feat: 完成 issue #8 ⑥ K8s/Helm 部署底座 + 昇腾适配层

2026-08-04 22:04:18 +08:00
parent 691a812fcf
commit e620d4937e
20 changed files with 1013 additions and 0 deletions
@@ -0,0 +1,16 @@
+# -*- coding: utf-8 -*-
+"""测试引导：把 `core/inference-backend` 以包名 `inference_backend` 挂载到 sys.modules。
+
+目录名 `inference-backend` 含连字符，无法直接以包名 import；挂载后模块内相对导入
+（`from inference_backend.base import ...`）在 unittest 发现机制下可正常解析。
+"""
+import os
+import sys
+import types
+
+BACKEND_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+sys.path.insert(0, BACKEND_DIR)
+if "inference_backend" not in sys.modules:
+    pkg = types.ModuleType("inference_backend")
+    pkg.__path__ = [BACKEND_DIR]
+    sys.modules["inference_backend"] = pkg
@@ -0,0 +1,127 @@
+# -*- coding: utf-8 -*-
+"""推理后端抽象层单元测试（EPIC #8 ⑥ 部署底座 / PRD 5.6）。
+
+覆盖：
+1. 模板配置资产可解析、含 backend 选择键；
+2. 工厂按配置构建 GPU / NPU 后端（可插拔）；
+3. 统一接口 load_model / infer / health / unload 行为；
+4. 切换后端仅改配置：同一业务调用面（接口方法）在两种后端下等价；
+5. 非法配置报错。
+"""
+import os
+import sys
+import unittest
+from unittest import mock
+
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+import _bootstrap  # noqa: F401
+
+from inference_backend.base import InferRequest, InferenceBackend  # noqa: E402
+from inference_backend.factory import (  # noqa: E402
+    BACKEND_REGISTRY,
+    build_backend,
+    default_config_path,
+    load_backend_config,
+)
+
+CONFIG_PATH = default_config_path()
+
+
+class TestTemplateConfig(unittest.TestCase):
+    """模板配置资产：可解析且包含后端选择配置点。"""
+
+    def test_template_parses_and_has_backend_key(self):
+        cfg = load_backend_config(CONFIG_PATH)
+        self.assertIn("backend", cfg)
+        self.assertIn(cfg["backend"].lower(), ("gpu", "npu"))
+        self.assertIn("inference", cfg)
+
+    def test_registry_covers_gpu_and_npu(self):
+        self.assertEqual(sorted(BACKEND_REGISTRY), ["gpu", "npu"])
+
+
+class TestFactory(unittest.TestCase):
+    """工厂：按配置选择后端实现（可插拔适配层）。"""
+
+    def test_build_gpu_from_template(self):
+        cfg = load_backend_config(CONFIG_PATH)
+        backend = build_backend(cfg)
+        self.assertIsInstance(backend, InferenceBackend)
+        self.assertEqual(backend.backend_name, cfg["backend"].lower())
+        self.assertEqual(backend.model, cfg["inference"]["model"])
+
+    def test_build_npu_by_config_switch(self):
+        cfg = load_backend_config(CONFIG_PATH)
+        cfg["backend"] = "npu"
+        backend = build_backend(cfg)
+        self.assertEqual(backend.backend_name, "npu")
+        # 业务调用面不变（接口等价，不感知硬件）
+        self.assertTrue(callable(backend.load_model))
+        self.assertTrue(callable(backend.infer))
+        self.assertTrue(callable(backend.health))
+        self.assertTrue(callable(backend.unload))
+
+    def test_build_rejects_unknown_backend(self):
+        with self.assertRaises(ValueError):
+            build_backend({"backend": "tpu"})
+
+    def test_build_rejects_missing_backend(self):
+        with self.assertRaises(ValueError):
+            build_backend({"inference": {"model": "x"}})
+
+
+class TestBackendBehavior(unittest.TestCase):
+    """统一接口行为（dry-run：未配置 endpoint 时适配层就绪）。"""
+
+    def setUp(self):
+        cfg = load_backend_config(CONFIG_PATH)
+        cfg["inference"]["endpoint"] = ""  # dry-run
+        self.gpu = build_backend(cfg)
+        cfg["backend"] = "npu"
+        self.npu = build_backend(cfg)
+
+    def test_dry_run_load_health_unload(self):
+        for backend in (self.gpu, self.npu):
+            loaded = backend.load_model()
+            self.assertEqual(loaded["status"], "ok")
+            health = backend.health()
+            self.assertEqual(health["status"], "ok")
+            self.assertEqual(health["backend"], backend.backend_name)
+            self.assertTrue(health["loaded"])
+            unloaded = backend.unload()
+            self.assertEqual(unloaded["status"], "ok")
+
+    def test_infer_extracts_text_from_openai_response(self):
+        fake_body = {
+            "choices": [{"message": {"content": "炉温偏高，建议降低加料比"}}],
+        }
+        for backend in (self.gpu, self.npu):
+            with mock.patch.object(backend, "_post_json",
+                                   return_value=fake_body) as post:
+                result = backend.infer(InferRequest(prompt="请解释炉温报警"))
+                post.assert_called_once_with(
+                    "/v1/chat/completions",
+                    mock.ANY,
+                )
+            self.assertEqual(result.text, "炉温偏高，建议降低加料比")
+            self.assertEqual(result.backend, backend.backend_name)
+            # mock 调用耗时可能为 0，仅断言字段类型与量级
+            self.assertIsInstance(result.latency_ms, float)
+            self.assertGreaterEqual(result.latency_ms, 0.0)
+
+    def test_infer_payload_carries_model(self):
+        with mock.patch.object(self.gpu, "_post_json",
+                               return_value={"choices": []}) as post:
+            self.gpu.infer(InferRequest(prompt="hi", max_tokens=64))
+            _, payload = post.call_args[0]
+            self.assertEqual(payload["model"], self.gpu.model)
+            self.assertEqual(payload["max_tokens"], 64)
+
+    def test_health_dry_run_notes_reason(self):
+        health = self.gpu.health()
+        self.assertIn("reason", health)
+        self.assertIn("dry-run", health["reason"])
+
+
+if __name__ == "__main__":
+    unittest.main()