Open kvndhrty opened 6 months ago
@bcharlier suggested I install via github to see if that solved my problem. I was running with the tagged 2.1.2 version of keops, installed via pip, but updated to commit dcec73cb066ecf4c13efe7612a585022123d6224 and my results are the same as before.
Hi @kvndhrty , the problems come from the kron product... I will test further how to fix that
after some investigations, this is a NotImplemented Error
: This is not possible to use ranges and batch dimension at the same time...
Hi KeOps team,
I'm moving our PyTorch layer QuadConv to keops to save memory overhead and boost speed. Thus far everything has been fairly smooth, and we see big speed increases on the CPU.
When I move this code to the GPU it throws a inscrutable CUDA error at runtime, but only if I use the ranges field of the LazyTensor before I perform the reduction. Otherwise, without the block sparsity feature this runs smoothly on the GPU.
Commenting in the second to last line makes this code die on my machine.
You can strip down G_ij to something like:
and it will run with the ranges argument, so the issue seems to be the shape of tensor G_ij and its interaction with ranges but this error only occurs on CUDA.
Any advice?