06. Infra & Local Ops¶

Requirements¶

Apple Silicon Mac with MLX GPU (mx.default_device())
Python ≥3.11, uv
OpenAI-compatible API for compression
Optional: Node 22 + pnpm for visualizer
Disk: on the order of ~10 GB noted in STEPS.md for full artifacts

Setup¶

uv sync
cp .env.example .env   # OPENAI_API_KEY, OPENAI_API_BASE, OPENAI_MODEL
uv run python -c "import mlx.core as mx; print(mx.default_device())"

Pull heavy artifacts from HF if data/ / adapters/ are empty.

Makefile targets¶

Target	Purpose
`train`	LoRA SFT (`ITERS`, `SAVE_EVERY` overridable)
`eval-base` / `eval-ft` / `eval-all`	GSM8K evals (`LIMIT`, `BATCH_SIZE`)
`plot`	Build report charts
`clean`	Remove generated local outputs (see Makefile)

Topology¶

graph TD
  classDef default fill:#1e293b,stroke:#38bdf8,stroke-width:2px,color:#f8fafc
  classDef highlight fill:#065f46,stroke:#34d399,stroke-width:2px,color:#f0fdf4

  Mac["Mac MLX scripts"]:::highlight
  API["Compression API"]
  HF["Hugging Face Hub"]
  Pages["GitHub Pages visualizer"]

  Mac --> API
  Mac --> HF
  Pages --> HF

  linkStyle default stroke:#64748b,stroke-width:2px

No Docker Compose for training. Only CI is visualizer deploy.

Interview Q&A¶

Q: Can this train on CUDA?
A: Not as written — the stack is MLX-specific. Porting would mean swapping generate/train/eval backends.