Reverting encoder_mask to None for T2V models

Author: eltsai

src/maxdiffusion/models/attention_flax.py

@@ -1193,6 +1193,11 @@ def __call__(
 
     is_i2v_cross_attention = self.added_kv_proj_dim is not None and not is_self_attention
 
+    # For T2V self-attention and cross-attention, we skip passing the mask
+    # to avoid overhead, as it should be all 1s for unpadded sequences.
+    if not is_i2v_cross_attention:
+      encoder_attention_mask = None
+
     if not is_i2v_cross_attention:
       with jax.named_scope("query_proj"):
         query_proj = self.query(hidden_states)