Ah, I see. There’s XLA.jl for TPUs and a bunch of stuff from the JuliaGPU organization for GPUs — I wonder how they compare with TF/PyTorch/JAX. I’m curious to

Ah, I see. There’s XLA.jl for TPUs and a bunch of stuff from the JuliaGPU organization for GPUs — I wonder how they compare with TF/PyTorch/JAX. I’m curious to get perspectives on what remains to be done to solve the missing piece, and what are the biggest challenges to overcome (assuming one gets access to the appropriate low level APIs on top of the hardware — like OpenAI’s Triton).