Using ML Production

Overview

This meta-skill routes you to the right production deployment skill based on your concern. Load this when you need to move ML models to production but aren't sure which specific aspect to address.

Core Principle: Production concerns fall into four categories. Identify the concern first, then route to the appropriate skill. Tools and infrastructure choices are implementation details, not routing criteria.

Capability tiers, not model IDs. Sheets describe model and hardware capabilities (e.g., "frontier reasoning model", "mid-tier serving GPU", "edge accelerator") rather than naming specific SKUs. Vendor model names and exact GPU SKUs change quarterly; capability tiers are stable.

Knowledge-cutoff acknowledgement. Tool inventory is calibrated to 2026-05. Fast-moving areas — LLM serving stacks (vLLM, SGLang, TensorRT-LLM, TGI, Triton), observability platforms (Phoenix, Langfuse, OTel GenAI), quantization toolchains (torch.ao.quantization, AWQ, GPTQ), and MLOps platforms — should be re-checked against vendor docs before architecting new systems. Treat sheet content as a structured starting point, not a substitute for current documentation.

When to Use