https://yennj12.js.org/yennj12_blog_V4/tags/cost-management/Recent content in Cost Management on YennJ12 Engineering BlogHugo -- gohugo.ioen-usTue, 26 May 2026 17:05:09 +0900https://yennj12.js.org/yennj12_blog_V4/posts/ai-fde-essential-guide-part4-zh/Tue, 26 May 2026 17:05:09 +0900https://yennj12.js.org/yennj12_blog_V4/posts/ai-fde-essential-guide-part4-zh/前言 生產環境 AI 系統的監控與最佳化是確保企業 AI 應用成功的關鍵。從模型效能追蹤、基礎設施監控到成本控制，AI FDE 需要建立全方位的可觀測性體系。本文將深入探討 LLM-native 指標設計、分散式監控架構、智能故障診斷與企業級成本最佳化策略。 1. LLM-native 指標與評估體系 核心效能指標設計 LLM 特定指標框架： 1from dataclasses import dataclass 2from typing import Dict, List, Optional, Union 3import numpy as np 4from collections import deque 5import time 6import asyncio 7from enum import Enum 8 9class MetricType(Enum): 10 LATENCY = "latency" 11 THROUGHPUT = "throughput" 12 QUALITY = "quality" 13 COST = "cost" 14 RELIABILITY = "reliability" 15 16@dataclass 17class LLMMetrics: 18 timestamp: float 19 request_id: str 20 model_name: str 21 22 # 效能指標 23 time_to_first_token: float # TTFT - 首個 token 延遲 24 time_per_output_token: float # TPOT - 每個輸出 token 時間 25 total_latency: float 26 tokens_per_second: float 27 28 # 品質指標 29 perplexity: Optional[float] = None 30 bleu_score: Optional[float] = None 31 rouge_score: Optional[Dict[str, float]] = None 32 human_feedback_score: Optional[float] = None 33 34 # 成本指標 35 input_tokens: int = 0 36 output_tokens: int = 0 37 compute_cost: float = 0.