fix!: retire the CLI flags llama.cpp's server parser rejects, and guard the contract

[bernardladenthin]

Summary

• Seven flags the Java layer emitted are rejected by llama.cpp's server argument parser, which makes the matching builder methods produce an unloadable model, not a merely ineffective one. LlamaModel.loadModel(parameters.toArray()) hands the ModelParameters map to common_params_parse(..., LLAMA_EXAMPLE_SERVER) as argv, and an unregistered option there is a hard error — arg.cpp throws, common_params_parse returns false, load_model_impl throws LlamaException("Failed to parse model parameters").
• Two are fresh breakage from the b10878 bump (feat!: upgrade llama.cpp from b10870 to b10878 #425): upstream deleted --mlock and --no-mmap, deprecated at b10092 — the whole deprecation window opened and closed inside eight tags. These have a faithful replacement, so no API is lost: a new args.LoadMode enum + ModelParameters.setLoadMode(LoadMode) expose upstream's -lm/--load-mode, and enableMlock() / disableMmap() keep working, re-pointed to LoadMode.MLOCK / LoadMode.NONE — the exact mapping upstream's own deprecation shim used (LLAMA_LOAD_MODE_MLOCK / LLAMA_LOAD_MODE_NONE), so behaviour is unchanged. Both are now @Deprecated.
• Five were already dead and are now @Deprecated no-ops that write nothing, keeping existing call sites compiling and loading. ModelFlag.MLOCK / NO_MMAP / DUMP_KV_CACHE are removed from the enum so the broken argv is not reachable through setFlag either — the same reasoning that already excluded FLASH_ATTN.
• A contract test now enforces this, because no Java test can: ModelFlagTest and ModelParametersExtendedTest assert the string mapping (hasKey("--mlock")), never that llama.cpp still accepts the string, so they stayed green for as long as the flags were dead.

The seven, and why each is dead

Flag	Java member	Status upstream	Now
--mlock	enableMlock(), ModelFlag.MLOCK	deprecated b10092, deleted b10878	emits --load-mode mlock
--no-mmap	disableMmap(), ModelFlag.NO_MMAP	deprecated b10092, deleted b10878	emits --load-mode none
--dump-kv-cache	enableDumpKvCache(), ModelFlag.DUMP_KV_CACHE	removed, no replacement	no-op
--hf-repo-v	setHfRepoV(String)	removed with the OuteTTS-era two-model TTS design	no-op
--hf-file-v	setHfFileV(String)	same	no-op
--grp-attn-n	setGrpAttnN(int)	present in arg.cpp, set_examples({COMPLETION, PASSKEY})	no-op
--grp-attn-w	setGrpAttnW(int)	present in arg.cpp, set_examples({COMPLETION})	no-op

The last two are the interesting ones: they exist in common/arg.cpp at every tag this project has pinned, so any textual sweep reports them alive. common_params_parser_init's add_opt filters by example at registration time, so they are never registered for LLAMA_EXAMPLE_SERVER — the example this binding parses with — and the parser rejects them exactly like a deleted flag. That is precisely why the guard below drives the real option table instead of grepping upstream sources.

The guard

• llama/cmake/extract-java-cli-flags.cmake — at configure time, extracts every "--flag" string literal ModelFlag.java + ModelParameters.java can emit into a generated header. Line-oriented so Javadoc mentions ({@code --flash-attn}, prose naming --mlock) are dropped, while enum constants, putScalar/putEnum keys, parameters.put keys and the private ARG_* constants all survive. The Java sources stay the single source of truth, so the two halves cannot drift; it fails loud below a 50-flag floor so a broken extractor cannot make the test vacuously pass.
• llama/src/test/cpp/test_model_flags.cpp (4 tests) — asserts each extracted flag is present in common_params_parser_init(params, LLAMA_EXAMPLE_SERVER).options. Hermetic: no model, no JVM, no network — common_params_parser_init only fills a struct. It also checks the oracle (the option table must be plausibly large and contain --model) and that the exemption list has not rotted, so neither an empty table nor a stale exemption can hide a real failure. --vocab-only is the one exemption, and deliberately so: it is a project pseudo-flag that jllama.cpp removes via strip_flag_from_argv before common_params_parse sees the argv, to select the vocab-only path.

Falsified before it was trusted. Built against the pre-fix Java sources, the test named exactly the seven flags above — including the two set_examples()-scoped ones my own grep-based audit had missed. After the fix it is green.

Test plan

• Affected unit / integration tests pass locally
• CI is green on this branch
• Docs / CHANGELOG updated where applicable

Run locally on Linux x86_64:

• ctest 531/531 (was 527; +4 from the new file)
• mvn test 1759 tests, 0 failures, 0 errors
• PIT 320/320 mutations killed, 0 NO_COVERAGE, test strength 100% (args.* is a gate target at mutationThreshold 100, so LoadMode needed LoadModeTest — added, mirroring LazyModeTest)
• spotbugs:check clean — setLoadMode added to the design-intent OCP_OVERLY_CONCRETE_PARAMETER list in llama/spotbugs-exclude.xml, the trap CLAUDE.md documents for any new enum-valued ModelParameters setter
• spotless:check, clang-format 22.1.8, and javadoc:jar all clean

Not verified locally, and the reason to watch CI here: the CMake extractor uses file(STRINGS) + while(... MATCHES ...) and has only been exercised on Linux with CMake 3.28. macOS, both Windows generators, aarch64, s390x (cross + qemu) and Android are unproven until this PR's C++ Tests matrix runs. That is the main risk in this change.

Also updated: CLAUDE.md (test-file table, total 527 → 531, the SpotBugs rename/addition note), TODO.md (the "five CLI flags" entry closed, corrected to seven), and the b10870–b10878 row in docs/history/llama-cpp-breaking-changes.md, which had explicitly deferred this public-API decision to a follow-up.

Related issues / PRs

Refs #425 (the b10878 bump that deleted --mlock / --no-mmap). Implements the fix prescribed by the "ModelParameters emits five CLI flags the server arg parser rejects" entry in TODO.md, added during the b10649 review (#403).

Checklist

• I have read CONTRIBUTING.md and CODE_OF_CONDUCT.md
• My commits follow Conventional Commits
• No security-sensitive changes

Breaking change

ModelFlag.MLOCK, ModelFlag.NO_MMAP and ModelFlag.DUMP_KV_CACHE are removed (source-breaking for anyone calling setFlag(ModelFlag.MLOCK)), and ModelFlag.values().length drops 34 → 31. Every affected constant guaranteed a failed model load, so there is no working behaviour being taken away. All seven builder methods are retained and still compile.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AnNYn8W1xuVxVJtyL34GyH

---

Generated by Claude Code

[sonarqubecloud]

Quality Gate failed

Failed conditions
B Maintainability Rating on New Code (required ≥ A)

See analysis details on SonarQube Cloud

Catch issues before they fail your Quality Gate with our IDE extension SonarQube for IDE