Open Cookiee235 opened 2 months ago
I can run the test case and reproduce the error, but the error message seems correct for the test case. The first argument to Module.multiply_by_two
is a tensor, but the first item of R.call_tir
's argument tuple is a tuple. This could be caught earlier by the well-formed checker, when updated to validate the R.call_tir
arguments.
(As a side-note, replacing (args,)
with args
would have the correct struct info, but wouldn't be an in-line relax Tuple as required by R.call_tir
. See the discussion in https://github.com/apache/tvm/pull/15916 for more detail on the requirement for an in-line tuple.)
It seems the provided Relax IRs are valid, however, it crashed when was compiled using
relax.build()
unexpectedly.Actual behavior
Steps to reproduce
cc @Lunderberg @junrushao