"""Decision Index engine for Dinah-1 (https://github.com/apolinario/decision-index).

    python -m decision_index run --engine decision_index_engine:Dinah01Engine \
        --option repo=<path or repo> --edition 0.3 --suite-dir suite-0.3 --out runs/dinah-1

Put this file next to `experts.py`, `dinah.py` and `tools.py` (all in the model repository) and on PYTHONPATH.
Kit rules: nothing is truncated, no option is dropped and the rendering is the same for every benchmark. The gate
reads only the request content. A request that does not fit in 8,192 tokens, or a question type the model does not
support, raises `Unsupported` (unless an expert settles all of its questions without the encoder).
"""
from __future__ import annotations

import sys
from pathlib import Path

from decision_index.engines import Engine, Unsupported

sys.path.insert(0, str(Path(__file__).resolve().parent))
from dinah import RequestTooLong  # noqa: E402
from experts import Dinah01  # noqa: E402


class Dinah01Engine(Engine):
    name = "dinah-1"
    latency = ("In-process request wall time: code and calculator experts (when they fire), then one encoder pass over the "
               "request (all questions batched together); excludes model loading.")

    def __init__(self, repo: str = ".", revision: str | None = None, device: str | None = None, dtype: str | None = None,
                 batch_size: int = 16, tools: str = "code,calculator", backend: str = "torch", calibrate: str | bool = True, **options):
        super().__init__(**options)
        self.model = Dinah01.from_pretrained(repo, device=device, dtype=dtype, revision=revision, tools=tools, backend=backend)
        self.batch_size = int(batch_size)
        enc = self.model.encoder
        self.calibrated = str(calibrate).lower() not in ("0", "false", "no", "off")
        if not self.calibrated:   # raw logits and raw confidence head, for comparison with the calibrated run
            enc.temperature, enc.platt = None, None
        self.provenance = {"repo": repo, "revision": revision or "main", "backend": backend,
                           "device": getattr(enc, "device", "cpu"), "dtype": getattr(enc, "dtype", "int8"), "calibrated": self.calibrated,
                           "max_length": self.model.max_length, "experts": self.model.tools,
                           "truncation": "none: requests above 8,192 tokens raise Unsupported"}

    def __call__(self, state, questions):
        for q in questions.values():
            if q["type"] not in ("choice", "noul"):
                raise Unsupported(f"question type {q['type']}")
        try:
            answers = self.model.predict(state, questions, batch_size=self.batch_size)
        except RequestTooLong as e:
            raise Unsupported(f"context window: {e}")
        return {"model": "dinah-1", "answers": answers}, None

    def synchronize(self):
        enc = self.model.encoder
        if getattr(enc, "device", "cpu").startswith("cuda"):
            enc.torch.cuda.synchronize()