llm-calc

Use when adding a new model (or model family) to the llm-calc database (src/data/models.ts). Walks through sourcing fields from HuggingFace config.json, reading papers for novel attention variants, writing tests first, and registering the entry. Invoke this whenever a model is being added or updated — even for a single SKU, since the failure modes (wrong KV bytes, mis-sized attention, silently broken roofline) all look fine until someone runs the calc against a real workload.

Use when adding a new accelerator (GPU/TPU/Trainium/Gaudi), interconnect fabric, or multi-accelerator system to the llm-calc database (src/data/{accelerators,interconnects,systems}.ts). Routes to the right sub-procedure based on whether the SKU is a new chip, a new fabric, or a new product composition. Invoke whenever hardware is being added — the failure modes (sparsity-inflated TFLOPS, confused per-direction vs aggregate BW, wrong variant in a system) all produce plausible-looking perf numbers that are silently wrong.

Create a feature branch, commit changes, push, and open a PR against iopsystems/llm-calc