AMD tagged the first release of Loom, part of its ROCm HRX (HIP Runtime Extended) stack.
What it does
Loom is a compiler and IR stack that compiles kernels and device programs, links model programs, and specializes them for AMD CPUs, GPUs and NPUs. The hrx-loom SDK bundles the compiler, runtime, Python APIs and CLI tools. loom-fetch installs a model’s programs and assets, and loom-serve serves requests over HTTP. It targets AMDGPU HSA for GPUs and SPIR-V/Vulkan for portable execution, with x86 and WebAssembly still maturing. It’s on GitHub and on pip.
My take
One toolchain that covers the CPU, GPU and NPU on an AMD box is exactly what local AI on APUs needs. It’s v0.1, so expect rough edges. I’ll be watching how it compares to the back-ends people use today.
God Bless, and May the Source Be with You!
