---

name: export-and-deployment description: "Export trained YOLOX checkpoints to ONNX or TorchScript and reason about optional deployment runtimes with backend caveats." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

YOLOX Export And Deployment

Use this sub-skill when a task starts from an existing YOLOX checkpoint and asks how to export it to ONNX or TorchScript, dry-run export settings, debug export failures, or choose among optional deployment runtimes such as ONNXRuntime, TensorRT, OpenVINO, ncnn, MegEngine, or nebullvm.

Do not use this sub-skill to produce a training checkpoint; route checkpoint production to ../training-and-data/SKILL.md. Do not use it for ordinary PyTorch .pth inference; route that to ../inference-and-api/SKILL.md.

Start here

1. Identify checkpoint path, model name or custom Exp, output format, input/test size, decode setting, and target runtime.
2. Read references/export-workflows.md for ONNX/TorchScript flags, checkpoint-loading rules, decode guidance, and dry-run patterns.
3. Run scripts/export_yolox_template.py for a safe dry-run or real ONNX/TorchScript export; real export requires an explicit checkpoint path.
4. Read references/deployment-backends.md before choosing ONNXRuntime, TensorRT, OpenVINO, ncnn, MegEngine, or nebullvm.
5. Read references/troubleshooting.md when export, checkpoint loading, dynamic axes, decode settings, or optional backend dependencies fail.

Minimal operating pattern

Dry-run first:

python scripts/export_yolox_template.py --format onnx --name yolox-s --dry-run

Then export with an explicit checkpoint:

python scripts/export_yolox_template.py --format onnx --name yolox-s --checkpoint yolox_s.pth --output yolox_s.onnx
python scripts/export_yolox_template.py --format torchscript --name yolox-s --checkpoint yolox_s.pth --output yolox_s.torchscript.pt

Keep model.head.decode_in_inference consistent with the downstream runtime. YOLOX deployment examples commonly expect raw head outputs and decode/postprocess outside the graph; custom runtimes may prefer decoded outputs in the graph.