Coelanox · Repository docsBrowse by category

COELANOX — Customer onboarding (model → `.cnox` → run)

Goal: Get from a trained model export to a running .cnox in your environment without vendor SSH access. Adjust paths and OS for your site.

Prerequisites

A supported source artifact (see PRODUCT_BRIEF.md and KNOWN_LIMITATIONS.md).
A x86_64 (or supported) host for the prebuilt CLI, or Rust stable to build from source.
Enough disk and RAM for packaging and inference (model-dependent).

Step 1 — Install the CLI

Option A — Binary release (no compiler):
Download the release archive for your OS from your vendor or project releases, extract, add coelanox to PATH.

Option B — Build from source:
From the repository: cd coelanox && cargo build --release -p coelanox-cli. Copy the coelanox binary to a directory on PATH.

Verify: coelanox --version

Step 2 — (Optional) Prepare signing keys

If you use Ed25519 signing:

coelanox keygen -o ./keys

Store keys/secret.bin in a secrets manager; distribute keys/public.bin only where verification should occur.

Step 3 — Package the model

ONNX example:

coelanox package -i /path/to/model.onnx -o /path/to/model.cnox --target cpu

If packaging fails with Custom ops: See ONNX_SUPPORTED_OPS.md and fix the graph or translator.

Scalar-only / portable container:

coelanox package -i model.onnx -o model.cnox --target cpu --fallback-only

Demo bundles (ResNet tiny, etc.) use --use-demo-translator and format flags per Quickstart.

Step 4 — Verify before you ship

coelanox verify -f /path/to/model.cnox

Expected: success message. If hash fails: do not deploy; re-copy or re-build.

With signing:

coelanox verify -f model.cnox --trusted-key keys/public.bin

Step 5 — Inspect the container

coelanox info -f model.cnox

Confirm input/output shapes, sizes, and flags (e.g. audit required).

Step 6 — Run inference locally

Prepare input JSON matching manifest shapes (flattened f32, row-major order as documented for your build).

coelanox run -f model.cnox -i input.json -o output.json

Synthetic input (testing only):

coelanox run -f model.cnox -o output.json

Step 7 — Production configuration

Do not rely on defaults in production.

Create a JSON config or set COELANOX_* environment variables (see Operations): limits for container size, input size, memory, timeout, COELANOX_ALLOW_ABSOLUTE_PATHS=false where policy requires.
Point COELANOX_CONFIG_FILE at your config file on the runtime host.
Set RUST_LOG / COELANOX_LOG_LEVEL so security-relevant events reach your log stack.

Step 8 — Airgapped transfer

Run verify on the build side.
Copy .cnox (and public key if signing) via approved transfer.
Run verify again on the isolated host before run.

See DATA_FLOW.md.

Step 9 — Long-lived serving (optional)

For process-to-process integration:

coelanox serve -f model.cnox -b scalar

(Or another backend name your build supports.) IPC is stdin/stdout framed binary; see CLI reference § serve. Wrap with your supervisor and health logic—there is no built-in HTTP server.

Step 10 — Get help without SSH

Problem	Self-serve resource
Command flags	CLI_REFERENCE.md
Ops not packaging	ONNX_SUPPORTED_OPS.md, troubleshooting
Slow inference	KNOWN_LIMITATIONS.md, CLF path in Operations
Integrity / policy	RUNTIME_SPECIFICATION.md, Operations

COELANOX — Customer onboarding (model → .cnox → run)