omnipkg 3.4.1

Python Runtime Hypervisor. Run infinite Python package & interpreter versions concurrently in one environment in milliseconds.

omnipkg — Python Runtime Hypervisor

Run infinite Python package and interpreter versions concurrently in one environment, in milliseconds.

Multi-version installation tests run every 3 hours. Live results.

---

⚡ Startup: 237µs

8pkg doesn't spawn a Python interpreter for every command. A custom C dispatcher built on uint64_t word-loads connects to the daemon socket in under 100µs.

$ hyperfine '8pkg --help' 'bun --help' 'uv --help'

8pkg --help    237.5 µs ± 30.9 µs    [2976 runs]
bun --help     683.2 µs ± 52.0 µs    [1799 runs]
uv --help        4.1 ms ±  0.1 ms     [656 runs]

  8pkg ran  2.88x faster than bun
  8pkg ran 17.12x faster than uv

--help is served from a compiled static header — zero Python. For real daemon work:

$ hyperfine '8pkg swap rich' 'uv pip install rich'  # both are no-ops, package already satisfied

8pkg swap rich       1.5 ms ± 0.1 ms    [1337 runs]
uv pip install rich  9.0 ms ± 0.2 ms    [305 runs]   (prints "Checked 1 package in 2ms" internally)

  8pkg ran 6.01x faster than uv

Both commands find the package already satisfied and do nothing. The 7.5ms gap is pure overhead that omnipkg eliminated: uv spawns a process, parses its CLI via clap, scans the environment, then tears everything down. omnipkg embeds uv as a persistent in-process library (uv-ffi) — no spawn, no CLI parse, warm site-packages cache.

For actual version switches, omnipkg intercepts uv's resolved install plan via a Rust callback before any file I/O runs. Python performs an atomic os.rename() to swap the package directories (~0.1ms), returns True to short-circuit uv's installer, then uv does ~2ms of cleanup. The version swap itself is essentially free once the plan arrives.

---

What omnipkg is

omnipkg is a persistent daemon that manages isolated Python worker processes ("bubbles"), each with its own package versions, ABI, and optionally a different Python interpreter. The C dispatcher connects to the daemon socket directly, bypassing Python startup entirely for hot paths.

You can run torch==1.13.1+cu116 and torch==2.2.0+cu121 simultaneously in the same script, pass GPU tensors between them without ever leaving VRAM, and route through Python 3.9 and 3.12 workers in the same pipeline — all without containers or separate virtualenvs.

→ Architecture · CLI reference · Getting started

---

The Impossible Pipeline: 3 PyTorch ABIs, One VRAM Buffer

The daemon's Universal CUDA IPC passes raw cudaIpcGetMemHandle pointers between workers via ctypes, using no PyTorch on the transport path. Data never crosses the PCIe bus.

$ 8pkg demo 11  # test 8: Grand Unified Benchmark

📦 Payload: 3.81 MB float32 tensor
🌊 Pipeline: torch 1.13.1+cu116 → 2.0.1+cu118 → 2.1.0+cu121

MODE 1 — Pickle + subprocess fork:          4540 ms   (baseline)
MODE 2 — CPU shared memory (zero-copy):       24 ms   (185x faster)
MODE 3 — Universal CUDA IPC (VRAM-only):    6.67 ms   (warm workers)

[WARM] GPU Stage 1: torch 1.13.1+cu116    2.40 ms
[WARM] GPU Stage 2: torch 2.0.1+cu118     2.10 ms
[WARM] GPU Stage 3: torch 2.2.0+cu121     2.17 ms
  ↳ Total: 6.67 ms  — TRUE ZERO-COPY across 3 CUDA ABIs

This also works across Python interpreter versions. Tested up to py3.9 (cu118) → py3.12 (cu130):

[WARM] py3.9  alloc + share    1.64 ms   torch=2.0.1+cu118
[WARM] py3.12 relu→norm→tanh   2.07 ms   torch=2.12.0+cu130
  ↳ Total: 3.71 ms  (757x faster than cold spawn)

Run it yourself:

8pkg demo 11

---

ABI-Aware Backend Injection

When a compiled-package worker boots, omnipkg queries its knowledge base for the numpy ABI range that was baked into that specific build, then pre-injects the correct numpy bubble automatically before any user code runs:

[DAEMON] numpy ABI range from KB: 1.20-1.23 (spec=torch==1.13.1+cu116)
[DAEMON] numpy bubble pre-injected: numpy-1.23.5  [KB-guided (1.23.5 in 1.20-1.23.*)]

No manual pinning. No ABI crash on first import. The KB accumulates this mapping for every compiled package omnipkg has ever managed, and the pattern generalizes beyond numpy — anything with a compiled ABI dependency gets the same treatment.

---

Concurrent Isolation

10 threads, 3 numpy versions, 0 corruptions:

🔥 10 threads × 3 swaps each — numpy 1.24.3 / 1.26.4 / 2.3.5 (500×500 float32)

✅ Success Rate:   30/30  (100%)
⚡ Throughput:     502 swaps/sec
🚀 Avg latency:    2.37 ms/swap
✅ Memory integrity verified — no cross-contamination

3 Python interpreters, 3 PyTorch versions, concurrent:

Sequential baseline (3.9 → 3.10 → 3.11):   3869 ms
Concurrent via daemon:                         21 ms
Speedup:                                      186x

Each universe gets its own interpreter binary and torch CUDA build. Data moves between them via shared memory with zero serialization.

---

Multi-Version Support

pip and uv can only have one version of a package active at a time. omnipkg runs conflicting versions simultaneously in isolated workers. The test matrix above runs every 3 hours against real packages on real PyPI.

---

Other Capabilities

Auto-healing runner — 8pkg run script.py detects import failures, version conflicts, and C-extension ABI errors at runtime, activates the correct bubble, and re-executes — without touching your main environment. → docs

Reproducible environments — 8pkg export snapshots your full interpreter registry + bubble state to a content-addressed TOML lock. 8pkg sync rebuilds from it. SHA is deterministic and path-free, safe to commit. → docs

Python interpreter management — omnipkg manages CPython 3.7–3.15 (including pre-releases) inside a single environment. Adopt, switch, and run them concurrently. → docs

Environment recovery — 8pkg revert restores from last-known-good snapshot when an external tool (pip, uv) damages your environment.

24-language i18n — all output is localized. Auto-detects from system locale, override with --lang or omnipkg config set language zh_CN.

---

Install

# pip / uv
pip install omnipkg
uv pip install omnipkg

# conda-forge
conda install -c conda-forge omnipkg
mamba install -c conda-forge omnipkg

# pixi
pixi add omnipkg

# Docker
docker pull 1minds3t/omnipkg:latest

Try it immediately:

8pkg demo        # interactive demo menu
8pkg demo 11     # full IPC showcase (all transports, warm/cold timing)
8pkg stress-test # concurrent chaos tests

---

Platform Support

omnipkg ships pre-compiled C dispatcher wheels using the stable ABI (cp37-abi3):

Platform	Wheel tag
Linux x86_64 (glibc 2.17+)	manylinux2014_x86_64
Linux aarch64 (glibc 2.17+)	manylinux2014_aarch64
Linux x86_64 (musl 1.2+)	musllinux_1_2_x86_64
Linux aarch64 (musl 1.2+)	musllinux_1_2_aarch64
Windows x64	win_amd64
Windows ARM64	win_arm64
macOS 11+ (Intel + Apple Silicon)	macosx_11_0_universal2
Raspberry Pi armv6l	piwheels

Python: 3.7 – 3.15, all on the same stable ABI wheel. A pure-Python fallback wheel (py3-none-any) covers anything else.

conda-forge: linux-64, linux-aarch64, linux-ppc64le, osx-64, osx-arm64, win-64 (Python 3.10–3.13)

Installation notes for specific distros

Ubuntu 24.04+ / Debian 12+ (PEP 668):

pip install --break-system-packages omnipkg
# or use a venv

Alpine Linux:

apk add --no-cache gcc python3-dev musl-dev linux-headers
pip install --break-system-packages omnipkg

Arch Linux:

pip install --break-system-packages omnipkg

Rocky/AlmaLinux 8 (ships Python 3.6):

sudo dnf install -y python39 python39-pip
python3.9 -m pip install omnipkg

---

Documentation & Interactive Console

The docs run against your local omnipkg daemon — every code block has a live "Run Command" button that executes on your machine via a local web bridge.

8pkg web start   # starts local bridge on port 5000

Then open: https://1minds3t.echo-universe.ts.net/omnipkg/

Chrome will prompt once to allow local network access — that's the CORS handshake to your daemon. No data leaves your machine.

The console includes 33+ runnable demos:

• Getting Started
• Architecture & Performance
• CLI Reference
• Platform Support
• Demos

Or skip the browser entirely:

8pkg demo        # same demos, terminal UI
8pkg demo 11     # IPC showcase directly

---

License

Dual-licensed:

• AGPLv3 — open source and academic use (LICENSE)
• Commercial — proprietary and enterprise deployment (COMMERCIAL_LICENSE.md)

Commercial inquiries: [e-mail]

---

Contributing

Bug reports, pull requests, and translation corrections are welcome.

→ Open an issue

 ________________________________________________________________
/                                                                \
| pip:    "Version conflicts? New env please!"                   |
| Docker: "Spin up containers for 45s each!"                     |
| venv:   "90s of setup for one Python version!"                 |
|                                                                |
| omnipkg: *runs 3 torch CUDA ABIs across 3 Pythons"             |
|          "through one VRAM buffer in 6.67ms*                   |
|          "Hold my multiverse!"                                 |
\________________________________________________________________/
        \   ^__^
         \  (🐍)\_______
            (__)\       )\/\
                ||----w |
                ||     ||

                ~ omnipkg: The Python Runtime Hypervisor ~