Share how to install successful "ctcdecode"

[jeonhuhuhu]

After much struggle with installation, I am now sharing a definitive environment. This is undoubtedly a solid environment where the ctcdecode module works reliably.

"I have confirmed that the installation was successful after repeating the installation process five times consecutively in the same environment."

The installation environment is as follows:"

• OS: Ubuntu 20.04 LTS (RTX GPU, Compute Capability = 8.9), CUDA v12.0
• CUDA version: https://developer.nvidia.com/cuda-gpus

[INSTALLATION]

1. conda create -n [name] python=3.8
2. conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 pytorch-cuda=12.1 -c pytorch -c nvidia

*But, torchdata 0.7.1 requires torch>=2, but you have torch 1.12.0 which is incompatible.

[qianzhouyi2]

That works!Thank you!

[nofreewill42]

Can confirm, thanks!

([name]) lol xD

[rigved-desai]

Can confirm, thanks!

([name]) lol xD

Hello, could you please let me know what were the exact steps you followed to install this? Is it necessary to use conda here or can I use pip as well? Currently using docker to create the same env as mentioned in op.

[nofreewill42]

Hey there! I've literally just copy pasted OP's step 1 and 2 from:

[INSTALLATION]

conda create -n [name] python=3.8
conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 pytorch-cuda=12.1 -c pytorch -c nvidia

And then ran

pip install .

because that's what I saw on their terminal.

You might want to try creating a very new conda environment for this. And btw if you update some of the requirements ctcdecode has, it probably won't work anymore.

[yulrio]

Is it possible to install ctcdecode without having a Cuda GPU? Thank You

[yeeching0571]

I changed the system image to Ubuntu 22.0+ version, then ran "pip install torch==2.0.0" It's solved.