Hi. What if we support a trace node on the TVM level? It could work as an identity operation from the TVM’s point of view but have a special effect of printing a message and a value to the console (if available).
The goal may be to make programs like this possible.
tvm.compute((10,10), lambda i,j: A[i] + trace("B[%d] = %f", j, B[j]) )