Giles Bathgate

If I apply this change I can run on a single GPU: ```diff --- a/train.py +++ b/train.py @@ -413,7 +413,7 @@ def subprocess_fn(rank, args, temp_dir): dnnlib.util.Logger(file_name=os.path.join(args.run_dir, 'log.txt'), file_mode='a', should_flush=True) #...

@jkla139 I actually used the copy from https://github.com/Zhendong-Wang/Diffusion-GAN . I cloned from the main branch of this repo and got: ``` :$ python train.py --gpus=1 ... ... File "/home/giles/projects/insgen/training/training_loop.py", line...

> I basically agree but I am wondering if we can go further. Usually we divide by the size of the pointee. @sloriot any idea how to achieve that. Because...

@afabri The trouble with this approach I think is that the hashmap doesn't check the entries for equality, only the "hashes" of the entries. This works for handle addresses because...

@sloriot I am unsure about the right approach. I can close this draft if you want but I would rather leave it until the correct solution is merged.

When you say "(too much)", is it worst case `O(n)` or something else? (I use Big-O to refer to memory not speed) Do you think it could be made more...

> I made it of size 2 * number of elements to store in case that each leaf node stores only one element, in order to be sure that no...

> I missed that. You are right that the `2 *` is then not needed. I think I should explain why I made that change. The K3_tree is mostly intended...

This commit [26571bf](https://github.com/CGAL/cgal/pull/6478/commits/26571bf3549aa553e7febae5c819c18d6dfc710a) causes a crash in my tests, I will provide an example to re-produce.

```c++ #include #include #include #include using K = CGAL::Exact_predicates_exact_constructions_kernel; using P = CGAL::Polyhedron_3; using N = CGAL::Nef_polyhedron_3; using V = CGAL::Vector_3; int main(int argc, char* argv[]) { std::ifstream cube("cube.off"); //...