nvFuser (MosaicML fusion compiler, now under NVIDIA)

Provenance matters here and is the reason this record exists separately from the existing torch-compile-torchinductor record: nvFuser is MosaicML's fusion compiler, announced by the PyTorch foundation in 2022 and later donated to NVIDIA, which now hosts NVIDIA/Fuser. It is a graph optimizer in the schema's sense - it fuses groups of eager PyTorch operations into generated CUDA kernels rather than exposing a language to write kernels in. The documented flow is analyze the ops running on GPU, plan parallelization, apply the strategies in generated code, runtime-compile the generated functions, then execute those CUDA kernels on subsequent iterations. autotuning is false: nvFuser applies its own planned strategies, it does not search a configuration space per kernel the way @triton.autotune does. Its optimisation is whole-subgraph, which is a different axis from per-kernel tuning, and conflating the two is the usual source of confused claims. For a planner the practical point is the coupling, not the codegen: the wheel-to-PyTorch-version binding and the [torch] extra's PyTorch replacement behaviour mean nvFuser is a build-time choice that has to be pinned deliberately, and a nightly-index install can change the PyTorch underneath an unrelated service. compile_time_seconds is null because no primary source gives a per-kernel figure for it.