Loom SDK

hrx-loom contains the Loom compiler, runtime, model tools, and Python APIs. It compiles programs for CPUs, AMD GPUs, and AMD NPUs, and serves models through loom-serve. C++ import includes its headers and compiler support.

Install

Use Python 3.12 or newer in a virtual environment:

python -m venv .venv
# Linux:
source .venv/bin/activate
# Windows PowerShell:
# .venv\Scripts\Activate.ps1
python -m pip install hrx-loom
loom-serve --help

Pip installs the native commands and their matching libraries in the environment's bin directory on Linux or Scripts directory on Windows. Activating the environment puts the commands on PATH. Building models from Loom source uses the included JIT compiler; it does not require Bazel or a separate toolchain.

The initial platforms are Linux x86-64 with glibc 2.39 or newer and Windows 11 x86-64. Linux also needs a C++ runtime providing GLIBCXX_3.4.31; Ubuntu 24.04 with its standard libstdc++6 package satisfies these requirements. Windows needs the Microsoft Visual C++ v14 x64 Redistributable. AMD execution uses libamdf and requires a supported operating-system device driver and device access. No ROCm SDK installation is required.

Fetch and serve a model

Hugging Face acquisition uses the optional hf extra. Local files and HTTP sources work with the base package:

python -m pip install 'hrx-loom[hf]'
loom-fetch --list
loom-fetch

The release's compatible model catalog supplies the list and terminal picker. A development SDK without an associated catalog requires --catalog or a model card path or URL. Cards describe the exact files, variants, and runnable presets. For example, install the library's Moonshine Streaming Tiny model on an AMD GPU:

loom-fetch https://raw.githubusercontent.com/ROCm/hrx-loom-library/21637a18079d1f1dd9b8e20d26f700a62e60db4f/model/moonshine-ai/moonshine-streaming-tiny/model.json --preset=streaming --output=./moonshine
loom-serve --model=./moonshine/model.json --device=amdgpu

Choose a device supported by the model: task for CPU, amdgpu for AMD GPU, or xdna for AMD NPU. loom-run --list_drivers lists the installed drivers. The fetched card preserves the selected preset and can be served directly. Model weights are downloaded separately; the SDK does not contain checkpoints. The fetch-to-serve walkthrough includes the audio request, expected transcript, and offline/cache behavior.

Tools and Python APIs

The installed commands are loom-benchmark, loom-check, loom-compile, loom-compile-report, loom-export-cxx, loom-fetch, loom-format, loom-import-cxx, loom-link, loom-lint, loom-opt, loom-run, loom-serve, and loom-test. Each provides --help.

The loom Python package provides program authoring and model-card APIs using the Python standard library. Versioned JSON schemas are included with the SDK:

from importlib.resources import files

schema = files("loom.model").joinpath("model-card.v1.schema.json").read_text()

The model format and acquisition guide describes origins, variants, presets, locks, offline use, and model publication. The Loom library contains model cards, programs, kernels, and their tests. The SDK build guide describes Bazel builds and explicit Tracy, debug, and AddressSanitizer wheel profiles.