https://hgpu.org/?p=28299
ACRoBat: Optimizing Auto-batching of Dynamic Deep Learning at Compile Time