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

tests/test_llama_server.py

@@ -0,0 +1,152 @@
+"""T2 llama-server 进程管理单测（封闭：假二进制 + 注入，D11）。"""
+import json
+import os
+import subprocess
+import sys
+import tempfile
+from pathlib import Path
+
+import pytest
+
+from runtime.llama_server import LlamaServerManager, LlamaServerError, build_llama_server
+from tests._ports import free_port
+
+FAKE_SCRIPT = Path(__file__).parent / "fixtures" / "fake_llama_server.py"
+PYTHON = sys.executable
+
+
+def _make_fake_binary(tmp: Path) -> Path:
+    """生成一个 .cmd 包装器：把 venv python + 假脚本当作"二进制"启动。"""
+    cmd = tmp / "fake-llama-server.cmd"
+    cmd.write_text(
+        f'@echo off\r\n"{PYTHON}" "{FAKE_SCRIPT}" %*\r\n',
+        encoding="utf-8",
+    )
+    return cmd
+
+
+def _make_manager(tmp, binary, port, model, **kw):
+    marker = tmp / "marker.json"
+    env = dict(os.environ)
+    env["FAKE_MARKER"] = str(marker)
+    return LlamaServerManager(
+        binary=str(binary),
+        model=str(model),
+        port=port,
+        hw={"tier": "cpu"},
+        health_timeout_s=15.0,
+        poll_interval_s=0.2,
+        log_dir=str(tmp / "runs"),
+        env=env,
+        **kw,
+    ), marker
+
+
+def test_build_command_uses_hw_tier():
+    m = LlamaServerManager(binary="bin/x.exe", model="models/m.gguf", port=8901,
+                           hw={"tier": "gpu12"})
+    cmd = m._build_command()
+    assert Path(cmd[0]) == Path("bin/x.exe")
+    assert cmd[1] == "-m" and Path(cmd[2]) == Path("models/m.gguf")
+    assert "--port" in cmd and "8901" in cmd
+    assert cmd[cmd.index("-ngl") + 1] == "99"
+    assert cmd[cmd.index("-c") + 1] == "32768"
+
+
+def test_build_command_extra_args_appended():
+    m = LlamaServerManager(binary="bin/x.exe", model="models/m.gguf", port=1,
+                           hw={"tier": "cpu"}, extra_args=["--cache-reuse", "256"])
+    cmd = m._build_command()
+    assert cmd[-2:] == ["--cache-reuse", "256"]
+
+
+def test_start_missing_binary_raises(tmp_path):
+    m = LlamaServerManager(binary=str(tmp_path / "nope.exe"), model=str(tmp_path / "m.gguf"),
+                           port=free_port())
+    with pytest.raises(LlamaServerError):
+        m.start()
+
+
+def test_start_missing_model_raises(tmp_path):
+    binary = _make_fake_binary(tmp_path)
+    m = LlamaServerManager(binary=str(binary), model=str(tmp_path / "missing.gguf"),
+                           port=free_port())
+    with pytest.raises(LlamaServerError):
+        m.start()
+
+
+def test_full_lifecycle(tmp_path):
+    binary = _make_fake_binary(tmp_path)
+    model = tmp_path / "model.gguf"
+    model.write_bytes(b"fake")
+    port = free_port()
+    m, marker = _make_manager(tmp_path, binary, port, model)
+
+    assert m.running is False
+    assert m.health() is False  # 无进程时不健康
+
+    ok = m.start()
+    assert ok is True
+    assert m.running is True
+    assert m.health() is True
+
+    # 假脚本确实收到了参数
+    data = json.loads(marker.read_text(encoding="utf-8"))
+    assert data["port"] == port
+    assert data["model"] == str(model)
+
+    # 再 start 幂等（已运行返回健康）
+    assert m.start() is True
+
+    m.stop()
+    assert m.running is False
+    assert m.health() is False
+
+
+def test_stop_idempotent(tmp_path):
+    binary = _make_fake_binary(tmp_path)
+    model = tmp_path / "model.gguf"
+    model.write_bytes(b"fake")
+    m, _ = _make_manager(tmp_path, binary, free_port(), model)
+    m.stop()  # 未启动时 stop 不抛
+    assert m.running is False
+
+
+def test_ensure_alive_healthy_no_restart(tmp_path):
+    binary = _make_fake_binary(tmp_path)
+    model = tmp_path / "model.gguf"
+    model.write_bytes(b"fake")
+    m, _ = _make_manager(tmp_path, binary, free_port(), model)
+    m.start()
+    assert m.ensure_alive() is True
+    # 不应触发重启
+    assert m._restart_count == 0
+    m.stop()
+
+
+def test_ensure_alive_restart_exhausted(tmp_path):
+    binary = _make_fake_binary(tmp_path)
+    model = tmp_path / "model.gguf"
+    model.write_bytes(b"fake")
+    m, _ = _make_manager(tmp_path, binary, free_port(), model, max_restarts=0)
+    # 未启动：restart 上限 0 -> False
+    assert m.ensure_alive() is False
+
+
+def test_endpoint_format():
+    m = LlamaServerManager(binary="x", model="m", port=8901, hw={"tier": "cpu"})
+    assert m.endpoint() == "http://127.0.0.1:8901"
+
+
+def test_build_from_config(tmp_path):
+    binary = _make_fake_binary(tmp_path)
+    cfg = {
+        "binary": str(binary),
+        "model": str(tmp_path / "m.gguf"),
+        "port": 8999,
+        "hw": {"tier": "cpu"},
+        "extra_args": ["-fa"],
+    }
+    m = build_llama_server(cfg)
+    assert m.port == 8999
+    assert "-fa" in m.extra_args