Module group_sizes 

Typical work-group sizes used by NVIDIA/AMD GPUs.

Constants

WG_64

64 threads — common on AMD RDNA and for register-heavy kernels.

WG_128

128 threads — common general-purpose choice.

WG_256

256 threads — default for many CUDA/Vulkan kernels.

WG_512

512 threads — useful for reduction passes.

WG_1024

1024 threads — maximum work-group size on most hardware.