requirements.txt

#
# XXX: ALWAYS save a lock file when modifying requirements:
#
#     pip3 freeze -r requirements.txt > requirements.lock
#
numpy
pyyaml
wandb
gymnasium
onnx-pytorch
pygit2

ray[tune]
ray[rllib]==2.38.0
ray[default]==2.38.0

# ray's DreamerV3 is tf-only
tensorflow
tensorflow_probability==0.22.1

# Nightly ray builds
#ray @ https://s3-us-west-2.amazonaws.com/ray-wheels/latest/ray-3.0.0.dev0-cp312-cp312-manylinux2014_x86_64.whl
#ray[tune] @ https://s3-us-west-2.amazonaws.com/ray-wheels/latest/ray-3.0.0.dev0-cp312-cp312-manylinux2014_x86_64.whl

# Manually uncomment one of the torch packages -- see comment below
#torch==2.4.1
#torch @ https://download.pytorch.org/whl/cpu-cxx11-abi/torch-2.3.1%2Bcpu.cxx11.abi-cp312-cp312-linux_x86_64.whl

#
# Python's "torch" package is *not* compiled with CXX11 ABI flag by default.
# - on macOS there is no issue (CXX11 ABI is not used there)
# - on linux, VCMI is CXX11 ABI, and torch are incompatible, so a special cxx11.abi version of torch is needed
#
# ref. https://github.com/pytorch/pytorch/issues/51039#issuecomment-2111849591
#
# Choose which torch you want based on your needs:
#
# a) default torch package (non-CXX11 ABI)
# - macOS users should always choose this option
# - linux users should choose this option if they need CUDA (for GPU training)
#   and won't be using VCMI for loading pre-trained models to play against.
#   VCMI must be compiled with ENABLE_LIBTORCH=0 in this case.
#
# b) URL-specified torch package (CXX11 ABI)
# - linux users should choose this option if they are ok with CPU-only training
#   and want to use VCMI for loading pre-trained models to play against.
#   URL must be modified based on your python version.
#   List of available wheels (documented officially by pytorch):
#   https://download.pytorch.org/whl/cpu-cxx11-abi/torch/
#