The Deployment Pipeline

Getting a model onto devices is a repeatable pipeline, not an event:

1. Freeze the graph — strip training machinery, fix shapes, produce a deployable format 2. Convert — to the target's format (TFLite, mlpackage, ONNX, GGUF) 3. Quantize — PTQ first; QAT only if PTQ fails; verify accuracy after quantization on the eval that matters 4. Delegate-test — check the partition: what actually landed on the accelerator vs. fell back 5. Benchmark on-device — real hardware, real conditions 6. Plan the fallback — for the devices where it doesn't fit or isn't fast enough

The step everyone skips is 5 — and it's the only one that tells the truth.