#!/usr/bin/env bash
# 作业12 · 步骤1：单次 curl /v1/chat/completions，记录延迟
# 用法：
#   bash hw12_01_单次延迟.sh
#   bash hw12_01_单次延迟.sh --via-higress
#   ROUNDS=5 bash hw12_01_单次延迟.sh
set -euo pipefail

VIA_HIGRESS=0
for a in "$@"; do
  [ "$a" = "--via-higress" ] && VIA_HIGRESS=1
done

ROUNDS="${ROUNDS:-3}"
MODEL="${MODEL:-qwen3.6-27b}"
KEY="${LITELLM_MASTER_KEY:-sk-litellm-bocheng-207}"
OUT_DIR="${HW12_OUT:-/opt/workbuddy/data/homework12}"
mkdir -p "$OUT_DIR"

if [ "$VIA_HIGRESS" = "1" ]; then
  URL="${HIGRESS_LLM_BASE:-http://192.168.110.207:18080/llm/v1}/chat/completions"
else
  URL="${LITELLM_BASE:-http://192.168.110.207:4000}/v1/chat/completions"
fi

BODY="$(mktemp)"
trap 'rm -f "$BODY"' EXIT
cat >"$BODY" <<EOF
{"model":"$MODEL","messages":[{"role":"user","content":"只回复一个词：pong"}],"max_tokens":16,"stream":false,"temperature":0}
EOF

CSV="$OUT_DIR/latency_single_$(date +%Y%m%d_%H%M%S).csv"
echo "round,http_code,time_total_s,url,model" >"$CSV"

echo "[INFO] URL=$URL"
echo "[INFO] ROUNDS=$ROUNDS  结果 → $CSV"
echo

for i in $(seq 1 "$ROUNDS"); do
  TMP_HDR="$(mktemp)"
  TMP_BODY="$(mktemp)"
  # curl -w 输出 timing；响应体另存
  CODE_TIME="$(
    curl -sS -o "$TMP_BODY" -D "$TMP_HDR" \
      -w "%{http_code} %{time_total}" \
      -H "Authorization: Bearer $KEY" \
      -H "Content-Type: application/json" \
      --data-binary @"$BODY" \
      --max-time 180 \
      "$URL" || echo "000 0"
  )"
  CODE="$(echo "$CODE_TIME" | awk '{print $1}')"
  T="$(echo "$CODE_TIME" | awk '{print $2}')"
  SNIP="$(python3 - <<PY
import json,sys
p="$TMP_BODY"
try:
  j=json.load(open(p,encoding="utf-8"))
  c=(j.get("choices") or [{}])[0].get("message",{}).get("content","")
  print((c or "")[:60].replace("\n"," "))
except Exception as e:
  print(open(p,encoding="utf-8",errors="replace").read()[:60].replace("\n"," "))
PY
)"
  echo "[R$i] http=$CODE time=${T}s  snippet=$SNIP"
  echo "$i,$CODE,$T,$URL,$MODEL" >>"$CSV"
  rm -f "$TMP_HDR" "$TMP_BODY"
done

echo
echo "[OK] CSV=$CSV"
python3 - <<PY
import csv
from pathlib import Path
p=Path("$CSV")
rows=list(csv.DictReader(p.open(encoding="utf-8")))
ts=[float(r["time_total_s"]) for r in rows if r["http_code"]=="200"]
oks=sum(1 for r in rows if r["http_code"]=="200")
print(f"成功 {oks}/{len(rows)}")
if ts:
  print(f"延迟 min={min(ts):.3f}s  avg={sum(ts)/len(ts):.3f}s  max={max(ts):.3f}s")
PY