Function run_completion_trajectory

refract/runner.py:548–667

v0.1.4: greedy-decode and capture model-token IDs at decode time. Drives the patched ``llama-completion`` binary with ``REFRACT_TRAJECTORY= `` set; the binary writes one JSONL record per sampled token (``{"step":N,"token_id":ID}``). We read the file back, return the ID seque

(
    model: Path,
    prompt: str,
    kv: KVConfig,
    n_predict: int = 128,
    ctx: int = 512,
    n_gpu_layers: int = 99,
    seed: int = 42,
    temperature: float = 0.0,
    timeout: float = 300.0,
    apply_chat_template: bool = True,
    system: Optional[str] = None,
    reasoning: str = "off",
)

def run_completion_trajectory(
    model: Path,
    prompt: str,
    kv: KVConfig,
    n_predict: int = 128,
    ctx: int = 512,
    n_gpu_layers: int = 99,
    seed: int = 42,
    temperature: float = 0.0,
    timeout: float = 300.0,
    apply_chat_template: bool = True,
    system: Optional[str] = None,
    reasoning: str = "off",
) -> tuple[list[int], dict]:
    """v0.1.4: greedy-decode and capture model-token IDs at decode time.

    Drives the patched ``llama-completion`` binary with
    ``REFRACT_TRAJECTORY=<tmpfile>`` set; the binary writes one JSONL record
    per sampled token (``{"step":N,"token_id":ID}``). We read the file back,
    return the ID sequence, and delete the file.

    Returns (token_ids, metadata). The token IDs are the model&#x27;s own
    sampled tokens — no detokenize→retokenize round-trip, no whitespace-vs-
    model-token unit mismatch. This is the v0.1.4 fix for GTM&#x27;s structural
    weakness (LIMITATIONS.md §1, §5).

    Requires the ``llama-completion`` binary to be built from the patched
    ``tools/completion/completion.cpp`` (REFRACT v0.1.4 patch). If the
    binary lacks the patch, the trajectory file will be empty and this
    function returns ``([], meta)``.
    """
    # v0.3.1: dispatch to active backend if non-llamacpp is set.
    if _ACTIVE_BACKEND is not None and getattr(_ACTIVE_BACKEND, "name", None) != "llamacpp":
        res = _ACTIVE_BACKEND.run_completion_trajectory(
            model=model, prompt=prompt, kv_config_str=kv.label(),
            n_predict=n_predict, ctx=ctx, n_gpu_layers=n_gpu_layers,
            seed=seed, temperature=temperature, timeout=timeout,
            apply_chat_template=apply_chat_template, system=system,
        )
        return res.token_ids, res.metadata

    bin_path = _bin("llama-completion")

    fd, traj_path = tempfile.mkstemp(prefix="refract-traj-", suffix=".jsonl")
    os.close(fd)
    os.unlink(traj_path)  # patched binary creates it itself

    cmd: list[str] = [
        str(bin_path),
        "-m", str(model),
        "-p", prompt,
        "-n", str(n_predict),
        "-c", str(ctx),
        "-ngl", str(n_gpu_layers),
        "--seed", str(seed),
        "--temp", str(temperature),
        "-no-cnv",
        "--no-display-prompt",