News › ai-infrastructure
By The Linux Camp News, October 29, 2026
ROCm HRX's first tagged Loom release packages a target-specializing compiler and model-serving tools. Linux x86-64 host support should not be confused with equal readiness across execution targets.
AMD's ROCm HRX project released Loom SDK 0.1.0 on October 29, 2026, packaging its compiler, runtime, command-line tools and Python APIs for Linux and Windows x86-64 hosts. Phoronix identifies it as the project's first tagged release following open-source development. The concrete change is a versioned SDK distribution, rather than simply another description of the compiler project. 1 2 4
The maintainer guide describes a source-first compiler: program structure and device-specific facts remain connected while Loom resolves compile-time choices and links reusable components. Only then does it lower the program toward native artifacts. This ordering keeps information needed for the selected target available during specialization, rather than discarding it before target-dependent decisions are made. 3 1
For kernel authors, the useful workflow change is keeping verification, target specialization, correctness tests, benchmarks and compile reports attached to the same source. The guide says this avoids requiring a separate test language or target-specific build graph. That is a development-workflow capability, not a measured speedup: embedded benchmarks provide a way to evaluate a workload, not proof that it runs faster. 3
For serving evaluations, the documented route is to fetch a model-card preset with loom-fetch, then pass the installed card to loom-serve. The card identifies programs, parameter variants and tokenizers; the compiler specializes those programs for the selected device, and the server exposes execution over HTTP. This makes the model card the connection between acquisition, compilation and serving. 3
The release instructions retrieved October 29 specify hrx-loom version 0.1.0 from PyPI after the release workflow completes. Attached wheels offer Release, Tracy, Debug and ASAN profiles, with exact qualified artifacts recorded in SHA256SUMS. Use that artifact record to identify the build being evaluated; the conditional installation instructions alone do not establish that PyPI publication had completed. 2
Host support is not an execution-target readiness verdict. Phoronix describes AMDGPU HSA and SPIR-V/Vulkan execution paths but calls x86 and WebAssembly support still maturing. Its documentation quotation includes AMD NPUs, yet its backend summary does not identify an NPU path. That leaves NPU compatibility unresolved rather than established by the release's broad targeting description. 1 2