feat(v2): T2 运维层 hw_profile + llama_server 进程管理

runtime/hw_profile.py

@@ -0,0 +1,140 @@
+"""硬件档位检测（runtime 运维层，纯标准库）。
+
+把真实机器映射到三档保守模板之一：
+
+- gpu12 : 约 ≥12GB 显存（NVIDIA / Vulkan 可探测）  -> ngl 99, ctx 32768
+- gpu8  : 约 ≥8GB 显存                            -> ngl 14, ctx 16384
+- cpu   : 无独显或探测失败（保守兜底）               -> ngl 0,  ctx 8192
+
+探测来源：nvidia-smi（NVIDIA 显存）优先；其次 vulkaninfo（AMD/Intel/通用，
+只能判断是否存在 Vulkan 设备，无法可靠拿到显存 -> 保守回退 cpu，并在结果标注
+probe:"conservative"）。总系统内存仅作为 cpu 档提示参考，不作为分档依据。
+
+任何探测失败都回退到 cpu 保守档，保证不崩、可离线运行（D5 / D8）。
+"""
+from __future__ import annotations
+
+import shutil
+import subprocess
+from typing import Any, Callable, Dict, List, Optional
+
+# 三档硬件模板（保守默认，可被 config.tiers 手动覆盖）
+TIER_SPECS: Dict[str, Dict[str, Any]] = {
+    "gpu12": {"tier": "gpu12", "ngl": 99, "ctx": 32768, "kv_quant": "q8_0"},
+    "gpu8":  {"tier": "gpu8",  "ngl": 14, "ctx": 16384, "kv_quant": "q8_0"},
+    "cpu":   {"tier": "cpu",   "ngl": 0,  "ctx": 8192,  "kv_quant": "q8_0"},
+}
+
+_GPU12_THRESHOLD_GB = 12.0
+_GPU8_THRESHOLD_GB = 8.0
+
+
+def _run(cmd: List[str], timeout: float = 10.0,
+         runner: Optional[Callable[[List[str], float], subprocess.CompletedProcess]] = None
+         ) -> Optional[subprocess.CompletedProcess]:
+    """执行命令并捕获输出；失败/超时返回 None（不抛异常）。"""
+    if runner is not None:
+        try:
+            return runner(cmd, timeout)
+        except Exception:
+            return None
+    try:
+        return subprocess.run(
+            cmd, capture_output=True, text=True, timeout=timeout,
+            creationflags=subprocess.CREATE_NO_WINDOW,
+        )
+    except Exception:
+        return None
+
+
+def nvidia_vram_gb(runner: Optional[Callable[..., subprocess.CompletedProcess]] = None) -> Optional[float]:
+    """通过 nvidia-smi 读取显存总量（GB）；无 NVIDIA 返回 None。"""
+    exe = shutil.which("nvidia-smi")
+    if not exe:
+        return None
+    out = _run([exe, "--query-gpu=memory.total", "--format=csv,noheader,nounits"], runner=runner)
+    if out is None or out.returncode != 0 or not out.stdout.strip():
+        return None
+    try:
+        # 多卡取最大值（第一行也接受，但保守起见取最大以保证模板内存够用）
+        vals = [float(v.strip()) for v in out.stdout.strip().splitlines() if v.strip().isdigit()]
+        if not vals:
+            return None
+        return max(vals) / 1024.0
+    except Exception:
+        return None
+
+
+def vulkan_present(runner: Optional[Callable[..., subprocess.CompletedProcess]] = None) -> bool:
+    """检测是否存在 Vulkan 设备（无法可靠拿显存 -> 只用于判定非 cpu 的候选）。"""
+    exe = shutil.which("vulkaninfo")
+    if not exe:
+        return False
+    out = _run([exe, "--summary"], timeout=15.0, runner=runner)
+    if out is None or out.returncode != 0:
+        return False
+    low = out.stdout.lower()
+    # 出现 deviceName 且非 "llvmpipe"/"software" 视为有真实设备
+    return ("devicename" in low or "gpu" in low) and "llvmpipe" not in low and "lavapipe" not in low
+
+
+def pick_tier(vram_gb: Optional[float]) -> str:
+    """按显存选择档位；None/未知 -> cpu 保守档。"""
+    if vram_gb is None:
+        return "cpu"
+    if vram_gb >= _GPU12_THRESHOLD_GB:
+        return "gpu12"
+    if vram_gb >= _GPU8_THRESHOLD_GB:
+        return "gpu8"
+    return "cpu"
+
+
+def detect(override: Optional[Dict[str, Any]] = None,
+           runner: Optional[Callable[..., subprocess.CompletedProcess]] = None) -> Dict[str, Any]:
+    """检测并返回当前档位规格。
+
+    override（可选）：{"tier": "gpu12"} 强制指定档位；或覆盖单个字段如 {"ctx": 16384}。
+
+    返回形如 {"tier": "cpu", "ngl": 0, "ctx": 8192, "kv_quant": "q8_0",
+              "probe": "nvidia|vulkan|cpu|override", "note": str}
+    """
+    if override and override.get("tier") in TIER_SPECS:
+        spec = dict(TIER_SPECS[override["tier"]])
+        spec.update({k: v for k, v in override.items() if k in spec})
+        spec["probe"] = "override"
+        spec["note"] = f"手动指定档位 {override['tier']}"
+        return spec
+
+    vram = nvidia_vram_gb(runner=runner)
+    probe = "nvidia"
+    if vram is None:
+        if vulkan_present(runner=runner):
+            probe = "vulkan"
+            note = "检测到 Vulkan 设备但无法读取显存，按保守档 cpu 运行（可在 config 手动覆盖 tier）"
+        else:
+            probe = "cpu"
+            note = "未检测到 GPU，按 cpu 档运行（-ngl 0，速度受限）"
+    else:
+        note = f"nvidia-smi 探测显存 {vram:.1f}GB"
+
+    tier = pick_tier(vram)
+    spec = dict(TIER_SPECS[tier])
+    spec["probe"] = probe
+    spec["note"] = note if probe != "nvidia" else f"{note} -> 档位 {tier}"
+    return spec
+
+
+def tier_spec(tier: str) -> Dict[str, Any]:
+    """返回指定档位的规格副本（供 config.tiers 兜底）。"""
+    if tier not in TIER_SPECS:
+        raise ValueError(f"未知硬件档位: {tier}（支持: {sorted(TIER_SPECS)}）")
+    return dict(TIER_SPECS[tier])
+
+
+def detect_summary() -> str:
+    """人类可读的检测摘要（setup_runtime / serve 启动时打印）。"""
+    spec = detect()
+    return (
+        f"硬件档位: {spec['tier']} (ngl={spec['ngl']}, ctx={spec['ctx']}, "
+        f"kv_quant={spec['kv_quant']})  [{spec.get('note', '')}]"
+    )