Skip to content

Platform support

The systems Altar runs on, what each kind of run needs, and how each execution backend, store and storage adapter was tested. Maturity labels says what each label means.

Operating systems and Python

Platform Support
Linux, x86-64 Supported. The test suites, a strict type check, and an install of every package together ran here for this release
Linux, other CPUs such as arm64 Not tested. Model images do not run there natively
macOS Supported for the Python packages. The release tests did not run on macOS. Model images need an x86-64 machine; see CPU architecture, Docker and GPUs
Windows Not supported. pysam, a core dependency, publishes no Windows wheels
Package Python
altar and every binding 3.12 or newer; 3.12 and 3.13 are tested
altar-identity 3.9 and later; tested on 3.9 to 3.13

altar-identity supports older Pythons because model runtime images import it, and most of them run Python 3.9 to 3.11. Each runtime image brings its own Python, so the Python on your machine matters only for altar and the bindings.

CPU architecture, Docker and GPUs

Task Needs
Annotate variants with annotation sources Python only. No Docker and no GPU
Score with LegNet Docker on an x86-64 (amd64) machine; runs on the CPU
Score with ChromBPNet, Cherimoya, Enformer, Borzoi or Sei Docker on an x86-64 machine, and an NVIDIA GPU
Score with AlphaGenome Python and network access to Google DeepMind's hosted API, with an API key. No Docker

The catalog lists each model's compute needs, and each model's page lists its memory and image size.

CPU architecture

Every model runtime image is built for linux/amd64 only. On Apple silicon, Docker Desktop can run these images under emulation; that has not been tested.

Docker

The local backend runs each task with the docker command, so Docker must be running and usable by your user: docker run --rm hello-world must work. On Linux, the container runs as the owner of the directory it writes to. Images are pulled from Docker Hub, pinned by digest, with no login.

GPUs

The GPU models need an NVIDIA GPU and the NVIDIA Container Toolkit. Altar's GPU validation runs used NVIDIA L40S GPUs; other NVIDIA GPUs have not been tested. On a machine without a GPU, [backend] cpu_only = true drops each task's GPU request and runs the model on the CPU, much more slowly. See backend.

Execution backends

An execution backend runs a model's container tasks. Each is named in a run spec's [backend] section.

Backend Name Label Tested with
Local Docker local Stable The execution-backend conformance suite with a fake Docker command; real Docker on Linux in the quickstart and in the model validation runs, on CPU and on NVIDIA L40S GPUs
Kubernetes kubernetes Preview The execution-backend conformance suite against a fake Kubernetes client only. Not run on a live cluster. Install altar[kubernetes]
Modal modal Preview The execution-backend conformance suite against a fake Modal client only. Not run on a live Modal account. Install altar[modal]

Run tasks on a GPU or a cluster shows how to choose one.

Stores and storage

A store keeps scores, details, artifacts or annotations. A storage adapter reads and writes files.

Adapter Name Label Tested with
SQLite score and detail stores sqlite Stable The score-store and detail-store conformance suites; the quickstart
SQLite annotation source sqlite Stable The annotation-source conformance suite
BigQuery score and detail stores, BigQuery annotation source bigquery Preview The conformance suites on DuckDB standing in for BigQuery; the read path on the BigQuery emulator. Not run against live BigQuery. Install altar[bigquery]
Artifact store on storage storage Stable The artifact-store conformance suite
Variant–gene link stores, in memory and SQLite — Stable The link-store conformance suite
Local files local Stable The storage conformance suite, including failed and interrupted writes
HTTPS downloads https Stable Its own tests against a local web server; real downloads of the LegNet checkpoint in the quickstart
fsspec filesystems fsspec Stable The storage conformance suite on memory:// and file://. Reads of the published datasets from gs:// in the quickstart, and from hf:// when each Hugging Face mirror was published. s3://, Azure and http(s):// datasets have not been tested against a live service

The fsspec adapter needs the extra for each provider: altar[gcs], altar[hf], altar[s3], altar[azure] or altar[http]. Install Altar lists them.

Install from PyPI or a source checkout

altar, altar-identity and every binding are released together, under one version number. Install and upgrade them together. Add bindings lists which bindings are on PyPI and which install from a source checkout.

Model runtime images are not Python packages. A model binding names its default image, and the local backend pulls it the first time a model runs.

Next

  • Install Altar: install the core package, the bindings and the extras a run needs.