I’m developing a compiler with 2 unique backends besides llvm. Now TVM BYOC is mainly referred。 I have partitioned my graph into several subgraphs，which correspond to unique hardware’s computing task. These subgraphs’ input & output tensor need to be aligned in W dim and C dim. Different target has different alignment requirement. In current TVM, these tensors’ shape will be stored in graph json, and GraphRuntime will allocate memory with this shape. In GraphRuntime, I can’t figure out a certain shape’s corresponding subgraph. So, I wonder if I can figure out alignment information when optimizing graph, and store it in relay?