#deployment

Merge adapters for single-tenant deployments. Keep them separate for multi-tenant. The serving architecture depends on how many customizations you're running.

Fine-tuning a model is the easy part. Running it in production with checkpoints, evals, rollback, and serving is the hard part. Here's the full picture.

If your optimization breaks an eval, the optimization is wrong. Evals are invariants, not suggestions. Ship nothing that fails them.

Models change. Prompts change. How do you update without breaking clients? Immutable versions and controlled rollout.

Model changes are high-risk deployments. 1% traffic to new, compare outputs, then gradually expand. Here's the playbook.

Raw PyTorch is 3-5x slower than optimized serving. Here's the gap and how to close it.

12 things to check before your LLM goes to production. Most teams skip at least half. That's how incidents happen.

Your API has rate limits. Your database has connection limits. Your LLM endpoints should have token limits. Here's how to add them without breaking production.