ml-serving-patterns

From checkpoint to production inference: safe model export and loading (state_dict with weights_only, safetensors for distribution, TorchScript/ONNX trade-offs), eval-mode and no-grad inference discipline, training-serving preprocessing parity via a shared transform module plus golden-batch tests, inference-server shape (health/readiness, versioned model registry, warmup), dynamic batching under an explicit latency budget, GPU memory measured at peak load before rollout, staged rollout (shadow, canary, full) with a pinned warm rollback, and drift plus quality monitoring wired as part of the deployment. Front-loads the production-breaking anti-patterns: pickle-loading untrusted checkpoints, reimplemented preprocessing skew, dropout left on in production, and rollouts whose only rollback plan is "redeploy the old container". Use when exporting, serving, deploying, rolling back, or monitoring a model in production.