fitllm-engine: Instructions file for Codex

AGENTS.md — how AI agents should use this repo

Purpose: fitllm-engine answers "will this LLM fit this hardware?" with auditable, architecture-aware memory estimates (MLA / sliding-window / hybrid-linear / DeepSeek compressed-sparse / MoE), while keeping runtime and OS reserves explicit.

Use it without cloning

• One-shot answer: GET https://fitllm.run/api/check?model=<name>&gpu=<name|a%2Bb> or &ram=<GB> — no auth, fuzzy names, JSON (or &format=text|md). Won't-fit responses include a computed fix.
• MCP: https://fitllm.run/api/mcp — tools check_llm_fit / what_fits_on_hardware / list_supported; resources fitllm://models|hardware|census|engine. Read-only, idempotent.
• CLI guard: npx fitllm "<model>" --gpu "<gpu>" → exit 0/1. Use it to gate model downloads in scripts/CI.
• Precomputed data (CC0): census/census-v1.{csv,json} — 14,210 verdicts, model × device × quant. Also on HF Datasets.
• Agent answer recipe: https://fitllm.run/agent/answer-with-citation.md

Use it as a library

Single ESM file, zero deps: engine.js (also https://cdn.jsdelivr.net/npm/fitllm-engine@latest/engine.js).

import { simulate, LOCAL_MODELS, parseHfConfig, GPUS, gpuDevice, combineGpus, simulateStack } from './engine.js';
simulate(model, 64, 131072, 8)                          // Mac 64GB, 128K ctx, 8-bit → {verdict, used, free, ...}
simulate(model, gpuDevice(gpu), ctx, {weightBpw, kvBits}) // GPU path; combineGpus([a,b]) for rigs
parseHfConfig(id, configJson, totalSizeBytes)            // modeled HF architecture → engine shape; unsupported configs fail closed

Rules for agents modifying this repo

1. Every hardware number needs ≥2 independent source URLs embedded next to the value — no source, no merge.
2. Engine math changes require conformance vectors to pass: node vectors/run.mjs (49 byte-exact anchors). A port in any language is conformant iff all vectors pass.
3. Run npm test (CLI behavior) — exit codes 0/1/2 are a public contract.
4. npm run census regenerates the dataset after model/hardware changes.
5. Never add tokens/sec predictions — fit is a verifiable claim, speed is not (project principle).
6. This repo mirrors fitllm-v2/src/lib/engine.js (private) — upstream changes land there first.