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

[Aharrypotter]

Summary

Moves the MiniCPM5-1B model-contract tests out of tests/cpu into tests/models/minicpm5, following the test placement introduced by #704 for Qwen3.5.

This PR:

• relocates MiniCPM5ConfigTest, MiniCPM5TokenizerTest, and MiniCPM5ModelTest to tests/models/minicpm5 with their own CMake targets and CTest registration (label minicpm5);
• removes the three model-specific targets from tests/cpu/CMakeLists.txt, which now holds only kernel and backend-op coverage;
• lets the config and model tests take an MLLM_MINICPM5_EXAMPLE_DIR override (same convention as Qwen35ConfigTest) so the same binaries can find the example config when executed on a device;
• adds GqaDecodeKernelTest to the unified CPU kernel suite: the native KV-head grouped-query-attention decode kernel shipped with MiniCPM5 previously had only public-op coverage in tests/nn, no kernel-level oracle in tests/cpu;
• registers the existing Mllm-Test-Nn-GroupedQueryAttention and Mllm-Test-Nn-KVHeadStaticCache executables with CTest (label nn-op); they were built but never run by ctest;
• changes nothing under mllm/.

Reviewer focus

1. Every MiniCPM5 test keeps its assertions; only its directory, CMake target location, and the example-directory lookup change.
2. tests/cpu no longer registers any model-named target; the retained CPU targets (Mllm-Test-CPUKernel, Mllm-Test-KaiW4A32Pack, Mllm-Test-CPUContiguousOp) are untouched.
3. The new kernel test enters through mllm::cpu::gqa_decode::fwdBhsdFp32 directly, with an independent double-accumulating scalar reference; it does not reconstruct the model or the public op.
4. This PR adds no operation, kernel, or model capability and makes no performance claim.

Test placement

Test	Protected contract	Old location	New location
MiniCPM5ConfigTest	official 1B runtime contract, config/checkpoint match	tests/cpu	tests/models/minicpm5
MiniCPM5TokenizerTest	digit pre-tokenizer, chat template, UTF-8 streaming, pinned official token IDs (env-gated)	tests/cpu	tests/models/minicpm5
MiniCPM5ModelTest	native KV-head cache geometry, logical slots, reset	tests/cpu	tests/models/minicpm5

Kernel coverage added to tests/cpu

Case	Contract proven
MatchesScalarReferenceAcrossFocusedMatrix	scalar path, exact NEON blocks, qk/value-dim tails, single KV head, group sizes 1/2/8, 128-dim heads at several cache fills (tolerance 1e-4 against the reference)
NativeCacheViewAndTransposedQueryMatchContiguousBitwise	a KV view inside a larger [B, Hkv, max_len, D] static cache plus a transposed [B, 1, Hq, D] query/output is addressed through the strides and is bitwise identical to the contiguous computation
GroupedSlicesMatchSingleHeadCallsBitwise	every (batch, kv-head) group reproduces the per-head single-KV call bitwise, so the grouped scratch rows do not leak across heads or batches
RepeatedCallsAreBitwiseStable	fixed accumulation order across repeats
RejectsInvalidGeometryAndStrides	head-count mismatch, zero extents, null buffers, non-unit dimension strides, and non-positive strides return false without touching the output

The cases are registered in tests/cpu/KernelTest.cpp and added to the CPUKernelFocused CTest filter. The fixture follows the neighbouring GatedDeltaRuleKernelTest.hpp shape: a gqa_decode_kernel_test namespace with the reference, a StridedView (storage plus the [B, H, S, D] strides under test) built by purpose-named helpers for the contiguous, static-cache, and transposed layouts, explicit fwdBhsdFp32 calls in every case, and a thin forwarding fixture class. Public-op behaviour (trace, serialization, fallback, product geometries) stays covered by the existing tests/nn/GroupedQueryAttentionTest.cpp.

The MiniCPM5 model graph in mllm/models/minicpm5/modeling_minicpm5.hpp already composes registered nn layers only (GroupedQueryAttention, KVHeadStaticCache, RoPE, RMSNorm, Linear) and includes no backend kernel header, so the operation-registration half of #704 does not apply here.

Review map

1. CMake ownership: tests/cpu/CMakeLists.txt (removals), tests/models/CMakeLists.txt, tests/models/minicpm5/CMakeLists.txt.
2. Example-directory override: exampleDir() in tests/models/minicpm5/MiniCPM5ConfigTest.cpp and MiniCPM5ModelTest.cpp.
3. Unchanged assertions: the three model-test bodies are byte-identical apart from the override helper.
4. Kernel oracle: tests/cpu/GqaDecodeKernelTest.hpp (reference, strided run builders, case functions) and the TEST_F entries in tests/cpu/KernelTest.cpp.
5. Public-op registration: the two add_test lines in tests/nn/CMakeLists.txt.

Validation

Validation is bound to commit 7c4223ce1debfacb0c1acee760352d3aecb43b17 (four commits on top of UbiquitousLearning/mllm:main@eef7dc2be28f54d5ee5d8079a8d3d1989878d31d). The model-test rows ran on the first commit d1a1a280, whose tests/models files are unchanged since; the ctest -L nn-op row ran on df85620d, whose registration lines are unchanged since; the kernel-test rows (host, NDK cross-build, device) were rerun on the exact 7c4223ce tree after the fixture restructure. All gates are ARM.

Validation (PASS) — macOS Apple Silicon host, Android NDK r28b cross-builds, and OnePlus 13T device runs for both the model tests and the kernel suite

Evidence class	Result	What it proves
Abstraction-boundary audit, git diff --check, clang-format --dry-run	PASS	0 errors, 0 warnings; the diff is formatted and touches only tests/
macOS Apple Silicon build (Release, ARM backend, GCD threads)	3/3 CTest	MiniCPM5{Config,Tokenizer,Model}Focused are registered from tests/models/minicpm5 and pass; ctest -N shows no MiniCPM5 target left under tests/cpu
macOS official-tokenizer oracle	4/4	with MLLM_MINICPM5_TOKENIZER_JSON set, pinned token IDs and the 200-token demo prompt match
Android NDK r28b arm64-v8a cross-build (API 28, -march=armv8.2-a+fp16+fp16fml+dotprod+i8mm, OpenMP)	PASS	the three test executables and mllm-minicpm5-runner configure, compile, and link; 8/8 artifacts are AArch64 ELF, 4/4 executables request /system/bin/linker64, NEEDED entries audited
OnePlus 13T (PKX110, Android 16, Snapdragon 8 Elite) focused device run	Config 3/3, Tokenizer 4/4, Model 1/1	the arm64 bundle passes on-device with MLLM_MINICPM5_EXAMPLE_DIR pointing at the shipped example config and the official tokenizer; an invalid override directory makes the config test fail, proving the override is honored
macOS Apple Silicon Mllm-Test-CPUKernel	GqaDecodeKernelTest.* 5/5; CPUKernelFocused CTest PASS	the new oracle passes on the host NEON path and is executed by CTest
Android NDK r28b arm64-v8a cross-build of Mllm-Test-CPUKernel	PASS	the unified kernel test plus libMllmRT.so/libMllmCPUBackend.so compile and link; 5/5 artifacts AArch64, executable requests /system/bin/linker64, ISA flags present in the compile database
OnePlus 13T Mllm-Test-CPUKernel	GqaDecodeKernelTest.* 5/5; focused filter 18/18 across 4 suites	the kernel oracle and the existing causal-conv/GDN kernel cases pass on the device NEON path with the exact H20-built bundle
macOS Apple Silicon ctest -L nn-op	2/2 (GroupedQueryAttention 12 cases, KVHeadStaticCache 4 cases)	the newly registered public-op tests are executed by CTest and pass; not rerun on Android, since the registration changes no binary

Dependencies for the cross-build came from a verified offline mirror (cache-assisted); product source, flags, and targets were unchanged.

Supported scope and limits

Changed by this PR: test file locations, CMake targets and CTest labels, the example-directory override in two tests, one new kernel-test fixture plus its registry entries, and CTest registration for two existing public-op tests.

Not changed: MiniCPM5 model graph, configuration, tokenizer, runner, conversion, CPU kernels, or any other model's tests.

Not claimed: new model capability, full-model generation on this commit (runtime bytes are unchanged from main), model-quality or performance results, or upstream CI (not run on this commit).

🤖 Generated with Claude Code

Summary by CodeRabbit

• Tests
  • Added focused validation for grouped-query attention decoding, including native cache layouts, strided inputs, grouped heads, repeatability, and invalid inputs.
  • Added dedicated MiniCPM5 coverage for configuration loading, tokenization, chat templates, UTF-8 streaming, and model behavior.
  • Added environment-based overrides for MiniCPM5 configuration and tokenizer test data.
  • Registered grouped-query attention and KV-cache neural-network tests.
  • Reorganized MiniCPM5 test targets into dedicated model test coverage.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Team

Run ID: 58633bad-2948-4422-a8b4-a7baf651704e

📥 Commits

Reviewing files that changed from the base of the PR and between df85620 and 7c4223c.

📒 Files selected for processing (1)

• tests/cpu/GqaDecodeKernelTest.hpp

Included review availability: Your plan provides up to 4 included reviews per hour; 2 remain after this review.

---

📝 Walkthrough

Walkthrough

The changes add grouped-query attention decode coverage, register MiniCPM5 configuration, tokenizer, and model tests, add environment-based configuration paths, and register two NN-focused CTest cases.

Changes

Grouped-query attention decode tests

Layer / File(s)	Summary
GQA decode oracle and validation cases tests/cpu/GqaDecodeKernelTest.hpp	Adds scalar-reference validation, contiguous and strided layout checks, grouped-slice comparisons, repeat-stability checks, and invalid-input checks.
GQA test registration and focused execution tests/cpu/KernelTest.cpp, tests/cpu/CMakeLists.txt	Adds five GoogleTest cases and includes them in the focused CPU test filter.

MiniCPM5 test suite

Layer / File(s)	Summary
MiniCPM5 build wiring and configuration paths tests/models/CMakeLists.txt, tests/models/minicpm5/CMakeLists.txt, tests/models/minicpm5/MiniCPM5ConfigTest.cpp, tests/models/minicpm5/MiniCPM5ModelTest.cpp	Registers three MiniCPM5 test targets and resolves example paths from MLLM_MINICPM5_EXAMPLE_DIR with a compile-time fallback.
MiniCPM5 tokenizer behavior cases tests/models/minicpm5/MiniCPM5TokenizerTest.cpp	Adds tests for digit chunking, chat templates, incremental UTF-8 decoding, and optional exact tokenizer outputs.

NN focused test registration

Layer / File(s)	Summary
NN CTest registrations tests/nn/CMakeLists.txt	Registers grouped-query attention and static KV-head cache tests with the nn-op label.

Estimated code review effort: 4 (Complex) | ~45 minutes

Merge Risk: ⚪ Minimal · up to 7c422

This change reorganizes and registers MiniCPM5 and NN test coverage while adding focused GQA decode validation; no current merge-blocking risk remains.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 33.33% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 39 functions across 5 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly and concisely describes the primary change: moving MiniCPM5 model-contract tests into tests/models. It is related to the main refactor, although it does not mention the additional ke…
Description check	✅ Passed	The description is complete and directly related to the changes. It explains the test relocation, new kernel coverage, CMake and CTest updates, environment overrides, scope limits, and validation resu…

• Fix all pre-merge checks with AI

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}