torch.compile fullgraph compatibility for Hunyuan Video (#11457)

udpate

Author: a-r-r-o-w

src/diffusers/models/transformers/transformer_hunyuan_video.py

@@ -1068,17 +1068,15 @@ def forward(
         latent_sequence_length = hidden_states.shape[1]
         condition_sequence_length = encoder_hidden_states.shape[1]
         sequence_length = latent_sequence_length + condition_sequence_length
-        attention_mask = torch.zeros(
+        attention_mask = torch.ones(
             batch_size, sequence_length, device=hidden_states.device, dtype=torch.bool
         )  # [B, N]
-
         effective_condition_sequence_length = encoder_attention_mask.sum(dim=1, dtype=torch.int)  # [B,]
         effective_sequence_length = latent_sequence_length + effective_condition_sequence_length
-
-        for i in range(batch_size):
-            attention_mask[i, : effective_sequence_length[i]] = True
-        # [B, 1, 1, N], for broadcasting across attention heads
-        attention_mask = attention_mask.unsqueeze(1).unsqueeze(1)
+        indices = torch.arange(sequence_length, device=hidden_states.device).unsqueeze(0)  # [1, N]
+        mask_indices = indices >= effective_sequence_length.unsqueeze(1)  # [B, N]
+        attention_mask = attention_mask.masked_fill(mask_indices, False)
+        attention_mask = attention_mask.unsqueeze(1).unsqueeze(1)  # [B, 1, 1, N]
 
         # 4. Transformer blocks
         if torch.is_grad_enabled() and self.gradient_checkpointing: