bg-ai-model-management 1.0.1

BAUER GROUP AI model management toolkit — a swiss-army knife for AI model development and operations, starting with failsafe Hugging Face to S3 backup, verification and restore

BAUER GROUP - AI Model Management Toolkit

Today, Tomorrow, Together | Building Better Software Together

A swiss-army knife for AI model development and operations. aimm is a toolkit, not a single script: one core provides configuration precedence, secret redaction, retry classification, hashing and a stable exit-code contract, and each tool mounts on top of it as a Typer sub-app.

Tool #1 is aimm hf-backup: it backs up Hugging Face repositories to S3-compatible storage and can prove afterwards that the bytes are still the right bytes.

The three names

This trips people up, so it is stated up front. They are deliberately different and none of them is a typo:

Thing	Value
Distribution name (what you pip install)	bg-ai-model-management
Import package (what you import)	bg_ai_model_management
CLI command (what you type)	aimm

The CLI name is also the plugin contract: tools mount through the setuptools entry-point group aimm.tools, every environment variable is prefixed AIMM_, and the profile file is aimm.yaml. Those five identifiers are public API and do not change.

What hf-backup does

• Pins before it reads. A revision is resolved to an immutable 40-character commit SHA once, and every subsequent enumeration and transfer runs against that SHA. A push to main mid-run cannot produce a torn snapshot.
• Three transfer paths, chosen per file. inline (a single PutObject from RAM), stream (Hugging Face straight into a hand-rolled multipart upload, no disk at all) and disk (staged download, for the cases where streaming is not safe). --mode forces either extreme.
• Integrity you can check, not assume. LFS files are anchored on Hugging Face's own sha256; non-LFS files on the git blob id. Both go into a per-revision manifest that is itself digest-protected.
• Restore never contacts Hugging Face. The manifest is the only authority, so the backup survives the upstream repo being deleted, renamed or gated.
• Retention that refuses to do the stupid thing. The newest complete revision is always kept, a revision a ref points at is protected, and a plan with no survivors is refused outright.
• Exit codes are a scheduling interface. 0 clean, 20 drift found, 6 corrupt. See below — this is the part that matters for cron.
• Backend-agnostic. MinIO, Ceph RGW, AWS S3, Cloudflare R2 and Wasabi are equal-ranked presets, and a runtime capability probe beats the preset table.

Install

pip install bg-ai-model-management

Python 3.12 or newer. See docs/installation.md.

60-second quickstart against MinIO

Start a MinIO, point aimm at it, back a real model up, verify it, restore it.

# 1. A throwaway MinIO on localhost:9000
docker run -d --name minio -p 9000:9000 \
  -e MINIO_ROOT_USER=minioadmin -e MINIO_ROOT_PASSWORD=minioadmin \
  minio/minio server /data

# 2. Credentials come from the environment, never from a flag
export AIMM_S3__ACCESS_KEY_ID=minioadmin
export AIMM_S3__SECRET_ACCESS_KEY=minioadmin
export AIMM_S3__ENDPOINT_URL=http://localhost:9000
export AIMM_S3__BUCKET=hf-backup

# 3. Check that everything is reachable before moving bytes
aimm hf-backup doctor --preset minio --ensure-bucket

# 4. Back up a real (tiny) model repository
aimm hf-backup sync hf-internal-testing/tiny-random-gpt2 --preset minio

# 5. Prove it landed intact
aimm hf-backup verify hf-internal-testing/tiny-random-gpt2 --preset minio --level deep

# 6. Bring it back to local disk — this step never talks to Hugging Face
aimm hf-backup restore hf-internal-testing/tiny-random-gpt2 \
  --dest ./restored --preset minio

--ensure-bucket is opt-in on purpose and is only appropriate for a scratch bucket. In production, buckets are provisioned as infrastructure; the tool should not be able to create them.

Longer walkthrough: docs/quickstart.md.

Commands

Global options apply to every command: --profile, --log-level, --log-format, --json, --run-id, --no-color, --version.

Command	What it does
aimm tools	List the mounted tools
aimm hf-backup sync	Back up one or more repositories at a pinned commit
aimm hf-backup verify	Check a stored revision against its manifest (quick, deep, upstream)
aimm hf-backup restore	Materialise a stored revision on local disk
aimm hf-backup prune	Delete revisions the retention policy no longer covers
aimm hf-backup catalog list	List repositories stored under the prefix
aimm hf-backup catalog revisions	List the stored revisions of one repository
aimm hf-backup catalog show	Show one manifest, after verifying its digest
aimm hf-backup doctor	Probe the environment; print resolved settings with secrets masked

Every flag is documented in docs/cli.md.

Exit codes

Every command returns a stable exit code. Three of them carry the operational meaning:

Code	Meaning
0	success — nothing to report
20	differences found. verify did its job and found drift or an incomplete revision. This is a finding, not a crash
6	integrity failure. Stored bytes do not hash to what the manifest says. Escalate

A monitoring job that treats "non-zero" as one bucket will page someone at 03:00 for a 20 that only needed a re-sync, and will under-react to a 6 that means data loss. The full table is in docs/troubleshooting.md.

Adding tool #2

The extension seam is a setuptools entry point that yields a typer.Typer. That is the whole mechanism — no registry, no base class, no plugin framework:

[project.entry-points."aimm.tools"]
hf-backup = "bg_ai_model_management.tools.hfbackup.cli:app"
my-tool   = "my_package.cli:app"

The loader mounts every registered entry point at startup, and a tool that fails to import is logged and skipped rather than taking aimm --help down with it. A tool can live in a different distribution package entirely and still mount. See docs/development.md.

Documentation

Document	Contents
Installation	supported Pythons, venv advice, verifying the install
Quickstart	first successful sync in under five minutes
Configuration	profile YAML, backends, every environment variable, the precedence chain
CLI reference	every command, every flag, every exit code
Transfer strategy	inline / stream / disk, how the path is chosen, tuning
Integrity	what is verified against what, and the honest limits
Backends	MinIO, Ceph RGW, AWS, R2, Wasabi — presets and probes
Operations	scheduling, retention, disaster recovery, monitoring
Troubleshooting	symptom to cause to fix, keyed on exit codes
Development	dev setup, the MinIO rig, how to add tool #2
Architecture	module map, data flows, concurrency, security model
Decision records	ADRs 0001-0006, including the reversals

License

MIT © BAUER GROUP