Compatibility between nightly build and ffmpeg

[w238liu]

🐛 Describe the bug

I am trying to use the nightly build to have a taste on this feature #3332 . However, I could not figure out which ffmpeg version is compatible with the nightly build.

According to issue #3269 , I first installed ffmpeg with conda install ffmpeg=5.1.2 -c conda-forge, and then installed torchaudio by conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch-nightly -c nvidia. Then I ran the following script

import torch
import torchaudio
from torchaudio.utils import ffmpeg_utils


print(torch.__version__)
print(torchaudio.__version__)
print(ffmpeg_utils.get_versions())
print(ffmpeg_utils.get_build_config())
print([k for k in ffmpeg_utils.get_video_decoders().keys() if 'cuvid' in k])

and got the following error message

2.1.0.dev20230606
2.1.0.dev20230606
Traceback (most recent call last):
  File "/home/ubuntu/.conda/envs/torchqa_nightly/lib/python3.10/site-packages/torchaudio/_extension/utils.py", line 134, in wrapped
    _init_ffmpeg()
  File "/home/ubuntu/.conda/envs/torchqa_nightly/lib/python3.10/site-packages/torchaudio/_extension/utils.py", line 91, in _init_ffmpeg
    torchaudio.lib._torchaudio_ffmpeg.init()
RuntimeError: Error in dlopen: /lib/x86_64-linux-gnu/libgobject-2.0.so.0: undefined symbol: ffi_type_uint32, version LIBFFI_BASE_7.0
Exception raised from DynamicLibrary at /opt/conda/conda-bld/pytorch_1686036062101/work/aten/src/ATen/DynamicLibrary.cpp:38 (most recent call first):
frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x57 (0x7fe50e5c3477 in /home/ubuntu/.conda/envs/torchqa_nightly/lib/python3.10/site-packages/torch/lib/libc10.so)
frame #1: <unknown function> + 0xd9699c (0x7fe557f1899c in /home/ubuntu/.conda/envs/torchqa_nightly/lib/python3.10/site-packages/torch/lib/libtorch_cpu.so)
frame #2: torchaudio::io::detail::ffmpeg_stub() + 0x94 (0x7fe4f3cf0054 in /home/ubuntu/.conda/envs/torchqa_nightly/lib/python3.10/site-packages/torchaudio/lib/libtorchaudio_ffmpeg.so)
frame #3: <unknown function> + 0xef49 (0x7fe4f3c93f49 in /home/ubuntu/.conda/envs/torchqa_nightly/lib/python3.10/site-packages/torchaudio/lib/_torchaudio_ffmpeg.so)
frame #4: <unknown function> + 0x2beb7 (0x7fe4f3cb0eb7 in /home/ubuntu/.conda/envs/torchqa_nightly/lib/python3.10/site-packages/torchaudio/lib/_torchaudio_ffmpeg.so)
frame #5: python() [0x4fc887]
<omitting python frames>
frame #12: python() [0x592592]
frame #14: python() [0x5c32c7]
frame #15: python() [0x5be400]
frame #16: python() [0x4598ca]
frame #21: __libc_start_main + 0xf3 (0x7fe5b46f1083 in /lib/x86_64-linux-gnu/libc.so.6)
frame #22: python() [0x5854ee]


The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "/home/ubuntu/git/ssimplus-library/research/TorchQA/tmp/test_torchaudio_sr/test_torchaudio.py", line 8, in <module>
    print(ffmpeg_utils.get_versions())
  File "/home/ubuntu/.conda/envs/torchqa_nightly/lib/python3.10/site-packages/torchaudio/_extension/utils.py", line 136, in wrapped
    raise RuntimeError(
RuntimeError: get_versions requires FFmpeg extension which is not available. Please refer to the stacktrace above for how to resolve this.

I then in a new conda env installed ffmpeg 4.4.2 by running conda install -y ffmpeg=4.4.2 -c conda-forge. This time, the test script above passed. However, when I try to decode real videos, the program stopped with a Segmentation fault. Specifically, I created three test video files

ffmpeg -f lavfi -i mandelbrot -t 3 -c:v libx265 -pix_fmt yuv420p10le -vtag hvc1 -y test_hevc_hdr.mp4
ffmpeg -f lavfi -i mandelbrot -t 3 -c:v libx265 -pix_fmt yuv420p -vtag hvc1 -y test_hevc_sdr.mp4
ffmpeg -f lavfi -i mandelbrot -t 3 -c:v libx264 -pix_fmt yuv420p -vtag avc1 -y test_h264_sdr.mp4

and ran the following script in the same folder

from torchaudio.io import StreamReader
from pathlib import Path


def test_func(src: str, decoder: str, device: str = 'cpu'):
    if device == 'cuda':
        decode_config = {
            'buffer_chunk_size': 50,
            'decoder': f'{decoder}_cuvid',
            'hw_accel': 'cuda',
            "format": None,
        }
    else:
        decode_config = {
            'buffer_chunk_size': 50,
            'decoder': decoder,
            "decoder_option": {"threads": "0"},
            "format": "yuv420p",
        }

    video = StreamReader(src=src)

    video.add_basic_video_stream(1, **decode_config)

    stream = video.stream()
    frame, = next(stream)

    print(frame.device, frame.shape, frame.dtype)
    return frame


if __name__ == "__main__":
    root_dir = Path('.')
    test_videos = [
        'test_hevc_hdr.mp4',
        'test_hevc_sdr.mp4',
        'test_h264_sdr.mp4'
    ]
    decoders = [
        'hevc',
        'hevc',
        'h264'
    ]
    devices = [
        'cpu',
        'cuda'
    ]

    for test_video, decoder in zip(test_videos, decoders):
        for device in devices:
            src_path = root_dir / test_video
            test_func(str(src_path), decoder, device)

The program stopped with the following message

[W conversion.cpp:210] Warning: The output format YUV420P is selected. This will be implicitly converted to YUV444P, in which all the color components Y, U, V have the same dimension. (function operator())
Segmentation fault (core dumped)

This error didn't happen with the latest stable release. I am not sure if it's just because nightly build is not built with full functionality or there are some new code changes that I am not aware of.

Versions

For FFmpeg 5.1.2 env

Collecting environment information...
PyTorch version: 2.1.0.dev20230606
Is debug build: False
CUDA used to build PyTorch: 11.8
ROCM used to build PyTorch: N/A

OS: Ubuntu 20.04.6 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0
Clang version: Could not collect
CMake version: version 3.16.3
Libc version: glibc-2.31

Python version: 3.10.11 (main, Apr 20 2023, 19:02:41) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.15.0-1017-aws-x86_64-with-glibc2.31
Is CUDA available: True
CUDA runtime version: 11.6.124
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA A10G
Nvidia driver version: 510.73.08
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
Address sizes: 48 bits physical, 48 bits virtual
CPU(s): 8
On-line CPU(s) list: 0-7
Thread(s) per core: 2
Core(s) per socket: 4
Socket(s): 1
NUMA node(s): 1
Vendor ID: AuthenticAMD
CPU family: 23
Model: 49
Model name: AMD EPYC 7R32
Stepping: 0
CPU MHz: 2799.946
BogoMIPS: 5599.89
Hypervisor vendor: KVM
Virtualization type: full
L1d cache: 128 KiB
L1i cache: 128 KiB
L2 cache: 2 MiB
L3 cache: 16 MiB
NUMA node0 CPU(s): 0-7
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Retbleed: Mitigation; untrained return thunk; SMT enabled with STIBP protection
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf tsc_known_freq pni pclmulqdq ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm cmp_legacy cr8_legacy abm sse4a misalignsse 3dnowprefetch topoext ssbd ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 clzero xsaveerptr rdpru wbnoinvd arat npt nrip_save rdpid

Versions of relevant libraries:
[pip3] flake8==6.0.0
[pip3] mypy==1.3.0
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.24.3
[pip3] pytorch-lightning==2.0.2
[pip3] torch==2.1.0.dev20230606
[pip3] torchaudio==2.1.0.dev20230606
[pip3] torchmetrics==0.11.4
[pip3] torchqa==0.2.1
[pip3] torchvision==0.16.0.dev20230606
[pip3] triton==2.1.0
[conda] blas 1.0 mkl
[conda] mkl 2023.1.0 h6d00ec8_46342
[conda] mkl-service 2.4.0 py310h5eee18b_1
[conda] mkl_fft 1.3.6 py310h1128e8f_1
[conda] mkl_random 1.2.2 py310h1128e8f_1
[conda] numpy 1.24.3 py310h5f9d8c6_1
[conda] numpy-base 1.24.3 py310hb5e798b_1
[conda] pytorch 2.1.0.dev20230606 py3.10_cuda11.8_cudnn8.7.0_0 pytorch-nightly
[conda] pytorch-cuda 11.8 h7e8668a_5 pytorch-nightly
[conda] pytorch-lightning 2.0.2 pypi_0 pypi
[conda] pytorch-mutex 1.0 cuda pytorch-nightly
[conda] torchaudio 2.1.0.dev20230606 py310_cu118 pytorch-nightly
[conda] torchmetrics 0.11.4 pypi_0 pypi
[conda] torchqa 0.2.1 pypi_0 pypi
[conda] torchtriton 2.1.0+9820899b38 py310 pytorch-nightly
[conda] torchvision 0.16.0.dev20230606 py310_cu118 pytorch-nightly

For FFmpeg 4.4.2 env

Collecting environment information...
PyTorch version: 2.1.0.dev20230606
Is debug build: False
CUDA used to build PyTorch: 11.8
ROCM used to build PyTorch: N/A

OS: Ubuntu 20.04.6 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0
Clang version: Could not collect
CMake version: version 3.16.3
Libc version: glibc-2.31

Python version: 3.10.11 (main, Apr 20 2023, 19:02:41) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.15.0-1017-aws-x86_64-with-glibc2.31
Is CUDA available: True
CUDA runtime version: 11.6.124
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA A10G
Nvidia driver version: 510.73.08
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
Address sizes: 48 bits physical, 48 bits virtual
CPU(s): 8
On-line CPU(s) list: 0-7
Thread(s) per core: 2
Core(s) per socket: 4
Socket(s): 1
NUMA node(s): 1
Vendor ID: AuthenticAMD
CPU family: 23
Model: 49
Model name: AMD EPYC 7R32
Stepping: 0
CPU MHz: 2799.946
BogoMIPS: 5599.89
Hypervisor vendor: KVM
Virtualization type: full
L1d cache: 128 KiB
L1i cache: 128 KiB
L2 cache: 2 MiB
L3 cache: 16 MiB
NUMA node0 CPU(s): 0-7
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Retbleed: Mitigation; untrained return thunk; SMT enabled with STIBP protection
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf tsc_known_freq pni pclmulqdq ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm cmp_legacy cr8_legacy abm sse4a misalignsse 3dnowprefetch topoext ssbd ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 clzero xsaveerptr rdpru wbnoinvd arat npt nrip_save rdpid

Versions of relevant libraries:
[pip3] numpy==1.24.3
[pip3] torch==2.1.0.dev20230606
[pip3] torchaudio==2.1.0.dev20230606
[pip3] torchvision==0.16.0.dev20230606
[pip3] triton==2.1.0
[conda] blas 1.0 mkl
[conda] mkl 2023.1.0 h6d00ec8_46342
[conda] mkl-service 2.4.0 py310h5eee18b_1
[conda] mkl_fft 1.3.6 py310h1128e8f_1
[conda] mkl_random 1.2.2 py310h1128e8f_1
[conda] numpy 1.24.3 py310h5f9d8c6_1
[conda] numpy-base 1.24.3 py310hb5e798b_1
[conda] pytorch 2.1.0.dev20230606 py3.10_cuda11.8_cudnn8.7.0_0 pytorch-nightly
[conda] pytorch-cuda 11.8 h7e8668a_5 pytorch-nightly
[conda] pytorch-mutex 1.0 cuda pytorch-nightly
[conda] torchaudio 2.1.0.dev20230606 py310_cu118 pytorch-nightly
[conda] torchtriton 2.1.0+9820899b38 py310 pytorch-nightly
[conda] torchvision 0.16.0.dev20230606 py310_cu118 pytorch-nightly

[mthrok]

The upgrade to FFmpeg 5 was reverted in #3377, due to inconsistent availability. I am still figuring out the best way to support FFmpeg.

The segfault could be a regression introduced in main branch. We don't have a good CI for GPU decoder so I might have missed something. I will try to look into it.

[w238liu]

CPU decoder also failed with the segfault, but seems like CPU decoder was tested in the CI pipeline without any error?

[mthrok]

CPU decoder also failed with the segfault, but seems like CPU decoder was tested in the CI pipeline without any error?

Yeah, and I tested it on my macbook pro, and it works fine. I have two hypothesis on this. 1 is some issue with FFmpeg you installed and 2 is the dlopen I introduced the last week.

To rule out 2, I made #3418. I will land it before tomorrow so that this feature is turned off in tomorrow's nightly build, and I would like ask you to try again and see if CPU decoder works and GPU decoder throws an error instead of segfault.

[w238liu]

CPU decoder also failed with the segfault, but seems like CPU decoder was tested in the CI pipeline without any error?

Yeah, and I tested it on my macbook pro, and it works fine. I have two hypothesis on this. 1 is some issue with FFmpeg you installed and 2 is the dlopen I introduced the last week.

To rule out 2, I made #3418. I will land it before tomorrow so that this feature is turned off in tomorrow's nightly build, and I would like ask you to try again and see if CPU decoder works and GPU decoder throws an error instead of segfault.

So I tested with the new nightly build and I still got the segfault for both both cpu and cuda decoders. See below for the test environment

Collecting environment information...
PyTorch version: 2.1.0.dev20230608
Is debug build: False
CUDA used to build PyTorch: 11.8
ROCM used to build PyTorch: N/A

OS: Ubuntu 20.04.6 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0
Clang version: Could not collect
CMake version: version 3.16.3
Libc version: glibc-2.31

Python version: 3.10.11 (main, Apr 20 2023, 19:02:41) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.15.0-1017-aws-x86_64-with-glibc2.31
Is CUDA available: True
CUDA runtime version: 11.6.124
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA A10G
Nvidia driver version: 510.73.08
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                    x86_64
CPU op-mode(s):                  32-bit, 64-bit
Byte Order:                      Little Endian
Address sizes:                   48 bits physical, 48 bits virtual
CPU(s):                          8
On-line CPU(s) list:             0-7
Thread(s) per core:              2
Core(s) per socket:              4
Socket(s):                       1
NUMA node(s):                    1
Vendor ID:                       AuthenticAMD
CPU family:                      23
Model:                           49
Model name:                      AMD EPYC 7R32
Stepping:                        0
CPU MHz:                         3002.855
BogoMIPS:                        5599.58
Hypervisor vendor:               KVM
Virtualization type:             full
L1d cache:                       128 KiB
L1i cache:                       128 KiB
L2 cache:                        2 MiB
L3 cache:                        16 MiB
NUMA node0 CPU(s):               0-7
Vulnerability Itlb multihit:     Not affected
Vulnerability L1tf:              Not affected
Vulnerability Mds:               Not affected
Vulnerability Meltdown:          Not affected
Vulnerability Mmio stale data:   Not affected
Vulnerability Retbleed:          Mitigation; untrained return thunk; SMT enabled with STIBP protection
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1:        Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:        Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling
Vulnerability Srbds:             Not affected
Vulnerability Tsx async abort:   Not affected
Flags:                           fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf tsc_known_freq pni pclmulqdq ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm cmp_legacy cr8_legacy abm sse4a misalignsse 3dnowprefetch topoext ssbd ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 clzero xsaveerptr rdpru wbnoinvd arat npt nrip_save rdpid

Versions of relevant libraries:
[pip3] numpy==1.24.3
[pip3] torch==2.1.0.dev20230608
[pip3] torchaudio==2.1.0.dev20230608
[pip3] torchvision==0.16.0.dev20230608
[pip3] triton==2.1.0
[conda] blas                      1.0                         mkl  
[conda] mkl                       2023.1.0         h6d00ec8_46342  
[conda] mkl-service               2.4.0           py310h5eee18b_1  
[conda] mkl_fft                   1.3.6           py310h1128e8f_1  
[conda] mkl_random                1.2.2           py310h1128e8f_1  
[conda] numpy                     1.24.3          py310h5f9d8c6_1  
[conda] numpy-base                1.24.3          py310hb5e798b_1  
[conda] pytorch                   2.1.0.dev20230608 py3.10_cuda11.8_cudnn8.7.0_0    pytorch-nightly
[conda] pytorch-cuda              11.8                 h7e8668a_5    pytorch-nightly
[conda] pytorch-mutex             1.0                        cuda    pytorch-nightly
[conda] torchaudio                2.1.0.dev20230608     py310_cu118    pytorch-nightly
[conda] torchtriton               2.1.0+9820899b38           py310    pytorch-nightly
[conda] torchvision               0.16.0.dev20230608     py310_cu118    pytorch-nightly

I also printed the version of ffmpeg by conda list ffmpeg, and the output is

# packages in environment at /home/ubuntu/.conda/envs/torchqa_nightly:
#
# Name                    Version                   Build  Channel
ffmpeg                    4.4.2           gpl_h8dda1f0_112    conda-forge

I also tested the latest stable release of torchaudio (2.0.2-py310_cu118) with the same version of ffmpeg, and the StreamReader works well.

[mthrok]

CPU decoder also failed with the segfault, but seems like CPU decoder was tested in the CI pipeline without any error?

Yeah, and I tested it on my macbook pro, and it works fine. I have two hypothesis on this. 1 is some issue with FFmpeg you installed and 2 is the dlopen I introduced the last week.
To rule out 2, I made #3418. I will land it before tomorrow so that this feature is turned off in tomorrow's nightly build, and I would like ask you to try again and see if CPU decoder works and GPU decoder throws an error instead of segfault.

So I tested with the new nightly build and I still got the segfault for both both cpu and cuda decoders. See below for the test environment

I also tested the latest stable release of torchaudio (2.0.2-py310_cu118) with the same version of ffmpeg, and the StreamReader works well.

Thanks for trying. That is strange. I tried your repro script (BTW thanks for the complete repro script), in Windows and it worked fine.

How did you install the ffmpeg? I see the particular build is listed as cf-staging, but I don't know how to install it. regular conda install -c conda-forge ffmpeg does not pick it.

[w238liu]

How did you install the ffmpeg? I see the particular build is listed as cf-staging, but I don't know how to install it. regular conda install -c conda-forge ffmpeg does not pick it.

I installed ffmpeg by conda install -y ffmpeg=4.4.2 -c conda-forge.

How do you normally install ffmpeg?

[mthrok]

How did you install the ffmpeg? I see the particular build is listed as cf-staging, but I don't know how to install it. regular conda install -c conda-forge ffmpeg does not pick it.

I installed ffmpeg by conda install -y ffmpeg=4.4.2 -c conda-forge.

How do you normally install ffmpeg?

I use the same command but it never picks up those packages from cf-staging.

[w238liu]

How did you install the ffmpeg? I see the particular build is listed as cf-staging, but I don't know how to install it. regular conda install -c conda-forge ffmpeg does not pick it.

I installed ffmpeg by conda install -y ffmpeg=4.4.2 -c conda-forge.
How do you normally install ffmpeg?

I use the same command but it never picks up those packages from cf-staging.

To pick up exactly the same ffmpeg, you could try conda install ffmpeg=4.4.2=gpl_h8dda1f0_112 -c conda-forge

BTW, I ran the following three commands to create my conda environemnt for the test.

conda create -n test python=3.10
conda install -y ffmpeg=4.4.2 -c conda-forge
conda install -y pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch-nightly -c nvidia

[w238liu]

@mthrok Hi, are you able to reproduce the error in the end? If no, do you need me to give you a dockerfile to reproduce the environment?

[mthrok]

@mthrok Hi, are you able to reproduce the error in the end? If no, do you need me to give you a dockerfile to reproduce the environment?

Hi - Sorry, I have not gotten the time to look into it yet. Yes, Docker-based repro would be nice. Thank you

[w238liu]

@mthrok Hi, I created a docker file named test.dockerfile as below to reproduce the segmentation fault I encountered.

FROM nvidia/cuda:11.8.0-cudnn8-runtime-ubuntu20.04

# install downloader
RUN apt update &&\
    apt -y upgrade &&\
    apt -y install wget

# install conda
ENV CONDA_DIR /opt/conda
RUN wget -P /media/ https://repo.anaconda.com/archive/Anaconda3-2023.07-1-Linux-x86_64.sh &&\
    /bin/bash /media/Anaconda3-2023.07-1-Linux-x86_64.sh -b -p ${CONDA_DIR}
ENV PATH=${CONDA_DIR}/bin:$PATH

# install torchaudio environment
RUN conda create -y -n torchenv python=3.10

RUN echo "source activate torchenv" > ~/.bashrc &&\
    echo "export PATH=/opt/conda/envs/torchenv/bin:$PATH"

SHELL [ "conda", "run", "-n", "torchenv", "/bin/bash", "-c"]
RUN conda install -y ffmpeg=4.4.2 -c conda-forge &&\
    conda install -y pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

# change working directory
WORKDIR /app

# copy test scripts and generate test data
COPY test_01.py test_02.py ./
RUN ffmpeg -f lavfi -i mandelbrot -t 3 -c:v libx265 -pix_fmt yuv420p10le -vtag hvc1 -y test_hevc_hdr.mp4 &&\
    ffmpeg -f lavfi -i mandelbrot -t 3 -c:v libx265 -pix_fmt yuv420p -vtag hvc1 -y test_hevc_sdr.mp4 &&\
    ffmpeg -f lavfi -i mandelbrot -t 3 -c:v libx264 -pix_fmt yuv420p -vtag avc1 -y test_h264_sdr.mp4

I also put the two following test scripts in the same folder of the docker file.

test_01.py

import torch
import torchaudio
from torchaudio.utils import ffmpeg_utils


print(torch.__version__)
print(torchaudio.__version__)
print(ffmpeg_utils.get_versions())
print(ffmpeg_utils.get_build_config())
print([k for k in ffmpeg_utils.get_video_decoders().keys() if 'cuvid' in k])

test_02.py

from torchaudio.io import StreamReader
from pathlib import Path


def test_func(src: str, decoder: str, device: str = 'cpu'):
    if device == 'cuda':
        decode_config = {
            'buffer_chunk_size': 50,
            'decoder': f'{decoder}_cuvid',
            'hw_accel': 'cuda',
            "format": None,
        }
    else:
        decode_config = {
            'buffer_chunk_size': 50,
            'decoder': decoder,
            "decoder_option": {"threads": "0"},
            "format": "yuv420p",
        }

    video = StreamReader(src=src)

    video.add_basic_video_stream(1, **decode_config)

    stream = video.stream()
    frame, = next(stream)

    print(frame.device, frame.shape, frame.dtype)
    return frame


if __name__ == "__main__":
    root_dir = Path(__file__).parent
    test_videos = [
        'test_hevc_hdr.mp4',
        'test_hevc_sdr.mp4',
        'test_h264_sdr.mp4'
    ]
    decoders = [
        'hevc',
        'hevc',
        'h264'
    ]
    devices = [
        'cpu',
        'cuda'
    ]

    for test_video, decoder in zip(test_videos, decoders):
        for device in devices:
            src_path = root_dir / test_video
            test_func(str(src_path), decoder, device)

Then I built the docker file in the same folder by running
docker build -t 11.8.0-cudnn8-ubuntu20.04:test -f test.dockerfile .

After the docker image is built, I started a container by
docker run -it --gpus all 11.8.0-cudnn8-ubuntu20.04:test /bin/bash

In the /app folder, I directly ran python test_01.py, and got the following

2.0.1
2.0.2
{'libavutil': (56, 70, 100), 'libavcodec': (58, 134, 100), 'libavformat': (58, 76, 100), 'libavfilter': (7, 110, 100), 'libavdevice': (58, 13, 100)}
--prefix=/opt/conda/envs/torchenv --cc=/home/conda/feedstock_root/build_artifacts/ffmpeg_1671040255947/_build_env/bin/x86_64-conda-linux-gnu-cc --cxx=/home/conda/feedstock_root/build_artifacts/ffmpeg_1671040255947/_build_env/bin/x86_64-conda-linux-gnu-c++ --nm=/home/conda/feedstock_root/build_artifacts/ffmpeg_1671040255947/_build_env/bin/x86_64-conda-linux-gnu-nm --ar=/home/conda/feedstock_root/build_artifacts/ffmpeg_1671040255947/_build_env/bin/x86_64-conda-linux-gnu-ar --disable-doc --disable-openssl --enable-avresample --enable-demuxer=dash --enable-hardcoded-tables --enable-libfreetype --enable-libfontconfig --enable-libopenh264 --enable-gnutls --enable-libmp3lame --enable-libvpx --enable-pthreads --enable-vaapi --enable-gpl --enable-libx264 --enable-libx265 --enable-libaom --enable-libsvtav1 --enable-libxml2 --enable-pic --enable-shared --disable-static --enable-version3 --enable-zlib --pkg-config=/home/conda/feedstock_root/build_artifacts/ffmpeg_1671040255947/_build_env/bin/pkg-config
['av1_cuvid', 'h264_cuvid', 'hevc_cuvid', 'mjpeg_cuvid', 'mpeg1_cuvid', 'mpeg2_cuvid', 'mpeg4_cuvid', 'vc1_cuvid', 'vp8_cuvid', 'vp9_cuvid']

which looks okay.

Then I ran python test_02.py, and it errored out with

cpu torch.Size([1, 3, 480, 640]) torch.uint8
Segmentation fault (core dumped)

This dockerfile installed the stable release torchaudio 2.0.2, and it also errored out. Probably it's not an issue of the nightly build's which I suspected before. And the same conda environment works perfectly on my ec2 machine. Probably I am missing some necessary libraries in the docker image? Do you see any problem in the dockerfile?