Coastline

A context-aware recommender system for fine-tuning LLMs.

Coastline makes context-, objective-, and policy-aware infrastructure recommendations for LLM fine-tuning workloads. It accounts for infrastructure constraints, workload demands, and user objectives, and recommends best-fit configurations as part of an LLM fine-tuning ecosystem. Coastline predicts performance with physics-driven simulation (Kavier) and machine-learning models. Every candidate configuration is cross-checked through a feasibility module (IBM AutoConf) before being output to the user.

Built with PriorLabs-TabPFN. See the TabPFN section below.

Install

pip install coastline-recommender          # Kavier, AutoConf, the CLI, and the dashboard
pip install "coastline-recommender[ml]"    # adds the data-driven predictors

Python 3.11 to 3.13. The import name is coastline.

Quick start

From a clone of this repository:

uv sync
uv run coastline recommend-job
uv run coastline explain --model mistral-7b-v0.1 --method lora --gpu-model NVIDIA-A100-SXM4-80GB \
  --tokens 2048 --batch-size 8
uv run coastline --help

coastline recommend-job recommends a configuration for the job declared in config/coastline_functionality/experiment.yaml. The subcommands are recommend-job, recommend-trace, simulate, explain, and utils. Each documents its flags with --help. uv run coastline-ui serves the dashboard at http://127.0.0.1:8000.

Documentation

https://atlarge-research.github.io/coastline-recommender/. Build it locally with uv run --group docs mkdocs serve.

Development

uv sync installs the dev tools. CI runs these gates on every pull request (ci.yml):

uv run pre-commit run --all-files --show-diff-on-failure      # ruff check, ruff format, whitespace
uv run mypy                                                   # strict, on the packages listed in pyproject.toml
uv run --all-extras pytest --cov
uv run --all-extras pytest -m ml_isolated -p no:cacheprovider FILE  # native ML backends, one file per process
uv run --group docs mkdocs build --strict

Run uv run pre-commit install once per clone to get the hooks on commit.

TabPFN

Built with PriorLabs-TabPFN.

The tabpfn predictor and the model file portfolio/tabpfn.pkl (in src/coastline/sdk/predictors/performance/data_driven/) contain TabPFN v2 weights. TabPFN v2 is described in Hollmann et al., "Accurate predictions on small data with a tabular foundation model", Nature 637, 319-326 (2025), doi:10.1038/s41586-024-08328-6. The weights are licensed under the Prior Labs License v1.2; a copy is in LICENSE-TabPFN.txt. The model file is in the repository and its Zenodo archive, and left out of the PyPI wheel.

Citation

See CITATION.cff.

License

MIT. See LICENSE. The TabPFN weights are under the Prior Labs License v1.2, in LICENSE-TabPFN.txt.