Skip to content

README.md

Latest commit

 

History

History
55 lines (45 loc) · 2.88 KB

README.md

File metadata and controls

55 lines (45 loc) · 2.88 KB
Raw

Machine image build and qualification

bootstrap-image.sh creates a reproducible ESMFold2-pipeline installation on a fresh Ubuntu 22.04 or 24.04 GPU machine. It installs a pinned pipeline revision, the selected model checkpoints, Protenix v2 and its verified checkpoint, CUDA build and inspection tools, rclone, pigz, sqlite3, tmux, and HMMER.

Set ESMFOLD2_PIPELINE_REF to an immutable commit before building a reusable image. Other optional inputs are:

  • ESMFOLD2_IMAGE_CUDA_VERSION: 12.8 (default) or 13.0;
  • ESMFOLD2_PRELOAD_MODELS: comma- or space-separated model aliases (default: fast);
  • ESMFOLD2_IMAGE_DISK_GB: optional declared size of the provisioned image;
  • ESMFOLD2_IMAGE_MIN_ROOT_GB: optional measured root-filesystem floor (use a lower value than the declared cloud size to account for GB/GiB and filesystem overhead); and
  • ESMFOLD2_IMAGE_CLEAN_USER_CACHES=1: remove duplicate package/model caches from non-root home directories without deleting user projects.

The bootstrap writes a sanitized evidence set under /etc/esmfold2-pipeline:

  • esmfold2-pipeline-image.json: source revisions, CUDA contract, selected models, disk information, utility list, and subordinate-manifest hashes;
  • esmfold2-pipeline-sbom.json: exact Python package names and versions without source URLs or editable paths;
  • esmfold2-os-packages.txt: exact installed Debian package inventory; and
  • esmfold2-models.sha256: content hashes for every preloaded Hugging Face model blob.

Run qualify-image.sh --full before capturing a reusable image. Qualification re-hashes the evidence, OS inventory, lockfile, bootstrap script, model blobs, and Protenix checkpoint; checks every model recorded by the image manifest; and performs a real ESMFold2 GPU smoke. Full qualification additionally launches one template-enabled Protenix v2 validation for each supported modality: miniprotein, VHH, and scFv. It verifies the compiled Protenix extension's CUDA runtime linkage and native architecture targets.

Optional qualification inputs are:

  • ESMFOLD2_QUALIFICATION_OUTPUT_ROOT: local evidence/output directory;
  • ESMFOLD2_QUALIFICATION_MIN_GPU_MEMORY_MIB: enforce a site-selected minimum memory size for every visible GPU; and
  • ESMFOLD2_QUALIFICATION_REMOTE: an rclone base path for a temporary, automatically removed object-storage round trip.

The result is written to qualification.json in the output directory and to /etc/esmfold2-pipeline/esmfold2-pipeline-qualification.json. Promote an image only after the required qualification mode reports passed: true. Storage credentials are runtime inputs and must not be persisted in captured images.

CUDA 12.8 uses the cu128 PyTorch backend and cuEquivariance cu12 operator. CUDA 13.0 uses cu130 and cu13. Bootstrap removes the opposite development profile and qualification rejects mixed compiler, PyTorch, and accelerator backends.