self_attn_mask_1 = attention_mask.view(batch_size, 1, 1, seq_len).expand(-1, -1, seq_len, -1)

diffusers_helper/models/hunyuan_video_packed.py

@@ -358,7 +358,7 @@ class HunyuanVideoIndividualTokenRefiner(nn.Module):
             batch_size = attention_mask.shape[0]
             seq_len = attention_mask.shape[1]
             attention_mask = attention_mask.to(hidden_states.device).bool()
-            self_attn_mask_1 = attention_mask.view(batch_size, 1, 1, seq_len).repeat(1, 1, seq_len, 1)
+            self_attn_mask_1 = attention_mask.view(batch_size, 1, 1, seq_len).expand(-1, -1, seq_len, -1)
             self_attn_mask_2 = self_attn_mask_1.transpose(2, 3)
             self_attn_mask = (self_attn_mask_1 & self_attn_mask_2).bool()
             self_attn_mask[:, :, :, 0] = True