From 61b5c5fc75eec4de5ec782b741880ef09410d2f9 Mon Sep 17 00:00:00 2001
From: Jedrzej Kosinski <kosinkadink1@gmail.com>
Date: Thu, 28 Aug 2025 20:34:06 -0700
Subject: [PATCH] Made Mochi work with optimized_attention_override

---
 comfy/ldm/genmo/joint_model/asymm_models_joint.py | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/comfy/ldm/genmo/joint_model/asymm_models_joint.py b/comfy/ldm/genmo/joint_model/asymm_models_joint.py
index 366a8b713..5c1bb4d42 100644
--- a/comfy/ldm/genmo/joint_model/asymm_models_joint.py
+++ b/comfy/ldm/genmo/joint_model/asymm_models_joint.py
@@ -109,6 +109,7 @@ class AsymmetricAttention(nn.Module):
         scale_x: torch.Tensor,  # (B, dim_x), modulation for pre-RMSNorm.
         scale_y: torch.Tensor,  # (B, dim_y), modulation for pre-RMSNorm.
         crop_y,
+        transformer_options={},
         **rope_rotation,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         rope_cos = rope_rotation.get("rope_cos")
@@ -143,7 +144,7 @@ class AsymmetricAttention(nn.Module):
 
         xy = optimized_attention(q,
                                  k,
-                                 v, self.num_heads, skip_reshape=True)
+                                 v, self.num_heads, skip_reshape=True, transformer_options=transformer_options)
 
         x, y = torch.tensor_split(xy, (q_x.shape[1],), dim=1)
         x = self.proj_x(x)
@@ -224,6 +225,7 @@ class AsymmetricJointBlock(nn.Module):
         x: torch.Tensor,
         c: torch.Tensor,
         y: torch.Tensor,
+        transformer_options={},
         **attn_kwargs,
     ):
         """Forward pass of a block.
@@ -256,6 +258,7 @@ class AsymmetricJointBlock(nn.Module):
             y,
             scale_x=scale_msa_x,
             scale_y=scale_msa_y,
+            transformer_options=transformer_options,
             **attn_kwargs,
         )
 
@@ -524,10 +527,11 @@ class AsymmDiTJoint(nn.Module):
                                                     args["txt"],
                                                     rope_cos=args["rope_cos"],
                                                     rope_sin=args["rope_sin"],
-                                                    crop_y=args["num_tokens"]
+                                                    crop_y=args["num_tokens"],
+                                                    transformer_options=args["transformer_options"]
                                                     )
                     return out
-                out = blocks_replace[("double_block", i)]({"img": x, "txt": y_feat, "vec": c, "rope_cos": rope_cos, "rope_sin": rope_sin, "num_tokens": num_tokens}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": y_feat, "vec": c, "rope_cos": rope_cos, "rope_sin": rope_sin, "num_tokens": num_tokens, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 y_feat = out["txt"]
                 x = out["img"]
             else:
@@ -538,6 +542,7 @@ class AsymmDiTJoint(nn.Module):
                     rope_cos=rope_cos,
                     rope_sin=rope_sin,
                     crop_y=num_tokens,
+                    transformer_options=transformer_options,
                 )  # (B, M, D), (B, L, D)
         del y_feat  # Final layers don't use dense text features.