Wheel files built for Linux. From the 8 repos listed in requirements_post.txt.
I only built for Python3.11+CUDA12.1. I guess most Linux users would be OK with that (by using Docker).
I wrote the steps in a repo for people who want to build themselves.
I noticed a warning message while building simple-knn:
simple-knn/spatial.cu:23:64: warning: ‘T* at::Tensor::data() const [with T = float]’ is deprecated: Tensor.data() is deprecated. Please use Tensor.data_ptr() instead.
So I replaced data() with data_ptr. But I guess there's no need to commit that change in this PR. It's just a deprecation warning after all.
For test, I run ComfyUI in a new container, and it works.
(The generated obj file is OK, I opened it up in blender and a random online viewer. But the preview node in ComfyUI didn't work. I guess it's not worth the effort to fix. Just a side note)
One thing I forgot to mention, I use TORCH_CUDA_ARCH_LIST="6.1+PTX", it should be fine in terms of compatibility. (Don't know about performance, guess won't make big difference.)
Wheel files built for Linux. From the 8 repos listed in
requirements_post.txt
.I only built for Python3.11+CUDA12.1. I guess most Linux users would be OK with that (by using Docker).
I wrote the steps in a repo for people who want to build themselves.
I noticed a warning message while building simple-knn:
So I replaced
data()
withdata_ptr
. But I guess there's no need to commit that change in this PR. It's just a deprecation warning after all.For test, I run ComfyUI in a new container, and it works. (The generated obj file is OK, I opened it up in blender and a random online viewer. But the preview node in ComfyUI didn't work. I guess it's not worth the effort to fix. Just a side note)