @@ def forward(
         #     height * width, batch_size, 1, time_context.shape[-1]
         # )
         # Rewrite for onediff SVD dynamic shape
-        broadcast_tensor = hidden_states.flatten(2, 3).permute(2, 0, 1)
         time_context = torch._C.broadcast_dim_like(
-            time_context_first_timestep[None, :], broadcast_tensor, dim=0
+            time_context_first_timestep[None, :],
+            hidden_states.flatten(2, 3),
+            dim=0,
+            like_dim=2,
         )
         # time_context = time_context.reshape(height * width * batch_size, 1, time_context.shape[-1])
         # Rewrite for onediff SVD dynamic shape
@@ -450,7 +451,9 @@ def forward(
         # hidden_states = hidden_states.reshape(batch_frames, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
         # Rewrite for onediff SVD dynamic shape
         hidden_states = (
-            hidden_states.permute(0, 2, 1).reshape_as(hidden_states_in).contiguous()
+            hidden_states.reshape_as(residual.permute(0, 2, 3, 1))
+            .permute(0, 3, 1, 2)
+            .contiguous()
         )
 
         output = hidden_states + residual