How to Package TVM into Another Python Wheel Package?

Questions
#1

Backgroud

After doing some development based on official TVM repository, now we need release our modified TVM to our users, we already have a bigger development tool chain, they are packaged to one Python wheel package, for example our top level tool is named “nn_builder”, because the our modified TVM is a member of “nn_builder”, so we want to put the TVM into “nn_builder”.

Problem

If the tvm directory is put inside “nn_builder”, then lots of TVM original python script will fail like something below.

----> 1 from nn_builder import tvm

~/.local/lib/python3.6/site-packages/nn_builder/tvm/__init__.py in <module>
     30 # top-level alias
     31 # tvm.runtime
---> 32 from .runtime.object import Object
     33 from .runtime.ndarray import device, cpu, cuda, gpu, opencl, cl, vulkan, metal, mtl
     34 from .runtime.ndarray import vpi, rocm, ext_dev, hexagon
~/.local/lib/python3.6/site-packages/nn_builder/tvm/runtime/__init__.py in <module>
     18 
     19 # class exposures
---> 20 from .packed_func import PackedFunc
     21 from .object import Object
     22 from .object_generic import ObjectGeneric, ObjectTypes

~/.local/lib/python3.6/site-packages/nn_builder/tvm/runtime/packed_func.py in <module>
     19 """Packed Function namespace."""
     20 import ctypes
---> 21 from tvm._ffi.base import _LIB, check_call, c_str, string_types, _FFI_MODE
     22 
     23 try:

ModuleNotFoundError: No module named 'tvm'

I think this issue is a common problem for company who use TVM as a part of their tools, is there a solution for it?

@tqchen @comaniac @junrushao Thanks.

#2

You should just need to add <path>/nn_builder/tvm/python/ to PYTHONPATH.

1 Like
#3

Thanks for replying, but it maybe unacceptable for our customers to set a specific value in PYTHONPATH after install our released python package.

The only method I can found to achieve the same goal is appending the path <path>/nn_builder to sys.path in file <path>/nn_builder/__init__.py.

Do you think there is a better way?

#4

Alternatively we could change all the

from tvm.xxx import xxx

to relative import:

from ..xxx import xxx

but it is heavy engineering time

#5

if maintaining one’s own conda channel, it won’t be a problem, because you may just release an own version of tvm there

#6

Yes, the relative import will make TVM distribute as a sub directory easy, but as you said, it really is a heavy work, in addition the disadvantage of relative(i.e., it is hard to find the real file location from the import statement) maybe will hurt TVM code.

#7

Hi @Johnson9009

had you find away to do that? i am facing the same issue , tried with python setup.py install but it actually install the tvmc to the current user env…

Thanks, Ebraheem