Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Small cosmos attention code refactor. (#8530)

comfyanonymous29596bd53fd1
diff --git a/comfy/ldm/cosmos/predict2.py b/comfy/ldm/cosmos/predict2.py
index 3b91b3f6e..316117f77 100644
--- a/comfy/ldm/cosmos/predict2.py
+++ b/comfy/ldm/cosmos/predict2.py
@@ -70,11 +70,7 @@ def torch_attention_op(q_B_S_H_D: torch.Tensor, k_B_S_H_D: torch.Tensor, v_B_S_H
q_B_H_S_D = rearrange(q_B_S_H_D, "b ... h k -> b h ... k").view(in_q_shape[0], in_q_shape[-2], -1, in_q_shape[-1])
k_B_H_S_D = rearrange(k_B_S_H_D, "b ... h v -> b h ... v").view(in_k_shape[0], in_k_shape[-2], -1, in_k_shape[-1])
v_B_H_S_D = rearrange(v_B_S_H_D, "b ... h v -> b h ... v").view(in_k_shape[0], in_k_shape[-2], -1, in_k_shape[-1])
- result_B_S_HD = rearrange(
- optimized_attention(q_B_H_S_D, k_B_H_S_D, v_B_H_S_D, in_q_shape[-2], skip_reshape=True, skip_output_reshape=True), "b h ... l -> b ... (h l)"
- )
-
- return result_B_S_HD
+ return optimized_attention(q_B_H_S_D, k_B_H_S_D, v_B_H_S_D, in_q_shape[-2], skip_reshape=True)
class Attention(nn.Module):