TVM forcing compilation with USE_MICRO ON

[philatoctoml]

@areusch

TVM is forcing compilation with USE_MICRO ON since a recent commit even though I have no reason to compile with it.

I am using the following to compile in this environment:

RUN sed -i "s|set(USE_CUDA OFF)|set(USE_CUDA ON)|g" /home/ubuntu/tvm/build/config.cmake && sed -i "s|set(USE_LLVM OFF)|set(USE_LLVM ON)|g" /home/ubuntu/tvm/build/config.cmake && sed -i "s|set(USE_CUDNN OFF)|set(USE_CUDNN ON)|g" /home/ubuntu/tvm/build/config.cmake && sed -i "s|set(USE_CUBLAS OFF)|set(USE_CUBLAS ON)|g" /home/ubuntu/tvm/build/config.cmake && sed -i "s|set(USE_GRAPH_EXECUTOR_CUDA_GRAPH OFF)|set(USE_GRAPH_EXECUTOR_CUDA_GRAPH ON)|g" /home/ubuntu/tvm/build/config.cmake && sed -i "s|set(USE_OPENMP none)|set(USE_OPENMP gnu)|g" /home/ubuntu/tvm/build/config.cmake

Traceback:

ubuntu@3c2bc3dd6016:/external$ python3.8 tvmcompile.py
Traceback (most recent call last):
  File "/home/ubuntu/tvm/python/tvm/micro/session.py", line 35, in <module>
    from .base import _rpc_connect
ImportError: cannot import name '_rpc_connect' from 'tvm.micro.base' (/home/ubuntu/tvm/python/tvm/micro/base.py)

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "tvmcompile.py", line 1, in <module>
    from tvm.driver import tvmc
  File "/home/ubuntu/tvm/python/tvm/driver/tvmc/__init__.py", line 22, in <module>
    from . import micro
  File "/home/ubuntu/tvm/python/tvm/driver/tvmc/micro.py", line 26, in <module>
    import tvm.micro.project as project
  File "/home/ubuntu/tvm/python/tvm/micro/__init__.py", line 28, in <module>
    from .session import (
  File "/home/ubuntu/tvm/python/tvm/micro/session.py", line 37, in <module>
    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

[ZQPei]

Can't reproduce your error.
If you didn't compile with USE_MICRO=ON, then don't from tvm import micro.
Would you mind to provide your tvmcompile.py code?

[philatoctoml]

@ZQPei thanks! Of course not. I just did this based on a model I had already tuned:

from tvm.driver import tvmc

model = tvmc.load('model.onnx')

log_file = "mymodel.json"
target = "cuda"

#tvmc.tune(
#    model,
#    target=target,
#    tuning_records=log_file,
#)

package = tvmc.compile(model, target=target, tuning_records=log_file)

result = tvmc.run(package, device='cuda')
print(result)

[ZQPei]

from tvm.driver import tvmc

I think I have found the root cause of your problem.
As we can see in

tvm/python/tvm/driver/tvmc/__init__.py

Line 22 in 32e8012

from . import micro

and

tvm/python/tvm/driver/tvmc/runner.py

Lines 37 to 39 in 32e8012

	import tvm.micro.project as project
	from tvm.micro.project import TemplateProjectError
	from tvm.micro.project_api.client import ProjectAPIServerNotFoundError

tvm.micro is imported in any case without a detection of USE_MICRO macro.
This bug was mistakenly introduced by #9229.

Thank you so much for reporting this bug. I will try to fix it as soon as possiable.

[ZQPei]

Thank you @philatoctoml. The bug has been fixed by #9632.

[leandron]

Thanks @ZQPei @philatoctoml and @gromero for fixing the issue!