Unable to import PytorchNvCodec

[Renzzauw]

Hi!

I'm trying to use the PyTorch plugin in my project. I've installed the latest version as follows (as partly described by the comments in the SamplePytorch.py example):

pip install git+https://github.com/NVIDIA/VideoProcessingFramework.git
pip install git+https://github.com/NVIDIA/VideoProcessingFramework#subdirectory=src/PytorchNvCodec

Upon trying to import as follows:

import PyNvCodec as nvc
import PytorchNvCodec as pnvc

I get the following exceptions:

Traceback (most recent call last):
  File "/home/renzo/repos/ML-VideoEnhancement/venv/lib/python3.8/site-packages/PytorchNvCodec/__init__.py", line 18, in <module>
    from _PytorchNvCodec import * 
ImportError: /home/renzo/repos/ML-VideoEnhancement/venv/lib/python3.8/site-packages/_PytorchNvCodec.cpython-38-x86_64-linux-gnu.so: undefined symbol: _ZN2at4_ops4full4callEN3c108ArrayRefINS2_6SymIntEEERKNS2_6ScalarENS2_8optionalINS2_10ScalarTypeEEENS9_INS2_6LayoutEEENS9_INS2_6DeviceEEENS9_IbEE

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/home/renzo/repos/ML-GameService/penalty_challenge/draw/penalty_challenge_attempt_renderer.py", line 15, in <module>
    from mml_graphics.layer import VideoLayer, ImageLayer, ImageSequenceLayer
  File "/home/renzo/repos/ML-VideoEnhancement/venv/lib/python3.8/site-packages/mml_graphics/layer/__init__.py", line 4, in <module>
    from mml_graphics.layer.video_layer import VideoLayer
  File "/home/renzo/repos/ML-VideoEnhancement/venv/lib/python3.8/site-packages/mml_graphics/layer/video_layer.py", line 4, in <module>
    import PytorchNvCodec as pnvc
  File "/home/renzo/repos/ML-VideoEnhancement/venv/lib/python3.8/site-packages/PytorchNvCodec/__init__.py", line 22, in <module>
    raise RuntimeError("Failed to import native module _PytorchNvCodec! "
RuntimeError: Failed to import native module _PytorchNvCodec! Please check whether "/home/renzo/repos/ML-VideoEnhancement/venv/lib/python3.8/site-packages/PytorchNvCodec/_PytorchNvCodec.cpython-38-x86_64-linux-gnu.so" exists and can find all library dependencies (CUDA, ffmpeg).
On Unix systems, you can use `ldd` on the file to see whether it can find all dependencies.
On Windows, you can use "dumpbin /dependents" in a Visual Studio command prompt or
https://github.com/lucasg/Dependencies/releases.

When I check the files in my virtual environment site packages, the folder only contains an __init__.py file, but no .so files, so some files seem to be missing here that are being referred to by the exception.

Perhaps this is a bug during the pip installation process?

Thanks in advance for the help!

[Renzzauw]

Update: This issue does not occur when I check out an older commit (e.g. commit associated with #477), perhaps this bug got introduced in a recent commit.

[theHamsta]

Hmm, so the culprit seems to be #478. Can you try to run pip install -U pip wheel setuptools and then repeat the installation? The only thing we changed in https://github.com/NVIDIA/VideoProcessingFramework/commit/ebbf718d8bfc19bd84c9ede342060ac6c1c3c2f5 is to add this snippet to PyTorchNVCodec https://github.com/NVIDIA/VideoProcessingFramework/blob/master/src/PytorchNvCodec/pyproject.toml.

[Renzzauw]

@theHamsta thanks for your reply, I just updated those pip packages and repeated the installation, but I still get the same exceptions.

[theHamsta]

Can you see anything suspicious in the diff I linked? Would for instance deleting https://github.com/NVIDIA/VideoProcessingFramework/blob/master/src/PytorchNvCodec/pyproject.toml make a difference for you. This would be one of the few differences between the commits you said that would work and that would not work

[Renzzauw]

Hi @theHamsta, I have my suspicions that my issues are not related to the changes made in the latest commit. When I started with a fresh virtual environment, I did manage to successfully install the latest commit and import the packages with no issues. I'll investigate further if I run into any issues, but I'll close the issue for now. Thanks for the help!

[theHamsta]

https://github.com/NVIDIA/VideoProcessingFramework/blob/master/src/PytorchNvCodec/pyproject.toml seems indeed cause a problem. Since the isolate installation of torch might lead to a version mismatch of the compile-torch and the runtime-torch. If both are updated to the newest version there is no problem.

[onmygame]

I think I found the reason. When I try to reinstall PytorchNvCodec and enter pip uninstall PytorchNvCodec and I saw this shown Found existing installation: PytorchNvCodec 0.0.0 Uninstalling PytorchNvCodec-0.0.0: Would remove: /home/aque/.conda/envs/3.9/lib/python3.9/site-packages/PytorchNvCodec-0.0.0.dist-info/* /home/aque/.conda/envs/3.9/lib/python3.9/site-packages/PytorchNvCodec/* /home/aque/.conda/envs/3.9/lib/python3.9/site-packages/_PytorchNvCodec.cpython-39-x86_64-linux-gnu.so

so actually it copied the .so file into a wrong path, and after I copy it to the PytorchNvCodec path the problem solved

[burningion]

Recreated this issue with the NVIDIA Pytorch container image.

You can reproduce with the following Dockerfile:

FROM nvcr.io/nvidia/pytorch:23.05-py3
RUN git clone https://git.videolan.org/git/ffmpeg/nv-codec-headers.git
RUN cd nv-codec-headers && make install && cd -
RUN git clone https://git.ffmpeg.org/ffmpeg.git ffmpeg/ && cd ffmpeg
RUN apt-get update && apt-get install -y build-essential yasm cmake libtool libc6 libc6-dev unzip wget libnuma1 libnuma-dev
RUN cd ffmpeg && ./configure --enable-nonfree --enable-cuda-nvcc --enable-libnpp --extra-cflags=-I/usr/local/cuda/include --extra-ldflags=-L/usr/local/cuda/lib64 --disable-static --enable-shared
RUN cd ffmpeg && make -j8
RUN cd ffmpeg && make install
RUN git clone https://github.com/NVIDIA/VideoProcessingFramework
RUN apt install -y libavfilter-dev libavformat-dev libavcodec-dev libswresample-dev libavutil-dev wget build-essential 
RUN pip install --upgrade pip
RUN cd VideoProcessingFramework && pip install . 
RUN cd VideoProcessingFramework && pip install src/PytorchNvCodec
RUN mkdir /app

Running python in the container and then trying to import PytorchNvCodec generates an error:

>>> import PytorchNvCodec
Traceback (most recent call last):
  File "/usr/local/lib/python3.10/dist-packages/PytorchNvCodec/__init__.py", line 18, in <module>
    from _PytorchNvCodec import * 
ImportError: /usr/local/lib/python3.10/dist-packages/_PytorchNvCodec.cpython-310-x86_64-linux-gnu.so: undefined symbol: _ZN3c106detail14torchCheckFailEPKcS2_jRKSs

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python3.10/dist-packages/PytorchNvCodec/__init__.py", line 22, in <module>
    raise RuntimeError("Failed to import native module _PytorchNvCodec! "
RuntimeError: Failed to import native module _PytorchNvCodec! Please check whether "/usr/local/lib/python3.10/dist-packages/PytorchNvCodec/_PytorchNvCodec.cpython-310-x86_64-linux-gnu.so" exists and can find all library dependencies (CUDA, ffmpeg).
On Unix systems, you can use `ldd` on the file to see whether it can find all dependencies.
On Windows, you can use "dumpbin /dependents" in a Visual Studio command prompt or
https://github.com/lucasg/Dependencies/releases.

Again, deleting the pyproject.toml fixes the build, and makes the import work. Change the line to delete pyproject.toml before the PytorchNvCodec buidl:

RUN cd VideoProcessingFramework && rm src/PytorchNvCodec/pyproject.toml && pip install src/PytorchNvCodec

[gzchenjiajun]

@onmygame same problem .. how to solve ...

[z-mahmud22]

I think I found the reason. When I try to reinstall PytorchNvCodec and enter pip uninstall PytorchNvCodec and I saw this shown Found existing installation: PytorchNvCodec 0.0.0 Uninstalling PytorchNvCodec-0.0.0: Would remove: /home/aque/.conda/envs/3.9/lib/python3.9/site-packages/PytorchNvCodec-0.0.0.dist-info/* /home/aque/.conda/envs/3.9/lib/python3.9/site-packages/PytorchNvCodec/* /home/aque/.conda/envs/3.9/lib/python3.9/site-packages/_PytorchNvCodec.cpython-39-x86_64-linux-gnu.so

so actually it copied the .so file into a wrong path, and after I copy it to the PytorchNvCodec path the problem solved

Hi, I happened to run into this same issue. I then copied _PytorchNvCodec.cpython-39-x86_64-linux-gnu.so under my site-packages/PytorchNvCodec folder, specifically here home/user/anaconda3/envs/test_env/lib/python3.10/site-packages/PytorchNvCodec but the issue still persists!

Here is my system info:

OS: Ubuntu 22.04 on WSL2 (Windows 11)
CUDA: 11.8
Cudnn: 9.0
Torch: 2.0.1+cu117

And when I run ldd _PytorchNvCodec.cpython-310-x86_64-linux-gnu.so this is what I see:

linux-vdso.so.1 (0x00007ffcd7ffe000)
libc10.so => not found
libtorch_cpu.so => not found
libtorch_python.so => not found
libcudart.so.11.0 => /usr/local/cuda/lib64/libcudart.so.11.0 (0x00007fc806400000)
libstdc++.so.6 => /home/z84294149/anaconda3/envs/gauss_view/lib/libstdc++.so.6 (0x00007fc8061ec000)
libgcc_s.so.1 => /home/z84294149/anaconda3/envs/gauss_view/lib/libgcc_s.so.1 (0x00007fc806800000)
libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007fc805fc3000)
/lib64/ld-linux-x86-64.so.2 (0x00007fc806849000)
libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x00007fc8067fb000)
libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007fc8067f6000)
librt.so.1 => /lib/x86_64-linux-gnu/librt.so.1 (0x00007fc8067f1000)
libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007fc806708000)

[TarioYX]

I ran into the same issue, too. Finally it was resolved by python3.10 and the newest version of torch downloaded in it.