This was built using the environment -
[workspace]
authors = ["Aakanksha Singh [email protected]"]
channels = ["conda-forge"]
name = "deep_ep_whl"
platforms = ["linux-64"]
version = "0.1.0"

[activation.env]
TORCH_CUDA_ARCH_LIST = "9.0"
CUDAARCHS="90-real"
CUDA_HOME="$CONDA_PREFIX/"
NCCL_HOME="$CONDA_PREFIX"
NVSHMEM_DIR="$CONDA_PREFIX"
CUDA_INCLUDE_DIRS="$CONDA_PREFIX/targets/x86_64-linux"
LD_LIBRARY_PATH = "$CONDA_PREFIX/lib:$LD_LIBRARY_PATH" # where pytorch uses CXXABI_1.3.15, so TE can use it also

[dependencies]
python = ">=3.12.11,<3.13"
pip = "*"
cudnn = ">=9.10,<9.13"
cuda-version = ">=12.8,<12.9"
cuda-compiler = ">=12.8.1,<12.9"
cuda-libraries-dev = ">=12.8.1,<12.9"
cuda-nvtx = ">=12.8,<12.9"
cuda-nvtx-dev = ">=12.8,<12.9"
libcudnn-dev = ">=9.10,<9.13"
libcusparse-dev = ">=12.5.8.93,<13"
libnvshmem-dev = "==3.3.24"
libnvshmem-static = "==3.3.24"
flax = ">=0.10.4,<0.11"
jax = ">=0.7.2,<0.8"
gh = ">=2.83.0,<3"

[pypi-dependencies]
torch = ">=2.8,<2.9"
setuptools = ">=61.0"
cmake = ">=3.21"
wheel = ""
pybind11 = { version = "", extras = ["global"] }
ninja = "*"
build = ">=1.3.0, <2"

And then just running -
python setup.py bdist_wheel