Local AI Engineering with Ollama

uv run mycli list
# NAME              SIZE      MODIFIED
# granite3.3:2b     1.5 GB    2 hours ago
# llama3.2:3b       2.0 GB    3 days ago

uv run mycli show granite3.3:2b
# Architecture:  granite
# Parameters:    2.5B
# Quantization:  Q4_K_M
# Context:       131072
# Capabilities:  completion, tools

uv run mycli pull qwen2.5:3b
# Pulling qwen2.5:3b
# downloading: 1.85 GB / 1.85 GB (100.0%)
# Done.

uv run mycli ps
# NAME              SIZE      PROCESSOR    CONTEXT    UNTIL
# granite3.3:2b     1.9 GB    100% CPU     4096       4 minutes from now

def cmd_list(client: Client, _args: argparse.Namespace) -> int:
    """List models on disk. Same data as `ollama list` and GET /api/tags."""
    resp = client.list()
    rows = [
        [
            m.model,                          # the tag, e.g. "granite3.3:2b"
            human_size(m.size),               # bytes -> "1.5 GB"
            human_age(m.modified_at),         # datetime -> "2h ago"
        ]
        for m in resp.models
    ]
    print_table(rows, ["NAME", "SIZE", "MODIFIED"])
    return 0

def cmd_show(client: Client, args: argparse.Namespace) -> int:
    """Print the metadata for one model. Same data as `ollama show`."""
    resp = client.show(args.model)

    # `details` is a small object with architecture/quantization/etc.
    # `model_info` is the full raw metadata dict from the GGUF; we pull the
    # context length out of it. The key is prefixed by architecture, e.g.
    # "granite.context_length" for granite, "llama.context_length" for llama.
    arch = resp.details.family
    ctx_key = f"{arch}.context_length"
    context = resp.modelinfo.get(ctx_key) if resp.modelinfo else None

    print(f"Architecture:  {arch}")
    print(f"Parameters:    {resp.details.parameter_size}")
    print(f"Quantization:  {resp.details.quantization_level}")
    print(f"Context:       {context if context is not None else 'unknown'}")
    if resp.capabilities:
        print(f"Capabilities:  {', '.join(resp.capabilities)}")
    return 0

def cmd_pull(client: Client, args: argparse.Namespace) -> int:
    """Pull a model, printing each progress event."""
    for event in client.pull(args.model, stream=True):
        if event.total and event.completed is not None:
            print(f"{event.status}: {event.completed}/{event.total}")
        else:
            print(event.status)
    return 0

def cmd_ps(client: Client, _args: argparse.Namespace) -> None:
    """List models currently loaded in memory."""
    print(f"{'NAME':<20} {'SIZE':<10} {'GPU':<6} EXPIRES")
    for m in client.ps().models:
        gpu_pct = (m.size_vram / m.size * 100) if m.size else 0
        print(
            f"{m.model:<20} {human_size(m.size):<10} {gpu_pct:>4.0f}% {m.expires_at.astimezone()}"
        )

def human_size