HRX Loom Library

The official collection of model implementations and reusable components for Loom. Use it to serve models, adapt an implementation to your workload, or build a model of your own. The catalog covers text generation, image generation, speech recognition and synthesis, and text embeddings.

Each model includes its Loom programs, a machine-readable model card, supported checkpoint variants, serving presets, and correctness checks. Shared programs, kernels, and motifs provide building blocks for other models and applications. The HRX System SDK supplies the compiler, runtime, and tools; this repository supplies the implementations. This is an inference library, not a training framework or checkpoint mirror. Weights are fetched from their publishers under their respective licenses. A checkpoint alone is not a Loom implementation: the model's programs must support its architecture, format, and execution target.

Run a model

Install the hrx-loom SDK with its optional [hf] extra using the installation guide, which covers Linux and Windows prerequisites. With a supported AMD GPU, fetch Moonshine Streaming Tiny and start its speech-recognition service:

loom-fetch https://raw.githubusercontent.com/ROCm/hrx-loom-library/main/model/moonshine-ai/moonshine-streaming-tiny/model.json --preset=streaming --output=./moonshine
loom-serve --model=./moonshine/model.json --device=amdgpu --listen=127.0.0.1:8080

No repository clone or Bazel build is needed. Fetch copies the selected programs, verifies locked assets, and reuses cached weights. The SDK compiles the programs when serving. AMD device access uses libamdf and the operating-system driver; it does not require a ROCm SDK installation.

Continue with the Moonshine request example to transcribe a recording, or choose another model. The fetch-to-serve guide covers discovery, gated downloads, offline use, and custom serving settings. Each model's overview owns its hardware requirements, supported inputs, and complete request examples.

Make it yours

Loom models can live anywhere. This repository is the supported official set, not a registry that every Loom model must join. Fork and remix the library, or publish an independent repository using the same model-card format and public Loom build rules. A model card and its programs can be hosted on GitHub, Hugging Face, or another HTTPS host and used directly with loom-fetch.

• Author or adapt a model: start from a fork, build and test a package, iterate with a local compiler, and publish your own card.
• Develop reusable components: placement, production/test declarations, dependencies, and local checks.
• Contribute to the official library: scope and review expectations. Model additions are coordinated with maintainers while the library's organization and contribution workflow are being established.