What are the tested dependency versions for installation?

[robertartress]

pip install . fails with dependency conflicts. What versions of torch, torchvision, torchaudio, xformers, and torchcodec are known to work together for this?

[lematt1991]

I'm not able to reproduce this issue. I ran the following:

conda create -n sam-audio-env python=3.11 ffmpeg -y
conda activate sam-audio-env
pip install .

And didn't get any version conflicts.

[harpreetsahota204]

i get the same issues, it would be helpful if you can pin versions in the requirements

[harpreetsahota204]

It took a while to figure out, but i got the env set up. hopefully this helps someone

 Create new with Python 3.11
conda create -n sam_audio python=3.11 -y
conda activate sam_audio

# Install FFmpeg
conda install -c conda-forge "ffmpeg<8" -y

# Install PyTorch with CUDA 12.8 (or whatever version works for you, check nvidia-smi)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

# Install core dependencies (note: transformers>=4.54.0)
pip install "numpy>=1.24.0,<2.0" protobuf==3.20.3 scipy scikit-learn "transformers>=4.54.0" einops pydub torchdiffeq safetensors

# Install torchcodec
pip install torchcodec

# Install git dependencies
pip install git+https://github.com/facebookresearch/dacvae.git git+https://github.com/facebookresearch/ImageBind.git git+https://github.com/lematt1991/CLAP.git git+https://github.com/facebookresearch/perception_models@unpin-deps

# Install final packages
pip install audiobox_aesthetics git+https://github.com/facebookresearch/sam-audio#egg=sam_audio