diff --git a/multitalk/multitalk.py b/multitalk/multitalk.py
index 581b43f..5922ac5 100644
--- a/multitalk/multitalk.py
+++ b/multitalk/multitalk.py
@@ -256,12 +256,13 @@ class SingleStreamAttention(nn.Module):
         N_t, N_h, N_w = shape
 
         x_extra = None
-        if x.shape[0] != encoder_hidden_states.shape[0]:
+        try:
+            x = rearrange(x, "B (N_t S) C -> (B N_t) S C", N_t=N_t)
+        except:
             x_extra = x[:, -N_h * N_w:, :]
             x = x[:, :-N_h * N_w, :]
             N_t = N_t - 1
-      
-        x = rearrange(x, "B (N_t S) C -> (B N_t) S C", N_t=N_t)
+            x = rearrange(x, "B (N_t S) C -> (B N_t) S C", N_t=N_t)
 
         # get q for hidden_state
         B, N, C = x.shape
@@ -367,7 +368,7 @@ class SingleStreamMultiAttention(SingleStreamAttention):
         if human_num is None or human_num <= 1:
             return super().forward(x, encoder_hidden_states, shape)
 
-        N_t, _, _ = shape
+        N_t, N_h, N_w = shape
         x = rearrange(x, "B (N_t S) C -> (B N_t) S C", N_t=N_t)
 
         x_extra = None
diff --git a/nodes.py b/nodes.py
index d67f7a6..7b967b7 100644
--- a/nodes.py
+++ b/nodes.py
@@ -2157,15 +2157,17 @@ class WanVideoSampler:
             rope_function = "default" #echoshot does not support comfy rope function
             log.info(f"Number of shots in prompt: {shot_num}, Shot token lengths: {shot_len}")
 
-        #region transformer settings
-
-        #blockswap init
         
         mm.unload_all_models()
         mm.soft_empty_cache()
         gc.collect()
 
-        if block_swap_args is not None and not weights_assigned:
+        #region transformer settings
+        if transformer_options is not None:
+            block_swap_args = transformer_options.get("block_swap_args", None)
+
+        #blockswap init
+        if block_swap_args is not None:
             transformer.use_non_blocking = block_swap_args.get("use_non_blocking", False)
             for name, param in transformer.named_parameters():
                 if "block" not in name:
@@ -2313,6 +2315,29 @@ class WanVideoSampler:
             for block in transformer.vace_blocks:
                 block.rope_func = rope_function
 
+        #rope
+        freqs = None
+        transformer.rope_embedder.k = None
+        transformer.rope_embedder.num_frames = None
+        if "default" in rope_function or bidirectional_sampling:
+            d = transformer.dim // transformer.num_heads
+            freqs = torch.cat([
+                rope_params(1024, d - 4 * (d // 6), L_test=latent_video_length, k=riflex_freq_index),
+                rope_params(1024, 2 * (d // 6)),
+                rope_params(1024, 2 * (d // 6))
+            ],
+            dim=1)
+        elif "comfy" in rope_function:
+            transformer.rope_embedder.k = riflex_freq_index
+            transformer.rope_embedder.num_frames = latent_video_length
+           
+        transformer.rope_func = rope_function
+        for block in transformer.blocks:
+            block.rope_func = rope_function
+        if transformer.vace_layers is not None:
+            for block in transformer.vace_blocks:
+                block.rope_func = rope_function
+
         #region model pred
         def predict_with_cfg(z, cfg_scale, positive_embeds, negative_embeds, timestep, idx, image_cond=None, clip_fea=None, 
                              control_latents=None, vace_data=None, unianim_data=None, audio_proj=None, control_camera_latents=None,