CipherFlowCipherFlow
CipherFlow · demo

Plan explorer

Input: a Transformer written in PyTorch, and a deployment (hardware, bandwidth, round-trip time). Output: CipherFlow's plan — for every operator of the model, whether it runs in homomorphic encryption (HE) or secure multi-party computation (MPC), and where the plan converts between them or refreshes a ciphertext. Every plan shown here was computed by cipherflow.compile on the PyTorch code below.

① Input

Transformer in PyTorch

Model

    

    

    

    

      

      
The traced operator graph. Each column is one encoder block (the client's input, ViT's patch projection and its final LayerNorm get their own columns); arcs and cross-column links are residual connections. Hover an operator for where it is in the code and its tensor shapes.

    

  


  

    

      
② Profile
Deployment

    

    

      
Hardware

    

    

    

      

        
        

        
        

      

      

        
        

        
        

      

    

  


  

    

      
③ Output
CipherFlow plan

      
    

    

    

      

        HE (H)
        MPC (M)
         domain change on an edge
        ⇄ conversion before this operator
        ↻ bootstrap before this operator
      

      

    

    

      

      

        

        

          Reproduce this plan
          

          
        

      

    

    

      

        
Operator by operator

        
      

      
Show

      

      Rows are in execution order. Conversions (⇄) and bootstraps (↻) run before the operator in their row, on the output of the operator they name.