Test: Split C++ unit tests for CI granularity (#3868)

Signed-off-by: Dom Brown <[email protected]>

Author: DomBrown

cpp/tests/CMakeLists.txt

@@ -73,8 +73,6 @@ add_subdirectory(unit_tests)
 
 add_gtest(mpiUtilsTest runtime/mpiUtilsTest.cpp)
 
-add_gtest(loraManagerTest runtime/loraManagerTest.cpp)
-add_gtest(loraCacheTest runtime/loraCacheTest.cpp)
 add_gtest(gptDecoderTest runtime/gptDecoderTest.cpp)
 add_gtest(gptDecoderBatchedTest runtime/gptDecoderBatchedTest.cpp)
 add_gtest(gptSessionTest runtime/gptSessionTest.cpp)

cpp/tests/unit_tests/runtime/CMakeLists.txt

@@ -14,6 +14,8 @@ add_gtest(cudaMemPoolTest cudaMemPoolTest.cpp)
 add_gtest(decodingLayerWorkspaceTest decodingLayerWorkspaceTest.cpp)
 add_gtest(iBufferTest iBufferTest.cpp)
 add_gtest(iTensorTest iTensorTest.cpp)
+add_gtest(loraCacheTest loraCacheTest.cpp)
+add_gtest(loraManagerTest loraManagerTest.cpp)
 add_gtest(loraUtilsTest loraUtilsTest.cpp)
 add_gtest(runtimeKernelTest runtimeKernelTest.cpp)
 add_gtest(samplingConfigTest samplingConfigTest.cpp)

cpp/tests/runtime/loraCacheTest.cpp renamed to cpp/tests/unit_tests/runtime/loraCacheTest.cpp

@@ -41,22 +41,6 @@
 }
 
 
-def generate_excluded_model_tests() -> Generator[str, None, None]:
-    yield "Gpt[^j]"
-    yield "GptExecutor"
-    yield "Gptj"
-    yield "Llama"
-    yield "ChatGlm"
-    yield "Medusa"
-    yield "Eagle"
-    yield "ExplicitDraftTokensDecoding"
-    yield "Mamba"
-    yield "RecurrentGemma"
-    yield "Encoder"
-    yield "EncDec"
-    yield "SpeculativeDecoding"
-
-
 def generate_included_model_tests(
         test_list: List[str]) -> Generator[str, None, None]:
 

cpp/tests/runtime/loraManagerTest.cpp renamed to cpp/tests/unit_tests/runtime/loraManagerTest.cpp

@@ -301,20 +301,24 @@ def _run(
 
 @pytest.mark.parametrize("build_google_tests", ["80", "86", "89", "90"],
                          indirect=True)
-def test_unit_tests(build_google_tests, build_dir, lora_setup):
+@pytest.mark.parametrize("test_group", [
+    "batch_manager", "common", "executor", "kernels", "layers", "runtime",
+    "thop", "utils"
+])
+def test_unit_tests(build_google_tests, test_group, build_dir, lora_setup):
+
+    xml_name = f"results-unit-tests-{test_group}.xml"
 
     # Discover and run the actual gtests
     ctest_command = [
         "ctest",
         "--output-on-failure",
+        "--test-dir",
+        f"{build_dir}/tests/unit_tests/{test_group}",
         "--output-junit",
-        "results-unit-tests.xml",
+        f"{build_dir}/{xml_name}",
     ]
 
-    excluded_tests = list(_cpp.generate_excluded_model_tests())
-
-    ctest_command.extend(["-E", "|".join(excluded_tests)])
-
     parallel = _cpp.default_test_parallel
     if parallel_override := _os.environ.get("LLM_TEST_PARALLEL_OVERRIDE", None):
         parallel = int(parallel_override)

tests/integration/defs/cpp_common.py

@@ -39,7 +39,14 @@ l0_a30:
       backend: cpp
   tests:
   # ------------- CPP tests ---------------
-  - test_cpp.py::test_unit_tests[80]
+  - test_cpp.py::test_unit_tests[batch_manager-80]
+  - test_cpp.py::test_unit_tests[common-80]
+  - test_cpp.py::test_unit_tests[executor-80]
+  - test_cpp.py::test_unit_tests[kernels-80]
+  - test_cpp.py::test_unit_tests[layers-80]
+  - test_cpp.py::test_unit_tests[runtime-80]
+  - test_cpp.py::test_unit_tests[thop-80]
+  - test_cpp.py::test_unit_tests[utils-80]
   - test_cpp.py::test_model[gpt-80]
   - test_cpp.py::test_model[gpt_executor-80]
   - test_cpp.py::test_model[gpt_tests-80]

tests/integration/defs/test_cpp.py

@@ -72,7 +72,14 @@ l0_h100:
       backend: cpp
   tests:
   # ------------- CPP tests ---------------
-  - test_cpp.py::test_unit_tests[90]
+  - test_cpp.py::test_unit_tests[batch_manager-90]
+  - test_cpp.py::test_unit_tests[common-90]
+  - test_cpp.py::test_unit_tests[executor-90]
+  - test_cpp.py::test_unit_tests[kernels-90]
+  - test_cpp.py::test_unit_tests[layers-90]
+  - test_cpp.py::test_unit_tests[runtime-90]
+  - test_cpp.py::test_unit_tests[thop-90]
+  - test_cpp.py::test_unit_tests[utils-90]
   - test_cpp.py::test_model[fp8-llama-90]
   - test_cpp.py::test_model[t5-90]
   - test_cpp.py::test_benchmarks[t5-90]