Taichi on linux with no CUDA fails

Description

With the nightly taichi build as of today (July 21, 2025) I crash at line 117 in taichi_modules/volume_train.py with a complaint that ti.init() needs to be called.

When I added "ti.init(arch=ti.cpu)"

python main.py --taichi_ray --text "hamburger" --workspace trial

fails in site-packages/torch/utils/cpp_extension.py at "if (not os.path.exists(_join_cuda_home(extra_lib_dir))" and whinges about CUDA_HOME environment variable not set.

My install taichi runs the Julia set demo fine.

Is there something else I need to do to run stable-dreamfusion on a linux box with no NVIDIA gpu?

Steps to Reproduce

Install without cuda. Install taichi python main.py --taichi_ray --text "Hobgoblin wizard with hat" --workspace trial

Expected Behavior

stable-diffusion runs using taichi instead of cuda

Environment

Ubuntu 24.04.2 LTS Python 3.12.3 [Taichi] version 1.8.0, llvm 15.0.4, commit a1d99c9c, linux, python 3.12.3 Torch: 2.7.1+cu126