refactor(minicpm5): move model-contract tests under tests/models

tests/cpu/CMakeLists.txt

@@ -6,22 +6,6 @@ add_executable(Mllm-Test-KaiW4A32Pack KaiW4A32PackTest.cpp)
 target_link_libraries(Mllm-Test-KaiW4A32Pack PRIVATE gtest_main MllmCPUBackend)
 target_include_directories(Mllm-Test-KaiW4A32Pack PRIVATE ${MLLM_INCLUDE_DIR})
 
-add_executable(Mllm-Test-MiniCPM5-Config MiniCPM5ConfigTest.cpp)
-target_link_libraries(Mllm-Test-MiniCPM5-Config PRIVATE gtest_main MllmCPUBackend)
-target_include_directories(Mllm-Test-MiniCPM5-Config PRIVATE ${MLLM_INCLUDE_DIR})
-target_compile_definitions(Mllm-Test-MiniCPM5-Config
-                           PRIVATE MINICPM5_EXAMPLE_DIR="${CMAKE_CURRENT_SOURCE_DIR}/../../examples/minicpm5")
-
-add_executable(Mllm-Test-MiniCPM5-Tokenizer MiniCPM5TokenizerTest.cpp)
-target_link_libraries(Mllm-Test-MiniCPM5-Tokenizer PRIVATE gtest_main MllmCPUBackend)
-target_include_directories(Mllm-Test-MiniCPM5-Tokenizer PRIVATE ${MLLM_INCLUDE_DIR})
-
-add_executable(Mllm-Test-MiniCPM5-Model MiniCPM5ModelTest.cpp)
-target_link_libraries(Mllm-Test-MiniCPM5-Model PRIVATE gtest_main MllmCPUBackend)
-target_include_directories(Mllm-Test-MiniCPM5-Model PRIVATE ${MLLM_INCLUDE_DIR})
-target_compile_definitions(Mllm-Test-MiniCPM5-Model
-                           PRIVATE MINICPM5_EXAMPLE_DIR="${CMAKE_CURRENT_SOURCE_DIR}/../../examples/minicpm5")
-
 add_executable(Mllm-Test-CPUContiguousOp ContiguousOpTest.cpp)
 target_link_libraries(Mllm-Test-CPUContiguousOp PRIVATE gtest_main MllmRT MllmCPUBackend)
 target_include_directories(Mllm-Test-CPUContiguousOp PRIVATE ${MLLM_INCLUDE_DIR})
@@ -31,5 +15,5 @@ add_test(
   NAME CPUKernelFocused
   COMMAND
     Mllm-Test-CPUKernel
-    --gtest_filter=CausalDepthwiseConvKernelTest.*:CausalDepthwiseConvCurrentFirstKernelTest.*:GatedDeltaRuleKernelTest.*)
+    --gtest_filter=CausalDepthwiseConvKernelTest.*:CausalDepthwiseConvCurrentFirstKernelTest.*:GatedDeltaRuleKernelTest.*:GqaDecodeKernelTest.*)
 set_tests_properties(CPUKernelFocused PROPERTIES LABELS cpu-kernel)

tests/cpu/KernelTest.cpp

@@ -630,6 +630,49 @@ TEST_F(GatedDeltaRuleKernelTest, ProductionGroupedHeadGeometry8LaneIsBitwiseStab
       /*repeats=*/24));
 }
 
+//===----------------------------------------------------------------------===//
+// Grouped-query attention decode (native KV-head cache views)
+//===----------------------------------------------------------------------===//
+#include "GqaDecodeKernelTest.hpp"
+TEST_F(GqaDecodeKernelTest, MatchesScalarReferenceAcrossFocusedMatrix) {
+  EXPECT_NO_FATAL_FAILURE(testMatchesScalarReference({
+      // Scalar path and single KV head.
+      {1, 1, 1, 1, 1, 1},
+      {1, 2, 1, 3, 3, 3},
+      {2, 4, 1, 5, 4, 4},
+      // Exact NEON vector blocks, group size 1 and 2.
+      {1, 2, 2, 7, 8, 8},
+      {1, 4, 2, 16, 16, 16},
+      // Vector loop with scalar tails in qk and value dims.
+      {1, 4, 2, 9, 5, 7},
+      {2, 6, 3, 11, 13, 6},
+      {1, 8, 2, 33, 127, 130},
+      // Production 128-dim heads at several cache fills.
+      {1, 8, 2, 1, 128, 128},
+      {1, 16, 2, 69, 128, 128},
+      {1, 16, 2, 517, 128, 128},
+      {2, 16, 2, 40, 128, 128},
+  }));
+}
+
+TEST_F(GqaDecodeKernelTest, NativeCacheViewAndTransposedQueryMatchContiguousBitwise) {
+  EXPECT_NO_FATAL_FAILURE(testNativeCacheAndTransposedQueryStrides({1, 16, 2, 69, 128, 128}, /*cache_capacity=*/2048));
+  EXPECT_NO_FATAL_FAILURE(testNativeCacheAndTransposedQueryStrides({2, 6, 3, 11, 13, 6}, /*cache_capacity=*/32));
+}
+
+TEST_F(GqaDecodeKernelTest, GroupedSlicesMatchSingleHeadCallsBitwise) {
+  EXPECT_NO_FATAL_FAILURE(testGroupedSlicesMatchSingleHeadCallsBitwise({2, 16, 2, 69, 128, 128}));
+  EXPECT_NO_FATAL_FAILURE(testGroupedSlicesMatchSingleHeadCallsBitwise({1, 6, 3, 11, 13, 6}));
+}
+
+TEST_F(GqaDecodeKernelTest, RepeatedCallsAreBitwiseStable) {
+  EXPECT_NO_FATAL_FAILURE(testRepeatedCallsAreBitwiseStable({1, 16, 2, 517, 128, 128}, /*repeats=*/8));
+}
+
+TEST_F(GqaDecodeKernelTest, RejectsInvalidGeometryAndStrides) {
+  EXPECT_NO_FATAL_FAILURE(testRejectsInvalidGeometryAndStrides());
+}
+
 //===----------------------------------------------------------------------===//
 // Parallel linear
 //===----------------------------------------------------------------------===//

tests/models/CMakeLists.txt

@@ -1 +1,2 @@
 add_subdirectory(qwen3_5)
+add_subdirectory(minicpm5)

tests/models/minicpm5/CMakeLists.txt

@@ -0,0 +1,20 @@
+add_executable(Mllm-Test-MiniCPM5-Config MiniCPM5ConfigTest.cpp)
+target_link_libraries(Mllm-Test-MiniCPM5-Config PRIVATE gtest_main MllmCPUBackend)
+target_include_directories(Mllm-Test-MiniCPM5-Config PRIVATE ${MLLM_INCLUDE_DIR})
+target_compile_definitions(Mllm-Test-MiniCPM5-Config
+                           PRIVATE MINICPM5_EXAMPLE_DIR="${CMAKE_CURRENT_SOURCE_DIR}/../../../examples/minicpm5")
+
+add_executable(Mllm-Test-MiniCPM5-Tokenizer MiniCPM5TokenizerTest.cpp)
+target_link_libraries(Mllm-Test-MiniCPM5-Tokenizer PRIVATE gtest_main MllmCPUBackend)
+target_include_directories(Mllm-Test-MiniCPM5-Tokenizer PRIVATE ${MLLM_INCLUDE_DIR})
+
+add_executable(Mllm-Test-MiniCPM5-Model MiniCPM5ModelTest.cpp)
+target_link_libraries(Mllm-Test-MiniCPM5-Model PRIVATE gtest_main MllmCPUBackend)
+target_include_directories(Mllm-Test-MiniCPM5-Model PRIVATE ${MLLM_INCLUDE_DIR})
+target_compile_definitions(Mllm-Test-MiniCPM5-Model
+                           PRIVATE MINICPM5_EXAMPLE_DIR="${CMAKE_CURRENT_SOURCE_DIR}/../../../examples/minicpm5")
+
+add_test(NAME MiniCPM5ConfigFocused COMMAND Mllm-Test-MiniCPM5-Config)
+add_test(NAME MiniCPM5TokenizerFocused COMMAND Mllm-Test-MiniCPM5-Tokenizer)
+add_test(NAME MiniCPM5ModelFocused COMMAND Mllm-Test-MiniCPM5-Model)
+set_tests_properties(MiniCPM5ConfigFocused MiniCPM5TokenizerFocused MiniCPM5ModelFocused PROPERTIES LABELS minicpm5)

tests/cpu/MiniCPM5ConfigTest.cpp → tests/models/minicpm5/MiniCPM5ConfigTest.cpp

@@ -3,15 +3,21 @@
 
 #include <gtest/gtest.h>
 
+#include <cstdlib>
 #include <string>
 #include <vector>
 
 #include "mllm/models/minicpm5/configuration_minicpm5.hpp"
 
 namespace {
 
+auto exampleDir() -> std::string {
+  const char* example_dir_override = std::getenv("MLLM_MINICPM5_EXAMPLE_DIR");
+  return example_dir_override == nullptr ? std::string(MINICPM5_EXAMPLE_DIR) : std::string(example_dir_override);
+}
+
 auto loadConfig() -> mllm::models::minicpm5::MiniCPM5Config {
-  return mllm::models::minicpm5::MiniCPM5Config(std::string(MINICPM5_EXAMPLE_DIR) + "/config_1B_w4a32_kai.json");
+  return mllm::models::minicpm5::MiniCPM5Config(exampleDir() + "/config_1B_w4a32_kai.json");
 }
 
 auto parameterFile(mllm::ModelFileVersion version, const std::vector<int32_t>& embedding_shape, bool include_lm_head = true)

tests/cpu/MiniCPM5ModelTest.cpp → tests/models/minicpm5/MiniCPM5ModelTest.cpp

@@ -3,14 +3,24 @@
 
 #include <gtest/gtest.h>
 
+#include <cstdlib>
 #include <string>
 
 #include "mllm/mllm.hpp"
 #include "mllm/models/minicpm5/modeling_minicpm5.hpp"
 
+namespace {
+
+auto exampleDir() -> std::string {
+  const char* example_dir_override = std::getenv("MLLM_MINICPM5_EXAMPLE_DIR");
+  return example_dir_override == nullptr ? std::string(MINICPM5_EXAMPLE_DIR) : std::string(example_dir_override);
+}
+
+}  // namespace
+
 TEST(MiniCPM5ModelTest, BuildsNativeKVHeadLogicalSlotsAndResetsThem) {
   mllm::initializeContext();
-  const auto config = mllm::models::minicpm5::MiniCPM5Config(std::string(MINICPM5_EXAMPLE_DIR) + "/config_1B_w4a32_kai.json");
+  const auto config = mllm::models::minicpm5::MiniCPM5Config(exampleDir() + "/config_1B_w4a32_kai.json");
   auto model = mllm::models::minicpm5::MiniCPM5ForCausalLM(config);
   auto& cache = model.kvCache();
 

tests/nn/CMakeLists.txt

@@ -28,3 +28,6 @@ target_include_directories(Mllm-Test-Nn-GatedDeltaRule PRIVATE ${MLLM_INCLUDE_DI
 
 include(GoogleTest)
 add_test(NAME GatedDeltaRuleFocused COMMAND Mllm-Test-Nn-GatedDeltaRule)
+add_test(NAME GroupedQueryAttentionFocused COMMAND Mllm-Test-Nn-GroupedQueryAttention)
+add_test(NAME KVHeadStaticCacheFocused COMMAND Mllm-Test-Nn-KVHeadStaticCache)
+set_tests_properties(GroupedQueryAttentionFocused KVHeadStaticCacheFocused PROPERTIES LABELS nn-op)