ROCm HRX Loom

Loom is the compiler and IR layer of AMD's ROCm HRX (HIP Runtime Extended). Its first tagged release, hrx-loom 0.1.0, landed on PyPI on October 29, 2026. One pip-installable SDK packages the Loom compiler, runtime, Python APIs and CLI tools, and can compile and serve AI models on CPUs, AMD GPUs and AMD XDNA NPUs — without installing the full ROCm SDK.

Key facts

• Targets: task (CPU), amdgpu (native AMD GPU via HSA code), xdna (AMD NPU) and vulkan (portable SPIR-V path). x86 and WebAssembly back ends are still maturing.
• Tools: loom-fetch installs a model's programs and assets from a catalog; loom-serve serves the model over HTTP; loom-benchmark measures it.
• No ROCm install: AMD execution goes through libamdf plus a supported OS driver.
• Platforms: Linux x86-64 (glibc 2.39+, e.g. Ubuntu 24.04) and Windows 11 x86-64; Python 3.12+.
• License: Apache-2.0 with LLVM exception. Model weights are downloaded separately.

Why it matters

ROCm was built mainly for the data center and has struggled on client PCs. HRX/Loom is a lighter stack aimed at client machines — especially unified-memory chips like Ryzen AI Max that have both a GPU and an XDNA NPU. It is early access and no independent benchmarks exist yet.

How to try it

python -m pip install hrx-loom
loom-run --list_drivers
loom-serve --model=<model.json> --device=amdgpu   # or task / xdna

• PyPI: hrx-loom
• Docs: Loom Programming Guide
• News: Phoronix

A model must ship programs for the chosen device — installing the SDK does not make every model run everywhere.