Hf Layer Package Jobs
Apps & Automationhf-layer-package-jobsHf Layer Package Jobs is an Agent Skill for apps & automation published in the Mesh-LLM/mesh-llm repository on GitHub: a SKILL.md instruction file an AI agent can load to do this task.
Use when changing mesh-llm automation or CLI flows that discover Hugging Face GGUF models, plan CPU Hugging Face Jobs for layer-package splitting, estimate max cost, or publish skippy layer packages/catalog entries.
How to use this skill
Bring this guide into your coding agent with a prompt tailored to the tool you use.
- Open your project in Codex.
- Copy and paste the prompt below.
- Review files and risks before approving any action.
Inspect the Agent Skill at https://github.com/Mesh-LLM/mesh-llm/blob/HEAD/.agents/skills/hf-layer-package-jobs/SKILL.md and all supporting files. Report provenance, permissions, scripts, dependencies, and security risks. Propose an exact installation into .agents/skills/hf-layer-package-jobs, but wait for my approval before writing files or running scripts.
Copying this prompt does not install or run the skill. Review third-party files before use.
HF Layer Package Jobs
Use this skill for the models package CLI, the model-package crate, and the
daily Unsloth queue workflow. This skill starts after a quantized GGUF artifact
exists. It does not quantize models; use hf-gguf-quant-jobs first or
hf-quant-and-layer-package-jobs when quantization and layer packaging should
run in one job.
Workflow
- Keep model refs in colon-selector form such as
unsloth/Qwen3-8B-GGUF:Q4_K_M; do not split the quant into a separate--quantargument for generated job inputs. - Treat package submission as spend-bearing. The default behavior must be a dry run that prints the resolved package plan, effective timeout, selected HF Jobs hardware, and maximum cost. Require
--confirmbefore submitting jobs. - Splitting is CPU and I/O bound. The HF Jobs hardware does not need enough RAM or VRAM to hold the full model; use CPU hardware suitable for running the splitter/build and scale timeout/cost estimates with model file size.
- If the bucket script is stale during a confirmed submission, update it automatically before queuing jobs. Dry runs should avoid side effects.
- The GitHub workflow should default to dry run. When confirmed, it should pass
--confirm, submit at most the requested number of jobs, wait for every submitted HF Job, and fail if any job finishes unsuccessfully. - Prefer family-diverse candidate ordering after ranking by selected quant size, so one run does not consume the whole queue on a single model family.
Commands
Preview a package job:
mesh-llm models package <gguf-repo>:<quant-selector> --dry-run
Submit and follow:
mesh-llm models package <gguf-repo>:<quant-selector> --confirm --follow
Inspect jobs:
mesh-llm models package --status <job-id>
mesh-llm models package --logs <job-id>
mesh-llm models package --list
For local package certification after the artifact exists:
mesh-llm models certify <layer-package-ref> --package-only --json
Local Package Workflow
When the quantized GGUF is already available on the local machine, build the
package locally with skippy-model-package, then publish the package directory
to a Hugging Face model repo:
just build
target/debug/skippy-model-package write-package \
<org>/<gguf-repo>:<quant-selector> \
--out-dir /tmp/<model>-layers
target/debug/skippy-model-package preflight \
/tmp/<model>-layers \
--verify-sha256
hf repo create <org>/<layer-package-repo> --type model --private
hf upload <org>/<layer-package-repo> /tmp/<model>-layers . --repo-type model
For local GGUF paths outside the Hugging Face cache, include explicit provenance
flags on write-package: --model-id, --source-repo, --source-revision,
and --source-file.
Validation
Run Rust formatting and the focused package checks before committing:
cargo fmt --all -- --check
cargo test -p model-package
cargo check -p mesh-llm-host-runtime
For behavior smoke tests, use a tiny dry run first:
cargo run -p model-package --bin queue-unsloth-layer-packages -- --max-jobs 1 --recent-limit 3 --popular-limit 3 --dry-run