f(x) = WX + WAX + BX, where A and B are low-rank matrices bypass_forward(x) = W(X+A(X)) + B(X) bypass_forward_diff(x) = W(A(X

Apply max norm regularization to weights.

Register to modules to the subclass so that torch sees them.

### boft_bypass_forward_diff Args: x (torch.Tensor): the input tensor for original model org_out (torch.Tensor): the output tenso

### bypass_forward_diff Args: weights (tuple[torch.Tensor]): (w1, w1a, w1b, w2, w2a, w2b, t) gamma (float, optional): scale facto

### bypass_forward_diff Args: x (torch.Tensor): input tensor weights (tuple[torch.Tensor]): (w1d, w2d, w1u, w2u[, t1, t2])

### bypass_forward_diff Args: x (torch.Tensor): input tensor weights (tuple[torch.Tensor]): (down, up[, mid]) gamma (floa

### bypass_forward_diff Args: x (torch.Tensor): the input tensor for original model org_out (torch.Tensor): the output tensor fro

Clear the timestep mask after forward pass.

Compute a binary rank mask based on timestep. At high noise (high timestep): fewer ranks active (structure-level adaptation) At low nois

Return state dict for saving.

### boft_diff_weight Args: org_weight (torch.Tensor): the weight tensor of original model weights (tuple[torch.Tensor]): (oft_blo

### diff_weight Args: weights (tuple[torch.Tensor]): (w1, w1a, w1b, w2, w2a, w2b, t) gamma (float, optional): scale factor, norma

### diff_weight Get ΔW = BA, where BA is low rank decomposition Args: weights (tuple[torch.Tensor]): (down, up[, mid]) gamma

### diff_weight Args: org_weight (torch.Tensor): the weight tensor of original model weights (tuple[torch.Tensor]): (oft_blocks[,

Forward pass with timestep-dependent rank masking.

Get original weight + LoRA delta.