Support pass kwargs to sd3 custom attention processor

Author: Matrix53

src/diffusers/models/attention.py

@@ -188,7 +188,11 @@ def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
         self._chunk_dim = dim
 
     def forward(
-        self, hidden_states: torch.FloatTensor, encoder_hidden_states: torch.FloatTensor, temb: torch.FloatTensor
+        self,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor,
+        temb: torch.FloatTensor,
+        joint_attention_kwargs: Dict[str, Any] = None,
     ):
         if self.use_dual_attention:
             norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp, norm_hidden_states2, gate_msa2 = self.norm1(
@@ -206,15 +210,17 @@ def forward(
 
         # Attention.
         attn_output, context_attn_output = self.attn(
-            hidden_states=norm_hidden_states, encoder_hidden_states=norm_encoder_hidden_states
+            hidden_states=norm_hidden_states,
+            encoder_hidden_states=norm_encoder_hidden_states,
+            **joint_attention_kwargs,
         )
 
         # Process attention outputs for the `hidden_states`.
         attn_output = gate_msa.unsqueeze(1) * attn_output
         hidden_states = hidden_states + attn_output
 
         if self.use_dual_attention:
-            attn_output2 = self.attn2(hidden_states=norm_hidden_states2)
+            attn_output2 = self.attn2(hidden_states=norm_hidden_states2, **joint_attention_kwargs)
             attn_output2 = gate_msa2.unsqueeze(1) * attn_output2
             hidden_states = hidden_states + attn_output2
 

src/diffusers/models/transformers/transformer_sd3.py

@@ -334,12 +334,16 @@ def custom_forward(*inputs):
                     hidden_states,
                     encoder_hidden_states,
                     temb,
+                    joint_attention_kwargs,
                     **ckpt_kwargs,
                 )
 
             else:
                 encoder_hidden_states, hidden_states = block(
-                    hidden_states=hidden_states, encoder_hidden_states=encoder_hidden_states, temb=temb
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    temb=temb,
+                    joint_attention_kwargs=joint_attention_kwargs,
                 )
 
             # controlnet residual