Import fast_rnnt is Failed

[jeff-lim]

Hi team,

I installed successfully with pip install fast_rnnt.
But can't import fast_rnnt (as attached img).

Here's my environment information.

• Win11 / Python 3.11
• CUDA 11.8 / torch 2.0.1

• Also I tried install from source, but The import fast_rnnt is not work with same error message.

[jeff-lim]

I think.. it is same issue to #17 . Is there're any updates or PR about this issue?

[csukuangfj]

Could you list the files in your python site-packages, especially the files whose name contains fast_rnnt?

[jeff-lim]

there're only directories... It is okay....?

[jeff-lim]

FYI tree of fast_rnnt, fast_rnnt-1.2.dist-info

fast_rnnt

fast_rnnt-1.2.dist-info

[csukuangfj]

Is there a file containing _fast_rnnt in its name?

[jeff-lim]

no.. not at all. Actually I wondering the reason why from/import _fast_rnnt in the _init_.py, mutual_information.py

init.py GitRepo
from _fast_rnnt import with_cuda

mutual_information.py GitRepo
import _fast_rnnt

[csukuangfj]

After checking pypi.org, I am afraid you cannot use "pip install fast_rnnt" on Windows at pesent.

[jeff-lim]

Oh my..
But, I also faced same issue, when I installed from source(git clone, python setup.py...), too.

Is it known issue to you/team?

• If I install and run the code on Ubuntu, I will be okay to using fast_rnnt with cuda?

[csukuangfj]

Actually I wondering the reason why from/import _fast_rnnt in the init.py, mutual_information.py

I bet you have seen torch._C. If you don't know what _C is, please google it.

Similarly, we have _fast_rnnt, which is a Python extension module, implemented in C++/CUDA.

---

Is it known issue to you/team?

I just looked at the file setup.py and find that it does not support Windows.

Please use the following commands to build fast_rnnt for windows:

git clone https://github.com/k2-fsa/fast_rnnt
cd fast_rnnt
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
cmake --build . --target _fast_rnnt --config Release

After running the above commands, please find the directory where _fast_rnnt*.pyd locates and add the directory to the environment variable PYTHONPATH.

Similarly, find the directory where mutual_information_core.dll locates and add it to the environment variable path.

Finally, add the directory /path/to/fast_rnnt/fast_rnnt/python to the environment variable PYTHONPATH.

Now you should be able to run python -c "import fast_rnnt".

[csukuangfj]

If I install and run the code on Ubuntu, I will be okay to using fast_rnnt with cuda?

Yes, I believe so.

[csukuangfj]

I think @pkufool will support installing fast_rnnt on Windows so that others won't have the same issue as yours.