Press release
RiftStack Releases Emmy, an Open-Source ML Compiler That Lowers PyTorch Graphs to CUDA in 5,000 Lines of Python
April 28, 2026
RiftStack, the team behind the CloudRift GPU cloud platform, today announced the open-source release of Emmy, the ML compiler behind its inference work. Emmy lowers PyTorch graphs to CUDA through six inspectable intermediate representations in roughly 5,000 lines of Python. It is published under the Apache-2.0 license and installs from PyPI with "pip install emmy-ml".
Most compiler stacks choose kernel schedules from fixed heuristics tuned for hardware the author had on hand. Emmy searches over Tile-IR rewrite rules instead, deriving a schedule for the specific shape and the specific GPU in front of it. Each of the six stages between Torch IR and generated CUDA can be printed and inspected, so the decisions the compiler makes are auditable rather than opaque.
“Kernel engineering is the bottleneck of the whole inference stack, and almost all of it happens behind closed doors. We are releasing Emmy because a compiler you can read is worth more to this field than another black box that claims a number. Every stage prints, so you can disagree with what it did and go look.”
Published benchmarks on consumer-class hardware are mixed by design. On FP32 SGEMM in batched mode on the RTX 5090, Emmy-generated kernels run 50 to 60 percent faster than cuBLAS, while cuBLAS still wins in non-batched mode. A fused GELU pointwise kernel reaches 4.87 times PyTorch eager. End to end on TinyLlama-1.1B, a matmul-dominated workload, Emmy remains behind eager at 0.56 times.
The source, the six-stage pipeline, and the engineering series documenting how it was built are available at https://github.com/cloudrift-ai/emmy.
About RiftStack
RiftStack builds the operating system for sovereign AI deployments: a control plane that turns datacenters into AI cloud offerings, with built-in LLM inference, GPU virtualization, and the Emmy ML compiler. RiftStack is SOC 2 certified and an NVIDIA Inception member.
Media contact: Heiko Polinski, heiko@cloudrift.ai.
All press releases