[microtvm] cannot import tvm.micro [fixed]

Troubleshooting
#1

Hi,

My tvm version is bd24133fce184c84539461b3dc19fdcf7ad60a9e (mid june this year). My tvm stack works fine (relay tir te …).

But I cannot import micro TVM in python see below:
import tvm.micro as micro

"… raise ImportError(“micro tvm is not enabled. Set USE_MICRO to ON in config.cmake”) ImportError: micro tvm is not enabled. Set USE_MICRO to ON in config.cmake "

However, I have built TVM stack with microtvm active (see this grep): ./tvm/build/config.cmake:set(USE_MICRO ON) ./tvm/build/config.cmake:set(USE_MICRO_STANDALONE_RUNTIME ON)

And I have tvm/python in my PYTHONPATH which should be sufficient as it points to parent folder of micro: python/tvm/micro/.

Any idea how to fix this issue?

Thank you very much for your help. Gilles

#2

Hi Gilles,

I was wondering why would you get this message when building with USE_MICRO ON and it seems micro also expects you to have USE_RPC ON.

Can you check you USE_RPC setting? Hope that helps.

Leandro

#3

Hi Leandro,
Thanks for quick reply. Unfortunatelly it seems enabled: tvm$ find . -type f -name “config.cmake” | xargs grep USE_RPC ./cmake/config.cmake:set(USE_RPC ON) ./build/config.cmake:set(USE_RPC ON) Gilles

#4

So here is another thought: What error you get when running that line from .base import _rpc_connect?

Maybe for a test would be something like:

>>> from tvm.micro.base import _rpc_connect
#5

from tvm.micro.base import _rpc_connect

Traceback (most recent call last): File “”, line 1, in ImportError: cannot import name ‘_rpc_connect’ from ‘tvm.micro.base’ (tvm/python/tvm/micro/base.py)

#6

I’ve just tried it with the upstream package apache-tvm:

$ pip list | grep apache-tvm
apache-tvm        0.12.0 

$ python
Python 3.8.10 (default, May 26 2023, 14:05:08) 
[GCC 9.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from tvm.micro.base import _rpc_connect
>>>

So definitely something to do with build flags.

These are the ones we use to generate the package:

I suggest cross-checking these with yours.

#8

Hi Leandro,
Thanks again for your help.
Except Cuda and EthosN, I used your settings but it did not go any better:
set(USE_LLVM “llvm-config --ignore-libllvm --link-static”)
set(HIDE_PRIVATE_SYMBOLS ON)
set(USE_RPC ON)
set(USE_SORT ON)
set(USE_GRAPH_RUNTIME ON)
#set(USE_ETHOSN /opt/arm/ethosn-driver)
set(USE_ARM_COMPUTE_LIB /opt/arm/acl)
set(USE_MICRO ON)
set(USE_MICRO_STANDALONE_RUNTIME ON)
set(USE_ETHOSU ON)
set(USE_CMSISNN ON)

I also tried-out other settings that worked for me in the past (several years ago).

Could there be a link issue that does not prevent from building but prevents microtvm access ? (objects are built but not linked which is not harming as long as we don’t call the .so)

Also this setup is done manually on Ubuntu following install from source w/o Docker.
Can it be an issue ?
(I mean is TVM mainly installed and tested via Docker so manual installation is less tested ?)
Gilles

#9

Reinstalling same tvm commit from scratch worked. No clue what went wrong. Thank you very much for your support Leandro !

1 Like