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

tests/test_hw_profile.py

@@ -0,0 +1,114 @@
+"""T2 硬件档位检测单测（封闭，纯逻辑 + 注入 runner）。"""
+import subprocess
+
+import pytest
+
+from runtime.hw_profile import (
+    TIER_SPECS,
+    detect,
+    nvidia_vram_gb,
+    pick_tier,
+    tier_spec,
+    vulkan_present,
+)
+
+
+class _FakeRun:
+    """注入 subprocess runner：按命令返回预置输出。"""
+
+    def __init__(self, mapping):
+        self.mapping = mapping  # {关键子串: CompletedProcess}
+
+    def __call__(self, cmd, timeout):
+        joined = " ".join(cmd)
+        for key, cp in self.mapping.items():
+            if key in joined:
+                return cp
+        raise FileNotFoundError(cmd)
+
+
+def _cp(stdout="", rc=0):
+    return subprocess.CompletedProcess(args=[], returncode=rc, stdout=stdout, stderr="")
+
+
+def test_pick_tier_thresholds():
+    assert pick_tier(None) == "cpu"
+    assert pick_tier(6.0) == "cpu"
+    assert pick_tier(8.0) == "gpu8"
+    assert pick_tier(11.9) == "gpu8"
+    assert pick_tier(12.0) == "gpu12"
+    assert pick_tier(24.0) == "gpu12"
+
+
+def test_tier_specs_have_required_fields():
+    for tier, spec in TIER_SPECS.items():
+        assert spec["tier"] == tier
+        assert "ngl" in spec and "ctx" in spec and "kv_quant" in spec
+    # gpu12 ngl 全量卸载，cpu ngl 0
+    assert TIER_SPECS["gpu12"]["ngl"] == 99
+    assert TIER_SPECS["cpu"]["ngl"] == 0
+
+
+def test_tier_spec_unknown_raises():
+    with pytest.raises(ValueError):
+        tier_spec("nonexistent")
+
+
+def test_nvidia_vram_gb_parses():
+    fake = _FakeRun({"nvidia-smi": _cp("24576\n")})
+    assert nvidia_vram_gb(runner=fake) == 24.0
+
+
+def test_nvidia_vram_gb_multi_gpu_takes_max():
+    fake = _FakeRun({"nvidia-smi": _cp("8192\n12288\n")})
+    assert nvidia_vram_gb(runner=fake) == 12.0
+
+
+def test_nvidia_vram_gb_missing_tool_returns_none(monkeypatch):
+    import shutil
+    monkeypatch.setattr(shutil, "which", lambda name: None)
+    assert nvidia_vram_gb() is None
+
+
+def test_vulkan_present_true():
+    fake = _FakeRun({"vulkaninfo": _cp("deviceName : NVIDIA GeForce RTX 4090")})
+    assert vulkan_present(runner=fake) is True
+
+
+def test_vulkan_present_software_false():
+    fake = _FakeRun({"vulkaninfo": _cp("deviceName : llvmpipe (LLVM)")})
+    assert vulkan_present(runner=fake) is False
+
+
+def test_detect_nvidia_gpu12(monkeypatch):
+    # 让 shutil.which 只对 nvidia-smi 生效
+    real_which = __import__("shutil").which
+    def fake_which(name):
+        return "C:/x/nvidia-smi.exe" if name == "nvidia-smi" else None
+    monkeypatch.setattr(__import__("shutil"), "which", fake_which)
+    fake = _FakeRun({"nvidia-smi": _cp("24576\n")})
+    spec = detect(runner=fake)
+    assert spec["tier"] == "gpu12"
+    assert spec["ngl"] == 99 and spec["ctx"] == 32768
+    assert spec["probe"] == "nvidia"
+
+
+def test_detect_cpu_fallback(monkeypatch):
+    monkeypatch.setattr(__import__("shutil"), "which", lambda name: None)
+    spec = detect()
+    assert spec["tier"] == "cpu"
+    assert spec["ngl"] == 0
+    assert spec["probe"] == "cpu"
+
+
+def test_detect_override_tier():
+    spec = detect(override={"tier": "gpu8"})
+    assert spec["tier"] == "gpu8"
+    assert spec["ngl"] == 14 and spec["ctx"] == 16384
+    assert spec["probe"] == "override"
+
+
+def test_detect_override_field():
+    spec = detect(override={"tier": "cpu", "ctx": 16384})
+    assert spec["ctx"] == 16384
+    assert spec["tier"] == "cpu"