port / publish

From a GGUF on Hugging Face to a Loom package.

Your coding agent does the work. loom-skills teach it to port the model to AMD Loom, validate it against llama.cpp ROCm and write the model card from the benchmark report. Then you can publish the package to your own Hugging Face account.

01install

Install loom-skills into your agent

The skills work in Claude Code, Codex and Pi. Install all of them together, since they share scripts, then start a fresh agent session. You don't install Loom by hand: the skills install hrx-loom (the Loom compiler, runtime and loom-serve, built from ROCm/hrx-system) from PyPI into ~/loom/.venv, and download the latest llama.cpp ROCm build as the reference.

shell
$ npx skills add loom-fun/loom-skills --skill '*' -g
No Node.js? Use the Python 3.12+ installer instead. It installs the skills for all three agents; add --agent codex to pick one.
shell
$ git clone https://github.com/loom-fun/loom-skills
$ python3 loom-skills/scripts/install_skills.py
02port

Port, optimize and benchmark in one prompt

The agent ports the checkpoint (loom-port-llm/port.sh), qualifies it against llama.cpp ROCm (qualify.sh: gates, baselines, roofline), optimizes it within the model's roofline budget and finalizes it (finalize.sh: full gates, audit, report). It never stops to ask, except to accept a gated model's license.

agent prompt
> /loom Port, validate and optimize <hf-gguf>.
Budget: add a time or a target at the end, like "…for 3 hours." or "…until decode is 50 % faster than llama.cpp.". The roofline step tells you up front if a target is out of reach. In Codex the skill is $loom, and in Pi it's /skill:loom.
03upload

Generate the card and publish to Hugging Face

The agent generates info.json and README.md from the run's own logs with the loom-skills card scripts, then uploads the package: the .loom programs, model.json and model.json.lock, the tokenizer and the card. The GGUF itself isn't copied, since loom-fetch pulls it from its original repo and checks its sha256. Log in to Hugging Face once first:

shell
$ pip install -U huggingface_hub
$ hf auth login
agent prompt
> /loom Generate the Hugging Face model card for this port and upload the package to <hf-org>/<model>-loom.
Every number on the card has to come from the report. A value without evidence renders as not measured or no evidence.