benchscope 1.0.0__py3-none-any.whl

benchscope/server/test_manager.py

@@ -0,0 +1,367 @@
+"""测试执行管理器：编排用例×并发，流式执行、实时推送、日志落盘、xlsx 汇总。"""
+from __future__ import annotations
+
+import json
+import logging
+import re
+import threading
+import time
+from dataclasses import dataclass, field
+from datetime import datetime
+from pathlib import Path
+from typing import Optional
+
+from benchscope.benches.base import BenchOptions
+from benchscope.benches.runner import BenchRunner, StopRequested
+from benchscope.benches import sglang_bench, vllm_bench
+from benchscope.constants import (
+    DATASET_RANDOM, DATASET_SHAREGPT, FRAMEWORK_NAMES,
+)
+from benchscope.summary import write_summary_csv, write_xlsx
+
+log = logging.getLogger("benchscope.test")
+
+
+def sanitize_name(name: str) -> str:
+    """把模型名/路径转成安全的文件名片段。"""
+    return re.sub(r"[^\w.-]", "_", name).strip("_")
+
+
+def build_cases(dataset: dict, model: str) -> list[dict]:
+    """根据数据集配置生成用例列表。"""
+    ds_type = dataset.get("type", DATASET_RANDOM)
+    cases: list[dict] = []
+    if ds_type == DATASET_RANDOM:
+        pairs = dataset.get("length_pairs") or []
+        for il, ol, label in pairs:
+            cases.append({
+                "label": label,
+                "input_len": il,
+                "output_len": ol,
+                "path": None,
+            })
+    else:
+        label = "ShareGPT" if ds_type == DATASET_SHAREGPT else "Custom"
+        if ds_type == DATASET_SHAREGPT and dataset.get("label"):
+            label = dataset["label"]
+        cases.append({
+            "label": label,
+            "input_len": None,
+            "output_len": None,
+            "path": dataset.get("path"),
+        })
+    return cases
+
+
+@dataclass
+class TestRun:
+    run_id: str
+    run_dir: Path
+    payload: dict
+    framework: str
+    model: str
+    gpu: dict
+    cases: list = field(default_factory=list)
+    rows: list = field(default_factory=list)
+    status: str = "running"          # running | done | stopped | error
+    error: Optional[str] = None
+    summary: Optional[dict] = None
+    started_at: str = ""
+    finished_at: str = ""
+    precision: str = ""
+
+    def snapshot(self, include_rows: bool = True) -> dict:
+        data = {
+            "run_id": self.run_id,
+            "run_dir": str(self.run_dir),
+            "framework": self.framework,
+            "framework_name": FRAMEWORK_NAMES.get(self.framework, self.framework),
+            "model": self.model,
+            "gpu": self.gpu,
+            "cases": self.cases,
+            "status": self.status,
+            "error": self.error,
+            "summary": self.summary,
+            "started_at": self.started_at,
+            "finished_at": self.finished_at,
+            "precision": self.precision,
+            "concurrency_list": self.payload.get("concurrency_list", []),
+            "request_rate": self.payload.get("request_rate", "inf"),
+            "tpot_threshold_ms": self.payload.get("tpot_threshold_ms"),
+            "dataset": self.payload.get("dataset", {}),
+        }
+        if include_rows:
+            data["rows"] = self.rows
+        return data
+
+
+def build_single_command(
+    framework: str, model: str, tokenizer: str, api: dict, dataset: dict,
+    concurrency: int, request_rate, curated: dict, extra_args: list,
+) -> list[str]:
+    """构建单条 bench 命令（供执行与预览共用）。"""
+    opts = BenchOptions(
+        framework=framework,
+        model=model,
+        api=api,
+        dataset=dataset,
+        concurrency=concurrency,
+        request_rate=request_rate,
+        curated=curated or {},
+        extra_args=extra_args or [],
+    )
+    opts.tokenizer = tokenizer or model
+    if framework == "sglang":
+        return sglang_bench.build_command(opts)
+    return vllm_bench.build_command(opts)
+
+
+def build_command_lines(payload: dict, config) -> list[dict]:
+    """为 payload 中的所有用例×并发构建命令列表（预览用）。"""
+    framework = payload.get("framework", "vllm")
+    model = payload.get("model", "")
+    tokenizer = payload.get("tokenizer", "")
+    dataset = dict(payload.get("dataset", {}))
+    cases = build_cases(dataset, model)
+    lines = []
+    for case in cases:
+        ds = dict(dataset)
+        ds.update({
+            "input_len": case.get("input_len"),
+            "output_len": case.get("output_len"),
+            "path": case.get("path"),
+        })
+        for conc in payload.get("concurrency_list", []):
+            if conc == "inf" or conc is None:
+                continue
+            cmd = build_single_command(
+                framework, model, tokenizer, dict(config.api), ds,
+                int(conc), payload.get("request_rate", "inf"),
+                payload.get("curated", {}), payload.get("extra_args", []),
+            )
+            lines.append({"case": case["label"], "concurrency": int(conc), "cmd": " ".join(cmd)})
+    return lines
+
+
+class TestManager:
+    def __init__(self, config, hub):
+        self.config = config
+        self.hub = hub
+        self._lock = threading.RLock()
+        self.current: Optional[TestRun] = None
+        self._thread: Optional[threading.Thread] = None
+        self._runner: Optional[BenchRunner] = None
+
+    # ------------------------------------------------------------------
+    @property
+    def running(self) -> bool:
+        with self._lock:
+            return self.current is not None and self.current.status == "running"
+
+    def start(self, payload: dict) -> TestRun:
+        with self._lock:
+            if self.running:
+                raise RuntimeError("已有测试正在运行，请先停止。")
+            # sharegpt 数据集：未指定路径时自动从 modelscope 下载
+            dataset = payload.get("dataset", {})
+            if dataset.get("type") == DATASET_SHAREGPT and not dataset.get("path"):
+                from benchscope.datasets import ensure_sharegpt
+
+                path = str(ensure_sharegpt(self.config.datasets_dir))
+                dataset = dict(dataset)
+                dataset["path"] = path
+                payload = dict(payload)
+                payload["dataset"] = dataset
+            run = self._create_run(payload)
+            self.current = run
+            self._thread = threading.Thread(
+                target=self._execute, args=(run,), name="bench-run", daemon=True
+            )
+            self._thread.start()
+            return run
+
+    def stop(self) -> None:
+        runner = self._runner
+        if runner:
+            runner.kill()
+
+    def _create_run(self, payload: dict) -> TestRun:
+        framework = payload.get("framework", "vllm")
+        model = payload.get("model", "")
+        gpu = payload.get("gpu", {})
+        run_id = datetime.now().strftime("%m%d-%H%M%S")
+        run_dir = self.config.logs_dir / run_id
+        run_dir.mkdir(parents=True, exist_ok=True)
+        run = TestRun(
+            run_id=run_id,
+            run_dir=run_dir,
+            payload=payload,
+            framework=framework,
+            model=model,
+            gpu=gpu,
+            precision=payload.get("precision", ""),
+            started_at=datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
+        )
+        run.cases = build_cases(payload.get("dataset", {}), model)
+        return run
+
+    # ------------------------------------------------------------------
+    def _execute(self, run: TestRun) -> None:
+        framework = run.framework
+        tmpl = (self.config.get("bench_commands") or {}).get(framework, "")
+        runner = BenchRunner(tmpl)
+        self._runner = runner
+
+        model_name = sanitize_name(Path(run.model).name or run.model)
+        gpu_count = run.gpu.get("count", "") if isinstance(run.gpu, dict) else run.gpu
+        gpu_label = f"{gpu_count}" if gpu_count else ""
+        if isinstance(run.gpu, dict) and run.gpu.get("name"):
+            gpu_label = f"{run.gpu['name']}×{gpu_count}" if gpu_count else run.gpu["name"]
+
+        meta = {
+            "model": run.model,
+            "model_name": model_name,
+            "framework": FRAMEWORK_NAMES.get(framework, framework),
+            "gpu": gpu_label,
+            "precision": run.precision,
+        }
+        mean_csv = run.run_dir / f"{model_name}_X{gpu_count}.log"
+        p99_csv = run.run_dir / f"{model_name}_X{gpu_count}_p99.log"
+
+        try:
+            self.hub.broadcast({"type": "run_started", "run": run.snapshot()})
+            cases_done = 0
+            for case in run.cases:
+                if runner._stop_flag.is_set():
+                    break
+                case_label = case["label"]
+                detail_path = run.run_dir / f"{model_name}_{case_label}_X{gpu_count}.log"
+                detail_fp = open(detail_path, "a", encoding="utf-8")
+                concurrency_ok = 0
+                for conc in run.payload.get("concurrency_list", []):
+                    if runner._stop_flag.is_set():
+                        break
+                    if conc == "inf" or conc is None:
+                        continue
+                    conc = int(conc)
+                    try:
+                        row = self._run_one(runner, run, case, conc, detail_fp, meta)
+                        run.rows.append(row)
+                        concurrency_ok += 1
+                        # 增量写汇总 CSV
+                        write_summary_csv(
+                            mean_csv, [row], p99=False, append=True, case_header=concurrency_ok == 1, case=case, meta=meta
+                        )
+                        write_summary_csv(
+                            p99_csv, [row], p99=True, append=True, case_header=concurrency_ok == 1, case=case, meta=meta
+                        )
+                        self.hub.broadcast({"type": "result", "run_id": run.run_id, "row": row})
+                    except StopRequested:
+                        break
+                    except Exception as e:
+                        log.exception("并发 %s 执行失败", conc)
+                        err_row = {
+                            "case": case_label, "label": case_label,
+                            "input_len": case.get("input_len"), "output_len": case.get("output_len"),
+                            "concurrency": conc, "error": str(e)[:500],
+                        }
+                        run.rows.append(err_row)
+                        self.hub.broadcast({"type": "result", "run_id": run.run_id, "row": err_row})
+                detail_fp.close()
+                cases_done += 1
+
+            # 汇总 xlsx
+            if run.rows:
+                rows_for_xlsx = [r for r in run.rows if "metrics" in r]
+                if rows_for_xlsx:
+                    annotated = self._annotate_best(rows_for_xlsx, run.payload.get("tpot_threshold_ms"))
+                    xlsx_path = run.run_dir / f"benchmark-{datetime.now().strftime('%d%m%y')}.xlsx"
+                    write_xlsx(xlsx_path, annotated, meta)
+                    run.summary = {"xlsx": str(xlsx_path), "rows": len(rows_for_xlsx)}
+
+            run.status = "stopped" if runner._stop_flag.is_set() else "done"
+            run.finished_at = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+            self._save_run_json(run)
+            self.hub.broadcast({"type": "run_done", "run_id": run.run_id, "run": run.snapshot()})
+        except StopRequested:
+            run.status = "stopped"
+            run.finished_at = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+            self._save_run_json(run)
+            self.hub.broadcast({"type": "run_done", "run_id": run.run_id, "run": run.snapshot()})
+        except Exception as e:
+            log.exception("测试执行失败")
+            run.status = "error"
+            run.error = str(e)
+            run.finished_at = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+            self._save_run_json(run)
+            self.hub.broadcast({"type": "run_error", "run_id": run.run_id, "error": str(e), "run": run.snapshot()})
+        finally:
+            self._runner = None
+
+    def _run_one(self, runner: BenchRunner, run: TestRun, case: dict,
+                 concurrency: int, detail_fp, meta: dict) -> dict:
+        ds = dict(run.payload.get("dataset", {}))
+        ds.update({
+            "input_len": case.get("input_len"),
+            "output_len": case.get("output_len"),
+            "path": case.get("path"),
+        })
+        cmd = build_single_command(
+            run.framework, run.model, run.payload.get("tokenizer", ""),
+            dict(self.config.api), ds, concurrency,
+            run.payload.get("request_rate", "inf"),
+            run.payload.get("curated", {}), run.payload.get("extra_args", []),
+        )
+
+        def stream(line: str):
+            detail_fp.write(line)
+            self.hub.broadcast({
+                "type": "log_line", "run_id": run.run_id,
+                "case": case["label"], "concurrency": concurrency, "line": line,
+            })
+
+        metrics = runner.run(cmd, stream_cb=stream)
+        return {
+            "case": case["label"], "label": case["label"],
+            "input_len": case.get("input_len"), "output_len": case.get("output_len"),
+            "concurrency": concurrency,
+            "cmd": " ".join(cmd),
+            "metrics": metrics,
+        }
+
+    def _annotate_best(self, rows: list[dict], threshold) -> list[dict]:
+        """为每个用例标记最接近且低于阈值的 TPOT 行为最佳行。"""
+        if threshold is None:
+            return rows
+        try:
+            threshold = float(threshold)
+        except (TypeError, ValueError):
+            return rows
+        by_case: dict = {}
+        for r in rows:
+            by_case.setdefault(r.get("label"), []).append(r)
+        for label, items in by_case.items():
+            valid = []
+            for r in items:
+                m = r.get("metrics", {})
+                tpot = m.get("tpot_mean")
+                if tpot is not None:
+                    valid.append((float(tpot), r))
+            if not valid:
+                continue
+            below = [(t, r) for t, r in valid if t < threshold]
+            if below:
+                best_t, best_r = max(below, key=lambda x: x[0])  # 最接近阈值（从下方）
+            else:
+                best_t, best_r = min(valid, key=lambda x: x[0])  # 无低于阈值则取最小
+            best_r["best"] = True
+            best_r["best_tpot"] = best_t
+        return rows
+
+    def _save_run_json(self, run: TestRun) -> None:
+        try:
+            (run.run_dir / "run.json").write_text(
+                json.dumps(run.snapshot(), ensure_ascii=False, indent=2), encoding="utf-8"
+            )
+        except Exception:
+            log.exception("保存 run.json 失败")

benchscope/server/ws.py

@@ -0,0 +1,56 @@
+"""WebSocket 客户端广播。"""
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import threading
+from typing import Any
+
+from fastapi import WebSocket
+
+log = logging.getLogger("benchscope.ws")
+
+
+class WebSocketHub:
+    """向所有已连接的前端客户端广播 JSON 消息（线程安全，可在后台线程调用）。"""
+
+    def __init__(self):
+        self._clients: set[tuple[WebSocket, asyncio.AbstractEventLoop]] = set()
+        self._lock = threading.Lock()
+
+    def register(self, ws: WebSocket, loop: asyncio.AbstractEventLoop) -> None:
+        with self._lock:
+            self._clients.add((ws, loop))
+
+    def unregister(self, ws: WebSocket) -> None:
+        with self._lock:
+            self._clients = {(w, l) for w, l in self._clients if w is not ws}
+
+    @property
+    def count(self) -> int:
+        with self._lock:
+            return len(self._clients)
+
+    def broadcast(self, payload: dict[str, Any]) -> None:
+        try:
+            msg = json.dumps(payload, ensure_ascii=False, default=str)
+        except TypeError:
+            log.exception("WS 广播序列化失败")
+            return
+        with self._lock:
+            clients = list(self._clients)
+        for ws, loop in clients:
+            try:
+                fut = asyncio.run_coroutine_threadsafe(ws.send_text(msg), loop)
+                fut.add_done_callback(self._on_sent(ws))
+            except Exception:
+                self.unregister(ws)
+
+    def _on_sent(self, ws: WebSocket):
+        def _cb(fut: asyncio.Future):
+            try:
+                fut.result()
+            except Exception:
+                self.unregister(ws)
+        return _cb

benchscope/summary.py

@@ -0,0 +1,148 @@
+"""日志汇总：CSV 汇总日志与 benchmark-*.xlsx 生成（mean / P99 双面板）。"""
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from typing import Iterable
+
+from openpyxl import Workbook
+from openpyxl.styles import Alignment, Font, PatternFill
+
+log = logging.getLogger("benchscope.summary")
+
+# xlsx 列定义（与 asserts/benchmark-260821.xlsx 对齐，末尾追加 单用户）
+XLSX_HEADERS = [
+    "GPU", "模型", "精度", "推理框架", "输入长度", "输出长度", "并发数",
+    "output", "peakoutput", "total", "ttft", "itl", "tpot", "单用户",
+]
+
+
+def _fmt(value, digits: int = 2) -> str:
+    if value is None or value == "":
+        return ""
+    try:
+        return f"{float(value):.{digits}f}"
+    except (TypeError, ValueError):
+        return str(value)
+
+
+def write_summary_csv(
+    path: Path,
+    rows: Iterable[dict],
+    p99: bool = False,
+    append: bool = False,
+    case_header: bool = False,
+    case: dict | None = None,
+    meta: dict | None = None,
+) -> Path:
+    """按用例分组写汇总 CSV（兼容 asserts/logs 旧格式）。
+
+    rows: [{case, label, input_len, output_len, concurrency, metrics}]
+    p99=False 时取 mean 指标；p99=True 时取 P99 指标。
+    append=True 时为增量追加模式（配合 case_header 控制块头写入）。
+    """
+    path.parent.mkdir(parents=True, exist_ok=True)
+    suffix = "_p99" if p99 else ""
+    key = "p99" if p99 else "mean"
+
+    def metric(r, name):
+        m = r.get("metrics", {})
+        return m.get(f"{name}_{key}", m.get(name, ""))
+
+    rows = list(rows)
+    if not append:
+        with open(path, "w", encoding="utf-8") as f:
+            pass  # 清空
+
+    with open(path, "a", encoding="utf-8") as f:
+        if append:
+            if case_header and case:
+                gpu_count = (meta or {}).get("gpu", "") or ""
+                f.write("=" * 60 + "\n")
+                f.write(
+                    f"测试条件：{case.get('label')} | 输入={case.get('input_len')} | "
+                    f"输出={case.get('output_len')} | 部署GPU={gpu_count}\n"
+                )
+                f.write("=" * 60 + "\n")
+                f.write("并发数,Output Token,Peak Output Token,Total Token,TTFT,TPOT,ITL\n")
+            for r in rows:
+                f.write(
+                    f"{r.get('concurrency')},{metric(r, 'output')},{metric(r, 'peakoutput')},"
+                    f"{metric(r, 'total')},{metric(r, 'ttft')},{metric(r, 'tpot')},{metric(r, 'itl')}\n"
+                )
+            return path
+
+        # 全量模式：按用例分组
+        groups: dict = {}
+        for r in rows:
+            groups.setdefault(r.get("label", ""), []).append(r)
+        for label, items in groups.items():
+            first = items[0]
+            f.write("=" * 60 + "\n")
+            f.write(
+                f"测试条件：{label} | 输入={first.get('input_len')} | "
+                f"输出={first.get('output_len')} | 部署GPU={(meta or {}).get('gpu', '')}\n"
+            )
+            f.write("=" * 60 + "\n")
+            f.write("并发数,Output Token,Peak Output Token,Total Token,TTFT,TPOT,ITL\n")
+            for r in items:
+                f.write(
+                    f"{r.get('concurrency')},{metric(r, 'output')},{metric(r, 'peakoutput')},"
+                    f"{metric(r, 'total')},{metric(r, 'ttft')},{metric(r, 'tpot')},{metric(r, 'itl')}\n"
+                )
+            f.write("\n")
+    return path
+
+
+def write_xlsx(path: Path, rows: Iterable[dict], meta: dict) -> Path:
+    """生成 benchmark-*.xlsx，含 均值 与 P99 两个 sheet。"""
+    rows = list(rows)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    wb = Workbook()
+    wb.remove(wb.active)
+
+    header_font = Font(bold=True, color="FFFFFF")
+    header_fill = PatternFill("solid", fgColor="409EFF")
+    best_fill = PatternFill("solid", fgColor="FFF3CD")
+
+    for sheet_name, key in (("均值 Mean", "mean"), ("P99", "p99")):
+        ws = wb.create_sheet(sheet_name)
+        ws.append(XLSX_HEADERS)
+        for cell in ws[1]:
+            cell.font = header_font
+            cell.fill = header_fill
+            cell.alignment = Alignment(horizontal="center")
+
+        for r in rows:
+            m = r.get("metrics", {})
+            concurrency = r.get("concurrency")
+            # 最佳行（tpot 最接近且低于阈值）加亮
+            best = r.get("best", False)
+            values = [
+                meta.get("gpu", ""),
+                meta.get("model", ""),
+                meta.get("precision", ""),
+                meta.get("framework", ""),
+                r.get("input_len", ""),
+                r.get("output_len", ""),
+                concurrency,
+                _fmt(m.get(f"output_{key}", m.get("output_mean", m.get("output")))),
+                _fmt(m.get(f"peakoutput_{key}", m.get("peakoutput_mean", m.get("peakoutput")))),
+                _fmt(m.get(f"total_{key}", m.get("total_mean", m.get("total")))),
+                _fmt(m.get(f"ttft_{key}", m.get("ttft"))),
+                _fmt(m.get(f"itl_{key}", m.get("itl"))),
+                _fmt(m.get(f"tpot_{key}", m.get("tpot"))),
+                _fmt(m.get("single_user")),
+            ]
+            ws.append(values)
+            if best:
+                for cell in ws[ws.max_row]:
+                    cell.fill = best_fill
+
+        # 列宽
+        for col, _ in enumerate(XLSX_HEADERS, start=1):
+            ws.column_dimensions[chr(64 + col)].width = 14
+        ws.freeze_panes = "A2"
+
+    wb.save(path)
+    return path

benchscope/webui/assets/LogView-BDFIduo7.css

@@ -0,0 +1 @@
+.record-list[data-v-85fc4f26]{padding:12px 8px;height:100%;display:flex;flex-direction:column}.record-head[data-v-85fc4f26]{display:flex;align-items:center;justify-content:space-between;margin-bottom:10px;padding:0 8px}.record-title[data-v-85fc4f26]{font-weight:600;font-size:14px}.record-selected[data-v-85fc4f26]{background:#e6f4ff}.run-detail[data-v-53d908f7]{padding:16px;height:100%;overflow:auto}.detail-meta[data-v-53d908f7]{margin-bottom:8px;border-radius:8px}.detail-tabs[data-v-53d908f7]{background:#fff;border:1px solid #f0f0f0;border-radius:8px;padding:0 12px 12px}.preview-pre[data-v-53d908f7]{max-height:520px;overflow:auto;background:#f6f8fa;padding:12px;font-size:12px;white-space:pre-wrap;margin:0}.page-layout[data-v-8fab23f1]{height:100%}.record-sider[data-v-8fab23f1]{border-right:1px solid #f0f0f0;height:100%;overflow:auto}.page-content[data-v-8fab23f1]{height:100%;background:#f5f5f5}

benchscope/webui/assets/LogView-BMsVPLVq.js

@@ -0,0 +1 @@
+import{_ as E,b as Z,a as F}from"./index-Dfr0hp72.js";import{f as x,w as G,o as ee,S as c,U as b,a6 as S,X as w,V as L,W as a,G as v,_ as i,k as t,Y as te,c as O,F as X,a8 as D,r as Y}from"./vue-Ch4zjUb1.js";import{M as H,_ as ae}from"./MetricsCharts-DHO93JrC.js";import{j as W}from"./antd-DWALckI0.js";import"./echarts-Bb6yjXMn.js";const ne={class:"record-list"},oe={class:"record-head"},le={style:{"font-weight":"600","font-size":"13px"}},se={style:{"font-size":"12px","line-height":"1.6"}},re={__name:"RunRecordList",props:{modelValue:{type:String,default:""},framework:{type:String,default:""}},emits:["update:modelValue"],setup(p,{emit:d}){const u=p,r=d,l=Z(),o=O(()=>u.framework),m=x(u.framework||""),T=x([]),f=x(!1),B=O(()=>{const s=o.value||m.value;if(!s)return T.value;const k=s==="vllm"?"vLLM":"SGLang";return T.value.filter(g=>{var e;return(((e=g.meta)==null?void 0:e.framework)||"")===k})});function j(s){return s==="vllm"?"vLLM":"SGLang"}function K(s){return s==="done"?"green":s==="error"?"red":s==="stopped"?"orange":"blue"}function A(s){return s==="done"?"完成":s==="error"?"失败":s==="stopped"?"已停止":"运行中"}async function M(){f.value=!0;try{const s=await F.listRuns();T.value=s.runs||[]}finally{f.value=!1}}return G(B,s=>{if(!s.length){u.modelValue&&r("update:modelValue","");return}if(!(u.modelValue&&s.some(g=>g.run_id===u.modelValue))){const g=l.lastRunId&&s.some(e=>e.run_id===l.lastRunId)?l.lastRunId:s[0].run_id;r("update:modelValue",g)}},{immediate:!0}),G(()=>l.lastRunId,M),G(m,()=>{}),ee(M),(s,k)=>{const g=c("a-select"),e=c("a-tag"),n=c("a-list-item-meta"),C=c("a-list-item"),$=c("a-list"),R=c("a-spin");return b(),S("div",ne,[w("div",oe,[k[1]||(k[1]=w("span",{class:"record-title"},"测试记录",-1)),o.value?(b(),L(e,{key:1,size:"small",color:"blue"},{default:a(()=>[v(i(j(o.value)),1)]),_:1})):(b(),L(g,{key:0,value:m.value,"onUpdate:value":k[0]||(k[0]=_=>m.value=_),size:"small",style:{width:"110px"},options:[{value:"",label:"全部"},{value:"vllm",label:"vLLM"},{value:"sglang",label:"SGLang"}]},null,8,["value"]))]),t(R,{spinning:f.value},{default:a(()=>[t($,{size:"small",bordered:"","data-source":B.value},{renderItem:a(({item:_})=>[t(C,{style:{cursor:"pointer"},class:te({"record-selected":_.run_id===p.modelValue}),onClick:z=>s.$emit("update:modelValue",_.run_id)},{default:a(()=>[t(n,null,{title:a(()=>{var z;return[w("span",le,i(_.run_id),1),t(e,{size:"small",color:K((z=_.meta)==null?void 0:z.status),style:{"margin-left":"6px"}},{default:a(()=>{var h;return[v(i(A((h=_.meta)==null?void 0:h.status)),1)]}),_:2},1032,["color"])]}),description:a(()=>{var z,h,N;return[w("div",se,[w("div",null,i(((z=_.meta)==null?void 0:z.framework)||"-")+" · "+i(((h=_.meta)==null?void 0:h.model)||"-"),1),w("div",null,i(((N=_.meta)==null?void 0:N.started_at)||"")+" · "+i(_.files.length)+" 文件",1)])]}),_:2},1024)]),_:2},1032,["class","onClick"])]),empty:a(()=>[...k[2]||(k[2]=[w("div",{style:{padding:"12px",color:"#999"}},"暂无日志记录",-1)])]),_:1},8,["data-source"])]),_:1},8,["spinning"])])}}},ue=E(re,[["__scopeId","data-v-85fc4f26"]]),de={__name:"RunSummaryBlock",props:{records:{type:Array,default:()=>[]}},setup(p){const d=p,u=O(()=>d.records.map(l=>({label:l.label,concurrency:l.concurrency,metrics:{output_mean:l.output_mean,total_mean:l.total_mean,ttft_mean:l.ttft_mean,tpot_mean:l.tpot_mean,ttft_p99:l.ttft_p99,tpot_p99:l.tpot_p99}}))),r=[{key:"output_mean",label:"Output 吞吐 (tok/s)"},{key:"total_mean",label:"Total 吞吐 (tok/s)"},{key:"ttft_mean",label:"TTFT mean 耗时 (ms)"},{key:"tpot_mean",label:"TPOT mean 耗时 (ms)"},{key:"ttft_p99",label:"TTFT P99 耗时 (ms)"},{key:"tpot_p99",label:"TPOT P99 耗时 (ms)"}];return(l,o)=>(b(),S("div",null,[t(H,{rows:u.value,"metric-defs":r},null,8,["rows"])]))}},ie={style:{"margin-bottom":"8px",color:"#999","font-size":"12px"}},q={__name:"AnalysisBlock",props:{rows:{type:Array,default:()=>[]},best:{type:Object,default:()=>({})},threshold:{type:[Number,String],default:null}},setup(p){const d=p,u=[{key:"output_mean",label:"Output 吞吐 (tok/s)"},{key:"peakoutput_mean",label:"Peak Output 吞吐 (tok/s)"},{key:"total_mean",label:"Total 吞吐 (tok/s)"},{key:"ttft_mean",label:"TTFT mean 耗时 (ms)"},{key:"itl_mean",label:"ITL mean 耗时 (ms)"},{key:"tpot_mean",label:"TPOT mean 耗时 (ms)"}],r=O(()=>{const l={};for(const o of Object.keys(d.best||{})){const m=d.best[o];m&&m.row&&(l[`${o}-${m.concurrency}`]=!0)}return d.rows.map(o=>({label:o.label,concurrency:o.concurrency,best:!!l[`${o.label}-${o.concurrency}`],metrics:{output_mean:o.output,peakoutput_mean:o.peakoutput,total_mean:o.total,ttft_mean:o.ttft,tpot_mean:o.tpot,itl_mean:o.itl}}))});return(l,o)=>(b(),S("div",null,[w("div",ie,i(p.threshold?`TPOT 阈值 ${p.threshold}ms，金色行为最佳并发`:"未设置 TPOT 阈值"),1),t(ae,{rows:r.value,threshold:p.threshold,pagination:{pageSize:20,showSizeChanger:!0}},null,8,["rows","threshold"]),o[0]||(o[0]=w("div",{style:{"margin-top":"12px","font-weight":"600"}},"曲线（横轴：并发数）",-1)),t(H,{rows:r.value,"metric-defs":u},null,8,["rows"])]))}},ce={class:"run-detail"},me={style:{color:"#999","font-size":"12px","margin-bottom":"6px"}},pe={class:"preview-pre"},_e={__name:"RunDetailPanel",props:{runId:{type:String,default:""}},setup(p){const d=p,u=Y({records:[],records_mean:[],records_p99:[],best_mean:{},best_p99:{},threshold:null}),r=x({}),l=x([]),o=x("summary"),m=x(!1),T=x(""),f=Y({content:"",total_lines:0,truncated:0}),B=[{title:"文件名",dataIndex:"name",key:"name"},{title:"大小",dataIndex:"size",key:"size",width:110,customRender:({text:e})=>K(e)},{title:"操作",key:"actions",width:140}],j=O(()=>{var e;return((e=l.value.find(n=>n.name.endsWith(".xlsx")))==null?void 0:e.name)||""});function K(e){return e==null?"-":e>1024*1024?(e/1024/1024).toFixed(1)+" MB":e>1024?(e/1024).toFixed(1)+" KB":e+" B"}function A(e){return e==="done"?"green":e==="error"?"red":e==="stopped"?"orange":"blue"}function M(e){return e==="done"?"完成":e==="error"?"失败":e==="stopped"?"已停止":"运行中"}async function s(){if(d.runId)try{const[e,n]=await Promise.all([F.runSummary(d.runId),F.getRun(d.runId).catch(()=>null)]);Object.assign(u,e),r.value=e.meta||{},l.value=((n==null?void 0:n.files)||[]).map(C=>({name:C[0],size:C[1]}))}catch(e){W.error(e.message)}}async function k(e){T.value=e,f.content="";try{const n=await F.previewFile(d.runId,e);f.content=n.content,f.total_lines=n.total_lines,f.truncated=n.truncated,m.value=!0}catch(n){W.error(n.message)}}function g(e){if(!e){W.warning("该运行暂无 xlsx 汇总（可能没有成功记录）");return}window.open(F.downloadUrl(d.runId,e),"_blank")}return G(()=>d.runId,()=>s(),{immediate:!0}),(e,n)=>{const C=c("a-empty"),$=c("a-tag"),R=c("a-button"),_=c("a-space"),z=c("a-card"),h=c("a-tab-pane"),N=c("a-table"),J=c("a-modal"),Q=c("a-tabs");return b(),S("div",ce,[p.runId?(b(),S(X,{key:1},[t(z,{size:"small",class:"detail-meta",bordered:!0},{default:a(()=>[t(_,{wrap:""},{default:a(()=>{var I,P,U;return[t($,{color:"blue"},{default:a(()=>{var y,V;return[v(i(((y=u.meta)==null?void 0:y.framework_name)||((V=r.value)==null?void 0:V.framework)||"-"),1)]}),_:1}),t($,null,{default:a(()=>{var y,V;return[v(i(((y=u.meta)==null?void 0:y.model)||((V=r.value)==null?void 0:V.model)||"-"),1)]}),_:1}),(I=r.value)!=null&&I.gpu?(b(),L($,{key:0},{default:a(()=>[v("GPU: "+i(r.value.gpu),1)]),_:1})):D("",!0),(P=r.value)!=null&&P.precision?(b(),L($,{key:1},{default:a(()=>[v("精度: "+i(r.value.precision),1)]),_:1})):D("",!0),t($,null,{default:a(()=>{var y,V;return[v(i(((y=r.value)==null?void 0:y.started_at)||"")+" → "+i(((V=r.value)==null?void 0:V.finished_at)||""),1)]}),_:1}),t($,{color:A((U=r.value)==null?void 0:U.status)},{default:a(()=>{var y;return[v(i(M((y=r.value)==null?void 0:y.status)),1)]}),_:1},8,["color"]),t(R,{size:"small",type:"primary",ghost:"",onClick:n[0]||(n[0]=y=>g(j.value))},{default:a(()=>[...n[3]||(n[3]=[v("下载 xlsx 汇总",-1)])]),_:1}),t(R,{size:"small",onClick:s},{default:a(()=>[...n[4]||(n[4]=[v("刷新",-1)])]),_:1})]}),_:1})]),_:1}),t(Q,{activeKey:o.value,"onUpdate:activeKey":n[2]||(n[2]=I=>o.value=I),class:"detail-tabs"},{default:a(()=>[t(h,{key:"summary",tab:"指标汇总"},{default:a(()=>[t(de,{records:u.records||[]},null,8,["records"])]),_:1}),t(h,{key:"mean",tab:"均值分析 Mean"},{default:a(()=>[t(q,{rows:u.records_mean||[],best:u.best_mean||{},threshold:u.threshold},null,8,["rows","best","threshold"])]),_:1}),t(h,{key:"p99",tab:"P99 分析"},{default:a(()=>[t(q,{rows:u.records_p99||[],best:u.best_p99||{},threshold:u.threshold},null,8,["rows","best","threshold"])]),_:1}),t(h,{key:"files",tab:"日志文件"},{default:a(()=>[t(N,{columns:B,"data-source":l.value,size:"small",pagination:!1,"row-key":"name"},{bodyCell:a(({column:I,record:P})=>[I.key==="actions"?(b(),S(X,{key:0},[t(R,{size:"small",type:"link",onClick:U=>k(P.name)},{default:a(()=>[...n[5]||(n[5]=[v("预览",-1)])]),_:1},8,["onClick"]),t(R,{size:"small",type:"link",onClick:U=>g(P.name)},{default:a(()=>[...n[6]||(n[6]=[v("下载",-1)])]),_:1},8,["onClick"])],64)):D("",!0)]),_:1},8,["data-source"]),t(J,{open:m.value,"onUpdate:open":n[1]||(n[1]=I=>m.value=I),title:`预览：${T.value}`,width:"900px",footer:null},{default:a(()=>[w("div",me," 共 "+i(f.total_lines)+" 行，显示末尾 "+i(f.truncated>0?`（省略前 ${f.truncated} 行）`:"全部"),1),w("pre",pe,i(f.content),1)]),_:1},8,["open","title"])]),_:1})]),_:1},8,["activeKey"])],64)):(b(),L(C,{key:0,description:"请选择左侧测试记录",style:{"padding-top":"80px"}}))])}}},fe=E(_e,[["__scopeId","data-v-53d908f7"]]),ye={__name:"LogView",setup(p){const d=x("");return(u,r)=>{const l=c("a-layout-sider"),o=c("a-layout-content"),m=c("a-layout");return b(),L(m,{class:"page-layout"},{default:a(()=>[t(l,{width:"260",theme:"light",class:"record-sider"},{default:a(()=>[t(ue,{modelValue:d.value,"onUpdate:modelValue":r[0]||(r[0]=T=>d.value=T),framework:""},null,8,["modelValue"])]),_:1}),t(o,{class:"page-content"},{default:a(()=>[t(fe,{"run-id":d.value},null,8,["run-id"])]),_:1})]),_:1})}}},he=E(ye,[["__scopeId","data-v-8fab23f1"]]);export{he as default};

benchscope/webui/assets/MetricsCharts-D1wU0LbK.css

@@ -0,0 +1 @@
+.row-near-threshold>td{background-color:#e6fffb!important}.row-best>td{background-color:#fffbe6!important;font-weight:600}.row-error>td{background-color:#fff1f0!important}.chart-title[data-v-99455d6f]{font-size:13px;color:#333;font-weight:600;margin-bottom:4px;padding-left:4px}

benchscope/webui/assets/MetricsCharts-DHO93JrC.js

@@ -0,0 +1 @@
+import{S as g,U as a,V as p,W as y,a6 as b,_ as k,F as T,G as x,a8 as v,c as R,w as z,o as C,j as O,a7 as N,X as S,a9 as I}from"./vue-Ch4zjUb1.js";import{i as P}from"./echarts-Bb6yjXMn.js";import{_ as F}from"./index-Dfr0hp72.js";const L={key:0,style:{fontWeight:600}},A={key:0,style:{color:"#999","font-size":"12px"}},E={__name:"MetricsTable",props:{rows:{type:Array,default:()=>[]},threshold:{type:Number,default:null},pagination:{type:[Boolean,Object],default:!1}},setup(d){const m=d,c=R(()=>[{title:"用例 Case",key:"label",width:180,fixed:"left"},{title:"并发数 Concurrency",key:"concurrency",width:110,fixed:"left"},{title:"Output 吞吐 Output tok/s",dataIndex:["metrics","output_mean"],key:"output",width:130,customRender:s},{title:"Peak Output 吞吐 Peak tok/s",dataIndex:["metrics","peakoutput_mean"],key:"peak",width:140,customRender:s},{title:"Total 吞吐 Total tok/s",dataIndex:["metrics","total_mean"],key:"total",width:130,customRender:s},{title:"TTFT mean (ms)",dataIndex:["metrics","ttft_mean"],key:"ttft_mean",width:120,customRender:s},{title:"TPOT mean (ms)",dataIndex:["metrics","tpot_mean"],key:"tpot_mean",width:120,customRender:s},{title:"ITL mean (ms)",dataIndex:["metrics","itl_mean"],key:"itl_mean",width:120,customRender:s},{title:"TTFT P99 (ms)",dataIndex:["metrics","ttft_p99"],key:"ttft_p99",width:120,customRender:s},{title:"TPOT P99 (ms)",dataIndex:["metrics","tpot_p99"],key:"tpot_p99",width:120,customRender:s},{title:"ITL P99 (ms)",dataIndex:["metrics","itl_p99"],key:"itl_p99",width:120,customRender:s},{title:"单用户 QPS (1000/TPOT)",key:"single_user",width:150,customRender:w},{title:"状态 Status",key:"status",width:90,fixed:"right"}]);function s({text:n}){return n==null?"-":Number(n).toFixed(2)}function w({record:n}){var o;const e=(o=n.metrics)==null?void 0:o.single_user;return e==null?"-":Number(e).toFixed(2)}function h(n){return`${n.label||n.case||""}-${n.concurrency}`}function r(n){var o;if(n.error)return"row-error";if(n.best)return"row-best";const e=(o=n.metrics)==null?void 0:o.tpot_mean;return e!=null&&m.threshold&&Number(e)<Number(m.threshold)?"row-near-threshold":""}return(n,e)=>{const o=g("a-tag"),u=g("a-table");return a(),p(u,{columns:c.value,"data-source":d.rows,"row-key":h,pagination:d.pagination,size:"small",scroll:{x:1400},"row-class-name":r,bordered:""},{bodyCell:y(({column:i,record:t})=>[i.key==="concurrency"?(a(),b("span",L,k(t.concurrency),1)):i.key==="status"?(a(),b(T,{key:1},[t.error?(a(),p(o,{key:0,color:"red"},{default:y(()=>[...e[0]||(e[0]=[x("失败",-1)])]),_:1})):t.best?(a(),p(o,{key:1,color:"gold"},{default:y(()=>[...e[1]||(e[1]=[x("最佳",-1)])]),_:1})):(a(),p(o,{key:2,color:"green"},{default:y(()=>[...e[2]||(e[2]=[x("成功",-1)])]),_:1}))],64)):i.key==="label"?(a(),b(T,{key:2},[x(k(t.label)+" ",1),t.input_len?(a(),b("span",A," ("+k(t.input_len)+"/"+k(t.output_len)+") ",1)):v("",!0)],64)):v("",!0)]),_:1},8,["columns","data-source","pagination"])}}},B={class:"chart-title"},D="260px",M={__name:"MetricsCharts",props:{rows:{type:Array,default:()=>[]},metricDefs:{type:Array,default:()=>[]},height:{type:String,default:"300px"}},setup(d){const m=d,c={},s=[];function w(r,n){n&&(c[r]||(c[r]=P(n),s.push(new ResizeObserver(()=>c[r]&&c[r].resize())),s[s.length-1].observe(n)))}function h(){for(const r of m.metricDefs){const n=c[r.key];if(!n)continue;const e={};for(const t of m.rows){const l=t.metrics||{},f=r.value?r.value(l):l[r.key];if(f==null)continue;const _=t.label||t.case||"unknown";e[_]||(e[_]=[]),e[_].push({concurrency:t.concurrency,value:Number(f)})}const o=Object.keys(e),u={color:["#1677ff","#52c41a","#faad14","#f5222d","#13c2c2","#722ed1","#eb2f96","#fa8c16"],tooltip:{trigger:"axis",backgroundColor:"rgba(255,255,255,0.96)",borderColor:"#f0f0f0",textStyle:{color:"rgba(0,0,0,0.88)",fontSize:12},valueFormatter:t=>t==null?"-":Number(t).toFixed(2)},legend:o.length>1?{data:o,type:"scroll",top:0,textStyle:{fontSize:12}}:void 0,grid:{left:54,right:16,top:o.length>1?32:20,bottom:28},xAxis:{type:"category",name:"并发数",nameTextStyle:{fontSize:11,color:"rgba(0,0,0,0.45)"},axisLine:{lineStyle:{color:"#d9d9d9"}},axisLabel:{fontSize:11}},yAxis:{type:"value",scale:!0,splitLine:{lineStyle:{type:"dashed",color:"#f0f0f0"}},axisLabel:{fontSize:11}},series:o.map(t=>{const l=e[t].sort((f,_)=>f.concurrency-_.concurrency);return{name:t,type:"line",smooth:!0,symbol:"circle",symbolSize:6,lineStyle:{width:2},data:l.map(f=>f.value),markPoint:{data:[{type:"max",name:"峰值"}],symbolSize:44,label:{fontSize:10}}}})},i=[];for(const t of o)for(const l of e[t])i.includes(l.concurrency)||i.push(l.concurrency);u.xAxis.data=i.sort((t,l)=>t-l),n.setOption(u,!0)}}return z(()=>m.rows,h,{deep:!0}),z(()=>m.metricDefs,h,{deep:!0}),C(h),O(()=>{s.forEach(r=>r.disconnect()),Object.values(c).forEach(r=>r.dispose())}),(r,n)=>{const e=g("a-col"),o=g("a-row");return a(),p(o,{gutter:12},{default:y(()=>[(a(!0),b(T,null,N(d.metricDefs,u=>(a(),p(e,{key:u.key,xs:24,sm:12,lg:8,style:{"margin-bottom":"12px"}},{default:y(()=>[S("div",{style:I({height:d.height,width:"100%"})},[S("div",B,k(u.label),1),S("div",{ref_for:!0,ref:i=>w(u.key,i),style:I({height:D,width:"100%"})},null,4)],4)]),_:2},1024))),128))]),_:1})}}},U=F(M,[["__scopeId","data-v-99455d6f"]]);export{U as M,E as _};

benchscope/webui/assets/SettingsView-7uuTFqXU.css

@@ -0,0 +1 @@
+.panel[data-v-a3dbcdbc]{max-width:1200px;margin:0 auto;border-radius:8px;box-shadow:0 1px 2px #0000000a}

benchscope/webui/assets/SettingsView-DJmFSQbU.js

@@ -0,0 +1 @@
+import{_ as F,u as H,a as K}from"./index-Dfr0hp72.js";import{i as j,t as W,u as Q,v as Y,j as I}from"./antd-DWALckI0.js";import{o as Z,V as ee,W as t,f as k,r as ae,S as r,U as h,k as e,u as w,G as p,a6 as V,_ as x,a8 as E,X as te}from"./vue-Ch4zjUb1.js";const le={key:0,style:{"margin-right":"12px"}},ne={key:0},oe={key:1,style:{color:"#999"}},se={key:0,style:{color:"#999","font-size":"12px","margin-top":"4px"}},ue={__name:"SettingsView",setup(re){const o=H(),A=k(!1),B=k(!1),_=k(null),U=k(null),l=ae({framework:"vllm",api:{base_url:"",endpoint:"/v1/chat/completions",api_key:"",extra_headers:{}},gpu:{auto:!0,name:"",count:8},logs_dir:"./logs",datasets_dir:"./datasets",tpot_threshold_ms:100,request_rate:"inf",bench_commands:{vllm:"vllm bench serve",sglang:"python -m sglang.bench_serving"}}),v=k("{}");Z(async()=>{var d,a,g,m,c,S,s,u,i,f,O,C,N,b,y,q,L,P,G,n,J,z,D,T,M,R;A.value=!0;try{await o.load(),Object.assign(l,{framework:((d=o.config)==null?void 0:d.framework)||"vllm",api:{base_url:((g=(a=o.config)==null?void 0:a.api)==null?void 0:g.base_url)||"",endpoint:((c=(m=o.config)==null?void 0:m.api)==null?void 0:c.endpoint)||"/v1/chat/completions",api_key:((s=(S=o.config)==null?void 0:S.api)==null?void 0:s.api_key)||"",extra_headers:((i=(u=o.config)==null?void 0:u.api)==null?void 0:i.extra_headers)||{}},gpu:{auto:((O=(f=o.config)==null?void 0:f.gpu)==null?void 0:O.auto)??!0,name:((N=(C=o.config)==null?void 0:C.gpu)==null?void 0:N.name)||"",count:((y=(b=o.config)==null?void 0:b.gpu)==null?void 0:y.count)||8},logs_dir:((q=o.config)==null?void 0:q.logs_dir)||"./logs",datasets_dir:((L=o.config)==null?void 0:L.datasets_dir)||"./datasets",tpot_threshold_ms:((P=o.config)==null?void 0:P.tpot_threshold_ms)??100,request_rate:((G=o.config)==null?void 0:G.request_rate)||"inf",bench_commands:{vllm:((J=(n=o.config)==null?void 0:n.bench_commands)==null?void 0:J.vllm)||"vllm bench serve",sglang:((D=(z=o.config)==null?void 0:z.bench_commands)==null?void 0:D.sglang)||"python -m sglang.bench_serving"}}),v.value=JSON.stringify(((M=(T=o.config)==null?void 0:T.api)==null?void 0:M.extra_headers)||{},null,2),(R=o.gpu)!=null&&R.auto_detected&&(U.value=o.gpu.auto_detected)}finally{A.value=!1}});async function X(){B.value=!0,_.value=null;try{let d={};try{d=JSON.parse(v.value||"{}")}catch{I.error("额外请求头不是合法 JSON");return}_.value=await K.testConnection({base_url:l.api.base_url,endpoint:l.api.endpoint,api_key:l.api.api_key,extra_headers:d})}finally{B.value=!1}}async function $(){try{let d={};try{d=JSON.parse(v.value||"{}")}catch{I.error("额外请求头不是合法 JSON");return}const a={framework:l.framework,api:{...l.api,extra_headers:d},gpu:l.gpu,logs_dir:l.logs_dir,datasets_dir:l.datasets_dir,tpot_threshold_ms:l.tpot_threshold_ms,request_rate:l.request_rate,bench_commands:l.bench_commands};await o.save(a),I.success("配置已保存"),o.refreshStatus()}catch(d){I.error(`保存失败：${d.message}`)}}return(d,a)=>{const g=r("a-button"),m=r("a-divider"),c=r("a-radio-button"),S=r("a-radio-group"),s=r("a-form-item"),u=r("a-col"),i=r("a-input"),f=r("a-row"),O=r("a-input-password"),C=r("a-tag"),N=r("a-switch"),b=r("a-input-number"),y=r("a-select-option"),q=r("a-select"),L=r("a-form"),P=r("a-spin"),G=r("a-card");return h(),ee(G,{size:"small",class:"panel",title:"服务设置 Service Settings"},{extra:t(()=>[e(g,{type:"link",size:"small",loading:B.value,onClick:X},{icon:t(()=>[e(w(j))]),default:t(()=>[a[14]||(a[14]=p(" 测试连接 ",-1))]),_:1},8,["loading"])]),default:t(()=>[e(P,{spinning:A.value},{default:t(()=>[e(L,{layout:"vertical"},{default:t(()=>[e(m,{orientation:"left"},{default:t(()=>[e(w(j),{style:{color:"#1677ff"}}),a[15]||(a[15]=p(" 推理服务 API 配置 ",-1))]),_:1}),e(f,{gutter:16},{default:t(()=>[e(u,{span:8},{default:t(()=>[e(s,{label:"默认框架 Default Framework"},{default:t(()=>[e(S,{value:l.framework,"onUpdate:value":a[0]||(a[0]=n=>l.framework=n),"button-style":"solid"},{default:t(()=>[e(c,{value:"vllm"},{default:t(()=>[...a[16]||(a[16]=[p("vLLM",-1)])]),_:1}),e(c,{value:"sglang"},{default:t(()=>[...a[17]||(a[17]=[p("SGLang",-1)])]),_:1})]),_:1},8,["value"])]),_:1})]),_:1}),e(u,{span:8},{default:t(()=>[e(s,{label:"Base URL（OpenAI 兼容）"},{default:t(()=>[e(i,{value:l.api.base_url,"onUpdate:value":a[1]||(a[1]=n=>l.api.base_url=n),placeholder:"http://192.168.1.67:8000"},null,8,["value"])]),_:1})]),_:1}),e(u,{span:8},{default:t(()=>[e(s,{label:"Endpoint"},{default:t(()=>[e(i,{value:l.api.endpoint,"onUpdate:value":a[2]||(a[2]=n=>l.api.endpoint=n),placeholder:"/v1/chat/completions"},null,8,["value"])]),_:1})]),_:1})]),_:1}),e(f,{gutter:16},{default:t(()=>[e(u,{span:8},{default:t(()=>[e(s,{label:"API Key（可选）"},{default:t(()=>[e(O,{value:l.api.api_key,"onUpdate:value":a[3]||(a[3]=n=>l.api.api_key=n),placeholder:"Bearer token"},null,8,["value"])]),_:1})]),_:1}),e(u,{span:16},{default:t(()=>[e(s,{label:"额外请求头（JSON，可选）"},{default:t(()=>[e(i,{value:v.value,"onUpdate:value":a[4]||(a[4]=n=>v.value=n),placeholder:'{"X-Custom": "value"}'},null,8,["value"])]),_:1})]),_:1})]),_:1}),e(s,null,{default:t(()=>{var n;return[_.value?(h(),V("span",le,[e(C,{color:_.value.ok?"green":"red"},{default:t(()=>[p(x(_.value.ok?"连接成功":"连接失败"),1)]),_:1},8,["color"]),_.value.ok?(h(),V("span",ne,x(((n=_.value.models)==null?void 0:n.length)||0)+" 个模型",1)):(h(),V("span",oe,x(_.value.error),1))])):E("",!0)]}),_:1}),e(m,{orientation:"left"},{default:t(()=>[e(w(W),{style:{color:"#1677ff"}}),a[18]||(a[18]=p(" GPU / 目录 / 阈值 ",-1))]),_:1}),e(f,{gutter:16},{default:t(()=>[e(u,{span:6},{default:t(()=>[e(s,{label:"GPU 自动检测"},{default:t(()=>[e(N,{checked:l.gpu.auto,"onUpdate:checked":a[5]||(a[5]=n=>l.gpu.auto=n)},null,8,["checked"]),U.value?(h(),V("div",se," 检测到："+x(U.value.name)+" × "+x(U.value.count),1)):E("",!0)]),_:1})]),_:1}),e(u,{span:5},{default:t(()=>[e(s,{label:"GPU 型号（回退）"},{default:t(()=>[e(i,{value:l.gpu.name,"onUpdate:value":a[6]||(a[6]=n=>l.gpu.name=n)},null,8,["value"])]),_:1})]),_:1}),e(u,{span:4},{default:t(()=>[e(s,{label:"GPU 数量"},{default:t(()=>[e(b,{value:l.gpu.count,"onUpdate:value":a[7]||(a[7]=n=>l.gpu.count=n),min:1,style:{width:"100%"}},null,8,["value"])]),_:1})]),_:1}),e(u,{span:9},{default:t(()=>[e(s,{label:"TPOT 阈值 (ms)（默认）"},{default:t(()=>[e(b,{value:l.tpot_threshold_ms,"onUpdate:value":a[8]||(a[8]=n=>l.tpot_threshold_ms=n),min:1,style:{width:"100%"}},null,8,["value"])]),_:1})]),_:1})]),_:1}),e(f,{gutter:16},{default:t(()=>[e(u,{span:8},{default:t(()=>[e(s,{label:"日志目录 logs_dir"},{default:t(()=>[e(i,{value:l.logs_dir,"onUpdate:value":a[9]||(a[9]=n=>l.logs_dir=n),placeholder:"./logs"},null,8,["value"])]),_:1})]),_:1}),e(u,{span:8},{default:t(()=>[e(s,{label:"数据集缓存目录 datasets_dir"},{default:t(()=>[e(i,{value:l.datasets_dir,"onUpdate:value":a[10]||(a[10]=n=>l.datasets_dir=n),placeholder:"./datasets"},null,8,["value"])]),_:1})]),_:1}),e(u,{span:8},{default:t(()=>[e(s,{label:"请求速率 Request rate（默认）"},{default:t(()=>[e(q,{value:l.request_rate,"onUpdate:value":a[11]||(a[11]=n=>l.request_rate=n)},{default:t(()=>[e(y,{value:"inf"},{default:t(()=>[...a[19]||(a[19]=[p("inf（不限速）",-1)])]),_:1}),e(y,{value:"custom"},{default:t(()=>[...a[20]||(a[20]=[p("自定义",-1)])]),_:1})]),_:1},8,["value"])]),_:1})]),_:1})]),_:1}),e(m,{orientation:"left"},{default:t(()=>[e(w(Q),{style:{color:"#1677ff"}}),a[21]||(a[21]=p(" bench 执行命令 ",-1))]),_:1}),e(f,{gutter:16},{default:t(()=>[e(u,{span:12},{default:t(()=>[e(s,{label:"vLLM bench 命令模板"},{default:t(()=>[e(i,{value:l.bench_commands.vllm,"onUpdate:value":a[12]||(a[12]=n=>l.bench_commands.vllm=n),placeholder:"vllm bench serve"},null,8,["value"])]),_:1})]),_:1}),e(u,{span:12},{default:t(()=>[e(s,{label:"SGLang bench 命令模板"},{default:t(()=>[e(i,{value:l.bench_commands.sglang,"onUpdate:value":a[13]||(a[13]=n=>l.bench_commands.sglang=n),placeholder:"python -m sglang.bench_serving"},null,8,["value"])]),_:1})]),_:1})]),_:1}),a[23]||(a[23]=te("div",{style:{color:"#999","font-size":"12px","margin-bottom":"12px"}}," 说明：bench 工具在 benchscope 所在机器以子进程运行（需安装 vllm / sglang CLI），推理服务端只需提供 OpenAI 兼容 API，无需安装插件。 ",-1)),e(s,null,{default:t(()=>[e(g,{type:"primary",size:"large",onClick:$},{icon:t(()=>[e(w(Y))]),default:t(()=>[a[22]||(a[22]=p(" 保存配置 ",-1))]),_:1})]),_:1})]),_:1})]),_:1},8,["spinning"])]),_:1})}}},_e=F(ue,[["__scopeId","data-v-a3dbcdbc"]]);export{_e as default};