nvFuser (NVIDIA/Fuser) - a fusion code generator for NVIDIA GPUs

nvFuser repo. A fusion code generator for NVIDIA GPUs: it JIT-compiles generated fusion kernels rather than executing eager framework ops, and NVIDIA's own announcement describes the flow as analyze ops on GPU, plan parallelization, apply strategies in generated code, runtime-compile the generated GPU functions, then execute those CUDA kernels on subsequent iterations. Shipping shape matters more than the algorithm: nightly wheels are built against PyTorch:nightly and published to pypi.nvidia.com as `pip install --pre nvfuser-cuXXY --extra-index-url https://pypi.nvidia.com`, stable wheels go to both PyPI and pypi.nvidia.com, and nvFuser supports CUDA 12.6+. Two operational facts from the README worth carrying: the `[torch]` extra will REPLACE the user's existing PyTorch installation with a compatible nightly build, and it requires CUDA 12.6+ while the existing triton record notes Triton works on CC 8.0+ - the two NVIDIA kernel toolchains do not share a minimum architecture. License is NOASSERTION on the GitHub API; the LICENSE file is NVIDIA's BSD-style with bundled third-party attributions, not a standard SPDX id. Last push 2026-05-31, 406 stars, no GitHub release object. Cited for the compiler record's shape and wheel story, not for any speed claim.