Thanks for providing the concise example - very easy to reproduce.

The root cause seems to be ilastik/lazyflow creating worker threads once imported. These will not be available to the forked processes.

Currently, there are a couple of workarounds, but I'm not sure if any of those are applicable in your case.

1. Never import anything from ilastik in your main process
   Your code example will run through if you put the import (from ilastik.experimental.api import from_project_file) into the predict_ilp function, and never call this function from your main process.
2. Don't fork, but spawn (https://docs.python.org/3/library/multiprocessing.html#contexts-and-start-methods). Your example works for me if I change it as follows (this is noticeably slower than forking):

   import numpy as np
   from multiprocessing import set_start_method
   from concurrent import futures
   from xarray import DataArray
   from ilastik.experimental.api import from_project_file   
   
   ILP = "vnc2d.ilp"
   
   
   def predict_ilp():
       ilp = from_project_file(ILP)
       inp = DataArray(np.random.rand(1, 128, 128), dims=("z", "y", "x"))
       pred = ilp.predict(inp)
       print("Passed", pred.shape)
   
   
   def predict_in_thread_or_process(pool):
       with pool(1) as ex:
           task = ex.submit(predict_ilp)
           task.result()
   
   
   if __name__ == "__main__":
       set_start_method("spawn")
   
       print("Main process")
       predict_ilp()
       print("In thread")
       predict_in_thread_or_process(futures.ThreadPoolExecutor)
       print("In process")
       predict_in_thread_or_process(futures.ProcessPoolExecutor)

3. Don't use worker threads in lazyflow. This might be what you want anyway if you're going multiprocessing...

   import numpy as np
   from concurrent import futures
   from xarray import DataArray
   import lazyflow
   from ilastik.experimental.api import from_project_file
   lazyflow.request.Request.reset_thread_pool(0)
   
   ILP = "vnc2d.ilp"
   
   
   def predict_ilp():
       ilp = from_project_file(ILP)
       inp = DataArray(np.random.rand(1, 128, 128), dims=("z", "y", "x"))
       pred = ilp.predict(inp)
       print("Passed", pred.shape)
   
   
   def predict_in_thread_or_process(pool):
       with pool(1) as ex:
           task = ex.submit(predict_ilp)
           task.result()
   
   
   print("Main process")
   predict_ilp()
   print("In thread")
   predict_in_thread_or_process(futures.ThreadPoolExecutor)
   print("In process")
   predict_in_thread_or_process(futures.ProcessPoolExecutor)

would any of those options work for you?

Thanks for looking into this @k-dominik. For my issue option 3 is the way to go and it's working now :).
It would be good to document this somehow, but feel free to close this issue.

I'd like to keep it open until we find a user friendly solution to this. I've to admit that I haven't thought about this before - so it's great that you brought it up!

This issue has been mentioned on Image.sc Forum. There might be relevant details there:

https://forum.image.sc/t/documentation-ilastik-api/83202/2