# -*- coding: utf-8 -*-
"""华为昇腾 NPU 推理后端实现（ACL/CANN）。

对应 PRD 5.6：昇腾实现（ACL/CANN）——通过昇腾推理服务
（MindIE / onnxruntime-ascend）的 OpenAI 兼容接口对外提供推理，
与 NVIDIA GPU 后端实现同一 `InferenceBackend` 接口：
**切换后端仅改适配层配置，业务代码零改动**。
"""
from inference_backend.base import InferenceBackend, InferRequest, InferResult


class AscendNpuBackend(InferenceBackend):
    """华为昇腾 NPU 后端：面向昇腾 310P/910B（CANN/MindIE）。"""

    backend_name = "npu"

    def __init__(self, endpoint: str = "", model: str = "iaop-default",
                 timeout_seconds: float = 10.0, runtime: str = "mindie",
                 device: str = "ascend-910b", cann_version: str = "8.0",
                 **kwargs):
        super().__init__(endpoint, model, timeout_seconds)
        self.runtime = runtime            # mindie | onnx-ascend
        self.device = device              # ascend-310p | ascend-910b
        self.cann_version = cann_version  # CANN 工具链版本
        self._loaded = False

    def load_model(self, model_name: str | None = None) -> dict:
        """加载模型到 NPU（ACL aclmdlLoadFromFile 语义）。"""
        model_name = model_name or self.model
        if not self.endpoint:
            self._loaded = True
            return {"status": "ok", "backend": self.backend_name,
                    "model": model_name, "device": self.device,
                    "reason": "dry-run（未配置 endpoint）"}
        body = self._post_json("/acl/models/load",
                               {"model": model_name, "device": self.device,
                                "cann_version": self.cann_version})
        self._loaded = body.get("status") in ("ok", "loaded", "ready")
        return body

    def infer(self, request: InferRequest) -> InferResult:
        """昇腾推理（MindIE OpenAI 兼容 /v1/chat/completions）。"""
        import time
        started = time.monotonic()
        payload = request.to_payload()
        payload["model"] = payload["model"] or self.model
        body = self._post_json("/v1/chat/completions", payload)
        latency_ms = round((time.monotonic() - started) * 1000, 2)
        try:
            text = body["choices"][0]["message"]["content"]
        except (KeyError, IndexError, TypeError):
            text = str(body)
        return InferResult(
            text=text,
            backend=self.backend_name,
            latency_ms=latency_ms,
            meta={"runtime": self.runtime, "device": self.device,
                  "cann_version": self.cann_version, "model": self.model,
                  "raw": body},
        )

    def health(self) -> dict:
        """健康巡检：探测 /health，返回后端/设备/CANN 版本信息。"""
        base = self._healthz()
        base.update({
            "backend": self.backend_name,
            "runtime": self.runtime,
            "device": self.device,
            "cann_version": self.cann_version,
            "model": self.model,
            "loaded": self._loaded,
        })
        return base

    def unload(self) -> dict:
        """卸载模型、释放 NPU 资源（ACL aclmdlUnload 语义）。"""
        self._loaded = False
        if not self.endpoint:
            return {"status": "ok", "backend": self.backend_name,
                    "reason": "dry-run（未配置 endpoint）"}
        return self._post_json("/acl/models/unload", {"model": self.model})