Omnismi

Cross-vendor accelerator observability and structured diagnostics for Python apps and AI agents. Read normalized metrics, explain error evidence offline, compare performance with recorded baselines, and discover topology and affinity.

Start here: Quickstart · 中文上手指南 · Hardware compatibility · 2.0 delivery status

Install 2.0.0

Version 2.0.0 includes the Python API, CLI, offline diagnosis, perf-doctor, topology, and optional PPU and Cambricon adapters. See the changelog for migration notes and the release for wheel and source downloads.

First result without a GPU

On Linux or macOS with Python 3.9+:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade 'omnismi==2.0.0'

From a source checkout, install . in the virtual environment instead. Core has no mandatory vendor dependency. Python 3.12 is a tested starting point; hardware collection and SDK builds target Linux.

python -m omnismi decode --vendor nvidia --namespace xid --code 48

Expected: JSON with "status":"FAIL" and source-backed findings for the supplied error; exit code 2. This is a successful interpretation of error evidence, not a crash or proof that the current machine has faulty hardware. No GPU, driver, LLM or runtime network access is needed. See the quickstart for sample logs, an 80% performance comparison and exit-code handling.

Read a real device

Choose the dependency for the hardware you already have:

Omnismi does not install drivers, SDKs or PyTorch. omnismi[all] includes NVIDIA, AMD and TPU Python dependencies; PPU/MLU still require their vendor setup. Missing devices or permissions are explained by omnismi doctor; unavailable metrics stay null.

The small Python API remains compatible with 1.0:

import omnismi as omi

print(omi.count())
for device in omi.gpus():
    print(device.info())
    print(device.metrics())

Units are bytes, percent, Celsius, Watts and MHz. See API for gpu(index), caching and GPU.realtime().

Pick a workflow

The suite allocates accelerator memory and runs workloads; use it only when you intend to test the selected runtime device. It requires a compatible compute runtime. A passing self-test covers the executed checks, not whole-device health.

Validation status

CI checks Python regressions, native protocol stand-ins, built-wheel installation, quickstart commands, package metadata and documentation. A green badge is software validation, not accelerator certification. Full evidence is in the compatibility matrix and delivery status.

Development

python -m pip install -e '.[dev,docs]' build twine
python -m pytest -q
python -m mkdocs build --strict
python -m build
python -m twine check dist/*

See Contributing for installed-wheel checks and hardware evidence, Why Omnismi for the design, and MIT license.