Gemlite: Towards Building Custom Low-Bit Fused CUDA Kernels (mobiusml.github.io)

This would be great to have for the Triton language as well.

Weird that they don't mention Triton? I only skimmed it, but I'm not sure what the pros and cons would be vs. Triton, which is the tool I'd use if I wanted custom quantized inference kernels.