Docker build failed with cuda118 due to error with pandas package

[caph1993]

Script Mode

• Docker

Script Version

• 25.8.18

Process Mode

• Gradio

Operating System:

• Linux

Command Line used to start ebook2audiobook

• docker build -t athomasson2/ebook2audiobook . --build-arg TORCH_VERSION=cuda118

Describe the bug
Build failed due to an issue with the package pandas: FAILED: [code=1] pandas/_libs/pandas_datetime.cpython-312-x86_64-linux-gnu.so.p/src_vendored_numpy_datetime_np_datetime.c.o (See log for full details)

Note: The error occurs after pytorch with cuda 118 is installed succesfully.

Terminal log

log.txt

[ROBERT-MCDOWELL]

1 ERROR: Cannot install torch and torchvision==0.23.0.dev20250609+cu118 because these package versions have conflicting dependencies.
30.61
30.61 The conflict is caused by:
30.61 The user requested torch
30.61 torchvision 0.23.0.dev20250609+cu118 depends on torch==2.8.0.dev20250609
30.61
30.61 To fix this you could try to:
30.61 1. loosen the range of package versions you've specified
30.61 2. remove package versions to allow pip to attempt to solve the dependency conflict

I think it needs to find a way from the docker-compose to 1) uninstall all torch/torchaudio/torchvision installed by default and let install the right torch/udio/vision for the GPU.

[DrewThomasson]

I could of sworn I already prevent the dockerfile from using a conflicting installation from the requirments.txt by first installing the correct PyTorch version based on the TORCH_VERSION argument, and then using grep to create a temporary requirements_no_torch.txt file that strips out the conflicting torch packages before installing everything else.....

[caph1993]

You're referring to this line of the Dockerfile: grep -v -E "^torch==|^torchvision==|^torchaudio==|^torchvision$" requirements.txt > requirements_no_torch.txt && \

Well... Since it's docker, we know there is a mistake in the Dockerfile unrelated to my particular machine or environment. Maybe it's related to torchcode? I say this because I tried without docker, ignoring depedency conflicts and the app runs except for the custom voice, which fails with an error related to torchcodec.

Note. Using pipdeptree tool on the non-docker version installation produces this:

Warning!!! Possibly conflicting dependencies found:
* pyannote-audio==4.0.1
 - torch [required: >=2.8.0, installed: 2.7.1+cu118]
 - torchaudio [required: >=2.8.0, installed: 2.7.1+cu118]
* pyannote-core==6.0.1
 - numpy [required: >=2.0, installed: 1.26.4]
* pyannote-metrics==4.0.0
 - numpy [required: >=2.2.2, installed: 1.26.4]

[caph1993]

I found a solution:

How I found it?

First I used pipdeptree to find the conflict. Then I looked at the pyannote-audio releases and at the pypi dependencies for versions 3.4.0, 4.0.0 and 4.0.1. It seems that the latest version of pyannote-audio that supports torch==2.7.1 is pyannote-audio=3.4.0.

What's the solution?**

If you need cuda118, like I do, replace pyannote-audio with pyannote-audio==3.4.0 in the requirements.txt file. I also added torch==2.7.1 at the end of the file*. Then, run docker build -t athomasson2/ebook2audiobook . --build-arg TORCH_VERSION=cuda118 and finally docker run --rm --gpus all -p 7860:7860 athomasson2/ebook2audiobook.

*I didn't check if both changes are strictly needed or maybe only the first one or only the second one.
**This is a manual solution. I'm not sure what's the cleanest way to integrate it into the Dockerfile.

[ROBERT-MCDOWELL]

the issue with new version of torch is since version 2.9 they removed completely torchaudio/torchvision to replace with torchcodec...

[DrewThomasson]

they changed the names?? :0

so do i have to use diffrent names for diffrent torch versions or....

[ROBERT-MCDOWELL]

yes :-.
https://github.com/meta-pytorch/torchcodec
https://pytorch.org/blog/pytorch-2-9/
pytorch/audio#3902

[ROBERT-MCDOWELL]

better to lock the torch/torchaudio version to 2.8

[caph1993]

But remember that for cuda118, torch 2.7.1 is needed.

[ROBERT-MCDOWELL]

But remember that for cuda118, torch 2.7.1 is needed.

yes we know...

[DrewThomasson]

.... This sounds... complex...
Perhaps we should just lock it as the build arguments to only [cuda121, cuda118, cuda128, rocm, xpu, cpu] with pre-set pip installs

and take away the All CUDA version numbers should work, Ex: CUDA 11.6-> cuda116.... thing in the dockerfile

[ROBERT-MCDOWELL]

it seems that numpy > 1 is installed. it should be < 2 since for now we are stuck with coqui-tts 0.26.0 requirin numpy 1.x

[ROBERT-MCDOWELL]

try to replace requires.txt with this:

regex
tqdm
cutlet
deep_translator
docker
ebooklib
fastapi
num2words
argostranslate
beautifulsoup4
fugashi
sudachipy
sudachidict_core
ray
unidic
pymupdf4llm
translate
hangul-romanize
indic-nlp-library
iso-639
jieba
pycantonese
soynlp
pypinyin
pythainlp
mutagen
PyOpenGL
nvidia-ml-py
phonemizer-fork
pydub
pyannote-audio==3.4.0
demucs==4.0.1
gradio>=5.49
transformers==4.51.3
coqui-tts[languages]==0.26.0
torch==2.7.1
torchaudio==2.7.1
torchvggish

[DrewThomasson]

How does this look?

#1105

I simplified it and made it be specific if its a cuda 11.8 torch