Pipeling the graph and use tvm.contrib.pipeline_executer VS. using relay.transorm.PartitionGrpah

Questions
#1

Hi all,

I want to convert the computational graph to subgraphs and then apply the optimizations. So far, I’ve seen two methods:

  1. Pipelining: from GitHub

         remove_bn_pass = tvm.transform.Sequential(
         [
             transform.InferType(),
             transform.SimplifyInference(),
             transform.FoldConstant(),
             transform.FoldScaleAxis(),
         ]
     )
     composite_partition = tvm.transform.Sequential(
         [
             remove_bn_pass,
             transform.MergeComposite(pattern_table),
             transform.AnnotateTarget("dnnl"),
             transform.PartitionGraph(bind_constants=bind_constants),
         ]
     )

 with tvm.transform.PassContext(opt_level=3, disabled_pass=["AlterOpLayout"]):
     return composite_partition(mod)

  2. Relay.transform.PartitionGrpah: from GitHub

I’m not sure what the difference is between relay.transform.PartitionGraph (partitioning the passes into regions for execution on different backends) and pipelining. In pipelining, are the partitions always interdependent?

I am learning TVM and would appreciate it if anyone could help me find the answer to my question.