CPU only torchcodec links to Nvidia Performance Primitive libs

[sandhawalia]

🐛 Describe the bug

torchcodec==0.2.1 CPU only

ldd .venv/lib/python3.12/site-packages/torchcodec/libtorchcodec7.so | grep libnpp
        libnppig.so.12 => /lib/x86_64-linux-gnu/libnppig.so.12 (0x000071ebb8800000)
        libnppicc.so.12 => /lib/x86_64-linux-gnu/libnppicc.so.12 (0x000071ebb7c00000)
        libnppidei.so.12 => /lib/x86_64-linux-gnu/libnppidei.so.12 (0x000071ebb6c00000)
        libnppif.so.12 => /lib/x86_64-linux-gnu/libnppif.so.12 (0x000071ebb0600000)
        libnppc.so.12 => /lib/x86_64-linux-gnu/libnppc.so.12 (0x000071ebb0000000)

This requires NPP should be installed through nvidia-cuda-toolkit is this expected ? This causes a failure on CI which uses a non cuda host

Versions

Ubuntu 22.04

[NicolasHug]

Hi @sandhawalia , can you clarify how you installed torchcodec?

[sandhawalia]

Hey @NicolasHug sure, We install using uv

uv add torchcodec

with this entry in the uv.lock file

name = "torchcodec"
version = "0.2.1"
source = { registry = "https://pypi.org/simple" }
wheels = [
    { url = "https://files.pythonhosted.org/packages/02/fc/ad0931351b084c1a9840e018543d1316a1dbf6cc8b776c1d81045c2284fc/TorchCodec-0.2.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:d4b194bfd3f8cc77986e327c8a13d4eb86ef1eba860096e81117cd6b9cc64960", size = 3035140 },
    { url = "https://files.pythonhosted.org/packages/2c/e8/16093552d6381bc943bd1bfb0f27aa5c0382d8787449124d78b06213d96b/TorchCodec-0.2.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0735917480efe7c7b7ce3f1a7ccc9832faf43d085cce75cdd031fd7f8f14cbb9", size = 766567 },
    { url = "https://files.pythonhosted.org/packages/7b/af/a689c3fbc878b6080d52fb109863931746ddd5d8285e9abf606510f839b1/TorchCodec-0.2.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:c3a001722d1e9f11d867a4ebf7e428efbb0f1f699909c22f99f91a5faba35e9f", size = 2972544 },
    { url = "https://files.pythonhosted.org/packages/e9/35/f19511614be88226bf805a6594a978190232cdb9e6948da08aea6edd6720/TorchCodec-0.2.1-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:deba4158f7be4744cfb9bd2b178230f20c45a5901c2c2b2fd08d0b2a180d3b65", size = 765962 },
]

[scotts]

Curious, when I install with pip install torchcodec, I get:

ldd ~/.conda/envs/tc_test/lib/python3.12/site-packages/torchcodec/libtorchcodec7.so 
	linux-vdso.so.1 (0x00007ffdebcb5000)
	libtorch.so => not found
	libc10.so => not found
	libpython3.12.so.1.0 => not found
	libavutil.so.59 => not found
	libavcodec.so.61 => not found
	libavformat.so.61 => not found
	libavdevice.so.61 => not found
	libavfilter.so.10 => not found
	libswscale.so.8 => not found
	libtorch_cpu.so => not found
	libstdc++.so.6 => /lib64/libstdc++.so.6 (0x00007f5237e00000)
	libm.so.6 => /lib64/libm.so.6 (0x00007f5237d25000)
	libgcc_s.so.1 => /lib64/libgcc_s.so.1 (0x00007f523806e000)
	libc.so.6 => /lib64/libc.so.6 (0x00007f5237a00000)
	/lib64/ld-linux-x86-64.so.2 (0x00007f52380f9000)

We don't maintain the uv.lock file for TorchCodec, and I'm not sure who does. If you install with pip directly, do you still see the dependency?

[sandhawalia]

Hey thanks for looking into this. Yes I still do

harsimrat@aboutblank:~/workspace/lerobot$ which python
/home/harsimrat/.pyenv/shims/python
harsimrat@aboutblank:~/workspace/lerobot$ python --version
Python 3.12.4
harsimrat@aboutblank:~/workspace/lerobot$ pip --version
pip 24.0 from /home/harsimrat/.pyenv/versions/3.12.4/lib/python3.12/site-packages/pip (python 3.12)

 pip install torchcodec
Collecting torchcodec
  Downloading TorchCodec-0.2.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (11 kB)
Downloading TorchCodec-0.2.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (766 kB)
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 766.6/766.6 kB 9.9 MB/s eta 0:00:00
Installing collected packages: torchcodec
Successfully installed torchcodec-0.2.1

[notice] A new release of pip is available: 24.0 -> 25.0.1
[notice] To update, run: pip install --upgrade pip
harsimrat@aboutblank:~/workspace/lerobot$ pip show torchcodec
Name: TorchCodec
Version: 0.2.1
Summary: A video decoder for PyTorch
Home-page:
Author:
Author-email: PyTorch Team <packages@pytorch.org>
License: BSD 3-Clause License

Copyright 2024 Meta

Redistribution and use in source and binary forms, with or without modification,
are permitted provided that the following conditions are met:

1. Redistributions of source code must retain the above copyright notice,this list
of conditions and the following disclaimer.

2. Redistributions in binary form must reproduce the above copyright notice, this
list of conditions and the following disclaimer in the documentation
and/or other materials provided with the distribution.

3. Neither the name of the copyright holder nor the names of its contributors may
be used to endorse or promote products derived from this software without specific
prior written permission.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS “AS IS” AND ANY
EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT
SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
DAMAGE.

Location: /home/harsimrat/.pyenv/versions/3.12.4/lib/python3.12/site-packages
Requires:
Required-by:
ldd /home/harsimrat/.pyenv/versions/3.12.4/lib/python3.12/site-packages/torchcodec/libtorchcodec7.so | grep libnpp
	libnppig.so.12 => /lib/x86_64-linux-gnu/libnppig.so.12 (0x00007fc48fc00000)
	libnppicc.so.12 => /lib/x86_64-linux-gnu/libnppicc.so.12 (0x00007fc48f000000)
	libnppidei.so.12 => /lib/x86_64-linux-gnu/libnppidei.so.12 (0x00007fc48e000000)
	libnppif.so.12 => /lib/x86_64-linux-gnu/libnppif.so.12 (0x00007fc487a00000)
	libnppc.so.12 => /lib/x86_64-linux-gnu/libnppc.so.12 (0x00007fc487400000)

Though this issue not there for ffmpeg !=7

ldd /home/harsimrat/.pyenv/versions/3.12.4/lib/python3.12/site-packages/torchcodec/libtorchcodec4.so | grep libnpp
ldd /home/harsimrat/.pyenv/versions/3.12.4/lib/python3.12/site-packages/torchcodec/libtorchcodec5.so | grep libnpp
ldd /home/harsimrat/.pyenv/versions/3.12.4/lib/python3.12/site-packages/torchcodec/libtorchcodec6.so | grep libnpp
ldd /home/harsimrat/.pyenv/versions/3.12.4/lib/python3.12/site-packages/torchcodec/libtorchcodec7.so | grep libnpp
	libnppig.so.12 => /lib/x86_64-linux-gnu/libnppig.so.12 (0x00007da0a8600000)
	libnppicc.so.12 => /lib/x86_64-linux-gnu/libnppicc.so.12 (0x00007da0a7a00000)
	libnppidei.so.12 => /lib/x86_64-linux-gnu/libnppidei.so.12 (0x00007da0a6a00000)
	libnppif.so.12 => /lib/x86_64-linux-gnu/libnppif.so.12 (0x00007da0a0400000)
	libnppc.so.12 => /lib/x86_64-linux-gnu/libnppc.so.12 (0x00007da09fe00000)

[sandhawalia]

Some more info re: the cpu wheel

wget https://download.pytorch.org/whl/cpu/TorchCodec-0.2.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
unzip TorchCodec-0.2.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
Archive:  TorchCodec-0.2.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
  inflating: torchcodec/__init__.py
  inflating: torchcodec/_frame.py
  inflating: torchcodec/_internally_replaced_utils.py
  inflating: torchcodec/libtorchcodec4.so
  inflating: torchcodec/libtorchcodec5.so
  inflating: torchcodec/libtorchcodec6.so
  inflating: torchcodec/libtorchcodec7.so
  inflating: torchcodec/version.py
  inflating: torchcodec/_samplers/__init__.py
  inflating: torchcodec/_samplers/video_clip_sampler.py
  inflating: torchcodec/decoders/__init__.py
  inflating: torchcodec/decoders/_audio_decoder.py
  inflating: torchcodec/decoders/_decoder_utils.py
  inflating: torchcodec/decoders/_video_decoder.py
  inflating: torchcodec/decoders/_core/CMakeLists.txt
  inflating: torchcodec/decoders/_core/CPUOnlyDevice.cpp
  inflating: torchcodec/decoders/_core/CudaDevice.cpp
  inflating: torchcodec/decoders/_core/DeviceInterface.h
  inflating: torchcodec/decoders/_core/FFMPEGCommon.cpp
  inflating: torchcodec/decoders/_core/FFMPEGCommon.h
  inflating: torchcodec/decoders/_core/VideoDecoder.cpp
  inflating: torchcodec/decoders/_core/VideoDecoder.h
  inflating: torchcodec/decoders/_core/VideoDecoderOps.cpp
  inflating: torchcodec/decoders/_core/VideoDecoderOps.h
  inflating: torchcodec/decoders/_core/__init__.py
  inflating: torchcodec/decoders/_core/_metadata.py
  inflating: torchcodec/decoders/_core/fetch_and_expose_non_gpl_ffmpeg_libs.cmake
  inflating: torchcodec/decoders/_core/video_decoder_ops.py
  inflating: torchcodec/samplers/__init__.py
  inflating: torchcodec/samplers/_common.py
  inflating: torchcodec/samplers/_index_based.py
  inflating: torchcodec/samplers/_time_based.py
  inflating: TorchCodec-0.2.1.dist-info/LICENSE
  inflating: TorchCodec-0.2.1.dist-info/METADATA
  inflating: TorchCodec-0.2.1.dist-info/WHEEL
  inflating: TorchCodec-0.2.1.dist-info/top_level.txt
  inflating: TorchCodec-0.2.1.dist-info/RECORD

ldd torchcodec/libtorchcodec7.so | grep libnpp
	libnppig.so.12 => /lib/x86_64-linux-gnu/libnppig.so.12 (0x00007db983000000)
	libnppicc.so.12 => /lib/x86_64-linux-gnu/libnppicc.so.12 (0x00007db982400000)
	libnppidei.so.12 => /lib/x86_64-linux-gnu/libnppidei.so.12 (0x00007db981400000)
	libnppif.so.12 => /lib/x86_64-linux-gnu/libnppif.so.12 (0x00007db97ae00000)
	libnppc.so.12 => /lib/x86_64-linux-gnu/libnppc.so.12 (0x00007db97a800000)