This holds half-true.

What we decided to do is essentially just hand-rolling PTX with the agents under a harness. If there is lacking some function, we write/take a CuteDSL kernel, export PTX on compilation, transform it with the harness and then feed it into the agents.

We beat cuBLAS, and we beat all of abstractions - Triton, Helion, TLX, Mojo, ThunderKittens, cuteDSL, etc on Hopper and Blackwell for all of our workloads over 3-4 weeks.

vLLM dropped torch.compile support because they realized that they programmers were good enough to just generate the Triton kernels directly for all the passes efficiently.

If you work with this for prod the writing is on the wall sadly. The abstraction layer is just really much lower if you want full perf.