Upgrade llama.cpp from b11476 to b11512 (7 steps); ModelParameters.setMoeCacheMib

[bernardladenthin]

Summary

• llama.cpp b11476 → b11512, in 7 reviewed steps (b11480, b11484, b11487, b11491, b11495, b11505, b11512; each step's upstream diff is under 100 KiB, each ends at a tag). No carried patch needed a refresh: all ten apply unchanged at every one of the seven tags, and every drop-check still finds its defect (the 0001 override, the unconditional 0002 assignment, no split_sum guard for 0012, no common_log hook for 0014, no stop_server for 0015, no kolibri under src/ for 0016). Upstream touched nothing under .github/ in the range: CCCL v3.4.3, OpenVINO 2026.4.1 and ROCm 10.0.0 stay as they are.
• One project source change, in the C++ tests only. #29876 (b11480) hands the chat parser the generated tokens next to the text: common_chat_input replaces the std::string of common_chat_parse(), of common_chat_parser_params::generation_prompt and of the server result types' content. Its string constructor is explicit, so test_server.cpp and test_utils.cpp wrap their literals. Nothing moves on the wire -- request fields, their bounds and the response keys are identical between b11476 and b11512 (the two greps from the runbook), and jllama.cpp reads results through to_json() only.
• One new feature wired: ModelParameters.setMoeCacheMib(int) (--moe-cache-mib, #29887 at b11480, split across GPUs by #30112 at b11512; EXPERIMENTAL upstream): a device-side cache for the MoE expert weights a setCpuMoeLayers / --cpu-moe setup keeps in host memory. Registered as ModelOption.MOE_CACHE_MIB, so test_model_flags.cpp holds it against the server parser; negative values are rejected like upstream's parser does. Three tests in ModelParametersTest, a CHANGELOG Added entry.
• Reviewed and deliberately not wired (nothing for a Java consumer to call, all exercised by the existing suites): the two LiquidAI d1 decision models (#30110, #30114 -- SystemOneIntegrationTest covers the endpoint, the models are not in the CI set), the reranker classifier activation as a model property (#29692, read from the GGUF), the context checkpoints kept across a slot save/restore (#26004 -- an appendix older state files simply lack, so they still restore), cpp-httplib 0.60.1 (b11505), and the Vulkan / CUDA / SYCL backend fixes.
• Also on the branch: the docs fix saying the CPU variants ship with 5.2.0, not 5.3.0 (the reactor is at 5.2.0-SNAPSHOT).

Per-step record: two rows per range in docs/history/llama-cpp-breaking-changes.md.

Test plan

• All ten llama/patches/ apply in order on each of the seven pristine tags (git apply), and at b11480 and b11512 a fresh cmake -B … -DBUILD_TESTING=ON configure through the real FetchContent path succeeds (stamp lists the patches, head a11f57ba = b11512).
• Full local build at b11480 and b11512; ctest: 598/598 pass at both.
• Java against the fresh library: LlamaLoaderTest, BackendLoadTest, NativeLibraryLoadSmokeTest (pins LLAMA_CPP_VERSION against the linked build), ModelParametersTest, WireNameRegistryTest, RpcServerTest, LlamaLoggerTest -- 167 tests, 0 failures.
• spotbugs:check, spotless:check, javadoc:jar, check-natives.py, check-release-gate.py, check-run-scripts.py, check-shared-files.py and the buildcheck unit tests report nothing.
• CI is green on this branch. Note: this is also the first Publish run that reaches the two manylinux_2_28 CPU-variant jobs and test-cpp-linux-aarch64 from Ship CPU backend variants for Linux x86-64 and aarch64 (manylinux_2_28) #485 (both runs for Ship CPU backend variants for Linux x86-64 and aarch64 (manylinux_2_28) #485 were cancelled at the start gate).
• Docs / CHANGELOG updated.

Related issues / PRs

Follows #482 (b11476). Refs #485 (the CPU variants whose CI jobs run for the first time here).

Checklist

• I have read CONTRIBUTING.md and CODE_OF_CONDUCT.md
• My commits follow Conventional Commits. They use the repository's usual imperative style instead (Upgrade llama.cpp from bX to bY).
• No security-sensitive changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AytmJF9faEiQEVt6eetQS2

---

Generated by Claude Code

[claude]

✅ Overall Assessment: This PR is well-structured and thoroughly tested. No quality, correctness, or security issues found.

Strengths:

• Comprehensive 7-step bump strategy with clear verification at each stage
• Wire format stability verified (API changes are internal-only)
• Proper validation of setMoeCacheMib() with negative value rejection
• Complete test coverage: C++ tests updated (11 instances), 3 new Java unit tests, wire-name registry extended
• Thorough documentation in CHANGELOG and breaking-changes history
• All version constants updated consistently

Code Quality:

• ModelParameters.setMoeCacheMib() follows established putScalar() pattern
• Clear, actionable error messages
• Javadoc properly documents EXPERIMENTAL upstream status

Test Plan: All items verified

• ✅ All 10 patches apply unchanged at every tag
• ✅ 598 C++ tests pass
• ✅ 167 model-free Java tests pass
• ✅ All style checks clean
• ⏳ CI pending (expected for development branch)

Security: No issues found - no new JNI entry points, proper input validation at boundary.

[sonarqubecloud]

Quality Gate passed

Issues
1 New issue
0 Accepted issues

Measures
0 Security Hotspots
100.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud