From 776f1176119667378c95721aac079012877aa505 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Wed, 1 Jul 2026 19:51:24 +0000
Subject: [PATCH 01/12] refactor: move core into llama/ submodule under an
 aggregator POM (Maven reactor)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Phase 2 structural move. The repo root becomes a thin aggregator/parent POM
(net.ladenthin:llama-parent, packaging=pom) with two modules: llama (the native
JNI core) and llama-langchain4j. Both inherit the single shared <version> from
the parent, so they ship in lockstep by construction (the CI version-lockstep
guard becomes redundant and is removed in the CI pass).

- git mv the entire core into llama/: src/, CMakeLists.txt, cmake/, patches/,
  pom.xml, spotbugs-exclude.xml, lombok.config, .clang-format, .clang-tidy.
  (All tracked as renames; the core pom's ${basedir}/… and relative
  spotbugs-exclude.xml auto-re-root; CMakeLists paths are relative to itself.)
- New root pom.xml = aggregator (modules: llama, llama-langchain4j).
- llama/pom.xml and llama-langchain4j/pom.xml become children of llama-parent
  (drop their own <version>; inherit it). The core keeps its published
  coordinates net.ladenthin:llama — consumers are unaffected by the move.

Verified locally: the reactor builds green (parent → llama → llama-langchain4j),
all three install at 5.0.4-SNAPSHOT, langchain4j resolves the core via the
inherited ${project.version}.

CI path re-rooting (workflows, build scripts, REUSE.toml, .gitignore) follows in
the next commit.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rt1paYztGJ2AKUuBuAGDXE
---
 llama-langchain4j/pom.xml                     |    9 +-
 .clang-format => llama/.clang-format          |    0
 .clang-tidy => llama/.clang-tidy              |    0
 CMakeLists.txt => llama/CMakeLists.txt        |    0
 .../cmake}/apply-llama-patches.cmake          |    0
 .../cmake}/generate-tts-upstream.cmake        |    0
 lombok.config => llama/lombok.config          |    0
 .../0001-win32-arg-parse-embed-guard.patch    |    0
 ...eserve-caller-load-progress-callback.patch |    0
 ...slot-prompt-similarity-getter-setter.patch |    0
 ...-per-request-reasoning-budget-tokens.patch |    0
 llama/pom.xml                                 | 1354 +++++++++++++++++
 .../spotbugs-exclude.xml                      |    0
 .../src}/main/cpp/compat/ggml_x86_compat.c    |    0
 {src => llama/src}/main/cpp/jllama.cpp        |    0
 {src => llama/src}/main/cpp/jllama.h          |    0
 {src => llama/src}/main/cpp/jni_helpers.hpp   |    0
 {src => llama/src}/main/cpp/json_helpers.hpp  |    0
 {src => llama/src}/main/cpp/log_helpers.hpp   |    0
 {src => llama/src}/main/cpp/tts_engine.cpp    |    0
 {src => llama/src}/main/cpp/tts_engine.h      |    0
 {src => llama/src}/main/cpp/tts_upstream.h    |    0
 {src => llama/src}/main/cpp/tts_wav.hpp       |    0
 {src => llama/src}/main/cpp/utils.hpp         |    0
 {src => llama/src}/main/cpp/webui_stub/ui.h   |    0
 {src => llama/src}/main/java/module-info.java |    0
 .../net/ladenthin/llama/LlamaIterable.java    |    0
 .../net/ladenthin/llama/LlamaIterator.java    |    0
 .../java/net/ladenthin/llama/LlamaModel.java  |    0
 .../java/net/ladenthin/llama/Session.java     |    0
 .../net/ladenthin/llama/SessionState.java     |    0
 .../net/ladenthin/llama/TextToSpeech.java     |    0
 .../net/ladenthin/llama/ToolCallingAgent.java |    0
 .../net/ladenthin/llama/args/CacheType.java   |    0
 .../java/net/ladenthin/llama/args/CliArg.java |    0
 .../llama/args/ContinuationMode.java          |    0
 .../ladenthin/llama/args/GpuSplitMode.java    |    0
 .../net/ladenthin/llama/args/LogFormat.java   |    0
 .../net/ladenthin/llama/args/MiroStat.java    |    0
 .../net/ladenthin/llama/args/ModelFlag.java   |    0
 .../ladenthin/llama/args/NumaStrategy.java    |    0
 .../net/ladenthin/llama/args/PoolingType.java |    0
 .../ladenthin/llama/args/ReasoningFormat.java |    0
 .../ladenthin/llama/args/RopeScalingType.java |    0
 .../net/ladenthin/llama/args/Sampler.java     |    0
 .../ladenthin/llama/args/package-info.java    |    0
 .../llama/callback/CancellationToken.java     |    0
 .../llama/callback/LoadProgressCallback.java  |    0
 .../ladenthin/llama/callback/ToolHandler.java |    0
 .../llama/callback/package-info.java          |    0
 .../llama/exception/LlamaException.java       |    0
 .../exception/ModelUnavailableException.java  |    0
 .../llama/exception/package-info.java         |    0
 .../llama/json/ChatResponseParser.java        |    0
 .../llama/json/ChatStreamChunkParser.java     |    0
 .../llama/json/CompletionResponseParser.java  |    0
 .../llama/json/RerankResponseParser.java      |    0
 .../ladenthin/llama/json/TimingsLogger.java   |    0
 .../ladenthin/llama/json/package-info.java    |    0
 .../loader/Java8CompatibilityHelper.java      |    0
 .../ladenthin/llama/loader/LlamaLoader.java   |    0
 .../llama/loader/LlamaSystemProperties.java   |    0
 .../loader/NativeLibraryPermissionSetter.java |    0
 .../net/ladenthin/llama/loader/OSInfo.java    |    0
 .../llama/loader/OfflineModelGuard.java       |    0
 .../ladenthin/llama/loader/ProcessRunner.java |    0
 .../ladenthin/llama/loader/package-info.java  |    0
 .../net/ladenthin/llama/package-info.java     |    0
 .../llama/parameters/ChatRequest.java         |    0
 .../llama/parameters/CliParameters.java       |    0
 .../llama/parameters/InferenceParameters.java |    0
 .../llama/parameters/JsonParameters.java      |    0
 .../llama/parameters/ModelParameters.java     |    0
 .../parameters/ParameterJsonSerializer.java   |    0
 .../llama/parameters/package-info.java        |    0
 .../llama/server/AnthropicApiSupport.java     |    0
 .../server/AnthropicStreamTranslator.java     |    0
 .../net/ladenthin/llama/server/ChunkSink.java |    0
 .../llama/server/LlamaModelBackend.java       |    0
 .../ladenthin/llama/server/NativeServer.java  |    0
 .../llama/server/OaiRerankSupport.java        |    0
 .../llama/server/OllamaApiSupport.java        |    0
 .../ladenthin/llama/server/OpenAiBackend.java |    0
 .../llama/server/OpenAiCompatServer.java      |    0
 .../llama/server/OpenAiRequestMapper.java     |    0
 .../llama/server/OpenAiServerCli.java         |    0
 .../llama/server/OpenAiServerConfig.java      |    0
 .../llama/server/OpenAiSseFormatter.java      |    0
 .../llama/server/ResponsesApiSupport.java     |    0
 .../server/ResponsesStreamTranslator.java     |    0
 .../server/ToolCallDeltaAccumulator.java      |    0
 .../ladenthin/llama/server/package-info.java  |    0
 .../net/ladenthin/llama/value/ChatChoice.java |    0
 .../ladenthin/llama/value/ChatMessage.java    |    0
 .../ladenthin/llama/value/ChatResponse.java   |    0
 .../ladenthin/llama/value/ChatTranscript.java |    0
 .../llama/value/CompletionResult.java         |    0
 .../ladenthin/llama/value/ContentPart.java    |    0
 .../ladenthin/llama/value/LlamaOutput.java    |    0
 .../net/ladenthin/llama/value/LogLevel.java   |    0
 .../net/ladenthin/llama/value/ModelMeta.java  |    0
 .../java/net/ladenthin/llama/value/Pair.java  |    0
 .../ladenthin/llama/value/ServerMetrics.java  |    0
 .../ladenthin/llama/value/SlotMetrics.java    |    0
 .../net/ladenthin/llama/value/StopReason.java |    0
 .../net/ladenthin/llama/value/Timings.java    |    0
 .../ladenthin/llama/value/TokenLogprob.java   |    0
 .../net/ladenthin/llama/value/ToolCall.java   |    0
 .../ladenthin/llama/value/ToolDefinition.java |    0
 .../java/net/ladenthin/llama/value/Usage.java |    0
 .../ladenthin/llama/value/package-info.java   |    0
 .../src}/test/cpp/test_jni_helpers.cpp        |    0
 .../src}/test/cpp/test_json_helpers.cpp       |    0
 .../src}/test/cpp/test_log_helpers.cpp        |    0
 {src => llama/src}/test/cpp/test_server.cpp   |    0
 {src => llama/src}/test/cpp/test_tts_wav.cpp  |    0
 {src => llama/src}/test/cpp/test_utils.cpp    |    0
 .../src}/test/java/examples/ChatExample.java  |    0
 .../test/java/examples/GrammarExample.java    |    0
 .../test/java/examples/InfillExample.java     |    0
 .../src}/test/java/examples/MainExample.java  |    0
 .../java/examples/OpenAiServerExample.java    |    0
 .../llama/AudioInputIntegrationTest.java      |    0
 .../net/ladenthin/llama/ChatAdvancedTest.java |    0
 .../net/ladenthin/llama/ChatScenarioTest.java |    0
 .../net/ladenthin/llama/ClaudeGenerated.java  |    0
 .../llama/ConfigureParallelInferenceTest.java |    0
 .../ladenthin/llama/ErrorHandlingTest.java    |    0
 .../llama/LlamaArchitectureTest.java          |    0
 .../ladenthin/llama/LlamaEmbeddingsTest.java  |    0
 .../llama/LlamaModelOfflineTest.java          |    0
 .../net/ladenthin/llama/LlamaModelTest.java   |    0
 .../llama/LlamaParameterProperties.java       |    0
 .../ladenthin/llama/MemoryManagementTest.java |    0
 .../llama/MultimodalIntegrationTest.java      |    0
 .../llama/MultimodalMessagesTest.java         |    0
 .../llama/ReactorIntegrationTest.java         |    0
 .../ladenthin/llama/ReasoningBudgetTest.java  |    0
 .../ladenthin/llama/RerankingModelTest.java   |    0
 .../llama/ResponseJsonStructureTest.java      |    0
 .../llama/SessionConcurrencyTest.java         |    0
 .../net/ladenthin/llama/SessionStateTest.java |    0
 .../java/net/ladenthin/llama/SessionTest.java |    0
 .../net/ladenthin/llama/TestConstants.java    |    0
 .../ladenthin/llama/ToolCallingAgentTest.java |    0
 .../llama/ToolCallingIntegrationTest.java     |    0
 .../ladenthin/llama/TtsIntegrationTest.java   |    0
 .../llama/args/AbstractCliArgEnumTest.java    |    0
 .../ladenthin/llama/args/CacheTypeTest.java   |    0
 .../llama/args/ContinuationModeTest.java      |    0
 .../llama/args/GpuSplitModeTest.java          |    0
 .../ladenthin/llama/args/LogFormatTest.java   |    0
 .../ladenthin/llama/args/MiroStatTest.java    |    0
 .../ladenthin/llama/args/ModelFlagTest.java   |    0
 .../llama/args/NumaStrategyTest.java          |    0
 .../ladenthin/llama/args/PoolingTypeTest.java |    0
 .../llama/args/ReasoningFormatTest.java       |    0
 .../llama/args/RopeScalingTypeTest.java       |    0
 .../net/ladenthin/llama/args/SamplerTest.java |    0
 .../InferenceParametersBenchmark.java         |    0
 .../CancellationTokenLincheckTest.java        |    0
 .../llama/callback/CancellationTokenTest.java |    0
 .../callback/LoadProgressCallbackTest.java    |    0
 .../llama/exception/LlamaExceptionTest.java   |    0
 .../ModelUnavailableExceptionTest.java        |    0
 .../llama/jcstress/CancellationTokenRace.java |    0
 .../llama/json/ChatResponseParserTest.java    |    0
 .../llama/json/ChatStreamChunkParserTest.java |    0
 .../json/CompletionResponseParserTest.java    |    0
 .../json/ParameterJsonSerializerTest.java     |    0
 .../llama/json/RerankResponseParserTest.java  |    0
 .../llama/json/TimingsLoggerTest.java         |    0
 .../llama/loader/LlamaLoaderTest.java         |    0
 .../llama/loader/LoggingSmokeTest.java        |    0
 .../loader/NativeLibraryLoadSmokeTest.java    |    0
 .../NativeLibraryPermissionSetterTest.java    |    0
 .../ladenthin/llama/loader/OSInfoTest.java    |    0
 .../llama/parameters/ChatRequestTest.java     |    0
 .../parameters/InferenceParametersTest.java   |    0
 .../JsonEndpointParametersTest.java           |    0
 .../llama/parameters/JsonParametersTest.java  |    0
 .../ModelParametersExtendedTest.java          |    0
 .../llama/parameters/ModelParametersTest.java |    0
 .../llama/server/AnthropicApiSupportTest.java |    0
 .../server/AnthropicStreamTranslatorTest.java |    0
 .../llama/server/NativeServerSmokeTest.java   |    0
 .../llama/server/OaiRerankSupportTest.java    |    0
 .../llama/server/OllamaApiSupportTest.java    |    0
 .../server/OpenAiCompatServerHttpTest.java    |    0
 .../OpenAiCompatServerIntegrationTest.java    |    0
 .../llama/server/OpenAiRequestMapperTest.java |    0
 .../llama/server/OpenAiServerCliTest.java     |    0
 ...OpenAiServerCompletionIntegrationTest.java |    0
 .../llama/server/OpenAiServerConfigTest.java  |    0
 ...OpenAiServerEmbeddingsIntegrationTest.java |    0
 .../OpenAiServerRerankIntegrationTest.java    |    0
 .../llama/server/OpenAiServerTestSupport.java |    0
 ...penAiServerToolCallingIntegrationTest.java |    0
 .../llama/server/OpenAiSseFormatterTest.java  |    0
 .../llama/server/ResponsesApiSupportTest.java |    0
 .../server/ResponsesStreamTranslatorTest.java |    0
 .../server/ToolCallDeltaAccumulatorTest.java  |    0
 .../ladenthin/llama/value/ChatChoiceTest.java |    0
 .../llama/value/ChatMessageTest.java          |    0
 .../llama/value/ChatResponseTest.java         |    0
 .../llama/value/ChatTranscriptTest.java       |    0
 .../llama/value/CompletionResultTest.java     |    0
 .../llama/value/ContentPartTest.java          |    0
 .../llama/value/LlamaOutputTest.java          |    0
 .../ladenthin/llama/value/LogLevelTest.java   |    0
 .../ladenthin/llama/value/ModelMetaTest.java  |    0
 .../net/ladenthin/llama/value/PairTest.java   |    0
 .../llama/value/ServerMetricsTest.java        |    0
 .../ladenthin/llama/value/StopReasonTest.java |    0
 .../ladenthin/llama/value/TimingsTest.java    |    0
 .../llama/value/TokenLogprobTest.java         |    0
 .../ladenthin/llama/value/ToolCallTest.java   |    0
 .../llama/value/ToolDefinitionTest.java       |    0
 .../net/ladenthin/llama/value/UsageTest.java  |    0
 .../vmlens/SessionStateInterleavingTest.java  |    0
 .../vmlens/VmlensInterleavingSmokeTest.java   |    0
 .../src}/test/resources/images/README.md      |    0
 .../src}/test/resources/images/test-image.jpg |  Bin
 pom.xml                                       | 1317 +---------------
 224 files changed, 1376 insertions(+), 1304 deletions(-)
 rename .clang-format => llama/.clang-format (100%)
 rename .clang-tidy => llama/.clang-tidy (100%)
 rename CMakeLists.txt => llama/CMakeLists.txt (100%)
 rename {cmake => llama/cmake}/apply-llama-patches.cmake (100%)
 rename {cmake => llama/cmake}/generate-tts-upstream.cmake (100%)
 rename lombok.config => llama/lombok.config (100%)
 rename {patches => llama/patches}/0001-win32-arg-parse-embed-guard.patch (100%)
 rename {patches => llama/patches}/0002-server-preserve-caller-load-progress-callback.patch (100%)
 rename {patches => llama/patches}/0003-pr22393-server-add-slot-prompt-similarity-getter-setter.patch (100%)
 rename {patches => llama/patches}/0004-pr23116-server-per-request-reasoning-budget-tokens.patch (100%)
 create mode 100644 llama/pom.xml
 rename spotbugs-exclude.xml => llama/spotbugs-exclude.xml (100%)
 rename {src => llama/src}/main/cpp/compat/ggml_x86_compat.c (100%)
 rename {src => llama/src}/main/cpp/jllama.cpp (100%)
 rename {src => llama/src}/main/cpp/jllama.h (100%)
 rename {src => llama/src}/main/cpp/jni_helpers.hpp (100%)
 rename {src => llama/src}/main/cpp/json_helpers.hpp (100%)
 rename {src => llama/src}/main/cpp/log_helpers.hpp (100%)
 rename {src => llama/src}/main/cpp/tts_engine.cpp (100%)
 rename {src => llama/src}/main/cpp/tts_engine.h (100%)
 rename {src => llama/src}/main/cpp/tts_upstream.h (100%)
 rename {src => llama/src}/main/cpp/tts_wav.hpp (100%)
 rename {src => llama/src}/main/cpp/utils.hpp (100%)
 rename {src => llama/src}/main/cpp/webui_stub/ui.h (100%)
 rename {src => llama/src}/main/java/module-info.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/LlamaIterable.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/LlamaIterator.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/LlamaModel.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/Session.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/SessionState.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/TextToSpeech.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/ToolCallingAgent.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/args/CacheType.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/args/CliArg.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/args/ContinuationMode.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/args/GpuSplitMode.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/args/LogFormat.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/args/MiroStat.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/args/ModelFlag.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/args/NumaStrategy.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/args/PoolingType.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/args/ReasoningFormat.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/args/RopeScalingType.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/args/Sampler.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/args/package-info.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/callback/CancellationToken.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/callback/LoadProgressCallback.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/callback/ToolHandler.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/callback/package-info.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/exception/LlamaException.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/exception/ModelUnavailableException.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/exception/package-info.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/json/ChatResponseParser.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/json/ChatStreamChunkParser.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/json/CompletionResponseParser.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/json/RerankResponseParser.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/json/TimingsLogger.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/json/package-info.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/loader/Java8CompatibilityHelper.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/loader/LlamaLoader.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/loader/LlamaSystemProperties.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/loader/NativeLibraryPermissionSetter.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/loader/OSInfo.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/loader/OfflineModelGuard.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/loader/ProcessRunner.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/loader/package-info.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/package-info.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/parameters/ChatRequest.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/parameters/CliParameters.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/parameters/InferenceParameters.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/parameters/JsonParameters.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/parameters/ModelParameters.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/parameters/ParameterJsonSerializer.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/parameters/package-info.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/server/AnthropicApiSupport.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/server/AnthropicStreamTranslator.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/server/ChunkSink.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/server/LlamaModelBackend.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/server/NativeServer.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/server/OaiRerankSupport.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/server/OllamaApiSupport.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/server/OpenAiBackend.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/server/OpenAiCompatServer.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/server/OpenAiRequestMapper.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/server/OpenAiServerCli.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/server/OpenAiServerConfig.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/server/OpenAiSseFormatter.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/server/ResponsesApiSupport.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/server/ResponsesStreamTranslator.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/server/ToolCallDeltaAccumulator.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/server/package-info.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/ChatChoice.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/ChatMessage.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/ChatResponse.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/ChatTranscript.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/CompletionResult.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/ContentPart.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/LlamaOutput.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/LogLevel.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/ModelMeta.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/Pair.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/ServerMetrics.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/SlotMetrics.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/StopReason.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/Timings.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/TokenLogprob.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/ToolCall.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/ToolDefinition.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/Usage.java (100%)
 rename {src => llama/src}/main/java/net/ladenthin/llama/value/package-info.java (100%)
 rename {src => llama/src}/test/cpp/test_jni_helpers.cpp (100%)
 rename {src => llama/src}/test/cpp/test_json_helpers.cpp (100%)
 rename {src => llama/src}/test/cpp/test_log_helpers.cpp (100%)
 rename {src => llama/src}/test/cpp/test_server.cpp (100%)
 rename {src => llama/src}/test/cpp/test_tts_wav.cpp (100%)
 rename {src => llama/src}/test/cpp/test_utils.cpp (100%)
 rename {src => llama/src}/test/java/examples/ChatExample.java (100%)
 rename {src => llama/src}/test/java/examples/GrammarExample.java (100%)
 rename {src => llama/src}/test/java/examples/InfillExample.java (100%)
 rename {src => llama/src}/test/java/examples/MainExample.java (100%)
 rename {src => llama/src}/test/java/examples/OpenAiServerExample.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/AudioInputIntegrationTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/ChatAdvancedTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/ChatScenarioTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/ClaudeGenerated.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/ConfigureParallelInferenceTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/ErrorHandlingTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/LlamaArchitectureTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/LlamaEmbeddingsTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/LlamaModelOfflineTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/LlamaModelTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/LlamaParameterProperties.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/MemoryManagementTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/MultimodalIntegrationTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/MultimodalMessagesTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/ReactorIntegrationTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/ReasoningBudgetTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/RerankingModelTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/ResponseJsonStructureTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/SessionConcurrencyTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/SessionStateTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/SessionTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/TestConstants.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/ToolCallingAgentTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/ToolCallingIntegrationTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/TtsIntegrationTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/args/AbstractCliArgEnumTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/args/CacheTypeTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/args/ContinuationModeTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/args/GpuSplitModeTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/args/LogFormatTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/args/MiroStatTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/args/ModelFlagTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/args/NumaStrategyTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/args/PoolingTypeTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/args/ReasoningFormatTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/args/RopeScalingTypeTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/args/SamplerTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/benchmark/InferenceParametersBenchmark.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/callback/CancellationTokenLincheckTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/callback/CancellationTokenTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/callback/LoadProgressCallbackTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/exception/LlamaExceptionTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/exception/ModelUnavailableExceptionTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/jcstress/CancellationTokenRace.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/json/ChatResponseParserTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/json/ChatStreamChunkParserTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/json/CompletionResponseParserTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/json/ParameterJsonSerializerTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/json/RerankResponseParserTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/json/TimingsLoggerTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/loader/LlamaLoaderTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/loader/LoggingSmokeTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/loader/NativeLibraryLoadSmokeTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/loader/NativeLibraryPermissionSetterTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/loader/OSInfoTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/parameters/ChatRequestTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/parameters/InferenceParametersTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/parameters/JsonEndpointParametersTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/parameters/JsonParametersTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/parameters/ModelParametersExtendedTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/parameters/ModelParametersTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/AnthropicApiSupportTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/AnthropicStreamTranslatorTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/NativeServerSmokeTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/OaiRerankSupportTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/OllamaApiSupportTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/OpenAiCompatServerHttpTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/OpenAiCompatServerIntegrationTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/OpenAiRequestMapperTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/OpenAiServerCliTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/OpenAiServerCompletionIntegrationTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/OpenAiServerConfigTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/OpenAiServerEmbeddingsIntegrationTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/OpenAiServerRerankIntegrationTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/OpenAiServerTestSupport.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/OpenAiServerToolCallingIntegrationTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/OpenAiSseFormatterTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/ResponsesApiSupportTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/ResponsesStreamTranslatorTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/server/ToolCallDeltaAccumulatorTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/value/ChatChoiceTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/value/ChatMessageTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/value/ChatResponseTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/value/ChatTranscriptTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/value/CompletionResultTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/value/ContentPartTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/value/LlamaOutputTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/value/LogLevelTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/value/ModelMetaTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/value/PairTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/value/ServerMetricsTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/value/StopReasonTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/value/TimingsTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/value/TokenLogprobTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/value/ToolCallTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/value/ToolDefinitionTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/value/UsageTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/vmlens/SessionStateInterleavingTest.java (100%)
 rename {src => llama/src}/test/java/net/ladenthin/llama/vmlens/VmlensInterleavingSmokeTest.java (100%)
 rename {src => llama/src}/test/resources/images/README.md (100%)
 rename {src => llama/src}/test/resources/images/test-image.jpg (100%)

diff --git a/llama-langchain4j/pom.xml b/llama-langchain4j/pom.xml
index ecb0b8bb..45ca9944 100644
--- a/llama-langchain4j/pom.xml
+++ b/llama-langchain4j/pom.xml
@@ -9,9 +9,14 @@ SPDX-License-Identifier: MIT
 	xsi:schemaLocation="http://maven.apache.org/POM/4.0.0 http://maven.apache.org/xsd/maven-4.0.0.xsd">
 	<modelVersion>4.0.0</modelVersion>
 
-	<groupId>net.ladenthin</groupId>
+	<parent>
+		<groupId>net.ladenthin</groupId>
+		<artifactId>llama-parent</artifactId>
+		<version>5.0.4-SNAPSHOT</version>
+		<relativePath>../pom.xml</relativePath>
+	</parent>
+
 	<artifactId>llama-langchain4j</artifactId>
-	<version>5.0.4-SNAPSHOT</version>
 	<packaging>jar</packaging>
 
 	<name>${project.groupId}:${project.artifactId}</name>
diff --git a/.clang-format b/llama/.clang-format
similarity index 100%
rename from .clang-format
rename to llama/.clang-format
diff --git a/.clang-tidy b/llama/.clang-tidy
similarity index 100%
rename from .clang-tidy
rename to llama/.clang-tidy
diff --git a/CMakeLists.txt b/llama/CMakeLists.txt
similarity index 100%
rename from CMakeLists.txt
rename to llama/CMakeLists.txt
diff --git a/cmake/apply-llama-patches.cmake b/llama/cmake/apply-llama-patches.cmake
similarity index 100%
rename from cmake/apply-llama-patches.cmake
rename to llama/cmake/apply-llama-patches.cmake
diff --git a/cmake/generate-tts-upstream.cmake b/llama/cmake/generate-tts-upstream.cmake
similarity index 100%
rename from cmake/generate-tts-upstream.cmake
rename to llama/cmake/generate-tts-upstream.cmake
diff --git a/lombok.config b/llama/lombok.config
similarity index 100%
rename from lombok.config
rename to llama/lombok.config
diff --git a/patches/0001-win32-arg-parse-embed-guard.patch b/llama/patches/0001-win32-arg-parse-embed-guard.patch
similarity index 100%
rename from patches/0001-win32-arg-parse-embed-guard.patch
rename to llama/patches/0001-win32-arg-parse-embed-guard.patch
diff --git a/patches/0002-server-preserve-caller-load-progress-callback.patch b/llama/patches/0002-server-preserve-caller-load-progress-callback.patch
similarity index 100%
rename from patches/0002-server-preserve-caller-load-progress-callback.patch
rename to llama/patches/0002-server-preserve-caller-load-progress-callback.patch
diff --git a/patches/0003-pr22393-server-add-slot-prompt-similarity-getter-setter.patch b/llama/patches/0003-pr22393-server-add-slot-prompt-similarity-getter-setter.patch
similarity index 100%
rename from patches/0003-pr22393-server-add-slot-prompt-similarity-getter-setter.patch
rename to llama/patches/0003-pr22393-server-add-slot-prompt-similarity-getter-setter.patch
diff --git a/patches/0004-pr23116-server-per-request-reasoning-budget-tokens.patch b/llama/patches/0004-pr23116-server-per-request-reasoning-budget-tokens.patch
similarity index 100%
rename from patches/0004-pr23116-server-per-request-reasoning-budget-tokens.patch
rename to llama/patches/0004-pr23116-server-per-request-reasoning-budget-tokens.patch
diff --git a/llama/pom.xml b/llama/pom.xml
new file mode 100644
index 00000000..4291fcaf
--- /dev/null
+++ b/llama/pom.xml
@@ -0,0 +1,1354 @@
+<!--
+SPDX-FileCopyrightText: 2026 Bernard Ladenthin <bernard.ladenthin@gmail.com>
+SPDX-FileCopyrightText: 2023-2025 Konstantin Herud
+
+SPDX-License-Identifier: MIT
+-->
+
+<project xmlns="http://maven.apache.org/POM/4.0.0"
+	xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
+	xsi:schemaLocation="http://maven.apache.org/POM/4.0.0 http://maven.apache.org/xsd/maven-4.0.0.xsd">
+	<modelVersion>4.0.0</modelVersion>
+
+	<parent>
+		<groupId>net.ladenthin</groupId>
+		<artifactId>llama-parent</artifactId>
+		<version>5.0.4-SNAPSHOT</version>
+		<relativePath>../pom.xml</relativePath>
+	</parent>
+
+	<artifactId>llama</artifactId>
+	<packaging>jar</packaging>
+
+	<name>${project.groupId}:${project.artifactId}</name>
+	<description>Java Bindings for llama.cpp - A Port of Facebook's LLaMA model
+		in C/C++.</description>
+	<url>https://github.com/bernardladenthin/java-llama.cpp</url>
+
+	<licenses>
+		<license>
+			<name>MIT License</name>
+			<url>https://www.opensource.org/licenses/mit-license.php</url>
+			<distribution>repo</distribution>
+		</license>
+	</licenses>
+
+	<developers>
+		<developer>
+			<name>Bernard Ladenthin</name>
+			<organizationUrl>https://github.com/bernardladenthin</organizationUrl>
+		</developer>
+	</developers>
+
+	<scm>
+		<connection>scm:git:https://github.com/bernardladenthin/java-llama.cpp.git</connection>
+		<developerConnection>scm:git:https://github.com/bernardladenthin/java-llama.cpp.git</developerConnection>
+		<url>https://github.com/bernardladenthin/java-llama.cpp/tree/main</url>
+	</scm>
+
+	<distributionManagement>
+		<snapshotRepository>
+			<id>central</id>
+			<url>https://central.sonatype.com/repository/maven-snapshots/</url>
+		</snapshotRepository>
+	</distributionManagement>
+
+	<properties>
+		<sonar.organization>bernardladenthin</sonar.organization>
+		<jna.version>5.19.0</jna.version>
+		<jspecify.version>1.0.0</jspecify.version>
+		<lombok.version>1.18.46</lombok.version>
+		<errorprone.version>2.50.0</errorprone.version>
+		<nullaway.version>0.13.7</nullaway.version>
+		<checker.version>4.2.0</checker.version>
+		<jackson.version>2.22.0</jackson.version>
+		<reactor.version>3.8.6</reactor.version>
+		<slf4j.version>2.0.18</slf4j.version>
+		<logback.version>1.5.37</logback.version>
+		<animal-sniffer.version>1.27</animal-sniffer.version>
+		<junit.version>6.1.1</junit.version>
+		<hamcrest.version>3.0</hamcrest.version>
+		<jmh.version>1.37</jmh.version>
+		<jcstress.version>0.16</jcstress.version>
+		<lincheck.version>3.6</lincheck.version>
+		<logcaptor.version>2.12.6</logcaptor.version>
+		<vmlens.version>1.2.28</vmlens.version>
+		<!-- DO NOT UPGRADE jqwik past 1.9.3. jqwik 1.10.0 added a deliberate
+		     anti-AI prompt-injection string to test stdout; the 1.10.1 user
+		     guide states the library "is not meant to be used by any 'AI'
+		     coding agents at all." 1.9.3 is the last pre-disclosure release
+		     and is the pinned version for this repo. Any CI / Dependabot /
+		     contributor PR that bumps this MUST be rejected. See CLAUDE.md
+		     section "jqwik prompt-injection in test output" for full context. -->
+		<jqwik.version>1.9.3</jqwik.version>
+		<archunit.version>1.4.2</archunit.version>
+		<spotbugs.version>4.10.2.0</spotbugs.version>
+		<fb-contrib.version>7.7.4</fb-contrib.version>
+		<findsecbugs.version>1.14.0</findsecbugs.version>
+		<spotless.version>3.7.0</spotless.version>
+		<palantir-java-format.version>2.94.0</palantir-java-format.version>
+		<project.build.sourceEncoding>UTF-8</project.build.sourceEncoding>
+		<project.build.outputTimestamp>${git.commit.time}</project.build.outputTimestamp>
+	</properties>
+
+	<!--
+	    Pin transitively-conflicting versions so maven-enforcer's
+	    dependencyConvergence rule passes. The direct slf4j-api version above
+	    must always win over the older 2.0.17 that logback-classic 1.5.32
+	    brings in transitively.
+	-->
+	<dependencyManagement>
+		<dependencies>
+			<dependency>
+				<groupId>org.slf4j</groupId>
+				<artifactId>slf4j-api</artifactId>
+				<version>${slf4j.version}</version>
+			</dependency>
+		</dependencies>
+	</dependencyManagement>
+
+	<dependencies>
+		<dependency>
+			<groupId>org.projectlombok</groupId>
+			<artifactId>lombok</artifactId>
+			<version>${lombok.version}</version>
+			<scope>provided</scope>
+		</dependency>
+		<dependency>
+			<groupId>org.junit.jupiter</groupId>
+			<artifactId>junit-jupiter</artifactId>
+			<version>${junit.version}</version>
+			<scope>test</scope>
+		</dependency>
+		<dependency>
+			<groupId>org.hamcrest</groupId>
+			<artifactId>hamcrest</artifactId>
+			<version>${hamcrest.version}</version>
+			<scope>test</scope>
+		</dependency>
+		<dependency>
+			<groupId>net.jqwik</groupId>
+			<artifactId>jqwik</artifactId>
+			<version>${jqwik.version}</version>
+			<scope>test</scope>
+		</dependency>
+		<dependency>
+			<groupId>com.tngtech.archunit</groupId>
+			<artifactId>archunit-junit5</artifactId>
+			<version>${archunit.version}</version>
+			<scope>test</scope>
+		</dependency>
+		<dependency>
+			<groupId>org.jspecify</groupId>
+			<artifactId>jspecify</artifactId>
+			<version>${jspecify.version}</version>
+		</dependency>
+		<dependency>
+			<groupId>org.checkerframework</groupId>
+			<artifactId>checker-qual</artifactId>
+			<version>${checker.version}</version>
+		</dependency>
+		<dependency>
+			<groupId>com.fasterxml.jackson.core</groupId>
+			<artifactId>jackson-databind</artifactId>
+			<version>${jackson.version}</version>
+		</dependency>
+		<!-- Required by OSInfo (vendored from xerial/sqlite-jdbc) for log emission. -->
+		<dependency>
+			<groupId>org.slf4j</groupId>
+			<artifactId>slf4j-api</artifactId>
+			<version>${slf4j.version}</version>
+		</dependency>
+		<!-- Default SLF4J binding shipped with this library. Runtime scope: not
+		     required on the compile classpath, only loaded at JVM startup. -->
+		<dependency>
+			<groupId>ch.qos.logback</groupId>
+			<artifactId>logback-classic</artifactId>
+			<version>${logback.version}</version>
+			<scope>runtime</scope>
+		</dependency>
+		<!-- @IgnoreJRERequirement marker used by OSInfo (vendored from xerial/sqlite-jdbc)
+		     on a few methods that touch APIs unavailable on Android. Retention=CLASS, so
+		     the dep is compile-only (provided scope). -->
+		<dependency>
+			<groupId>org.codehaus.mojo</groupId>
+			<artifactId>animal-sniffer-annotations</artifactId>
+			<version>${animal-sniffer.version}</version>
+			<scope>provided</scope>
+		</dependency>
+		<dependency>
+			<groupId>org.openjdk.jmh</groupId>
+			<artifactId>jmh-core</artifactId>
+			<version>${jmh.version}</version>
+			<scope>test</scope>
+		</dependency>
+		<dependency>
+			<groupId>org.openjdk.jmh</groupId>
+			<artifactId>jmh-generator-annprocess</artifactId>
+			<version>${jmh.version}</version>
+			<scope>test</scope>
+		</dependency>
+		<dependency>
+			<groupId>org.openjdk.jcstress</groupId>
+			<artifactId>jcstress-core</artifactId>
+			<version>${jcstress.version}</version>
+			<scope>test</scope>
+		</dependency>
+		<dependency>
+			<groupId>org.jetbrains.lincheck</groupId>
+			<artifactId>lincheck</artifactId>
+			<version>${lincheck.version}</version>
+			<scope>test</scope>
+		</dependency>
+		<!-- Test-only SLF4J/Logback log-capture for the LoggingSmokeTest. -->
+		<dependency>
+			<groupId>io.github.hakky54</groupId>
+			<artifactId>logcaptor</artifactId>
+			<version>${logcaptor.version}</version>
+			<scope>test</scope>
+		</dependency>
+		<!-- Test-only Project Reactor used by ReactorIntegrationTest to prove the
+		     "wrap LlamaIterable with Flux.fromIterable + subscribeOn" pattern works
+		     end-to-end (backpressure, AutoCloseable cancel propagation). NOT a runtime
+		     dependency — consumers bring whichever reactive lib they already use
+		     (Reactor / RxJava / Kotlin Flow / Akka — see README "Reactive integration"). -->
+		<dependency>
+			<groupId>io.projectreactor</groupId>
+			<artifactId>reactor-core</artifactId>
+			<version>${reactor.version}</version>
+			<scope>test</scope>
+		</dependency>
+		<dependency>
+			<groupId>io.projectreactor</groupId>
+			<artifactId>reactor-test</artifactId>
+			<version>${reactor.version}</version>
+			<scope>test</scope>
+		</dependency>
+		<!--
+			vmlens interleaving-analysis API. Test-scoped and transitive-dependency-free
+			(its own deps are all test-scope, hence not propagated), so it is safe on the
+			default test classpath under dependencyConvergence. Needed so the
+			VmlensInterleavingSmoke* test compiles in every build; it is only *executed*
+			under the `vmlens` profile (see that profile + the vmlens CI job). The default
+			surefire run excludes it (managed surefire <excludes> below) so the vmlens
+			agent's JDK 9 API class is never loaded outside the agent-driven run.
+		-->
+		<dependency>
+			<groupId>com.vmlens</groupId>
+			<artifactId>api</artifactId>
+			<version>${vmlens.version}</version>
+			<scope>test</scope>
+		</dependency>
+	</dependencies>
+
+	<build>
+		<pluginManagement>
+			<plugins>
+				<plugin>
+					<groupId>com.diffplug.spotless</groupId>
+					<artifactId>spotless-maven-plugin</artifactId>
+					<version>${spotless.version}</version>
+				</plugin>
+				<plugin>
+					<groupId>com.github.spotbugs</groupId>
+					<artifactId>spotbugs-maven-plugin</artifactId>
+					<version>${spotbugs.version}</version>
+				</plugin>
+				<plugin>
+					<groupId>com.vmlens</groupId>
+					<artifactId>vmlens-maven-plugin</artifactId>
+					<version>${vmlens.version}</version>
+				</plugin>
+				<plugin>
+					<groupId>io.github.git-commit-id</groupId>
+					<artifactId>git-commit-id-maven-plugin</artifactId>
+					<version>10.0.0</version>
+				</plugin>
+				<plugin>
+					<groupId>org.apache.maven.plugins</groupId>
+					<artifactId>maven-assembly-plugin</artifactId>
+					<version>3.8.0</version>
+				</plugin>
+				<plugin>
+					<groupId>org.apache.maven.plugins</groupId>
+					<artifactId>maven-compiler-plugin</artifactId>
+					<version>3.15.0</version>
+				</plugin>
+				<plugin>
+					<groupId>org.apache.maven.plugins</groupId>
+					<artifactId>maven-gpg-plugin</artifactId>
+					<version>3.2.8</version>
+				</plugin>
+				<plugin>
+					<groupId>org.apache.maven.plugins</groupId>
+					<artifactId>maven-jar-plugin</artifactId>
+					<version>3.5.0</version>
+				</plugin>
+				<plugin>
+					<groupId>org.apache.maven.plugins</groupId>
+					<artifactId>maven-javadoc-plugin</artifactId>
+					<version>3.12.0</version>
+				</plugin>
+				<plugin>
+					<groupId>org.apache.maven.plugins</groupId>
+					<artifactId>maven-resources-plugin</artifactId>
+					<version>3.5.0</version>
+				</plugin>
+				<plugin>
+					<groupId>org.apache.maven.plugins</groupId>
+					<artifactId>maven-source-plugin</artifactId>
+					<version>3.4.0</version>
+				</plugin>
+				<plugin>
+					<groupId>org.apache.maven.plugins</groupId>
+					<artifactId>maven-surefire-plugin</artifactId>
+					<version>3.5.6</version>
+					<configuration>
+						<!--
+							Tests in the `vmlens` package are meaningful only when run
+							under the vmlens agent (the `vmlens` profile / CI job). Without the
+							agent its AllInterleavings loop body never executes (a vacuous pass
+							that also prints an "agent not configured" warning), so exclude it
+							from the ordinary surefire run. The vmlens job re-includes them with
+							`-Dtest=...`, which overrides this exclude.
+						-->
+						<excludes>
+							<exclude>**/vmlens/*.java</exclude>
+						</excludes>
+					</configuration>
+				</plugin>
+				<plugin>
+					<groupId>org.apache.maven.plugins</groupId>
+					<artifactId>maven-enforcer-plugin</artifactId>
+					<version>3.6.3</version>
+				</plugin>
+				<plugin>
+					<groupId>org.codehaus.mojo</groupId>
+					<artifactId>exec-maven-plugin</artifactId>
+					<version>3.6.3</version>
+				</plugin>
+				<plugin>
+					<groupId>org.jacoco</groupId>
+					<artifactId>jacoco-maven-plugin</artifactId>
+					<version>0.8.15</version>
+				</plugin>
+				<plugin>
+					<groupId>org.pitest</groupId>
+					<artifactId>pitest-maven</artifactId>
+					<version>1.25.5</version>
+				</plugin>
+				<plugin>
+					<groupId>org.sonatype.central</groupId>
+					<artifactId>central-publishing-maven-plugin</artifactId>
+					<version>0.11.0</version>
+				</plugin>
+			</plugins>
+		</pluginManagement>
+		<plugins>
+			<plugin>
+				<groupId>org.apache.maven.plugins</groupId>
+				<artifactId>maven-enforcer-plugin</artifactId>
+				<executions>
+					<execution>
+						<id>enforce</id>
+						<goals>
+							<goal>enforce</goal>
+						</goals>
+						<configuration>
+							<rules>
+								<requireMavenVersion>
+									<version>[3.6.3,)</version>
+								</requireMavenVersion>
+								<requireJavaVersion>
+									<version>[1.8,)</version>
+								</requireJavaVersion>
+								<dependencyConvergence/>
+								<bannedDependencies>
+									<excludes>
+										<!-- Use SLF4J directly; never commons-logging. -->
+										<exclude>commons-logging:commons-logging</exclude>
+										<!-- Legacy / EoL logging frameworks. -->
+										<exclude>log4j:log4j</exclude>
+										<!-- Old hamcrest split artifacts; use hamcrest:hamcrest 2.x+. -->
+										<exclude>org.hamcrest:hamcrest-core</exclude>
+										<exclude>org.hamcrest:hamcrest-library</exclude>
+										<exclude>org.hamcrest:hamcrest-all</exclude>
+										<!-- Legacy JUnit 3 / 4; we use Jupiter (junit:junit-jupiter). -->
+										<exclude>junit:junit</exclude>
+										<exclude>junit:junit-dep</exclude>
+									</excludes>
+								</bannedDependencies>
+							</rules>
+						</configuration>
+					</execution>
+				</executions>
+			</plugin>
+			<plugin>
+				<groupId>io.github.git-commit-id</groupId>
+				<artifactId>git-commit-id-maven-plugin</artifactId>
+				<executions>
+					<execution>
+						<id>get-git-properties</id>
+						<goals>
+							<goal>revision</goal>
+						</goals>
+						<phase>initialize</phase>
+					</execution>
+				</executions>
+				<configuration>
+					<dateFormat>yyyy-MM-dd'T'HH:mm:ss'Z'</dateFormat>
+					<dateFormatTimeZone>UTC</dateFormatTimeZone>
+					<generateGitPropertiesFile>false</generateGitPropertiesFile>
+					<failOnNoGitDirectory>false</failOnNoGitDirectory>
+				</configuration>
+			</plugin>
+			<plugin>
+				<groupId>org.apache.maven.plugins</groupId>
+				<artifactId>maven-compiler-plugin</artifactId>
+				<configuration>
+					<release>8</release>
+					<testRelease>21</testRelease>
+					<showWarnings>true</showWarnings>
+					<parameters>true</parameters>
+					<compilerArgs>
+						<!--
+						  -Xlint:all surfaces every category of javac warning. Excluded:
+						    -serial    : serialVersionUID warnings on non-Serializable classes
+						    -options   : bootclasspath / target-mismatch chatter under release=8
+						    -classfile : noise when javac reads our own .class output mid-build
+						    -processing: "No processor claimed any of these annotations" — the
+						                 animal-sniffer @IgnoreJRERequirement is consumed by the
+						                 maven-plugin (not a JSR-269 processor), so javac flags
+						                 it every build
+						  -Werror is on: the ElementType.MODULE blocker is gone now that
+						  @NullMarked lives only in module-info.java (compiled at release 9 in
+						  a separate execution); see module-info.java's javadoc.
+						-->
+						<arg>-Xlint:all,-serial,-options,-classfile,-processing</arg>
+						<arg>-Werror</arg>
+						<!--
+						  Checker Framework Nullness Checker runs as a second nullness pass
+						  alongside NullAway. CF is generics-aware and stricter than NullAway,
+						  so it acts as a second-opinion verifier on the same JSpecify
+						  annotations.
+						-->
+						<!-- Explicit processor list: Lombok (two SPI classes) + Checker Framework's
+							 Nullness Checker as a 2nd nullness pass alongside NullAway. The explicit
+							 -processor argument overrides SPI discovery, so every processor we want
+							 to run must be listed here. -->
+						<arg>-processor</arg>
+						<arg>lombok.launch.AnnotationProcessorHider$AnnotationProcessor,lombok.launch.AnnotationProcessorHider$ClaimingProcessor,org.checkerframework.checker.nullness.NullnessChecker</arg>
+						<arg>-XDaddTypeAnnotationsToSymbol=true</arg>
+						<arg>-XDcompilePolicy=simple</arg>
+						<arg>--should-stop=ifError=FLOW</arg>
+						<arg>-Xplugin:ErrorProne -Xep:NullAway:ERROR -XepOpt:NullAway:OnlyNullMarked=true -XepOpt:NullAway:JSpecifyMode=true -XepOpt:NullAway:CheckOptionalEmptiness=true -XepOpt:NullAway:AcknowledgeRestrictiveAnnotations=true -XepOpt:NullAway:AcknowledgeAndroidRecent=true -XepOpt:NullAway:AssertsEnabled=true -Xep:BoxedPrimitiveEquality:ERROR -Xep:EqualsHashCode:ERROR -Xep:EqualsIncompatibleType:ERROR -Xep:IdentityBinaryExpression:ERROR -Xep:SelfAssignment:ERROR -Xep:SelfComparison:ERROR -Xep:SelfEquals:ERROR -Xep:DeadException:ERROR -Xep:FormatString:ERROR -Xep:InvalidPatternSyntax:ERROR -Xep:OptionalEquality:ERROR -Xep:ImpossibleNullComparison:ERROR</arg>
+					</compilerArgs>
+					<annotationProcessorPaths>
+						<path>
+							<groupId>org.projectlombok</groupId>
+							<artifactId>lombok</artifactId>
+							<version>${lombok.version}</version>
+						</path>
+						<path>
+							<groupId>com.google.errorprone</groupId>
+							<artifactId>error_prone_core</artifactId>
+							<version>${errorprone.version}</version>
+						</path>
+						<path>
+							<groupId>com.uber.nullaway</groupId>
+							<artifactId>nullaway</artifactId>
+							<version>${nullaway.version}</version>
+						</path>
+						<path>
+							<groupId>org.checkerframework</groupId>
+							<artifactId>checker</artifactId>
+							<version>${checker.version}</version>
+						</path>
+					</annotationProcessorPaths>
+				</configuration>
+				<executions>
+					<execution>
+						<id>default-compile</id>
+						<configuration>
+							<!-- The default compile pass targets Java 8 bytecode for the library
+							     sources. module-info.java cannot be compiled at release 8 (it is a
+							     Java 9+ construct) and is handled by the separate execution below. -->
+							<excludes>
+								<exclude>module-info.java</exclude>
+							</excludes>
+						</configuration>
+					</execution>
+					<execution>
+						<id>module-info-compile</id>
+						<phase>compile</phase>
+						<goals>
+							<goal>compile</goal>
+						</goals>
+						<configuration>
+							<!-- module-info.java is the only Java 9+ source in this project; compile
+							     it at release 9 (the minimum version that understands modules) without
+							     the Error Prone / NullAway / Checker Framework processors that target
+							     ordinary source files only. Java 8 runtimes silently ignore the
+							     resulting module-info.class at the JAR root, so existing classpath
+							     consumers continue to work unchanged. LlamaLoader loads native libs
+							     via Class.getResourceAsStream on its own class object, which resolves
+							     to this module and therefore does not require any opens directives. -->
+							<release>9</release>
+							<includes>
+								<include>module-info.java</include>
+							</includes>
+							<compilerArgs combine.self="override"/>
+							<annotationProcessorPaths combine.self="override"/>
+						</configuration>
+					</execution>
+					<execution>
+						<id>default-testCompile</id>
+						<configuration>
+							<!-- Drop -parameters for test compile so jcstress reflection doesn't
+							     pick up parameter names from MethodParameters. -->
+							<parameters>false</parameters>
+							<compilerArgs combine.self="override">
+								<arg>-XDaddTypeAnnotationsToSymbol=true</arg>
+								<arg>-XDcompilePolicy=simple</arg>
+								<arg>--should-stop=ifError=FLOW</arg>
+								<arg>-Xplugin:ErrorProne -Xep:NullAway:OFF -Xep:GuardedBy:OFF</arg>
+							</compilerArgs>
+							<annotationProcessorPaths combine.children="append">
+								<path>
+									<groupId>org.openjdk.jcstress</groupId>
+									<artifactId>jcstress-core</artifactId>
+									<version>${jcstress.version}</version>
+								</path>
+								<path>
+									<groupId>org.openjdk.jmh</groupId>
+									<artifactId>jmh-generator-annprocess</artifactId>
+									<version>${jmh.version}</version>
+								</path>
+							</annotationProcessorPaths>
+						</configuration>
+					</execution>
+				</executions>
+			</plugin>
+			<plugin>
+				<artifactId>maven-resources-plugin</artifactId>
+			</plugin>
+			<plugin>
+				<groupId>org.apache.maven.plugins</groupId>
+				<artifactId>maven-source-plugin</artifactId>
+				<executions>
+					<execution>
+						<id>attach-sources</id>
+						<goals>
+							<goal>jar-no-fork</goal>
+						</goals>
+					</execution>
+				</executions>
+			</plugin>
+			<plugin>
+				<groupId>org.apache.maven.plugins</groupId>
+				<artifactId>maven-javadoc-plugin</artifactId>
+				<configuration>
+					<source>1.8</source>
+					<failOnError>true</failOnError>
+					<failOnWarnings>true</failOnWarnings>
+					<doclint>all</doclint>
+				</configuration>
+				<executions>
+					<execution>
+						<id>attach-javadocs</id>
+						<goals>
+							<goal>jar</goal>
+						</goals>
+					</execution>
+				</executions>
+			</plugin>
+			<plugin>
+				<groupId>org.jacoco</groupId>
+				<artifactId>jacoco-maven-plugin</artifactId>
+				<executions>
+					<execution>
+						<id>prepare-agent</id>
+						<goals>
+							<goal>prepare-agent</goal>
+						</goals>
+					</execution>
+					<execution>
+						<id>report</id>
+						<phase>test</phase>
+						<goals>
+							<goal>report</goal>
+						</goals>
+					</execution>
+				</executions>
+			</plugin>
+			<plugin>
+				<!--
+				  Surefire is configured here only to control native-crash diagnostics:
+				  - argLine inherits jacoco's @{argLine} (so the jacoco runtime agent stays
+				    attached) and appends -XX:ErrorFile so the JVM crash log (hs_err_pid<pid>.log)
+				    always lands in the project basedir, which is the same directory the CI
+				    upload-artifact globs (${{ github.workspace }}/hs_err_pid*.log).
+				  - HeapDumpOnOutOfMemoryError + HeapDumpPath catch OOMs that would otherwise
+				    just print a stack and exit.
+				  - The version is pinned to the same release Maven would otherwise auto-select
+				    (visible in CI error logs as 3.5.4), so we don't drift independently of
+				    Maven's plugin-defaults bump.
+				-->
+				<groupId>org.apache.maven.plugins</groupId>
+				<artifactId>maven-surefire-plugin</artifactId>
+				<configuration>
+					<argLine>@{argLine} -Xmx2g -XX:ErrorFile=hs_err_pid%p.log -XX:+HeapDumpOnOutOfMemoryError -XX:HeapDumpPath=.</argLine>
+					<!--
+					  Capture each test class's stdout/stderr into
+					  target/surefire-reports/<class>-output.txt. When a native crash
+					  takes down the forked JVM, this file is the ONLY record of
+					  which @Test method was last running (without it the surefire
+					  *.txt summary just says "VM terminated", which we already saw
+					  in CI without telling us which test was on fire).
+					-->
+					<redirectTestOutputToFile>true</redirectTestOutputToFile>
+				</configuration>
+			</plugin>
+			<plugin>
+				<groupId>com.diffplug.spotless</groupId>
+				<artifactId>spotless-maven-plugin</artifactId>
+				<configuration>
+					<java>
+						<includes>
+							<include>src/main/java/**/*.java</include>
+							<include>src/test/java/**/*.java</include>
+						</includes>
+						<palantirJavaFormat>
+							<version>${palantir-java-format.version}</version>
+						</palantirJavaFormat>
+						<removeUnusedImports/>
+						<trimTrailingWhitespace/>
+						<endWithNewline/>
+					</java>
+				</configuration>
+				<executions>
+					<execution>
+						<id>spotless-check</id>
+						<phase>verify</phase>
+						<goals>
+							<goal>check</goal>
+						</goals>
+					</execution>
+				</executions>
+			</plugin>
+			<plugin>
+				<groupId>com.github.spotbugs</groupId>
+				<artifactId>spotbugs-maven-plugin</artifactId>
+				<configuration>
+					<effort>Max</effort>
+					<threshold>Low</threshold>
+					<failOnError>true</failOnError>
+					<includeTests>false</includeTests>
+					<excludeFilterFile>spotbugs-exclude.xml</excludeFilterFile>
+					<plugins>
+						<plugin>
+							<groupId>com.mebigfatguy.fb-contrib</groupId>
+							<artifactId>fb-contrib</artifactId>
+							<version>${fb-contrib.version}</version>
+						</plugin>
+						<plugin>
+							<groupId>com.h3xstream.findsecbugs</groupId>
+							<artifactId>findsecbugs-plugin</artifactId>
+							<version>${findsecbugs.version}</version>
+						</plugin>
+					</plugins>
+				</configuration>
+				<executions>
+					<execution>
+						<id>spotbugs-check</id>
+						<phase>verify</phase>
+						<goals>
+							<goal>check</goal>
+						</goals>
+					</execution>
+				</executions>
+			</plugin>
+			<plugin>
+				<groupId>org.codehaus.mojo</groupId>
+				<artifactId>exec-maven-plugin</artifactId>
+				<configuration>
+					<mainClass>org.openjdk.jmh.Main</mainClass>
+					<classpathScope>test</classpathScope>
+				</configuration>
+			</plugin>
+			<plugin>
+				<!--
+				  Generic PIT policy (version, 100% gate, invocation, expansion rule):
+				  see ../workspace/policies/pit-mutation-testing.md.
+				  Repo-specific: the gated value/exception/args/json classes are pure-Java
+				  (no native libjllama or model file needed); targetClasses/targetTests are
+				  kept in lock-step. NOTE: value.ContentPart.audioFile(Path) mutants are only
+				  killed with the audio fixture present — see TODO.md and policy §4.
+				-->
+				<groupId>org.pitest</groupId>
+				<artifactId>pitest-maven</artifactId>
+				<dependencies>
+					<dependency>
+						<groupId>org.pitest</groupId>
+						<artifactId>pitest-junit5-plugin</artifactId>
+						<version>1.2.3</version>
+					</dependency>
+				</dependencies>
+				<configuration>
+					<targetClasses>
+						<param>net.ladenthin.llama.value.*</param>
+						<param>net.ladenthin.llama.exception.*</param>
+						<param>net.ladenthin.llama.args.*</param>
+						<param>net.ladenthin.llama.json.TimingsLogger</param>
+						<param>net.ladenthin.llama.json.RerankResponseParser</param>
+						<param>net.ladenthin.llama.json.ChatResponseParser</param>
+						<param>net.ladenthin.llama.json.CompletionResponseParser</param>
+					</targetClasses>
+					<targetTests>
+						<param>net.ladenthin.llama.value.*</param>
+						<param>net.ladenthin.llama.exception.*</param>
+						<param>net.ladenthin.llama.args.*</param>
+						<param>net.ladenthin.llama.json.*</param>
+					</targetTests>
+					<mutationThreshold>100</mutationThreshold>
+					<timeoutConstant>30000</timeoutConstant>
+				</configuration>
+			</plugin>
+		</plugins>
+	</build>
+
+	<profiles>
+		<profile>
+			<id>release</id>
+			<build>
+				<plugins>
+					<plugin>
+						<groupId>org.apache.maven.plugins</groupId>
+						<artifactId>maven-gpg-plugin</artifactId>
+						<executions>
+							<execution>
+								<id>sign-artifacts</id>
+								<phase>verify</phase>
+								<goals><goal>sign</goal></goals>
+								<configuration>
+									<keyname>${gpg.keyname}</keyname>
+									<gpgArguments>
+										<arg>--pinentry-mode</arg>
+										<arg>loopback</arg>
+									</gpgArguments>
+								</configuration>
+							</execution>
+						</executions>
+					</plugin>
+
+					<plugin>
+						<groupId>org.sonatype.central</groupId>
+						<artifactId>central-publishing-maven-plugin</artifactId>
+						<extensions>true</extensions>
+						<configuration>
+							<publishingServerId>central</publishingServerId>
+							<autoPublish>true</autoPublish>
+							<waitUntil>published</waitUntil>
+						</configuration>
+					</plugin>
+				</plugins>
+			</build>
+		</profile>
+
+		<profile>
+			<id>cuda</id>
+			<build>
+				<plugins>
+					<plugin>
+						<groupId>org.apache.maven.plugins</groupId>
+						<artifactId>maven-compiler-plugin</artifactId>
+						<executions>
+						<!-- We have to perform a separate build pass for cuda
+						classifier -->
+						<execution>
+							<id>gpu</id>
+							<phase>compile</phase>
+							<goals>
+								<goal>compile</goal>
+							</goals>
+							<configuration>
+								<!-- Same rationale as default-compile: this pass targets release 8;
+								     module-info.java is Java 9+ and is handled by module-info-compile. -->
+								<excludes>
+									<exclude>module-info.java</exclude>
+								</excludes>
+								<compilerArgs>
+									<arg>-h</arg>
+									<arg>src/main/cpp</arg>
+								</compilerArgs>
+								<outputDirectory>
+									${project.build.outputDirectory}_cuda</outputDirectory>
+							</configuration>
+						</execution>
+						</executions>
+					</plugin>
+					<plugin>
+						<artifactId>maven-resources-plugin</artifactId>
+						<executions>
+						<!-- Copy custom cuda libs to the output directory -->
+						<execution>
+							<id>copy-resources</id>
+							<phase>process-classes</phase>
+							<goals>
+								<goal>copy-resources</goal>
+							</goals>
+							<configuration>
+								<outputDirectory>
+									${project.build.outputDirectory}_cuda</outputDirectory>
+								<resources>
+									<resource>
+										<directory>
+											${basedir}/src/main/resources_linux_cuda/</directory>
+										<includes>
+											<include>**/*.*</include>
+										</includes>
+									</resource>
+								</resources>
+							</configuration>
+						</execution>
+						</executions>
+					</plugin>
+					<plugin>
+						<groupId>org.apache.maven.plugins</groupId>
+						<artifactId>maven-jar-plugin</artifactId>
+						<executions>
+							<!-- Pick class files AND libs from custom output
+							directory -->
+							<execution>
+								<id>cuda</id>
+								<phase>package</phase>
+								<goals>
+									<goal>jar</goal>
+								</goals>
+								<configuration>
+									<classifier>cuda13-linux-x86-64</classifier>
+									<classesDirectory>
+										${project.build.outputDirectory}_cuda</classesDirectory>
+								</configuration>
+							</execution>
+						</executions>
+					</plugin>
+				</plugins>
+			</build>
+		</profile>
+
+		<profile>
+			<id>opencl-android</id>
+			<build>
+				<plugins>
+					<plugin>
+						<groupId>org.apache.maven.plugins</groupId>
+						<artifactId>maven-compiler-plugin</artifactId>
+						<executions>
+						<!-- Separate build pass for opencl-android classifier -->
+						<execution>
+							<id>opencl-android</id>
+							<phase>compile</phase>
+							<goals>
+								<goal>compile</goal>
+							</goals>
+							<configuration>
+								<!-- Same rationale as default-compile: this pass targets release 8;
+								     module-info.java is Java 9+ and is handled by module-info-compile. -->
+								<excludes>
+									<exclude>module-info.java</exclude>
+								</excludes>
+								<compilerArgs>
+									<arg>-h</arg>
+									<arg>src/main/cpp</arg>
+								</compilerArgs>
+								<outputDirectory>
+									${project.build.outputDirectory}_opencl_android</outputDirectory>
+							</configuration>
+						</execution>
+						</executions>
+					</plugin>
+					<plugin>
+						<artifactId>maven-resources-plugin</artifactId>
+						<executions>
+						<!-- Copy custom opencl-android libs to the output directory -->
+						<execution>
+							<id>copy-resources-opencl-android</id>
+							<phase>process-classes</phase>
+							<goals>
+								<goal>copy-resources</goal>
+							</goals>
+							<configuration>
+								<outputDirectory>
+									${project.build.outputDirectory}_opencl_android</outputDirectory>
+								<resources>
+									<resource>
+										<directory>
+											${basedir}/src/main/resources_android_opencl/</directory>
+										<includes>
+											<include>**/*.*</include>
+										</includes>
+									</resource>
+								</resources>
+							</configuration>
+						</execution>
+						</executions>
+					</plugin>
+					<plugin>
+						<groupId>org.apache.maven.plugins</groupId>
+						<artifactId>maven-jar-plugin</artifactId>
+						<executions>
+							<!-- Pick class files AND libs from custom output
+							directory -->
+							<execution>
+								<id>opencl-android</id>
+								<phase>package</phase>
+								<goals>
+									<goal>jar</goal>
+								</goals>
+								<configuration>
+									<classifier>opencl-android-aarch64</classifier>
+									<classesDirectory>
+										${project.build.outputDirectory}_opencl_android</classesDirectory>
+								</configuration>
+							</execution>
+						</executions>
+					</plugin>
+				</plugins>
+			</build>
+		</profile>
+		<profile>
+			<!-- MSVC (Visual Studio generator) Windows CPU natives, shipped as the
+			     `msvc-windows` classifier JAR. After the default flip the DEFAULT JAR
+			     ships the Ninja Multi-Config natives (sccache-cached, merged into
+			     src/main/resources/ by CI); this profile packages the alternative
+			     MSVC-generator build for anyone who wants it. Both use the same MSVC
+			     toolchain (static /MT CRT) so they are functionally equivalent. Mirrors
+			     the cuda / opencl-android classifier pattern: a separate compile pass +
+			     resource copy + classified jar. The MSVC DLLs are staged into
+			     src/main/resources_windows_msvc/ by CI before this profile runs. -->
+			<id>windows-msvc</id>
+			<build>
+				<plugins>
+					<plugin>
+						<groupId>org.apache.maven.plugins</groupId>
+						<artifactId>maven-compiler-plugin</artifactId>
+						<executions>
+						<!-- Separate build pass for the windows-msvc classifier -->
+						<execution>
+							<id>windows-msvc</id>
+							<phase>compile</phase>
+							<goals>
+								<goal>compile</goal>
+							</goals>
+							<configuration>
+								<!-- Same rationale as default-compile: this pass targets release 8;
+								     module-info.java is Java 9+ and is handled by module-info-compile. -->
+								<excludes>
+									<exclude>module-info.java</exclude>
+								</excludes>
+								<compilerArgs>
+									<arg>-h</arg>
+									<arg>src/main/cpp</arg>
+								</compilerArgs>
+								<outputDirectory>
+									${project.build.outputDirectory}_windows_msvc</outputDirectory>
+							</configuration>
+						</execution>
+						</executions>
+					</plugin>
+					<plugin>
+						<artifactId>maven-resources-plugin</artifactId>
+						<executions>
+						<!-- Copy the Ninja-built Windows libs to the output directory -->
+						<execution>
+							<id>copy-resources-windows-msvc</id>
+							<phase>process-classes</phase>
+							<goals>
+								<goal>copy-resources</goal>
+							</goals>
+							<configuration>
+								<outputDirectory>
+									${project.build.outputDirectory}_windows_msvc</outputDirectory>
+								<resources>
+									<resource>
+										<directory>
+											${basedir}/src/main/resources_windows_msvc/</directory>
+										<includes>
+											<include>**/*.*</include>
+										</includes>
+									</resource>
+								</resources>
+							</configuration>
+						</execution>
+						</executions>
+					</plugin>
+					<plugin>
+						<groupId>org.apache.maven.plugins</groupId>
+						<artifactId>maven-jar-plugin</artifactId>
+						<executions>
+							<!-- Pick class files AND libs from custom output
+							directory -->
+							<execution>
+								<id>windows-msvc</id>
+								<phase>package</phase>
+								<goals>
+									<goal>jar</goal>
+								</goals>
+								<configuration>
+									<classifier>msvc-windows</classifier>
+									<classesDirectory>
+										${project.build.outputDirectory}_windows_msvc</classesDirectory>
+								</configuration>
+							</execution>
+						</executions>
+					</plugin>
+				</plugins>
+			</build>
+		</profile>
+
+		<profile>
+			<!-- Windows CUDA GPU natives (x86_64 only), shipped as the
+			     `cuda13-windows-x86-64` classifier JAR. Built with -DGGML_CUDA=ON and
+			     routed by CMakeLists.txt to src/main/resources_windows_cuda/. The CUDA
+			     runtime libs (cudart64_13 / cublas64_13 / cublasLt64_13) are NOT bundled
+			     — the consumer's installed CUDA Toolkit provides them. Mirrors the
+			     cuda (Linux) classifier pattern. Staged by CI before this profile runs. -->
+			<id>cuda-windows</id>
+			<build>
+				<plugins>
+					<plugin>
+						<groupId>org.apache.maven.plugins</groupId>
+						<artifactId>maven-compiler-plugin</artifactId>
+						<executions>
+						<execution>
+							<id>cuda-windows</id>
+							<phase>compile</phase>
+							<goals>
+								<goal>compile</goal>
+							</goals>
+							<configuration>
+								<excludes>
+									<exclude>module-info.java</exclude>
+								</excludes>
+								<compilerArgs>
+									<arg>-h</arg>
+									<arg>src/main/cpp</arg>
+								</compilerArgs>
+								<outputDirectory>
+									${project.build.outputDirectory}_windows_cuda</outputDirectory>
+							</configuration>
+						</execution>
+						</executions>
+					</plugin>
+					<plugin>
+						<artifactId>maven-resources-plugin</artifactId>
+						<executions>
+						<execution>
+							<id>copy-resources-cuda-windows</id>
+							<phase>process-classes</phase>
+							<goals>
+								<goal>copy-resources</goal>
+							</goals>
+							<configuration>
+								<outputDirectory>
+									${project.build.outputDirectory}_windows_cuda</outputDirectory>
+								<resources>
+									<resource>
+										<directory>
+											${basedir}/src/main/resources_windows_cuda/</directory>
+										<includes>
+											<include>**/*.*</include>
+										</includes>
+									</resource>
+								</resources>
+							</configuration>
+						</execution>
+						</executions>
+					</plugin>
+					<plugin>
+						<groupId>org.apache.maven.plugins</groupId>
+						<artifactId>maven-jar-plugin</artifactId>
+						<executions>
+							<execution>
+								<id>cuda-windows</id>
+								<phase>package</phase>
+								<goals>
+									<goal>jar</goal>
+								</goals>
+								<configuration>
+									<classifier>cuda13-windows-x86-64</classifier>
+									<classesDirectory>
+										${project.build.outputDirectory}_windows_cuda</classesDirectory>
+								</configuration>
+							</execution>
+						</executions>
+					</plugin>
+				</plugins>
+			</build>
+		</profile>
+
+		<profile>
+			<!-- Windows Vulkan GPU natives (x86_64 only), shipped as the
+			     `vulkan-windows-x86-64` classifier JAR. Built with -DGGML_VULKAN=ON and
+			     routed by CMakeLists.txt to src/main/resources_windows_vulkan/. The Vulkan
+			     runtime (vulkan-1.dll) ships with the GPU driver — nothing is bundled.
+			     Vulkan is the most portable Windows GPU backend (NVIDIA/AMD/Intel).
+			     Staged by CI before this profile runs. -->
+			<id>vulkan-windows</id>
+			<build>
+				<plugins>
+					<plugin>
+						<groupId>org.apache.maven.plugins</groupId>
+						<artifactId>maven-compiler-plugin</artifactId>
+						<executions>
+						<execution>
+							<id>vulkan-windows</id>
+							<phase>compile</phase>
+							<goals>
+								<goal>compile</goal>
+							</goals>
+							<configuration>
+								<excludes>
+									<exclude>module-info.java</exclude>
+								</excludes>
+								<compilerArgs>
+									<arg>-h</arg>
+									<arg>src/main/cpp</arg>
+								</compilerArgs>
+								<outputDirectory>
+									${project.build.outputDirectory}_windows_vulkan</outputDirectory>
+							</configuration>
+						</execution>
+						</executions>
+					</plugin>
+					<plugin>
+						<artifactId>maven-resources-plugin</artifactId>
+						<executions>
+						<execution>
+							<id>copy-resources-vulkan-windows</id>
+							<phase>process-classes</phase>
+							<goals>
+								<goal>copy-resources</goal>
+							</goals>
+							<configuration>
+								<outputDirectory>
+									${project.build.outputDirectory}_windows_vulkan</outputDirectory>
+								<resources>
+									<resource>
+										<directory>
+											${basedir}/src/main/resources_windows_vulkan/</directory>
+										<includes>
+											<include>**/*.*</include>
+										</includes>
+									</resource>
+								</resources>
+							</configuration>
+						</execution>
+						</executions>
+					</plugin>
+					<plugin>
+						<groupId>org.apache.maven.plugins</groupId>
+						<artifactId>maven-jar-plugin</artifactId>
+						<executions>
+							<execution>
+								<id>vulkan-windows</id>
+								<phase>package</phase>
+								<goals>
+									<goal>jar</goal>
+								</goals>
+								<configuration>
+									<classifier>vulkan-windows-x86-64</classifier>
+									<classesDirectory>
+										${project.build.outputDirectory}_windows_vulkan</classesDirectory>
+								</configuration>
+							</execution>
+						</executions>
+					</plugin>
+				</plugins>
+			</build>
+		</profile>
+
+		<profile>
+			<!-- Windows OpenCL GPU natives (x86_64 only), shipped as the
+			     `opencl-windows-x86-64` classifier JAR. Built with -DGGML_OPENCL=ON and
+			     routed by CMakeLists.txt to src/main/resources_windows_opencl/. The OpenCL
+			     ICD (System32\OpenCL.dll) ships with the GPU driver — nothing is bundled.
+			     NOTE: the GGML OpenCL backend is Adreno-tuned; on desktop GPUs CUDA/Vulkan
+			     are better supported. Staged by CI before this profile runs. -->
+			<id>opencl-windows</id>
+			<build>
+				<plugins>
+					<plugin>
+						<groupId>org.apache.maven.plugins</groupId>
+						<artifactId>maven-compiler-plugin</artifactId>
+						<executions>
+						<execution>
+							<id>opencl-windows</id>
+							<phase>compile</phase>
+							<goals>
+								<goal>compile</goal>
+							</goals>
+							<configuration>
+								<excludes>
+									<exclude>module-info.java</exclude>
+								</excludes>
+								<compilerArgs>
+									<arg>-h</arg>
+									<arg>src/main/cpp</arg>
+								</compilerArgs>
+								<outputDirectory>
+									${project.build.outputDirectory}_windows_opencl</outputDirectory>
+							</configuration>
+						</execution>
+						</executions>
+					</plugin>
+					<plugin>
+						<artifactId>maven-resources-plugin</artifactId>
+						<executions>
+						<execution>
+							<id>copy-resources-opencl-windows</id>
+							<phase>process-classes</phase>
+							<goals>
+								<goal>copy-resources</goal>
+							</goals>
+							<configuration>
+								<outputDirectory>
+									${project.build.outputDirectory}_windows_opencl</outputDirectory>
+								<resources>
+									<resource>
+										<directory>
+											${basedir}/src/main/resources_windows_opencl/</directory>
+										<includes>
+											<include>**/*.*</include>
+										</includes>
+									</resource>
+								</resources>
+							</configuration>
+						</execution>
+						</executions>
+					</plugin>
+					<plugin>
+						<groupId>org.apache.maven.plugins</groupId>
+						<artifactId>maven-jar-plugin</artifactId>
+						<executions>
+							<execution>
+								<id>opencl-windows</id>
+								<phase>package</phase>
+								<goals>
+									<goal>jar</goal>
+								</goals>
+								<configuration>
+									<classifier>opencl-windows-x86-64</classifier>
+									<classesDirectory>
+										${project.build.outputDirectory}_windows_opencl</classesDirectory>
+								</configuration>
+							</execution>
+						</executions>
+					</plugin>
+				</plugins>
+			</build>
+		</profile>
+
+		<profile>
+			<id>vmlens</id>
+			<build>
+				<plugins>
+					<plugin>
+						<groupId>com.vmlens</groupId>
+						<artifactId>vmlens-maven-plugin</artifactId>
+						<configuration>
+							<!--
+								Run vmlens interleaving analysis over the whole `vmlens` test
+								package (smoke test + SessionStateInterleavingTest). The `com.vmlens:api`
+								test dependency lives in the main <dependencies> block.
+								Expand <includes> as more concurrency tests are added (the
+								streambuffer repo runs vmlens over its whole suite).
+							-->
+							<includes>
+								<include>**/vmlens/*.java</include>
+							</includes>
+						</configuration>
+						<executions>
+							<execution>
+								<id>vmlens-test</id>
+								<goals>
+									<goal>test</goal>
+								</goals>
+							</execution>
+						</executions>
+					</plugin>
+				</plugins>
+			</build>
+		</profile>
+		<profile>
+			<id>jcstress</id>
+			<build>
+				<plugins>
+					<plugin>
+						<groupId>org.codehaus.mojo</groupId>
+						<artifactId>exec-maven-plugin</artifactId>
+						<executions>
+							<execution>
+								<id>jcstress</id>
+								<phase>test</phase>
+								<goals><goal>exec</goal></goals>
+								<configuration>
+									<executable>${java.home}/bin/java</executable>
+									<classpathScope>test</classpathScope>
+									<arguments>
+										<argument>-classpath</argument>
+										<classpath/>
+										<argument>org.openjdk.jcstress.Main</argument>
+										<argument>-v</argument>
+										<argument>-m</argument>
+										<argument>default</argument>
+									</arguments>
+								</configuration>
+							</execution>
+						</executions>
+					</plugin>
+				</plugins>
+			</build>
+		</profile>
+		<profile>
+			<!--
+				Builds the fat jar-with-dependencies uber JAR: the library classes, the
+				default-platform native libs from src/main/resources, and all runtime Java
+				dependencies in one drop-on-classpath JAR, runnable via the OpenAiCompatServer
+				Main-Class (set below) to start the OpenAI-compatible HTTP server. Off by
+				default; the CI `package` job activates it so the uber JAR rides along in the
+				`llama-jars` upload-artifact bundle. Documented in CLAUDE.md "Build Commands"
+				as `mvn -P assembly package`.
+			-->
+			<id>assembly</id>
+			<build>
+				<plugins>
+					<plugin>
+						<groupId>org.apache.maven.plugins</groupId>
+						<artifactId>maven-assembly-plugin</artifactId>
+						<configuration>
+							<descriptorRefs>
+								<descriptorRef>jar-with-dependencies</descriptorRef>
+							</descriptorRefs>
+							<archive>
+								<manifest>
+									<mainClass>net.ladenthin.llama.server.OpenAiCompatServer</mainClass>
+								</manifest>
+							</archive>
+						</configuration>
+						<executions>
+							<execution>
+								<id>build-fat-jar</id>
+								<phase>package</phase>
+								<goals>
+									<goal>single</goal>
+								</goals>
+							</execution>
+						</executions>
+					</plugin>
+				</plugins>
+			</build>
+		</profile>
+	</profiles>
+</project>
diff --git a/spotbugs-exclude.xml b/llama/spotbugs-exclude.xml
similarity index 100%
rename from spotbugs-exclude.xml
rename to llama/spotbugs-exclude.xml
diff --git a/src/main/cpp/compat/ggml_x86_compat.c b/llama/src/main/cpp/compat/ggml_x86_compat.c
similarity index 100%
rename from src/main/cpp/compat/ggml_x86_compat.c
rename to llama/src/main/cpp/compat/ggml_x86_compat.c
diff --git a/src/main/cpp/jllama.cpp b/llama/src/main/cpp/jllama.cpp
similarity index 100%
rename from src/main/cpp/jllama.cpp
rename to llama/src/main/cpp/jllama.cpp
diff --git a/src/main/cpp/jllama.h b/llama/src/main/cpp/jllama.h
similarity index 100%
rename from src/main/cpp/jllama.h
rename to llama/src/main/cpp/jllama.h
diff --git a/src/main/cpp/jni_helpers.hpp b/llama/src/main/cpp/jni_helpers.hpp
similarity index 100%
rename from src/main/cpp/jni_helpers.hpp
rename to llama/src/main/cpp/jni_helpers.hpp
diff --git a/src/main/cpp/json_helpers.hpp b/llama/src/main/cpp/json_helpers.hpp
similarity index 100%
rename from src/main/cpp/json_helpers.hpp
rename to llama/src/main/cpp/json_helpers.hpp
diff --git a/src/main/cpp/log_helpers.hpp b/llama/src/main/cpp/log_helpers.hpp
similarity index 100%
rename from src/main/cpp/log_helpers.hpp
rename to llama/src/main/cpp/log_helpers.hpp
diff --git a/src/main/cpp/tts_engine.cpp b/llama/src/main/cpp/tts_engine.cpp
similarity index 100%
rename from src/main/cpp/tts_engine.cpp
rename to llama/src/main/cpp/tts_engine.cpp
diff --git a/src/main/cpp/tts_engine.h b/llama/src/main/cpp/tts_engine.h
similarity index 100%
rename from src/main/cpp/tts_engine.h
rename to llama/src/main/cpp/tts_engine.h
diff --git a/src/main/cpp/tts_upstream.h b/llama/src/main/cpp/tts_upstream.h
similarity index 100%
rename from src/main/cpp/tts_upstream.h
rename to llama/src/main/cpp/tts_upstream.h
diff --git a/src/main/cpp/tts_wav.hpp b/llama/src/main/cpp/tts_wav.hpp
similarity index 100%
rename from src/main/cpp/tts_wav.hpp
rename to llama/src/main/cpp/tts_wav.hpp
diff --git a/src/main/cpp/utils.hpp b/llama/src/main/cpp/utils.hpp
similarity index 100%
rename from src/main/cpp/utils.hpp
rename to llama/src/main/cpp/utils.hpp
diff --git a/src/main/cpp/webui_stub/ui.h b/llama/src/main/cpp/webui_stub/ui.h
similarity index 100%
rename from src/main/cpp/webui_stub/ui.h
rename to llama/src/main/cpp/webui_stub/ui.h
diff --git a/src/main/java/module-info.java b/llama/src/main/java/module-info.java
similarity index 100%
rename from src/main/java/module-info.java
rename to llama/src/main/java/module-info.java
diff --git a/src/main/java/net/ladenthin/llama/LlamaIterable.java b/llama/src/main/java/net/ladenthin/llama/LlamaIterable.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/LlamaIterable.java
rename to llama/src/main/java/net/ladenthin/llama/LlamaIterable.java
diff --git a/src/main/java/net/ladenthin/llama/LlamaIterator.java b/llama/src/main/java/net/ladenthin/llama/LlamaIterator.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/LlamaIterator.java
rename to llama/src/main/java/net/ladenthin/llama/LlamaIterator.java
diff --git a/src/main/java/net/ladenthin/llama/LlamaModel.java b/llama/src/main/java/net/ladenthin/llama/LlamaModel.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/LlamaModel.java
rename to llama/src/main/java/net/ladenthin/llama/LlamaModel.java
diff --git a/src/main/java/net/ladenthin/llama/Session.java b/llama/src/main/java/net/ladenthin/llama/Session.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/Session.java
rename to llama/src/main/java/net/ladenthin/llama/Session.java
diff --git a/src/main/java/net/ladenthin/llama/SessionState.java b/llama/src/main/java/net/ladenthin/llama/SessionState.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/SessionState.java
rename to llama/src/main/java/net/ladenthin/llama/SessionState.java
diff --git a/src/main/java/net/ladenthin/llama/TextToSpeech.java b/llama/src/main/java/net/ladenthin/llama/TextToSpeech.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/TextToSpeech.java
rename to llama/src/main/java/net/ladenthin/llama/TextToSpeech.java
diff --git a/src/main/java/net/ladenthin/llama/ToolCallingAgent.java b/llama/src/main/java/net/ladenthin/llama/ToolCallingAgent.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/ToolCallingAgent.java
rename to llama/src/main/java/net/ladenthin/llama/ToolCallingAgent.java
diff --git a/src/main/java/net/ladenthin/llama/args/CacheType.java b/llama/src/main/java/net/ladenthin/llama/args/CacheType.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/args/CacheType.java
rename to llama/src/main/java/net/ladenthin/llama/args/CacheType.java
diff --git a/src/main/java/net/ladenthin/llama/args/CliArg.java b/llama/src/main/java/net/ladenthin/llama/args/CliArg.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/args/CliArg.java
rename to llama/src/main/java/net/ladenthin/llama/args/CliArg.java
diff --git a/src/main/java/net/ladenthin/llama/args/ContinuationMode.java b/llama/src/main/java/net/ladenthin/llama/args/ContinuationMode.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/args/ContinuationMode.java
rename to llama/src/main/java/net/ladenthin/llama/args/ContinuationMode.java
diff --git a/src/main/java/net/ladenthin/llama/args/GpuSplitMode.java b/llama/src/main/java/net/ladenthin/llama/args/GpuSplitMode.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/args/GpuSplitMode.java
rename to llama/src/main/java/net/ladenthin/llama/args/GpuSplitMode.java
diff --git a/src/main/java/net/ladenthin/llama/args/LogFormat.java b/llama/src/main/java/net/ladenthin/llama/args/LogFormat.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/args/LogFormat.java
rename to llama/src/main/java/net/ladenthin/llama/args/LogFormat.java
diff --git a/src/main/java/net/ladenthin/llama/args/MiroStat.java b/llama/src/main/java/net/ladenthin/llama/args/MiroStat.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/args/MiroStat.java
rename to llama/src/main/java/net/ladenthin/llama/args/MiroStat.java
diff --git a/src/main/java/net/ladenthin/llama/args/ModelFlag.java b/llama/src/main/java/net/ladenthin/llama/args/ModelFlag.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/args/ModelFlag.java
rename to llama/src/main/java/net/ladenthin/llama/args/ModelFlag.java
diff --git a/src/main/java/net/ladenthin/llama/args/NumaStrategy.java b/llama/src/main/java/net/ladenthin/llama/args/NumaStrategy.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/args/NumaStrategy.java
rename to llama/src/main/java/net/ladenthin/llama/args/NumaStrategy.java
diff --git a/src/main/java/net/ladenthin/llama/args/PoolingType.java b/llama/src/main/java/net/ladenthin/llama/args/PoolingType.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/args/PoolingType.java
rename to llama/src/main/java/net/ladenthin/llama/args/PoolingType.java
diff --git a/src/main/java/net/ladenthin/llama/args/ReasoningFormat.java b/llama/src/main/java/net/ladenthin/llama/args/ReasoningFormat.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/args/ReasoningFormat.java
rename to llama/src/main/java/net/ladenthin/llama/args/ReasoningFormat.java
diff --git a/src/main/java/net/ladenthin/llama/args/RopeScalingType.java b/llama/src/main/java/net/ladenthin/llama/args/RopeScalingType.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/args/RopeScalingType.java
rename to llama/src/main/java/net/ladenthin/llama/args/RopeScalingType.java
diff --git a/src/main/java/net/ladenthin/llama/args/Sampler.java b/llama/src/main/java/net/ladenthin/llama/args/Sampler.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/args/Sampler.java
rename to llama/src/main/java/net/ladenthin/llama/args/Sampler.java
diff --git a/src/main/java/net/ladenthin/llama/args/package-info.java b/llama/src/main/java/net/ladenthin/llama/args/package-info.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/args/package-info.java
rename to llama/src/main/java/net/ladenthin/llama/args/package-info.java
diff --git a/src/main/java/net/ladenthin/llama/callback/CancellationToken.java b/llama/src/main/java/net/ladenthin/llama/callback/CancellationToken.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/callback/CancellationToken.java
rename to llama/src/main/java/net/ladenthin/llama/callback/CancellationToken.java
diff --git a/src/main/java/net/ladenthin/llama/callback/LoadProgressCallback.java b/llama/src/main/java/net/ladenthin/llama/callback/LoadProgressCallback.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/callback/LoadProgressCallback.java
rename to llama/src/main/java/net/ladenthin/llama/callback/LoadProgressCallback.java
diff --git a/src/main/java/net/ladenthin/llama/callback/ToolHandler.java b/llama/src/main/java/net/ladenthin/llama/callback/ToolHandler.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/callback/ToolHandler.java
rename to llama/src/main/java/net/ladenthin/llama/callback/ToolHandler.java
diff --git a/src/main/java/net/ladenthin/llama/callback/package-info.java b/llama/src/main/java/net/ladenthin/llama/callback/package-info.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/callback/package-info.java
rename to llama/src/main/java/net/ladenthin/llama/callback/package-info.java
diff --git a/src/main/java/net/ladenthin/llama/exception/LlamaException.java b/llama/src/main/java/net/ladenthin/llama/exception/LlamaException.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/exception/LlamaException.java
rename to llama/src/main/java/net/ladenthin/llama/exception/LlamaException.java
diff --git a/src/main/java/net/ladenthin/llama/exception/ModelUnavailableException.java b/llama/src/main/java/net/ladenthin/llama/exception/ModelUnavailableException.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/exception/ModelUnavailableException.java
rename to llama/src/main/java/net/ladenthin/llama/exception/ModelUnavailableException.java
diff --git a/src/main/java/net/ladenthin/llama/exception/package-info.java b/llama/src/main/java/net/ladenthin/llama/exception/package-info.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/exception/package-info.java
rename to llama/src/main/java/net/ladenthin/llama/exception/package-info.java
diff --git a/src/main/java/net/ladenthin/llama/json/ChatResponseParser.java b/llama/src/main/java/net/ladenthin/llama/json/ChatResponseParser.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/json/ChatResponseParser.java
rename to llama/src/main/java/net/ladenthin/llama/json/ChatResponseParser.java
diff --git a/src/main/java/net/ladenthin/llama/json/ChatStreamChunkParser.java b/llama/src/main/java/net/ladenthin/llama/json/ChatStreamChunkParser.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/json/ChatStreamChunkParser.java
rename to llama/src/main/java/net/ladenthin/llama/json/ChatStreamChunkParser.java
diff --git a/src/main/java/net/ladenthin/llama/json/CompletionResponseParser.java b/llama/src/main/java/net/ladenthin/llama/json/CompletionResponseParser.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/json/CompletionResponseParser.java
rename to llama/src/main/java/net/ladenthin/llama/json/CompletionResponseParser.java
diff --git a/src/main/java/net/ladenthin/llama/json/RerankResponseParser.java b/llama/src/main/java/net/ladenthin/llama/json/RerankResponseParser.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/json/RerankResponseParser.java
rename to llama/src/main/java/net/ladenthin/llama/json/RerankResponseParser.java
diff --git a/src/main/java/net/ladenthin/llama/json/TimingsLogger.java b/llama/src/main/java/net/ladenthin/llama/json/TimingsLogger.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/json/TimingsLogger.java
rename to llama/src/main/java/net/ladenthin/llama/json/TimingsLogger.java
diff --git a/src/main/java/net/ladenthin/llama/json/package-info.java b/llama/src/main/java/net/ladenthin/llama/json/package-info.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/json/package-info.java
rename to llama/src/main/java/net/ladenthin/llama/json/package-info.java
diff --git a/src/main/java/net/ladenthin/llama/loader/Java8CompatibilityHelper.java b/llama/src/main/java/net/ladenthin/llama/loader/Java8CompatibilityHelper.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/loader/Java8CompatibilityHelper.java
rename to llama/src/main/java/net/ladenthin/llama/loader/Java8CompatibilityHelper.java
diff --git a/src/main/java/net/ladenthin/llama/loader/LlamaLoader.java b/llama/src/main/java/net/ladenthin/llama/loader/LlamaLoader.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/loader/LlamaLoader.java
rename to llama/src/main/java/net/ladenthin/llama/loader/LlamaLoader.java
diff --git a/src/main/java/net/ladenthin/llama/loader/LlamaSystemProperties.java b/llama/src/main/java/net/ladenthin/llama/loader/LlamaSystemProperties.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/loader/LlamaSystemProperties.java
rename to llama/src/main/java/net/ladenthin/llama/loader/LlamaSystemProperties.java
diff --git a/src/main/java/net/ladenthin/llama/loader/NativeLibraryPermissionSetter.java b/llama/src/main/java/net/ladenthin/llama/loader/NativeLibraryPermissionSetter.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/loader/NativeLibraryPermissionSetter.java
rename to llama/src/main/java/net/ladenthin/llama/loader/NativeLibraryPermissionSetter.java
diff --git a/src/main/java/net/ladenthin/llama/loader/OSInfo.java b/llama/src/main/java/net/ladenthin/llama/loader/OSInfo.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/loader/OSInfo.java
rename to llama/src/main/java/net/ladenthin/llama/loader/OSInfo.java
diff --git a/src/main/java/net/ladenthin/llama/loader/OfflineModelGuard.java b/llama/src/main/java/net/ladenthin/llama/loader/OfflineModelGuard.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/loader/OfflineModelGuard.java
rename to llama/src/main/java/net/ladenthin/llama/loader/OfflineModelGuard.java
diff --git a/src/main/java/net/ladenthin/llama/loader/ProcessRunner.java b/llama/src/main/java/net/ladenthin/llama/loader/ProcessRunner.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/loader/ProcessRunner.java
rename to llama/src/main/java/net/ladenthin/llama/loader/ProcessRunner.java
diff --git a/src/main/java/net/ladenthin/llama/loader/package-info.java b/llama/src/main/java/net/ladenthin/llama/loader/package-info.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/loader/package-info.java
rename to llama/src/main/java/net/ladenthin/llama/loader/package-info.java
diff --git a/src/main/java/net/ladenthin/llama/package-info.java b/llama/src/main/java/net/ladenthin/llama/package-info.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/package-info.java
rename to llama/src/main/java/net/ladenthin/llama/package-info.java
diff --git a/src/main/java/net/ladenthin/llama/parameters/ChatRequest.java b/llama/src/main/java/net/ladenthin/llama/parameters/ChatRequest.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/parameters/ChatRequest.java
rename to llama/src/main/java/net/ladenthin/llama/parameters/ChatRequest.java
diff --git a/src/main/java/net/ladenthin/llama/parameters/CliParameters.java b/llama/src/main/java/net/ladenthin/llama/parameters/CliParameters.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/parameters/CliParameters.java
rename to llama/src/main/java/net/ladenthin/llama/parameters/CliParameters.java
diff --git a/src/main/java/net/ladenthin/llama/parameters/InferenceParameters.java b/llama/src/main/java/net/ladenthin/llama/parameters/InferenceParameters.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/parameters/InferenceParameters.java
rename to llama/src/main/java/net/ladenthin/llama/parameters/InferenceParameters.java
diff --git a/src/main/java/net/ladenthin/llama/parameters/JsonParameters.java b/llama/src/main/java/net/ladenthin/llama/parameters/JsonParameters.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/parameters/JsonParameters.java
rename to llama/src/main/java/net/ladenthin/llama/parameters/JsonParameters.java
diff --git a/src/main/java/net/ladenthin/llama/parameters/ModelParameters.java b/llama/src/main/java/net/ladenthin/llama/parameters/ModelParameters.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/parameters/ModelParameters.java
rename to llama/src/main/java/net/ladenthin/llama/parameters/ModelParameters.java
diff --git a/src/main/java/net/ladenthin/llama/parameters/ParameterJsonSerializer.java b/llama/src/main/java/net/ladenthin/llama/parameters/ParameterJsonSerializer.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/parameters/ParameterJsonSerializer.java
rename to llama/src/main/java/net/ladenthin/llama/parameters/ParameterJsonSerializer.java
diff --git a/src/main/java/net/ladenthin/llama/parameters/package-info.java b/llama/src/main/java/net/ladenthin/llama/parameters/package-info.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/parameters/package-info.java
rename to llama/src/main/java/net/ladenthin/llama/parameters/package-info.java
diff --git a/src/main/java/net/ladenthin/llama/server/AnthropicApiSupport.java b/llama/src/main/java/net/ladenthin/llama/server/AnthropicApiSupport.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/server/AnthropicApiSupport.java
rename to llama/src/main/java/net/ladenthin/llama/server/AnthropicApiSupport.java
diff --git a/src/main/java/net/ladenthin/llama/server/AnthropicStreamTranslator.java b/llama/src/main/java/net/ladenthin/llama/server/AnthropicStreamTranslator.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/server/AnthropicStreamTranslator.java
rename to llama/src/main/java/net/ladenthin/llama/server/AnthropicStreamTranslator.java
diff --git a/src/main/java/net/ladenthin/llama/server/ChunkSink.java b/llama/src/main/java/net/ladenthin/llama/server/ChunkSink.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/server/ChunkSink.java
rename to llama/src/main/java/net/ladenthin/llama/server/ChunkSink.java
diff --git a/src/main/java/net/ladenthin/llama/server/LlamaModelBackend.java b/llama/src/main/java/net/ladenthin/llama/server/LlamaModelBackend.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/server/LlamaModelBackend.java
rename to llama/src/main/java/net/ladenthin/llama/server/LlamaModelBackend.java
diff --git a/src/main/java/net/ladenthin/llama/server/NativeServer.java b/llama/src/main/java/net/ladenthin/llama/server/NativeServer.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/server/NativeServer.java
rename to llama/src/main/java/net/ladenthin/llama/server/NativeServer.java
diff --git a/src/main/java/net/ladenthin/llama/server/OaiRerankSupport.java b/llama/src/main/java/net/ladenthin/llama/server/OaiRerankSupport.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/server/OaiRerankSupport.java
rename to llama/src/main/java/net/ladenthin/llama/server/OaiRerankSupport.java
diff --git a/src/main/java/net/ladenthin/llama/server/OllamaApiSupport.java b/llama/src/main/java/net/ladenthin/llama/server/OllamaApiSupport.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/server/OllamaApiSupport.java
rename to llama/src/main/java/net/ladenthin/llama/server/OllamaApiSupport.java
diff --git a/src/main/java/net/ladenthin/llama/server/OpenAiBackend.java b/llama/src/main/java/net/ladenthin/llama/server/OpenAiBackend.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/server/OpenAiBackend.java
rename to llama/src/main/java/net/ladenthin/llama/server/OpenAiBackend.java
diff --git a/src/main/java/net/ladenthin/llama/server/OpenAiCompatServer.java b/llama/src/main/java/net/ladenthin/llama/server/OpenAiCompatServer.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/server/OpenAiCompatServer.java
rename to llama/src/main/java/net/ladenthin/llama/server/OpenAiCompatServer.java
diff --git a/src/main/java/net/ladenthin/llama/server/OpenAiRequestMapper.java b/llama/src/main/java/net/ladenthin/llama/server/OpenAiRequestMapper.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/server/OpenAiRequestMapper.java
rename to llama/src/main/java/net/ladenthin/llama/server/OpenAiRequestMapper.java
diff --git a/src/main/java/net/ladenthin/llama/server/OpenAiServerCli.java b/llama/src/main/java/net/ladenthin/llama/server/OpenAiServerCli.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/server/OpenAiServerCli.java
rename to llama/src/main/java/net/ladenthin/llama/server/OpenAiServerCli.java
diff --git a/src/main/java/net/ladenthin/llama/server/OpenAiServerConfig.java b/llama/src/main/java/net/ladenthin/llama/server/OpenAiServerConfig.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/server/OpenAiServerConfig.java
rename to llama/src/main/java/net/ladenthin/llama/server/OpenAiServerConfig.java
diff --git a/src/main/java/net/ladenthin/llama/server/OpenAiSseFormatter.java b/llama/src/main/java/net/ladenthin/llama/server/OpenAiSseFormatter.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/server/OpenAiSseFormatter.java
rename to llama/src/main/java/net/ladenthin/llama/server/OpenAiSseFormatter.java
diff --git a/src/main/java/net/ladenthin/llama/server/ResponsesApiSupport.java b/llama/src/main/java/net/ladenthin/llama/server/ResponsesApiSupport.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/server/ResponsesApiSupport.java
rename to llama/src/main/java/net/ladenthin/llama/server/ResponsesApiSupport.java
diff --git a/src/main/java/net/ladenthin/llama/server/ResponsesStreamTranslator.java b/llama/src/main/java/net/ladenthin/llama/server/ResponsesStreamTranslator.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/server/ResponsesStreamTranslator.java
rename to llama/src/main/java/net/ladenthin/llama/server/ResponsesStreamTranslator.java
diff --git a/src/main/java/net/ladenthin/llama/server/ToolCallDeltaAccumulator.java b/llama/src/main/java/net/ladenthin/llama/server/ToolCallDeltaAccumulator.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/server/ToolCallDeltaAccumulator.java
rename to llama/src/main/java/net/ladenthin/llama/server/ToolCallDeltaAccumulator.java
diff --git a/src/main/java/net/ladenthin/llama/server/package-info.java b/llama/src/main/java/net/ladenthin/llama/server/package-info.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/server/package-info.java
rename to llama/src/main/java/net/ladenthin/llama/server/package-info.java
diff --git a/src/main/java/net/ladenthin/llama/value/ChatChoice.java b/llama/src/main/java/net/ladenthin/llama/value/ChatChoice.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/ChatChoice.java
rename to llama/src/main/java/net/ladenthin/llama/value/ChatChoice.java
diff --git a/src/main/java/net/ladenthin/llama/value/ChatMessage.java b/llama/src/main/java/net/ladenthin/llama/value/ChatMessage.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/ChatMessage.java
rename to llama/src/main/java/net/ladenthin/llama/value/ChatMessage.java
diff --git a/src/main/java/net/ladenthin/llama/value/ChatResponse.java b/llama/src/main/java/net/ladenthin/llama/value/ChatResponse.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/ChatResponse.java
rename to llama/src/main/java/net/ladenthin/llama/value/ChatResponse.java
diff --git a/src/main/java/net/ladenthin/llama/value/ChatTranscript.java b/llama/src/main/java/net/ladenthin/llama/value/ChatTranscript.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/ChatTranscript.java
rename to llama/src/main/java/net/ladenthin/llama/value/ChatTranscript.java
diff --git a/src/main/java/net/ladenthin/llama/value/CompletionResult.java b/llama/src/main/java/net/ladenthin/llama/value/CompletionResult.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/CompletionResult.java
rename to llama/src/main/java/net/ladenthin/llama/value/CompletionResult.java
diff --git a/src/main/java/net/ladenthin/llama/value/ContentPart.java b/llama/src/main/java/net/ladenthin/llama/value/ContentPart.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/ContentPart.java
rename to llama/src/main/java/net/ladenthin/llama/value/ContentPart.java
diff --git a/src/main/java/net/ladenthin/llama/value/LlamaOutput.java b/llama/src/main/java/net/ladenthin/llama/value/LlamaOutput.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/LlamaOutput.java
rename to llama/src/main/java/net/ladenthin/llama/value/LlamaOutput.java
diff --git a/src/main/java/net/ladenthin/llama/value/LogLevel.java b/llama/src/main/java/net/ladenthin/llama/value/LogLevel.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/LogLevel.java
rename to llama/src/main/java/net/ladenthin/llama/value/LogLevel.java
diff --git a/src/main/java/net/ladenthin/llama/value/ModelMeta.java b/llama/src/main/java/net/ladenthin/llama/value/ModelMeta.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/ModelMeta.java
rename to llama/src/main/java/net/ladenthin/llama/value/ModelMeta.java
diff --git a/src/main/java/net/ladenthin/llama/value/Pair.java b/llama/src/main/java/net/ladenthin/llama/value/Pair.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/Pair.java
rename to llama/src/main/java/net/ladenthin/llama/value/Pair.java
diff --git a/src/main/java/net/ladenthin/llama/value/ServerMetrics.java b/llama/src/main/java/net/ladenthin/llama/value/ServerMetrics.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/ServerMetrics.java
rename to llama/src/main/java/net/ladenthin/llama/value/ServerMetrics.java
diff --git a/src/main/java/net/ladenthin/llama/value/SlotMetrics.java b/llama/src/main/java/net/ladenthin/llama/value/SlotMetrics.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/SlotMetrics.java
rename to llama/src/main/java/net/ladenthin/llama/value/SlotMetrics.java
diff --git a/src/main/java/net/ladenthin/llama/value/StopReason.java b/llama/src/main/java/net/ladenthin/llama/value/StopReason.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/StopReason.java
rename to llama/src/main/java/net/ladenthin/llama/value/StopReason.java
diff --git a/src/main/java/net/ladenthin/llama/value/Timings.java b/llama/src/main/java/net/ladenthin/llama/value/Timings.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/Timings.java
rename to llama/src/main/java/net/ladenthin/llama/value/Timings.java
diff --git a/src/main/java/net/ladenthin/llama/value/TokenLogprob.java b/llama/src/main/java/net/ladenthin/llama/value/TokenLogprob.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/TokenLogprob.java
rename to llama/src/main/java/net/ladenthin/llama/value/TokenLogprob.java
diff --git a/src/main/java/net/ladenthin/llama/value/ToolCall.java b/llama/src/main/java/net/ladenthin/llama/value/ToolCall.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/ToolCall.java
rename to llama/src/main/java/net/ladenthin/llama/value/ToolCall.java
diff --git a/src/main/java/net/ladenthin/llama/value/ToolDefinition.java b/llama/src/main/java/net/ladenthin/llama/value/ToolDefinition.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/ToolDefinition.java
rename to llama/src/main/java/net/ladenthin/llama/value/ToolDefinition.java
diff --git a/src/main/java/net/ladenthin/llama/value/Usage.java b/llama/src/main/java/net/ladenthin/llama/value/Usage.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/Usage.java
rename to llama/src/main/java/net/ladenthin/llama/value/Usage.java
diff --git a/src/main/java/net/ladenthin/llama/value/package-info.java b/llama/src/main/java/net/ladenthin/llama/value/package-info.java
similarity index 100%
rename from src/main/java/net/ladenthin/llama/value/package-info.java
rename to llama/src/main/java/net/ladenthin/llama/value/package-info.java
diff --git a/src/test/cpp/test_jni_helpers.cpp b/llama/src/test/cpp/test_jni_helpers.cpp
similarity index 100%
rename from src/test/cpp/test_jni_helpers.cpp
rename to llama/src/test/cpp/test_jni_helpers.cpp
diff --git a/src/test/cpp/test_json_helpers.cpp b/llama/src/test/cpp/test_json_helpers.cpp
similarity index 100%
rename from src/test/cpp/test_json_helpers.cpp
rename to llama/src/test/cpp/test_json_helpers.cpp
diff --git a/src/test/cpp/test_log_helpers.cpp b/llama/src/test/cpp/test_log_helpers.cpp
similarity index 100%
rename from src/test/cpp/test_log_helpers.cpp
rename to llama/src/test/cpp/test_log_helpers.cpp
diff --git a/src/test/cpp/test_server.cpp b/llama/src/test/cpp/test_server.cpp
similarity index 100%
rename from src/test/cpp/test_server.cpp
rename to llama/src/test/cpp/test_server.cpp
diff --git a/src/test/cpp/test_tts_wav.cpp b/llama/src/test/cpp/test_tts_wav.cpp
similarity index 100%
rename from src/test/cpp/test_tts_wav.cpp
rename to llama/src/test/cpp/test_tts_wav.cpp
diff --git a/src/test/cpp/test_utils.cpp b/llama/src/test/cpp/test_utils.cpp
similarity index 100%
rename from src/test/cpp/test_utils.cpp
rename to llama/src/test/cpp/test_utils.cpp
diff --git a/src/test/java/examples/ChatExample.java b/llama/src/test/java/examples/ChatExample.java
similarity index 100%
rename from src/test/java/examples/ChatExample.java
rename to llama/src/test/java/examples/ChatExample.java
diff --git a/src/test/java/examples/GrammarExample.java b/llama/src/test/java/examples/GrammarExample.java
similarity index 100%
rename from src/test/java/examples/GrammarExample.java
rename to llama/src/test/java/examples/GrammarExample.java
diff --git a/src/test/java/examples/InfillExample.java b/llama/src/test/java/examples/InfillExample.java
similarity index 100%
rename from src/test/java/examples/InfillExample.java
rename to llama/src/test/java/examples/InfillExample.java
diff --git a/src/test/java/examples/MainExample.java b/llama/src/test/java/examples/MainExample.java
similarity index 100%
rename from src/test/java/examples/MainExample.java
rename to llama/src/test/java/examples/MainExample.java
diff --git a/src/test/java/examples/OpenAiServerExample.java b/llama/src/test/java/examples/OpenAiServerExample.java
similarity index 100%
rename from src/test/java/examples/OpenAiServerExample.java
rename to llama/src/test/java/examples/OpenAiServerExample.java
diff --git a/src/test/java/net/ladenthin/llama/AudioInputIntegrationTest.java b/llama/src/test/java/net/ladenthin/llama/AudioInputIntegrationTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/AudioInputIntegrationTest.java
rename to llama/src/test/java/net/ladenthin/llama/AudioInputIntegrationTest.java
diff --git a/src/test/java/net/ladenthin/llama/ChatAdvancedTest.java b/llama/src/test/java/net/ladenthin/llama/ChatAdvancedTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/ChatAdvancedTest.java
rename to llama/src/test/java/net/ladenthin/llama/ChatAdvancedTest.java
diff --git a/src/test/java/net/ladenthin/llama/ChatScenarioTest.java b/llama/src/test/java/net/ladenthin/llama/ChatScenarioTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/ChatScenarioTest.java
rename to llama/src/test/java/net/ladenthin/llama/ChatScenarioTest.java
diff --git a/src/test/java/net/ladenthin/llama/ClaudeGenerated.java b/llama/src/test/java/net/ladenthin/llama/ClaudeGenerated.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/ClaudeGenerated.java
rename to llama/src/test/java/net/ladenthin/llama/ClaudeGenerated.java
diff --git a/src/test/java/net/ladenthin/llama/ConfigureParallelInferenceTest.java b/llama/src/test/java/net/ladenthin/llama/ConfigureParallelInferenceTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/ConfigureParallelInferenceTest.java
rename to llama/src/test/java/net/ladenthin/llama/ConfigureParallelInferenceTest.java
diff --git a/src/test/java/net/ladenthin/llama/ErrorHandlingTest.java b/llama/src/test/java/net/ladenthin/llama/ErrorHandlingTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/ErrorHandlingTest.java
rename to llama/src/test/java/net/ladenthin/llama/ErrorHandlingTest.java
diff --git a/src/test/java/net/ladenthin/llama/LlamaArchitectureTest.java b/llama/src/test/java/net/ladenthin/llama/LlamaArchitectureTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/LlamaArchitectureTest.java
rename to llama/src/test/java/net/ladenthin/llama/LlamaArchitectureTest.java
diff --git a/src/test/java/net/ladenthin/llama/LlamaEmbeddingsTest.java b/llama/src/test/java/net/ladenthin/llama/LlamaEmbeddingsTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/LlamaEmbeddingsTest.java
rename to llama/src/test/java/net/ladenthin/llama/LlamaEmbeddingsTest.java
diff --git a/src/test/java/net/ladenthin/llama/LlamaModelOfflineTest.java b/llama/src/test/java/net/ladenthin/llama/LlamaModelOfflineTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/LlamaModelOfflineTest.java
rename to llama/src/test/java/net/ladenthin/llama/LlamaModelOfflineTest.java
diff --git a/src/test/java/net/ladenthin/llama/LlamaModelTest.java b/llama/src/test/java/net/ladenthin/llama/LlamaModelTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/LlamaModelTest.java
rename to llama/src/test/java/net/ladenthin/llama/LlamaModelTest.java
diff --git a/src/test/java/net/ladenthin/llama/LlamaParameterProperties.java b/llama/src/test/java/net/ladenthin/llama/LlamaParameterProperties.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/LlamaParameterProperties.java
rename to llama/src/test/java/net/ladenthin/llama/LlamaParameterProperties.java
diff --git a/src/test/java/net/ladenthin/llama/MemoryManagementTest.java b/llama/src/test/java/net/ladenthin/llama/MemoryManagementTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/MemoryManagementTest.java
rename to llama/src/test/java/net/ladenthin/llama/MemoryManagementTest.java
diff --git a/src/test/java/net/ladenthin/llama/MultimodalIntegrationTest.java b/llama/src/test/java/net/ladenthin/llama/MultimodalIntegrationTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/MultimodalIntegrationTest.java
rename to llama/src/test/java/net/ladenthin/llama/MultimodalIntegrationTest.java
diff --git a/src/test/java/net/ladenthin/llama/MultimodalMessagesTest.java b/llama/src/test/java/net/ladenthin/llama/MultimodalMessagesTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/MultimodalMessagesTest.java
rename to llama/src/test/java/net/ladenthin/llama/MultimodalMessagesTest.java
diff --git a/src/test/java/net/ladenthin/llama/ReactorIntegrationTest.java b/llama/src/test/java/net/ladenthin/llama/ReactorIntegrationTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/ReactorIntegrationTest.java
rename to llama/src/test/java/net/ladenthin/llama/ReactorIntegrationTest.java
diff --git a/src/test/java/net/ladenthin/llama/ReasoningBudgetTest.java b/llama/src/test/java/net/ladenthin/llama/ReasoningBudgetTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/ReasoningBudgetTest.java
rename to llama/src/test/java/net/ladenthin/llama/ReasoningBudgetTest.java
diff --git a/src/test/java/net/ladenthin/llama/RerankingModelTest.java b/llama/src/test/java/net/ladenthin/llama/RerankingModelTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/RerankingModelTest.java
rename to llama/src/test/java/net/ladenthin/llama/RerankingModelTest.java
diff --git a/src/test/java/net/ladenthin/llama/ResponseJsonStructureTest.java b/llama/src/test/java/net/ladenthin/llama/ResponseJsonStructureTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/ResponseJsonStructureTest.java
rename to llama/src/test/java/net/ladenthin/llama/ResponseJsonStructureTest.java
diff --git a/src/test/java/net/ladenthin/llama/SessionConcurrencyTest.java b/llama/src/test/java/net/ladenthin/llama/SessionConcurrencyTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/SessionConcurrencyTest.java
rename to llama/src/test/java/net/ladenthin/llama/SessionConcurrencyTest.java
diff --git a/src/test/java/net/ladenthin/llama/SessionStateTest.java b/llama/src/test/java/net/ladenthin/llama/SessionStateTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/SessionStateTest.java
rename to llama/src/test/java/net/ladenthin/llama/SessionStateTest.java
diff --git a/src/test/java/net/ladenthin/llama/SessionTest.java b/llama/src/test/java/net/ladenthin/llama/SessionTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/SessionTest.java
rename to llama/src/test/java/net/ladenthin/llama/SessionTest.java
diff --git a/src/test/java/net/ladenthin/llama/TestConstants.java b/llama/src/test/java/net/ladenthin/llama/TestConstants.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/TestConstants.java
rename to llama/src/test/java/net/ladenthin/llama/TestConstants.java
diff --git a/src/test/java/net/ladenthin/llama/ToolCallingAgentTest.java b/llama/src/test/java/net/ladenthin/llama/ToolCallingAgentTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/ToolCallingAgentTest.java
rename to llama/src/test/java/net/ladenthin/llama/ToolCallingAgentTest.java
diff --git a/src/test/java/net/ladenthin/llama/ToolCallingIntegrationTest.java b/llama/src/test/java/net/ladenthin/llama/ToolCallingIntegrationTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/ToolCallingIntegrationTest.java
rename to llama/src/test/java/net/ladenthin/llama/ToolCallingIntegrationTest.java
diff --git a/src/test/java/net/ladenthin/llama/TtsIntegrationTest.java b/llama/src/test/java/net/ladenthin/llama/TtsIntegrationTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/TtsIntegrationTest.java
rename to llama/src/test/java/net/ladenthin/llama/TtsIntegrationTest.java
diff --git a/src/test/java/net/ladenthin/llama/args/AbstractCliArgEnumTest.java b/llama/src/test/java/net/ladenthin/llama/args/AbstractCliArgEnumTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/args/AbstractCliArgEnumTest.java
rename to llama/src/test/java/net/ladenthin/llama/args/AbstractCliArgEnumTest.java
diff --git a/src/test/java/net/ladenthin/llama/args/CacheTypeTest.java b/llama/src/test/java/net/ladenthin/llama/args/CacheTypeTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/args/CacheTypeTest.java
rename to llama/src/test/java/net/ladenthin/llama/args/CacheTypeTest.java
diff --git a/src/test/java/net/ladenthin/llama/args/ContinuationModeTest.java b/llama/src/test/java/net/ladenthin/llama/args/ContinuationModeTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/args/ContinuationModeTest.java
rename to llama/src/test/java/net/ladenthin/llama/args/ContinuationModeTest.java
diff --git a/src/test/java/net/ladenthin/llama/args/GpuSplitModeTest.java b/llama/src/test/java/net/ladenthin/llama/args/GpuSplitModeTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/args/GpuSplitModeTest.java
rename to llama/src/test/java/net/ladenthin/llama/args/GpuSplitModeTest.java
diff --git a/src/test/java/net/ladenthin/llama/args/LogFormatTest.java b/llama/src/test/java/net/ladenthin/llama/args/LogFormatTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/args/LogFormatTest.java
rename to llama/src/test/java/net/ladenthin/llama/args/LogFormatTest.java
diff --git a/src/test/java/net/ladenthin/llama/args/MiroStatTest.java b/llama/src/test/java/net/ladenthin/llama/args/MiroStatTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/args/MiroStatTest.java
rename to llama/src/test/java/net/ladenthin/llama/args/MiroStatTest.java
diff --git a/src/test/java/net/ladenthin/llama/args/ModelFlagTest.java b/llama/src/test/java/net/ladenthin/llama/args/ModelFlagTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/args/ModelFlagTest.java
rename to llama/src/test/java/net/ladenthin/llama/args/ModelFlagTest.java
diff --git a/src/test/java/net/ladenthin/llama/args/NumaStrategyTest.java b/llama/src/test/java/net/ladenthin/llama/args/NumaStrategyTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/args/NumaStrategyTest.java
rename to llama/src/test/java/net/ladenthin/llama/args/NumaStrategyTest.java
diff --git a/src/test/java/net/ladenthin/llama/args/PoolingTypeTest.java b/llama/src/test/java/net/ladenthin/llama/args/PoolingTypeTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/args/PoolingTypeTest.java
rename to llama/src/test/java/net/ladenthin/llama/args/PoolingTypeTest.java
diff --git a/src/test/java/net/ladenthin/llama/args/ReasoningFormatTest.java b/llama/src/test/java/net/ladenthin/llama/args/ReasoningFormatTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/args/ReasoningFormatTest.java
rename to llama/src/test/java/net/ladenthin/llama/args/ReasoningFormatTest.java
diff --git a/src/test/java/net/ladenthin/llama/args/RopeScalingTypeTest.java b/llama/src/test/java/net/ladenthin/llama/args/RopeScalingTypeTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/args/RopeScalingTypeTest.java
rename to llama/src/test/java/net/ladenthin/llama/args/RopeScalingTypeTest.java
diff --git a/src/test/java/net/ladenthin/llama/args/SamplerTest.java b/llama/src/test/java/net/ladenthin/llama/args/SamplerTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/args/SamplerTest.java
rename to llama/src/test/java/net/ladenthin/llama/args/SamplerTest.java
diff --git a/src/test/java/net/ladenthin/llama/benchmark/InferenceParametersBenchmark.java b/llama/src/test/java/net/ladenthin/llama/benchmark/InferenceParametersBenchmark.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/benchmark/InferenceParametersBenchmark.java
rename to llama/src/test/java/net/ladenthin/llama/benchmark/InferenceParametersBenchmark.java
diff --git a/src/test/java/net/ladenthin/llama/callback/CancellationTokenLincheckTest.java b/llama/src/test/java/net/ladenthin/llama/callback/CancellationTokenLincheckTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/callback/CancellationTokenLincheckTest.java
rename to llama/src/test/java/net/ladenthin/llama/callback/CancellationTokenLincheckTest.java
diff --git a/src/test/java/net/ladenthin/llama/callback/CancellationTokenTest.java b/llama/src/test/java/net/ladenthin/llama/callback/CancellationTokenTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/callback/CancellationTokenTest.java
rename to llama/src/test/java/net/ladenthin/llama/callback/CancellationTokenTest.java
diff --git a/src/test/java/net/ladenthin/llama/callback/LoadProgressCallbackTest.java b/llama/src/test/java/net/ladenthin/llama/callback/LoadProgressCallbackTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/callback/LoadProgressCallbackTest.java
rename to llama/src/test/java/net/ladenthin/llama/callback/LoadProgressCallbackTest.java
diff --git a/src/test/java/net/ladenthin/llama/exception/LlamaExceptionTest.java b/llama/src/test/java/net/ladenthin/llama/exception/LlamaExceptionTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/exception/LlamaExceptionTest.java
rename to llama/src/test/java/net/ladenthin/llama/exception/LlamaExceptionTest.java
diff --git a/src/test/java/net/ladenthin/llama/exception/ModelUnavailableExceptionTest.java b/llama/src/test/java/net/ladenthin/llama/exception/ModelUnavailableExceptionTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/exception/ModelUnavailableExceptionTest.java
rename to llama/src/test/java/net/ladenthin/llama/exception/ModelUnavailableExceptionTest.java
diff --git a/src/test/java/net/ladenthin/llama/jcstress/CancellationTokenRace.java b/llama/src/test/java/net/ladenthin/llama/jcstress/CancellationTokenRace.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/jcstress/CancellationTokenRace.java
rename to llama/src/test/java/net/ladenthin/llama/jcstress/CancellationTokenRace.java
diff --git a/src/test/java/net/ladenthin/llama/json/ChatResponseParserTest.java b/llama/src/test/java/net/ladenthin/llama/json/ChatResponseParserTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/json/ChatResponseParserTest.java
rename to llama/src/test/java/net/ladenthin/llama/json/ChatResponseParserTest.java
diff --git a/src/test/java/net/ladenthin/llama/json/ChatStreamChunkParserTest.java b/llama/src/test/java/net/ladenthin/llama/json/ChatStreamChunkParserTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/json/ChatStreamChunkParserTest.java
rename to llama/src/test/java/net/ladenthin/llama/json/ChatStreamChunkParserTest.java
diff --git a/src/test/java/net/ladenthin/llama/json/CompletionResponseParserTest.java b/llama/src/test/java/net/ladenthin/llama/json/CompletionResponseParserTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/json/CompletionResponseParserTest.java
rename to llama/src/test/java/net/ladenthin/llama/json/CompletionResponseParserTest.java
diff --git a/src/test/java/net/ladenthin/llama/json/ParameterJsonSerializerTest.java b/llama/src/test/java/net/ladenthin/llama/json/ParameterJsonSerializerTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/json/ParameterJsonSerializerTest.java
rename to llama/src/test/java/net/ladenthin/llama/json/ParameterJsonSerializerTest.java
diff --git a/src/test/java/net/ladenthin/llama/json/RerankResponseParserTest.java b/llama/src/test/java/net/ladenthin/llama/json/RerankResponseParserTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/json/RerankResponseParserTest.java
rename to llama/src/test/java/net/ladenthin/llama/json/RerankResponseParserTest.java
diff --git a/src/test/java/net/ladenthin/llama/json/TimingsLoggerTest.java b/llama/src/test/java/net/ladenthin/llama/json/TimingsLoggerTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/json/TimingsLoggerTest.java
rename to llama/src/test/java/net/ladenthin/llama/json/TimingsLoggerTest.java
diff --git a/src/test/java/net/ladenthin/llama/loader/LlamaLoaderTest.java b/llama/src/test/java/net/ladenthin/llama/loader/LlamaLoaderTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/loader/LlamaLoaderTest.java
rename to llama/src/test/java/net/ladenthin/llama/loader/LlamaLoaderTest.java
diff --git a/src/test/java/net/ladenthin/llama/loader/LoggingSmokeTest.java b/llama/src/test/java/net/ladenthin/llama/loader/LoggingSmokeTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/loader/LoggingSmokeTest.java
rename to llama/src/test/java/net/ladenthin/llama/loader/LoggingSmokeTest.java
diff --git a/src/test/java/net/ladenthin/llama/loader/NativeLibraryLoadSmokeTest.java b/llama/src/test/java/net/ladenthin/llama/loader/NativeLibraryLoadSmokeTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/loader/NativeLibraryLoadSmokeTest.java
rename to llama/src/test/java/net/ladenthin/llama/loader/NativeLibraryLoadSmokeTest.java
diff --git a/src/test/java/net/ladenthin/llama/loader/NativeLibraryPermissionSetterTest.java b/llama/src/test/java/net/ladenthin/llama/loader/NativeLibraryPermissionSetterTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/loader/NativeLibraryPermissionSetterTest.java
rename to llama/src/test/java/net/ladenthin/llama/loader/NativeLibraryPermissionSetterTest.java
diff --git a/src/test/java/net/ladenthin/llama/loader/OSInfoTest.java b/llama/src/test/java/net/ladenthin/llama/loader/OSInfoTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/loader/OSInfoTest.java
rename to llama/src/test/java/net/ladenthin/llama/loader/OSInfoTest.java
diff --git a/src/test/java/net/ladenthin/llama/parameters/ChatRequestTest.java b/llama/src/test/java/net/ladenthin/llama/parameters/ChatRequestTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/parameters/ChatRequestTest.java
rename to llama/src/test/java/net/ladenthin/llama/parameters/ChatRequestTest.java
diff --git a/src/test/java/net/ladenthin/llama/parameters/InferenceParametersTest.java b/llama/src/test/java/net/ladenthin/llama/parameters/InferenceParametersTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/parameters/InferenceParametersTest.java
rename to llama/src/test/java/net/ladenthin/llama/parameters/InferenceParametersTest.java
diff --git a/src/test/java/net/ladenthin/llama/parameters/JsonEndpointParametersTest.java b/llama/src/test/java/net/ladenthin/llama/parameters/JsonEndpointParametersTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/parameters/JsonEndpointParametersTest.java
rename to llama/src/test/java/net/ladenthin/llama/parameters/JsonEndpointParametersTest.java
diff --git a/src/test/java/net/ladenthin/llama/parameters/JsonParametersTest.java b/llama/src/test/java/net/ladenthin/llama/parameters/JsonParametersTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/parameters/JsonParametersTest.java
rename to llama/src/test/java/net/ladenthin/llama/parameters/JsonParametersTest.java
diff --git a/src/test/java/net/ladenthin/llama/parameters/ModelParametersExtendedTest.java b/llama/src/test/java/net/ladenthin/llama/parameters/ModelParametersExtendedTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/parameters/ModelParametersExtendedTest.java
rename to llama/src/test/java/net/ladenthin/llama/parameters/ModelParametersExtendedTest.java
diff --git a/src/test/java/net/ladenthin/llama/parameters/ModelParametersTest.java b/llama/src/test/java/net/ladenthin/llama/parameters/ModelParametersTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/parameters/ModelParametersTest.java
rename to llama/src/test/java/net/ladenthin/llama/parameters/ModelParametersTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/AnthropicApiSupportTest.java b/llama/src/test/java/net/ladenthin/llama/server/AnthropicApiSupportTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/AnthropicApiSupportTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/AnthropicApiSupportTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/AnthropicStreamTranslatorTest.java b/llama/src/test/java/net/ladenthin/llama/server/AnthropicStreamTranslatorTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/AnthropicStreamTranslatorTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/AnthropicStreamTranslatorTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/NativeServerSmokeTest.java b/llama/src/test/java/net/ladenthin/llama/server/NativeServerSmokeTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/NativeServerSmokeTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/NativeServerSmokeTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/OaiRerankSupportTest.java b/llama/src/test/java/net/ladenthin/llama/server/OaiRerankSupportTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/OaiRerankSupportTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/OaiRerankSupportTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/OllamaApiSupportTest.java b/llama/src/test/java/net/ladenthin/llama/server/OllamaApiSupportTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/OllamaApiSupportTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/OllamaApiSupportTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/OpenAiCompatServerHttpTest.java b/llama/src/test/java/net/ladenthin/llama/server/OpenAiCompatServerHttpTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/OpenAiCompatServerHttpTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/OpenAiCompatServerHttpTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/OpenAiCompatServerIntegrationTest.java b/llama/src/test/java/net/ladenthin/llama/server/OpenAiCompatServerIntegrationTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/OpenAiCompatServerIntegrationTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/OpenAiCompatServerIntegrationTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/OpenAiRequestMapperTest.java b/llama/src/test/java/net/ladenthin/llama/server/OpenAiRequestMapperTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/OpenAiRequestMapperTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/OpenAiRequestMapperTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/OpenAiServerCliTest.java b/llama/src/test/java/net/ladenthin/llama/server/OpenAiServerCliTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/OpenAiServerCliTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/OpenAiServerCliTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/OpenAiServerCompletionIntegrationTest.java b/llama/src/test/java/net/ladenthin/llama/server/OpenAiServerCompletionIntegrationTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/OpenAiServerCompletionIntegrationTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/OpenAiServerCompletionIntegrationTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/OpenAiServerConfigTest.java b/llama/src/test/java/net/ladenthin/llama/server/OpenAiServerConfigTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/OpenAiServerConfigTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/OpenAiServerConfigTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/OpenAiServerEmbeddingsIntegrationTest.java b/llama/src/test/java/net/ladenthin/llama/server/OpenAiServerEmbeddingsIntegrationTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/OpenAiServerEmbeddingsIntegrationTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/OpenAiServerEmbeddingsIntegrationTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/OpenAiServerRerankIntegrationTest.java b/llama/src/test/java/net/ladenthin/llama/server/OpenAiServerRerankIntegrationTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/OpenAiServerRerankIntegrationTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/OpenAiServerRerankIntegrationTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/OpenAiServerTestSupport.java b/llama/src/test/java/net/ladenthin/llama/server/OpenAiServerTestSupport.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/OpenAiServerTestSupport.java
rename to llama/src/test/java/net/ladenthin/llama/server/OpenAiServerTestSupport.java
diff --git a/src/test/java/net/ladenthin/llama/server/OpenAiServerToolCallingIntegrationTest.java b/llama/src/test/java/net/ladenthin/llama/server/OpenAiServerToolCallingIntegrationTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/OpenAiServerToolCallingIntegrationTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/OpenAiServerToolCallingIntegrationTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/OpenAiSseFormatterTest.java b/llama/src/test/java/net/ladenthin/llama/server/OpenAiSseFormatterTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/OpenAiSseFormatterTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/OpenAiSseFormatterTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/ResponsesApiSupportTest.java b/llama/src/test/java/net/ladenthin/llama/server/ResponsesApiSupportTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/ResponsesApiSupportTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/ResponsesApiSupportTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/ResponsesStreamTranslatorTest.java b/llama/src/test/java/net/ladenthin/llama/server/ResponsesStreamTranslatorTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/ResponsesStreamTranslatorTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/ResponsesStreamTranslatorTest.java
diff --git a/src/test/java/net/ladenthin/llama/server/ToolCallDeltaAccumulatorTest.java b/llama/src/test/java/net/ladenthin/llama/server/ToolCallDeltaAccumulatorTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/server/ToolCallDeltaAccumulatorTest.java
rename to llama/src/test/java/net/ladenthin/llama/server/ToolCallDeltaAccumulatorTest.java
diff --git a/src/test/java/net/ladenthin/llama/value/ChatChoiceTest.java b/llama/src/test/java/net/ladenthin/llama/value/ChatChoiceTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/value/ChatChoiceTest.java
rename to llama/src/test/java/net/ladenthin/llama/value/ChatChoiceTest.java
diff --git a/src/test/java/net/ladenthin/llama/value/ChatMessageTest.java b/llama/src/test/java/net/ladenthin/llama/value/ChatMessageTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/value/ChatMessageTest.java
rename to llama/src/test/java/net/ladenthin/llama/value/ChatMessageTest.java
diff --git a/src/test/java/net/ladenthin/llama/value/ChatResponseTest.java b/llama/src/test/java/net/ladenthin/llama/value/ChatResponseTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/value/ChatResponseTest.java
rename to llama/src/test/java/net/ladenthin/llama/value/ChatResponseTest.java
diff --git a/src/test/java/net/ladenthin/llama/value/ChatTranscriptTest.java b/llama/src/test/java/net/ladenthin/llama/value/ChatTranscriptTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/value/ChatTranscriptTest.java
rename to llama/src/test/java/net/ladenthin/llama/value/ChatTranscriptTest.java
diff --git a/src/test/java/net/ladenthin/llama/value/CompletionResultTest.java b/llama/src/test/java/net/ladenthin/llama/value/CompletionResultTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/value/CompletionResultTest.java
rename to llama/src/test/java/net/ladenthin/llama/value/CompletionResultTest.java
diff --git a/src/test/java/net/ladenthin/llama/value/ContentPartTest.java b/llama/src/test/java/net/ladenthin/llama/value/ContentPartTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/value/ContentPartTest.java
rename to llama/src/test/java/net/ladenthin/llama/value/ContentPartTest.java
diff --git a/src/test/java/net/ladenthin/llama/value/LlamaOutputTest.java b/llama/src/test/java/net/ladenthin/llama/value/LlamaOutputTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/value/LlamaOutputTest.java
rename to llama/src/test/java/net/ladenthin/llama/value/LlamaOutputTest.java
diff --git a/src/test/java/net/ladenthin/llama/value/LogLevelTest.java b/llama/src/test/java/net/ladenthin/llama/value/LogLevelTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/value/LogLevelTest.java
rename to llama/src/test/java/net/ladenthin/llama/value/LogLevelTest.java
diff --git a/src/test/java/net/ladenthin/llama/value/ModelMetaTest.java b/llama/src/test/java/net/ladenthin/llama/value/ModelMetaTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/value/ModelMetaTest.java
rename to llama/src/test/java/net/ladenthin/llama/value/ModelMetaTest.java
diff --git a/src/test/java/net/ladenthin/llama/value/PairTest.java b/llama/src/test/java/net/ladenthin/llama/value/PairTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/value/PairTest.java
rename to llama/src/test/java/net/ladenthin/llama/value/PairTest.java
diff --git a/src/test/java/net/ladenthin/llama/value/ServerMetricsTest.java b/llama/src/test/java/net/ladenthin/llama/value/ServerMetricsTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/value/ServerMetricsTest.java
rename to llama/src/test/java/net/ladenthin/llama/value/ServerMetricsTest.java
diff --git a/src/test/java/net/ladenthin/llama/value/StopReasonTest.java b/llama/src/test/java/net/ladenthin/llama/value/StopReasonTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/value/StopReasonTest.java
rename to llama/src/test/java/net/ladenthin/llama/value/StopReasonTest.java
diff --git a/src/test/java/net/ladenthin/llama/value/TimingsTest.java b/llama/src/test/java/net/ladenthin/llama/value/TimingsTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/value/TimingsTest.java
rename to llama/src/test/java/net/ladenthin/llama/value/TimingsTest.java
diff --git a/src/test/java/net/ladenthin/llama/value/TokenLogprobTest.java b/llama/src/test/java/net/ladenthin/llama/value/TokenLogprobTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/value/TokenLogprobTest.java
rename to llama/src/test/java/net/ladenthin/llama/value/TokenLogprobTest.java
diff --git a/src/test/java/net/ladenthin/llama/value/ToolCallTest.java b/llama/src/test/java/net/ladenthin/llama/value/ToolCallTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/value/ToolCallTest.java
rename to llama/src/test/java/net/ladenthin/llama/value/ToolCallTest.java
diff --git a/src/test/java/net/ladenthin/llama/value/ToolDefinitionTest.java b/llama/src/test/java/net/ladenthin/llama/value/ToolDefinitionTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/value/ToolDefinitionTest.java
rename to llama/src/test/java/net/ladenthin/llama/value/ToolDefinitionTest.java
diff --git a/src/test/java/net/ladenthin/llama/value/UsageTest.java b/llama/src/test/java/net/ladenthin/llama/value/UsageTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/value/UsageTest.java
rename to llama/src/test/java/net/ladenthin/llama/value/UsageTest.java
diff --git a/src/test/java/net/ladenthin/llama/vmlens/SessionStateInterleavingTest.java b/llama/src/test/java/net/ladenthin/llama/vmlens/SessionStateInterleavingTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/vmlens/SessionStateInterleavingTest.java
rename to llama/src/test/java/net/ladenthin/llama/vmlens/SessionStateInterleavingTest.java
diff --git a/src/test/java/net/ladenthin/llama/vmlens/VmlensInterleavingSmokeTest.java b/llama/src/test/java/net/ladenthin/llama/vmlens/VmlensInterleavingSmokeTest.java
similarity index 100%
rename from src/test/java/net/ladenthin/llama/vmlens/VmlensInterleavingSmokeTest.java
rename to llama/src/test/java/net/ladenthin/llama/vmlens/VmlensInterleavingSmokeTest.java
diff --git a/src/test/resources/images/README.md b/llama/src/test/resources/images/README.md
similarity index 100%
rename from src/test/resources/images/README.md
rename to llama/src/test/resources/images/README.md
diff --git a/src/test/resources/images/test-image.jpg b/llama/src/test/resources/images/test-image.jpg
similarity index 100%
rename from src/test/resources/images/test-image.jpg
rename to llama/src/test/resources/images/test-image.jpg
diff --git a/pom.xml b/pom.xml
index d8c1492c..b136d4d8 100644
--- a/pom.xml
+++ b/pom.xml
@@ -10,14 +10,21 @@ SPDX-License-Identifier: MIT
 	xsi:schemaLocation="http://maven.apache.org/POM/4.0.0 http://maven.apache.org/xsd/maven-4.0.0.xsd">
 	<modelVersion>4.0.0</modelVersion>
 
+	<!--
+		Aggregator / parent POM. Carries no code — it only groups the modules and holds the
+		single shared <version>, which every module inherits (so the core net.ladenthin:llama
+		and net.ladenthin:llama-langchain4j always ship in lockstep by construction). The core
+		artifact keeps its coordinates net.ladenthin:llama (the `llama` module); consumers are
+		unaffected by the directory move.
+	-->
 	<groupId>net.ladenthin</groupId>
-	<artifactId>llama</artifactId>
+	<artifactId>llama-parent</artifactId>
 	<version>5.0.4-SNAPSHOT</version>
-	<packaging>jar</packaging>
+	<packaging>pom</packaging>
 
 	<name>${project.groupId}:${project.artifactId}</name>
-	<description>Java Bindings for llama.cpp - A Port of Facebook's LLaMA model
-		in C/C++.</description>
+	<description>Aggregator POM for java-llama.cpp: the native JNI binding
+		(net.ladenthin:llama) and its sibling integrations.</description>
 	<url>https://github.com/bernardladenthin/java-llama.cpp</url>
 
 	<licenses>
@@ -48,1302 +55,8 @@ SPDX-License-Identifier: MIT
 		</snapshotRepository>
 	</distributionManagement>
 
-	<properties>
-		<sonar.organization>bernardladenthin</sonar.organization>
-		<jna.version>5.19.0</jna.version>
-		<jspecify.version>1.0.0</jspecify.version>
-		<lombok.version>1.18.46</lombok.version>
-		<errorprone.version>2.50.0</errorprone.version>
-		<nullaway.version>0.13.7</nullaway.version>
-		<checker.version>4.2.0</checker.version>
-		<jackson.version>2.22.0</jackson.version>
-		<reactor.version>3.8.6</reactor.version>
-		<slf4j.version>2.0.18</slf4j.version>
-		<logback.version>1.5.37</logback.version>
-		<animal-sniffer.version>1.27</animal-sniffer.version>
-		<junit.version>6.1.1</junit.version>
-		<hamcrest.version>3.0</hamcrest.version>
-		<jmh.version>1.37</jmh.version>
-		<jcstress.version>0.16</jcstress.version>
-		<lincheck.version>3.6</lincheck.version>
-		<logcaptor.version>2.12.6</logcaptor.version>
-		<vmlens.version>1.2.28</vmlens.version>
-		<!-- DO NOT UPGRADE jqwik past 1.9.3. jqwik 1.10.0 added a deliberate
-		     anti-AI prompt-injection string to test stdout; the 1.10.1 user
-		     guide states the library "is not meant to be used by any 'AI'
-		     coding agents at all." 1.9.3 is the last pre-disclosure release
-		     and is the pinned version for this repo. Any CI / Dependabot /
-		     contributor PR that bumps this MUST be rejected. See CLAUDE.md
-		     section "jqwik prompt-injection in test output" for full context. -->
-		<jqwik.version>1.9.3</jqwik.version>
-		<archunit.version>1.4.2</archunit.version>
-		<spotbugs.version>4.10.2.0</spotbugs.version>
-		<fb-contrib.version>7.7.4</fb-contrib.version>
-		<findsecbugs.version>1.14.0</findsecbugs.version>
-		<spotless.version>3.7.0</spotless.version>
-		<palantir-java-format.version>2.94.0</palantir-java-format.version>
-		<project.build.sourceEncoding>UTF-8</project.build.sourceEncoding>
-		<project.build.outputTimestamp>${git.commit.time}</project.build.outputTimestamp>
-	</properties>
-
-	<!--
-	    Pin transitively-conflicting versions so maven-enforcer's
-	    dependencyConvergence rule passes. The direct slf4j-api version above
-	    must always win over the older 2.0.17 that logback-classic 1.5.32
-	    brings in transitively.
-	-->
-	<dependencyManagement>
-		<dependencies>
-			<dependency>
-				<groupId>org.slf4j</groupId>
-				<artifactId>slf4j-api</artifactId>
-				<version>${slf4j.version}</version>
-			</dependency>
-		</dependencies>
-	</dependencyManagement>
-
-	<dependencies>
-		<dependency>
-			<groupId>org.projectlombok</groupId>
-			<artifactId>lombok</artifactId>
-			<version>${lombok.version}</version>
-			<scope>provided</scope>
-		</dependency>
-		<dependency>
-			<groupId>org.junit.jupiter</groupId>
-			<artifactId>junit-jupiter</artifactId>
-			<version>${junit.version}</version>
-			<scope>test</scope>
-		</dependency>
-		<dependency>
-			<groupId>org.hamcrest</groupId>
-			<artifactId>hamcrest</artifactId>
-			<version>${hamcrest.version}</version>
-			<scope>test</scope>
-		</dependency>
-		<dependency>
-			<groupId>net.jqwik</groupId>
-			<artifactId>jqwik</artifactId>
-			<version>${jqwik.version}</version>
-			<scope>test</scope>
-		</dependency>
-		<dependency>
-			<groupId>com.tngtech.archunit</groupId>
-			<artifactId>archunit-junit5</artifactId>
-			<version>${archunit.version}</version>
-			<scope>test</scope>
-		</dependency>
-		<dependency>
-			<groupId>org.jspecify</groupId>
-			<artifactId>jspecify</artifactId>
-			<version>${jspecify.version}</version>
-		</dependency>
-		<dependency>
-			<groupId>org.checkerframework</groupId>
-			<artifactId>checker-qual</artifactId>
-			<version>${checker.version}</version>
-		</dependency>
-		<dependency>
-			<groupId>com.fasterxml.jackson.core</groupId>
-			<artifactId>jackson-databind</artifactId>
-			<version>${jackson.version}</version>
-		</dependency>
-		<!-- Required by OSInfo (vendored from xerial/sqlite-jdbc) for log emission. -->
-		<dependency>
-			<groupId>org.slf4j</groupId>
-			<artifactId>slf4j-api</artifactId>
-			<version>${slf4j.version}</version>
-		</dependency>
-		<!-- Default SLF4J binding shipped with this library. Runtime scope: not
-		     required on the compile classpath, only loaded at JVM startup. -->
-		<dependency>
-			<groupId>ch.qos.logback</groupId>
-			<artifactId>logback-classic</artifactId>
-			<version>${logback.version}</version>
-			<scope>runtime</scope>
-		</dependency>
-		<!-- @IgnoreJRERequirement marker used by OSInfo (vendored from xerial/sqlite-jdbc)
-		     on a few methods that touch APIs unavailable on Android. Retention=CLASS, so
-		     the dep is compile-only (provided scope). -->
-		<dependency>
-			<groupId>org.codehaus.mojo</groupId>
-			<artifactId>animal-sniffer-annotations</artifactId>
-			<version>${animal-sniffer.version}</version>
-			<scope>provided</scope>
-		</dependency>
-		<dependency>
-			<groupId>org.openjdk.jmh</groupId>
-			<artifactId>jmh-core</artifactId>
-			<version>${jmh.version}</version>
-			<scope>test</scope>
-		</dependency>
-		<dependency>
-			<groupId>org.openjdk.jmh</groupId>
-			<artifactId>jmh-generator-annprocess</artifactId>
-			<version>${jmh.version}</version>
-			<scope>test</scope>
-		</dependency>
-		<dependency>
-			<groupId>org.openjdk.jcstress</groupId>
-			<artifactId>jcstress-core</artifactId>
-			<version>${jcstress.version}</version>
-			<scope>test</scope>
-		</dependency>
-		<dependency>
-			<groupId>org.jetbrains.lincheck</groupId>
-			<artifactId>lincheck</artifactId>
-			<version>${lincheck.version}</version>
-			<scope>test</scope>
-		</dependency>
-		<!-- Test-only SLF4J/Logback log-capture for the LoggingSmokeTest. -->
-		<dependency>
-			<groupId>io.github.hakky54</groupId>
-			<artifactId>logcaptor</artifactId>
-			<version>${logcaptor.version}</version>
-			<scope>test</scope>
-		</dependency>
-		<!-- Test-only Project Reactor used by ReactorIntegrationTest to prove the
-		     "wrap LlamaIterable with Flux.fromIterable + subscribeOn" pattern works
-		     end-to-end (backpressure, AutoCloseable cancel propagation). NOT a runtime
-		     dependency — consumers bring whichever reactive lib they already use
-		     (Reactor / RxJava / Kotlin Flow / Akka — see README "Reactive integration"). -->
-		<dependency>
-			<groupId>io.projectreactor</groupId>
-			<artifactId>reactor-core</artifactId>
-			<version>${reactor.version}</version>
-			<scope>test</scope>
-		</dependency>
-		<dependency>
-			<groupId>io.projectreactor</groupId>
-			<artifactId>reactor-test</artifactId>
-			<version>${reactor.version}</version>
-			<scope>test</scope>
-		</dependency>
-		<!--
-			vmlens interleaving-analysis API. Test-scoped and transitive-dependency-free
-			(its own deps are all test-scope, hence not propagated), so it is safe on the
-			default test classpath under dependencyConvergence. Needed so the
-			VmlensInterleavingSmoke* test compiles in every build; it is only *executed*
-			under the `vmlens` profile (see that profile + the vmlens CI job). The default
-			surefire run excludes it (managed surefire <excludes> below) so the vmlens
-			agent's JDK 9 API class is never loaded outside the agent-driven run.
-		-->
-		<dependency>
-			<groupId>com.vmlens</groupId>
-			<artifactId>api</artifactId>
-			<version>${vmlens.version}</version>
-			<scope>test</scope>
-		</dependency>
-	</dependencies>
-
-	<build>
-		<pluginManagement>
-			<plugins>
-				<plugin>
-					<groupId>com.diffplug.spotless</groupId>
-					<artifactId>spotless-maven-plugin</artifactId>
-					<version>${spotless.version}</version>
-				</plugin>
-				<plugin>
-					<groupId>com.github.spotbugs</groupId>
-					<artifactId>spotbugs-maven-plugin</artifactId>
-					<version>${spotbugs.version}</version>
-				</plugin>
-				<plugin>
-					<groupId>com.vmlens</groupId>
-					<artifactId>vmlens-maven-plugin</artifactId>
-					<version>${vmlens.version}</version>
-				</plugin>
-				<plugin>
-					<groupId>io.github.git-commit-id</groupId>
-					<artifactId>git-commit-id-maven-plugin</artifactId>
-					<version>10.0.0</version>
-				</plugin>
-				<plugin>
-					<groupId>org.apache.maven.plugins</groupId>
-					<artifactId>maven-assembly-plugin</artifactId>
-					<version>3.8.0</version>
-				</plugin>
-				<plugin>
-					<groupId>org.apache.maven.plugins</groupId>
-					<artifactId>maven-compiler-plugin</artifactId>
-					<version>3.15.0</version>
-				</plugin>
-				<plugin>
-					<groupId>org.apache.maven.plugins</groupId>
-					<artifactId>maven-gpg-plugin</artifactId>
-					<version>3.2.8</version>
-				</plugin>
-				<plugin>
-					<groupId>org.apache.maven.plugins</groupId>
-					<artifactId>maven-jar-plugin</artifactId>
-					<version>3.5.0</version>
-				</plugin>
-				<plugin>
-					<groupId>org.apache.maven.plugins</groupId>
-					<artifactId>maven-javadoc-plugin</artifactId>
-					<version>3.12.0</version>
-				</plugin>
-				<plugin>
-					<groupId>org.apache.maven.plugins</groupId>
-					<artifactId>maven-resources-plugin</artifactId>
-					<version>3.5.0</version>
-				</plugin>
-				<plugin>
-					<groupId>org.apache.maven.plugins</groupId>
-					<artifactId>maven-source-plugin</artifactId>
-					<version>3.4.0</version>
-				</plugin>
-				<plugin>
-					<groupId>org.apache.maven.plugins</groupId>
-					<artifactId>maven-surefire-plugin</artifactId>
-					<version>3.5.6</version>
-					<configuration>
-						<!--
-							Tests in the `vmlens` package are meaningful only when run
-							under the vmlens agent (the `vmlens` profile / CI job). Without the
-							agent its AllInterleavings loop body never executes (a vacuous pass
-							that also prints an "agent not configured" warning), so exclude it
-							from the ordinary surefire run. The vmlens job re-includes them with
-							`-Dtest=...`, which overrides this exclude.
-						-->
-						<excludes>
-							<exclude>**/vmlens/*.java</exclude>
-						</excludes>
-					</configuration>
-				</plugin>
-				<plugin>
-					<groupId>org.apache.maven.plugins</groupId>
-					<artifactId>maven-enforcer-plugin</artifactId>
-					<version>3.6.3</version>
-				</plugin>
-				<plugin>
-					<groupId>org.codehaus.mojo</groupId>
-					<artifactId>exec-maven-plugin</artifactId>
-					<version>3.6.3</version>
-				</plugin>
-				<plugin>
-					<groupId>org.jacoco</groupId>
-					<artifactId>jacoco-maven-plugin</artifactId>
-					<version>0.8.15</version>
-				</plugin>
-				<plugin>
-					<groupId>org.pitest</groupId>
-					<artifactId>pitest-maven</artifactId>
-					<version>1.25.5</version>
-				</plugin>
-				<plugin>
-					<groupId>org.sonatype.central</groupId>
-					<artifactId>central-publishing-maven-plugin</artifactId>
-					<version>0.11.0</version>
-				</plugin>
-			</plugins>
-		</pluginManagement>
-		<plugins>
-			<plugin>
-				<groupId>org.apache.maven.plugins</groupId>
-				<artifactId>maven-enforcer-plugin</artifactId>
-				<executions>
-					<execution>
-						<id>enforce</id>
-						<goals>
-							<goal>enforce</goal>
-						</goals>
-						<configuration>
-							<rules>
-								<requireMavenVersion>
-									<version>[3.6.3,)</version>
-								</requireMavenVersion>
-								<requireJavaVersion>
-									<version>[1.8,)</version>
-								</requireJavaVersion>
-								<dependencyConvergence/>
-								<bannedDependencies>
-									<excludes>
-										<!-- Use SLF4J directly; never commons-logging. -->
-										<exclude>commons-logging:commons-logging</exclude>
-										<!-- Legacy / EoL logging frameworks. -->
-										<exclude>log4j:log4j</exclude>
-										<!-- Old hamcrest split artifacts; use hamcrest:hamcrest 2.x+. -->
-										<exclude>org.hamcrest:hamcrest-core</exclude>
-										<exclude>org.hamcrest:hamcrest-library</exclude>
-										<exclude>org.hamcrest:hamcrest-all</exclude>
-										<!-- Legacy JUnit 3 / 4; we use Jupiter (junit:junit-jupiter). -->
-										<exclude>junit:junit</exclude>
-										<exclude>junit:junit-dep</exclude>
-									</excludes>
-								</bannedDependencies>
-							</rules>
-						</configuration>
-					</execution>
-				</executions>
-			</plugin>
-			<plugin>
-				<groupId>io.github.git-commit-id</groupId>
-				<artifactId>git-commit-id-maven-plugin</artifactId>
-				<executions>
-					<execution>
-						<id>get-git-properties</id>
-						<goals>
-							<goal>revision</goal>
-						</goals>
-						<phase>initialize</phase>
-					</execution>
-				</executions>
-				<configuration>
-					<dateFormat>yyyy-MM-dd'T'HH:mm:ss'Z'</dateFormat>
-					<dateFormatTimeZone>UTC</dateFormatTimeZone>
-					<generateGitPropertiesFile>false</generateGitPropertiesFile>
-					<failOnNoGitDirectory>false</failOnNoGitDirectory>
-				</configuration>
-			</plugin>
-			<plugin>
-				<groupId>org.apache.maven.plugins</groupId>
-				<artifactId>maven-compiler-plugin</artifactId>
-				<configuration>
-					<release>8</release>
-					<testRelease>21</testRelease>
-					<showWarnings>true</showWarnings>
-					<parameters>true</parameters>
-					<compilerArgs>
-						<!--
-						  -Xlint:all surfaces every category of javac warning. Excluded:
-						    -serial    : serialVersionUID warnings on non-Serializable classes
-						    -options   : bootclasspath / target-mismatch chatter under release=8
-						    -classfile : noise when javac reads our own .class output mid-build
-						    -processing: "No processor claimed any of these annotations" — the
-						                 animal-sniffer @IgnoreJRERequirement is consumed by the
-						                 maven-plugin (not a JSR-269 processor), so javac flags
-						                 it every build
-						  -Werror is on: the ElementType.MODULE blocker is gone now that
-						  @NullMarked lives only in module-info.java (compiled at release 9 in
-						  a separate execution); see module-info.java's javadoc.
-						-->
-						<arg>-Xlint:all,-serial,-options,-classfile,-processing</arg>
-						<arg>-Werror</arg>
-						<!--
-						  Checker Framework Nullness Checker runs as a second nullness pass
-						  alongside NullAway. CF is generics-aware and stricter than NullAway,
-						  so it acts as a second-opinion verifier on the same JSpecify
-						  annotations.
-						-->
-						<!-- Explicit processor list: Lombok (two SPI classes) + Checker Framework's
-							 Nullness Checker as a 2nd nullness pass alongside NullAway. The explicit
-							 -processor argument overrides SPI discovery, so every processor we want
-							 to run must be listed here. -->
-						<arg>-processor</arg>
-						<arg>lombok.launch.AnnotationProcessorHider$AnnotationProcessor,lombok.launch.AnnotationProcessorHider$ClaimingProcessor,org.checkerframework.checker.nullness.NullnessChecker</arg>
-						<arg>-XDaddTypeAnnotationsToSymbol=true</arg>
-						<arg>-XDcompilePolicy=simple</arg>
-						<arg>--should-stop=ifError=FLOW</arg>
-						<arg>-Xplugin:ErrorProne -Xep:NullAway:ERROR -XepOpt:NullAway:OnlyNullMarked=true -XepOpt:NullAway:JSpecifyMode=true -XepOpt:NullAway:CheckOptionalEmptiness=true -XepOpt:NullAway:AcknowledgeRestrictiveAnnotations=true -XepOpt:NullAway:AcknowledgeAndroidRecent=true -XepOpt:NullAway:AssertsEnabled=true -Xep:BoxedPrimitiveEquality:ERROR -Xep:EqualsHashCode:ERROR -Xep:EqualsIncompatibleType:ERROR -Xep:IdentityBinaryExpression:ERROR -Xep:SelfAssignment:ERROR -Xep:SelfComparison:ERROR -Xep:SelfEquals:ERROR -Xep:DeadException:ERROR -Xep:FormatString:ERROR -Xep:InvalidPatternSyntax:ERROR -Xep:OptionalEquality:ERROR -Xep:ImpossibleNullComparison:ERROR</arg>
-					</compilerArgs>
-					<annotationProcessorPaths>
-						<path>
-							<groupId>org.projectlombok</groupId>
-							<artifactId>lombok</artifactId>
-							<version>${lombok.version}</version>
-						</path>
-						<path>
-							<groupId>com.google.errorprone</groupId>
-							<artifactId>error_prone_core</artifactId>
-							<version>${errorprone.version}</version>
-						</path>
-						<path>
-							<groupId>com.uber.nullaway</groupId>
-							<artifactId>nullaway</artifactId>
-							<version>${nullaway.version}</version>
-						</path>
-						<path>
-							<groupId>org.checkerframework</groupId>
-							<artifactId>checker</artifactId>
-							<version>${checker.version}</version>
-						</path>
-					</annotationProcessorPaths>
-				</configuration>
-				<executions>
-					<execution>
-						<id>default-compile</id>
-						<configuration>
-							<!-- The default compile pass targets Java 8 bytecode for the library
-							     sources. module-info.java cannot be compiled at release 8 (it is a
-							     Java 9+ construct) and is handled by the separate execution below. -->
-							<excludes>
-								<exclude>module-info.java</exclude>
-							</excludes>
-						</configuration>
-					</execution>
-					<execution>
-						<id>module-info-compile</id>
-						<phase>compile</phase>
-						<goals>
-							<goal>compile</goal>
-						</goals>
-						<configuration>
-							<!-- module-info.java is the only Java 9+ source in this project; compile
-							     it at release 9 (the minimum version that understands modules) without
-							     the Error Prone / NullAway / Checker Framework processors that target
-							     ordinary source files only. Java 8 runtimes silently ignore the
-							     resulting module-info.class at the JAR root, so existing classpath
-							     consumers continue to work unchanged. LlamaLoader loads native libs
-							     via Class.getResourceAsStream on its own class object, which resolves
-							     to this module and therefore does not require any opens directives. -->
-							<release>9</release>
-							<includes>
-								<include>module-info.java</include>
-							</includes>
-							<compilerArgs combine.self="override"/>
-							<annotationProcessorPaths combine.self="override"/>
-						</configuration>
-					</execution>
-					<execution>
-						<id>default-testCompile</id>
-						<configuration>
-							<!-- Drop -parameters for test compile so jcstress reflection doesn't
-							     pick up parameter names from MethodParameters. -->
-							<parameters>false</parameters>
-							<compilerArgs combine.self="override">
-								<arg>-XDaddTypeAnnotationsToSymbol=true</arg>
-								<arg>-XDcompilePolicy=simple</arg>
-								<arg>--should-stop=ifError=FLOW</arg>
-								<arg>-Xplugin:ErrorProne -Xep:NullAway:OFF -Xep:GuardedBy:OFF</arg>
-							</compilerArgs>
-							<annotationProcessorPaths combine.children="append">
-								<path>
-									<groupId>org.openjdk.jcstress</groupId>
-									<artifactId>jcstress-core</artifactId>
-									<version>${jcstress.version}</version>
-								</path>
-								<path>
-									<groupId>org.openjdk.jmh</groupId>
-									<artifactId>jmh-generator-annprocess</artifactId>
-									<version>${jmh.version}</version>
-								</path>
-							</annotationProcessorPaths>
-						</configuration>
-					</execution>
-				</executions>
-			</plugin>
-			<plugin>
-				<artifactId>maven-resources-plugin</artifactId>
-			</plugin>
-			<plugin>
-				<groupId>org.apache.maven.plugins</groupId>
-				<artifactId>maven-source-plugin</artifactId>
-				<executions>
-					<execution>
-						<id>attach-sources</id>
-						<goals>
-							<goal>jar-no-fork</goal>
-						</goals>
-					</execution>
-				</executions>
-			</plugin>
-			<plugin>
-				<groupId>org.apache.maven.plugins</groupId>
-				<artifactId>maven-javadoc-plugin</artifactId>
-				<configuration>
-					<source>1.8</source>
-					<failOnError>true</failOnError>
-					<failOnWarnings>true</failOnWarnings>
-					<doclint>all</doclint>
-				</configuration>
-				<executions>
-					<execution>
-						<id>attach-javadocs</id>
-						<goals>
-							<goal>jar</goal>
-						</goals>
-					</execution>
-				</executions>
-			</plugin>
-			<plugin>
-				<groupId>org.jacoco</groupId>
-				<artifactId>jacoco-maven-plugin</artifactId>
-				<executions>
-					<execution>
-						<id>prepare-agent</id>
-						<goals>
-							<goal>prepare-agent</goal>
-						</goals>
-					</execution>
-					<execution>
-						<id>report</id>
-						<phase>test</phase>
-						<goals>
-							<goal>report</goal>
-						</goals>
-					</execution>
-				</executions>
-			</plugin>
-			<plugin>
-				<!--
-				  Surefire is configured here only to control native-crash diagnostics:
-				  - argLine inherits jacoco's @{argLine} (so the jacoco runtime agent stays
-				    attached) and appends -XX:ErrorFile so the JVM crash log (hs_err_pid<pid>.log)
-				    always lands in the project basedir, which is the same directory the CI
-				    upload-artifact globs (${{ github.workspace }}/hs_err_pid*.log).
-				  - HeapDumpOnOutOfMemoryError + HeapDumpPath catch OOMs that would otherwise
-				    just print a stack and exit.
-				  - The version is pinned to the same release Maven would otherwise auto-select
-				    (visible in CI error logs as 3.5.4), so we don't drift independently of
-				    Maven's plugin-defaults bump.
-				-->
-				<groupId>org.apache.maven.plugins</groupId>
-				<artifactId>maven-surefire-plugin</artifactId>
-				<configuration>
-					<argLine>@{argLine} -Xmx2g -XX:ErrorFile=hs_err_pid%p.log -XX:+HeapDumpOnOutOfMemoryError -XX:HeapDumpPath=.</argLine>
-					<!--
-					  Capture each test class's stdout/stderr into
-					  target/surefire-reports/<class>-output.txt. When a native crash
-					  takes down the forked JVM, this file is the ONLY record of
-					  which @Test method was last running (without it the surefire
-					  *.txt summary just says "VM terminated", which we already saw
-					  in CI without telling us which test was on fire).
-					-->
-					<redirectTestOutputToFile>true</redirectTestOutputToFile>
-				</configuration>
-			</plugin>
-			<plugin>
-				<groupId>com.diffplug.spotless</groupId>
-				<artifactId>spotless-maven-plugin</artifactId>
-				<configuration>
-					<java>
-						<includes>
-							<include>src/main/java/**/*.java</include>
-							<include>src/test/java/**/*.java</include>
-						</includes>
-						<palantirJavaFormat>
-							<version>${palantir-java-format.version}</version>
-						</palantirJavaFormat>
-						<removeUnusedImports/>
-						<trimTrailingWhitespace/>
-						<endWithNewline/>
-					</java>
-				</configuration>
-				<executions>
-					<execution>
-						<id>spotless-check</id>
-						<phase>verify</phase>
-						<goals>
-							<goal>check</goal>
-						</goals>
-					</execution>
-				</executions>
-			</plugin>
-			<plugin>
-				<groupId>com.github.spotbugs</groupId>
-				<artifactId>spotbugs-maven-plugin</artifactId>
-				<configuration>
-					<effort>Max</effort>
-					<threshold>Low</threshold>
-					<failOnError>true</failOnError>
-					<includeTests>false</includeTests>
-					<excludeFilterFile>spotbugs-exclude.xml</excludeFilterFile>
-					<plugins>
-						<plugin>
-							<groupId>com.mebigfatguy.fb-contrib</groupId>
-							<artifactId>fb-contrib</artifactId>
-							<version>${fb-contrib.version}</version>
-						</plugin>
-						<plugin>
-							<groupId>com.h3xstream.findsecbugs</groupId>
-							<artifactId>findsecbugs-plugin</artifactId>
-							<version>${findsecbugs.version}</version>
-						</plugin>
-					</plugins>
-				</configuration>
-				<executions>
-					<execution>
-						<id>spotbugs-check</id>
-						<phase>verify</phase>
-						<goals>
-							<goal>check</goal>
-						</goals>
-					</execution>
-				</executions>
-			</plugin>
-			<plugin>
-				<groupId>org.codehaus.mojo</groupId>
-				<artifactId>exec-maven-plugin</artifactId>
-				<configuration>
-					<mainClass>org.openjdk.jmh.Main</mainClass>
-					<classpathScope>test</classpathScope>
-				</configuration>
-			</plugin>
-			<plugin>
-				<!--
-				  Generic PIT policy (version, 100% gate, invocation, expansion rule):
-				  see ../workspace/policies/pit-mutation-testing.md.
-				  Repo-specific: the gated value/exception/args/json classes are pure-Java
-				  (no native libjllama or model file needed); targetClasses/targetTests are
-				  kept in lock-step. NOTE: value.ContentPart.audioFile(Path) mutants are only
-				  killed with the audio fixture present — see TODO.md and policy §4.
-				-->
-				<groupId>org.pitest</groupId>
-				<artifactId>pitest-maven</artifactId>
-				<dependencies>
-					<dependency>
-						<groupId>org.pitest</groupId>
-						<artifactId>pitest-junit5-plugin</artifactId>
-						<version>1.2.3</version>
-					</dependency>
-				</dependencies>
-				<configuration>
-					<targetClasses>
-						<param>net.ladenthin.llama.value.*</param>
-						<param>net.ladenthin.llama.exception.*</param>
-						<param>net.ladenthin.llama.args.*</param>
-						<param>net.ladenthin.llama.json.TimingsLogger</param>
-						<param>net.ladenthin.llama.json.RerankResponseParser</param>
-						<param>net.ladenthin.llama.json.ChatResponseParser</param>
-						<param>net.ladenthin.llama.json.CompletionResponseParser</param>
-					</targetClasses>
-					<targetTests>
-						<param>net.ladenthin.llama.value.*</param>
-						<param>net.ladenthin.llama.exception.*</param>
-						<param>net.ladenthin.llama.args.*</param>
-						<param>net.ladenthin.llama.json.*</param>
-					</targetTests>
-					<mutationThreshold>100</mutationThreshold>
-					<timeoutConstant>30000</timeoutConstant>
-				</configuration>
-			</plugin>
-		</plugins>
-	</build>
-
-	<profiles>
-		<profile>
-			<id>release</id>
-			<build>
-				<plugins>
-					<plugin>
-						<groupId>org.apache.maven.plugins</groupId>
-						<artifactId>maven-gpg-plugin</artifactId>
-						<executions>
-							<execution>
-								<id>sign-artifacts</id>
-								<phase>verify</phase>
-								<goals><goal>sign</goal></goals>
-								<configuration>
-									<keyname>${gpg.keyname}</keyname>
-									<gpgArguments>
-										<arg>--pinentry-mode</arg>
-										<arg>loopback</arg>
-									</gpgArguments>
-								</configuration>
-							</execution>
-						</executions>
-					</plugin>
-
-					<plugin>
-						<groupId>org.sonatype.central</groupId>
-						<artifactId>central-publishing-maven-plugin</artifactId>
-						<extensions>true</extensions>
-						<configuration>
-							<publishingServerId>central</publishingServerId>
-							<autoPublish>true</autoPublish>
-							<waitUntil>published</waitUntil>
-						</configuration>
-					</plugin>
-				</plugins>
-			</build>
-		</profile>
-
-		<profile>
-			<id>cuda</id>
-			<build>
-				<plugins>
-					<plugin>
-						<groupId>org.apache.maven.plugins</groupId>
-						<artifactId>maven-compiler-plugin</artifactId>
-						<executions>
-						<!-- We have to perform a separate build pass for cuda
-						classifier -->
-						<execution>
-							<id>gpu</id>
-							<phase>compile</phase>
-							<goals>
-								<goal>compile</goal>
-							</goals>
-							<configuration>
-								<!-- Same rationale as default-compile: this pass targets release 8;
-								     module-info.java is Java 9+ and is handled by module-info-compile. -->
-								<excludes>
-									<exclude>module-info.java</exclude>
-								</excludes>
-								<compilerArgs>
-									<arg>-h</arg>
-									<arg>src/main/cpp</arg>
-								</compilerArgs>
-								<outputDirectory>
-									${project.build.outputDirectory}_cuda</outputDirectory>
-							</configuration>
-						</execution>
-						</executions>
-					</plugin>
-					<plugin>
-						<artifactId>maven-resources-plugin</artifactId>
-						<executions>
-						<!-- Copy custom cuda libs to the output directory -->
-						<execution>
-							<id>copy-resources</id>
-							<phase>process-classes</phase>
-							<goals>
-								<goal>copy-resources</goal>
-							</goals>
-							<configuration>
-								<outputDirectory>
-									${project.build.outputDirectory}_cuda</outputDirectory>
-								<resources>
-									<resource>
-										<directory>
-											${basedir}/src/main/resources_linux_cuda/</directory>
-										<includes>
-											<include>**/*.*</include>
-										</includes>
-									</resource>
-								</resources>
-							</configuration>
-						</execution>
-						</executions>
-					</plugin>
-					<plugin>
-						<groupId>org.apache.maven.plugins</groupId>
-						<artifactId>maven-jar-plugin</artifactId>
-						<executions>
-							<!-- Pick class files AND libs from custom output
-							directory -->
-							<execution>
-								<id>cuda</id>
-								<phase>package</phase>
-								<goals>
-									<goal>jar</goal>
-								</goals>
-								<configuration>
-									<classifier>cuda13-linux-x86-64</classifier>
-									<classesDirectory>
-										${project.build.outputDirectory}_cuda</classesDirectory>
-								</configuration>
-							</execution>
-						</executions>
-					</plugin>
-				</plugins>
-			</build>
-		</profile>
-
-		<profile>
-			<id>opencl-android</id>
-			<build>
-				<plugins>
-					<plugin>
-						<groupId>org.apache.maven.plugins</groupId>
-						<artifactId>maven-compiler-plugin</artifactId>
-						<executions>
-						<!-- Separate build pass for opencl-android classifier -->
-						<execution>
-							<id>opencl-android</id>
-							<phase>compile</phase>
-							<goals>
-								<goal>compile</goal>
-							</goals>
-							<configuration>
-								<!-- Same rationale as default-compile: this pass targets release 8;
-								     module-info.java is Java 9+ and is handled by module-info-compile. -->
-								<excludes>
-									<exclude>module-info.java</exclude>
-								</excludes>
-								<compilerArgs>
-									<arg>-h</arg>
-									<arg>src/main/cpp</arg>
-								</compilerArgs>
-								<outputDirectory>
-									${project.build.outputDirectory}_opencl_android</outputDirectory>
-							</configuration>
-						</execution>
-						</executions>
-					</plugin>
-					<plugin>
-						<artifactId>maven-resources-plugin</artifactId>
-						<executions>
-						<!-- Copy custom opencl-android libs to the output directory -->
-						<execution>
-							<id>copy-resources-opencl-android</id>
-							<phase>process-classes</phase>
-							<goals>
-								<goal>copy-resources</goal>
-							</goals>
-							<configuration>
-								<outputDirectory>
-									${project.build.outputDirectory}_opencl_android</outputDirectory>
-								<resources>
-									<resource>
-										<directory>
-											${basedir}/src/main/resources_android_opencl/</directory>
-										<includes>
-											<include>**/*.*</include>
-										</includes>
-									</resource>
-								</resources>
-							</configuration>
-						</execution>
-						</executions>
-					</plugin>
-					<plugin>
-						<groupId>org.apache.maven.plugins</groupId>
-						<artifactId>maven-jar-plugin</artifactId>
-						<executions>
-							<!-- Pick class files AND libs from custom output
-							directory -->
-							<execution>
-								<id>opencl-android</id>
-								<phase>package</phase>
-								<goals>
-									<goal>jar</goal>
-								</goals>
-								<configuration>
-									<classifier>opencl-android-aarch64</classifier>
-									<classesDirectory>
-										${project.build.outputDirectory}_opencl_android</classesDirectory>
-								</configuration>
-							</execution>
-						</executions>
-					</plugin>
-				</plugins>
-			</build>
-		</profile>
-		<profile>
-			<!-- MSVC (Visual Studio generator) Windows CPU natives, shipped as the
-			     `msvc-windows` classifier JAR. After the default flip the DEFAULT JAR
-			     ships the Ninja Multi-Config natives (sccache-cached, merged into
-			     src/main/resources/ by CI); this profile packages the alternative
-			     MSVC-generator build for anyone who wants it. Both use the same MSVC
-			     toolchain (static /MT CRT) so they are functionally equivalent. Mirrors
-			     the cuda / opencl-android classifier pattern: a separate compile pass +
-			     resource copy + classified jar. The MSVC DLLs are staged into
-			     src/main/resources_windows_msvc/ by CI before this profile runs. -->
-			<id>windows-msvc</id>
-			<build>
-				<plugins>
-					<plugin>
-						<groupId>org.apache.maven.plugins</groupId>
-						<artifactId>maven-compiler-plugin</artifactId>
-						<executions>
-						<!-- Separate build pass for the windows-msvc classifier -->
-						<execution>
-							<id>windows-msvc</id>
-							<phase>compile</phase>
-							<goals>
-								<goal>compile</goal>
-							</goals>
-							<configuration>
-								<!-- Same rationale as default-compile: this pass targets release 8;
-								     module-info.java is Java 9+ and is handled by module-info-compile. -->
-								<excludes>
-									<exclude>module-info.java</exclude>
-								</excludes>
-								<compilerArgs>
-									<arg>-h</arg>
-									<arg>src/main/cpp</arg>
-								</compilerArgs>
-								<outputDirectory>
-									${project.build.outputDirectory}_windows_msvc</outputDirectory>
-							</configuration>
-						</execution>
-						</executions>
-					</plugin>
-					<plugin>
-						<artifactId>maven-resources-plugin</artifactId>
-						<executions>
-						<!-- Copy the Ninja-built Windows libs to the output directory -->
-						<execution>
-							<id>copy-resources-windows-msvc</id>
-							<phase>process-classes</phase>
-							<goals>
-								<goal>copy-resources</goal>
-							</goals>
-							<configuration>
-								<outputDirectory>
-									${project.build.outputDirectory}_windows_msvc</outputDirectory>
-								<resources>
-									<resource>
-										<directory>
-											${basedir}/src/main/resources_windows_msvc/</directory>
-										<includes>
-											<include>**/*.*</include>
-										</includes>
-									</resource>
-								</resources>
-							</configuration>
-						</execution>
-						</executions>
-					</plugin>
-					<plugin>
-						<groupId>org.apache.maven.plugins</groupId>
-						<artifactId>maven-jar-plugin</artifactId>
-						<executions>
-							<!-- Pick class files AND libs from custom output
-							directory -->
-							<execution>
-								<id>windows-msvc</id>
-								<phase>package</phase>
-								<goals>
-									<goal>jar</goal>
-								</goals>
-								<configuration>
-									<classifier>msvc-windows</classifier>
-									<classesDirectory>
-										${project.build.outputDirectory}_windows_msvc</classesDirectory>
-								</configuration>
-							</execution>
-						</executions>
-					</plugin>
-				</plugins>
-			</build>
-		</profile>
-
-		<profile>
-			<!-- Windows CUDA GPU natives (x86_64 only), shipped as the
-			     `cuda13-windows-x86-64` classifier JAR. Built with -DGGML_CUDA=ON and
-			     routed by CMakeLists.txt to src/main/resources_windows_cuda/. The CUDA
-			     runtime libs (cudart64_13 / cublas64_13 / cublasLt64_13) are NOT bundled
-			     — the consumer's installed CUDA Toolkit provides them. Mirrors the
-			     cuda (Linux) classifier pattern. Staged by CI before this profile runs. -->
-			<id>cuda-windows</id>
-			<build>
-				<plugins>
-					<plugin>
-						<groupId>org.apache.maven.plugins</groupId>
-						<artifactId>maven-compiler-plugin</artifactId>
-						<executions>
-						<execution>
-							<id>cuda-windows</id>
-							<phase>compile</phase>
-							<goals>
-								<goal>compile</goal>
-							</goals>
-							<configuration>
-								<excludes>
-									<exclude>module-info.java</exclude>
-								</excludes>
-								<compilerArgs>
-									<arg>-h</arg>
-									<arg>src/main/cpp</arg>
-								</compilerArgs>
-								<outputDirectory>
-									${project.build.outputDirectory}_windows_cuda</outputDirectory>
-							</configuration>
-						</execution>
-						</executions>
-					</plugin>
-					<plugin>
-						<artifactId>maven-resources-plugin</artifactId>
-						<executions>
-						<execution>
-							<id>copy-resources-cuda-windows</id>
-							<phase>process-classes</phase>
-							<goals>
-								<goal>copy-resources</goal>
-							</goals>
-							<configuration>
-								<outputDirectory>
-									${project.build.outputDirectory}_windows_cuda</outputDirectory>
-								<resources>
-									<resource>
-										<directory>
-											${basedir}/src/main/resources_windows_cuda/</directory>
-										<includes>
-											<include>**/*.*</include>
-										</includes>
-									</resource>
-								</resources>
-							</configuration>
-						</execution>
-						</executions>
-					</plugin>
-					<plugin>
-						<groupId>org.apache.maven.plugins</groupId>
-						<artifactId>maven-jar-plugin</artifactId>
-						<executions>
-							<execution>
-								<id>cuda-windows</id>
-								<phase>package</phase>
-								<goals>
-									<goal>jar</goal>
-								</goals>
-								<configuration>
-									<classifier>cuda13-windows-x86-64</classifier>
-									<classesDirectory>
-										${project.build.outputDirectory}_windows_cuda</classesDirectory>
-								</configuration>
-							</execution>
-						</executions>
-					</plugin>
-				</plugins>
-			</build>
-		</profile>
-
-		<profile>
-			<!-- Windows Vulkan GPU natives (x86_64 only), shipped as the
-			     `vulkan-windows-x86-64` classifier JAR. Built with -DGGML_VULKAN=ON and
-			     routed by CMakeLists.txt to src/main/resources_windows_vulkan/. The Vulkan
-			     runtime (vulkan-1.dll) ships with the GPU driver — nothing is bundled.
-			     Vulkan is the most portable Windows GPU backend (NVIDIA/AMD/Intel).
-			     Staged by CI before this profile runs. -->
-			<id>vulkan-windows</id>
-			<build>
-				<plugins>
-					<plugin>
-						<groupId>org.apache.maven.plugins</groupId>
-						<artifactId>maven-compiler-plugin</artifactId>
-						<executions>
-						<execution>
-							<id>vulkan-windows</id>
-							<phase>compile</phase>
-							<goals>
-								<goal>compile</goal>
-							</goals>
-							<configuration>
-								<excludes>
-									<exclude>module-info.java</exclude>
-								</excludes>
-								<compilerArgs>
-									<arg>-h</arg>
-									<arg>src/main/cpp</arg>
-								</compilerArgs>
-								<outputDirectory>
-									${project.build.outputDirectory}_windows_vulkan</outputDirectory>
-							</configuration>
-						</execution>
-						</executions>
-					</plugin>
-					<plugin>
-						<artifactId>maven-resources-plugin</artifactId>
-						<executions>
-						<execution>
-							<id>copy-resources-vulkan-windows</id>
-							<phase>process-classes</phase>
-							<goals>
-								<goal>copy-resources</goal>
-							</goals>
-							<configuration>
-								<outputDirectory>
-									${project.build.outputDirectory}_windows_vulkan</outputDirectory>
-								<resources>
-									<resource>
-										<directory>
-											${basedir}/src/main/resources_windows_vulkan/</directory>
-										<includes>
-											<include>**/*.*</include>
-										</includes>
-									</resource>
-								</resources>
-							</configuration>
-						</execution>
-						</executions>
-					</plugin>
-					<plugin>
-						<groupId>org.apache.maven.plugins</groupId>
-						<artifactId>maven-jar-plugin</artifactId>
-						<executions>
-							<execution>
-								<id>vulkan-windows</id>
-								<phase>package</phase>
-								<goals>
-									<goal>jar</goal>
-								</goals>
-								<configuration>
-									<classifier>vulkan-windows-x86-64</classifier>
-									<classesDirectory>
-										${project.build.outputDirectory}_windows_vulkan</classesDirectory>
-								</configuration>
-							</execution>
-						</executions>
-					</plugin>
-				</plugins>
-			</build>
-		</profile>
-
-		<profile>
-			<!-- Windows OpenCL GPU natives (x86_64 only), shipped as the
-			     `opencl-windows-x86-64` classifier JAR. Built with -DGGML_OPENCL=ON and
-			     routed by CMakeLists.txt to src/main/resources_windows_opencl/. The OpenCL
-			     ICD (System32\OpenCL.dll) ships with the GPU driver — nothing is bundled.
-			     NOTE: the GGML OpenCL backend is Adreno-tuned; on desktop GPUs CUDA/Vulkan
-			     are better supported. Staged by CI before this profile runs. -->
-			<id>opencl-windows</id>
-			<build>
-				<plugins>
-					<plugin>
-						<groupId>org.apache.maven.plugins</groupId>
-						<artifactId>maven-compiler-plugin</artifactId>
-						<executions>
-						<execution>
-							<id>opencl-windows</id>
-							<phase>compile</phase>
-							<goals>
-								<goal>compile</goal>
-							</goals>
-							<configuration>
-								<excludes>
-									<exclude>module-info.java</exclude>
-								</excludes>
-								<compilerArgs>
-									<arg>-h</arg>
-									<arg>src/main/cpp</arg>
-								</compilerArgs>
-								<outputDirectory>
-									${project.build.outputDirectory}_windows_opencl</outputDirectory>
-							</configuration>
-						</execution>
-						</executions>
-					</plugin>
-					<plugin>
-						<artifactId>maven-resources-plugin</artifactId>
-						<executions>
-						<execution>
-							<id>copy-resources-opencl-windows</id>
-							<phase>process-classes</phase>
-							<goals>
-								<goal>copy-resources</goal>
-							</goals>
-							<configuration>
-								<outputDirectory>
-									${project.build.outputDirectory}_windows_opencl</outputDirectory>
-								<resources>
-									<resource>
-										<directory>
-											${basedir}/src/main/resources_windows_opencl/</directory>
-										<includes>
-											<include>**/*.*</include>
-										</includes>
-									</resource>
-								</resources>
-							</configuration>
-						</execution>
-						</executions>
-					</plugin>
-					<plugin>
-						<groupId>org.apache.maven.plugins</groupId>
-						<artifactId>maven-jar-plugin</artifactId>
-						<executions>
-							<execution>
-								<id>opencl-windows</id>
-								<phase>package</phase>
-								<goals>
-									<goal>jar</goal>
-								</goals>
-								<configuration>
-									<classifier>opencl-windows-x86-64</classifier>
-									<classesDirectory>
-										${project.build.outputDirectory}_windows_opencl</classesDirectory>
-								</configuration>
-							</execution>
-						</executions>
-					</plugin>
-				</plugins>
-			</build>
-		</profile>
-
-		<profile>
-			<id>vmlens</id>
-			<build>
-				<plugins>
-					<plugin>
-						<groupId>com.vmlens</groupId>
-						<artifactId>vmlens-maven-plugin</artifactId>
-						<configuration>
-							<!--
-								Run vmlens interleaving analysis over the whole `vmlens` test
-								package (smoke test + SessionStateInterleavingTest). The `com.vmlens:api`
-								test dependency lives in the main <dependencies> block.
-								Expand <includes> as more concurrency tests are added (the
-								streambuffer repo runs vmlens over its whole suite).
-							-->
-							<includes>
-								<include>**/vmlens/*.java</include>
-							</includes>
-						</configuration>
-						<executions>
-							<execution>
-								<id>vmlens-test</id>
-								<goals>
-									<goal>test</goal>
-								</goals>
-							</execution>
-						</executions>
-					</plugin>
-				</plugins>
-			</build>
-		</profile>
-		<profile>
-			<id>jcstress</id>
-			<build>
-				<plugins>
-					<plugin>
-						<groupId>org.codehaus.mojo</groupId>
-						<artifactId>exec-maven-plugin</artifactId>
-						<executions>
-							<execution>
-								<id>jcstress</id>
-								<phase>test</phase>
-								<goals><goal>exec</goal></goals>
-								<configuration>
-									<executable>${java.home}/bin/java</executable>
-									<classpathScope>test</classpathScope>
-									<arguments>
-										<argument>-classpath</argument>
-										<classpath/>
-										<argument>org.openjdk.jcstress.Main</argument>
-										<argument>-v</argument>
-										<argument>-m</argument>
-										<argument>default</argument>
-									</arguments>
-								</configuration>
-							</execution>
-						</executions>
-					</plugin>
-				</plugins>
-			</build>
-		</profile>
-		<profile>
-			<!--
-				Builds the fat jar-with-dependencies uber JAR: the library classes, the
-				default-platform native libs from src/main/resources, and all runtime Java
-				dependencies in one drop-on-classpath JAR, runnable via the OpenAiCompatServer
-				Main-Class (set below) to start the OpenAI-compatible HTTP server. Off by
-				default; the CI `package` job activates it so the uber JAR rides along in the
-				`llama-jars` upload-artifact bundle. Documented in CLAUDE.md "Build Commands"
-				as `mvn -P assembly package`.
-			-->
-			<id>assembly</id>
-			<build>
-				<plugins>
-					<plugin>
-						<groupId>org.apache.maven.plugins</groupId>
-						<artifactId>maven-assembly-plugin</artifactId>
-						<configuration>
-							<descriptorRefs>
-								<descriptorRef>jar-with-dependencies</descriptorRef>
-							</descriptorRefs>
-							<archive>
-								<manifest>
-									<mainClass>net.ladenthin.llama.server.OpenAiCompatServer</mainClass>
-								</manifest>
-							</archive>
-						</configuration>
-						<executions>
-							<execution>
-								<id>build-fat-jar</id>
-								<phase>package</phase>
-								<goals>
-									<goal>single</goal>
-								</goals>
-							</execution>
-						</executions>
-					</plugin>
-				</plugins>
-			</build>
-		</profile>
-	</profiles>
+	<modules>
+		<module>llama</module>
+		<module>llama-langchain4j</module>
+	</modules>
 </project>

From a2857c5b55f3c4b4b14e1374a0a9f12ac8890081 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Wed, 1 Jul 2026 20:09:04 +0000
Subject: [PATCH 02/12] ci: re-root the pipeline for the llama/ reactor module
 (build scripts, workflows, configs)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Follows the structural move: everything that assumed the core lived at the repo
root now points into the llama/ module. No behavioural change to the native
build — only paths.

- build.sh / build.bat: cd into llama/ (relative to the script's own location)
  before cmake, so every native build (incl. the build_cuda_linux.sh /
  build_opencl_*.sh wrappers and the dockcross containers, whose workdir stays the
  repo root) configures the module. This single hook re-roots all native builds.
- publish.yml: mvn compile/test/PIT/vmlens/spotbugs/spotless now target
  -f llama/pom.xml; ctest uses llama/build; all artifact upload/download +
  webui-generated paths prefixed with llama/; GIT_TAG read from llama/CMakeLists.txt;
  jacoco/pit/vmlens/jar produce-paths -> llama/target (report-job download dirs
  stay at root). Removed the now-redundant version-lockstep guard (inheritance
  guarantees it) and the separate llama-langchain4j deploy steps (one reactor
  `mvn -P release deploy` publishes parent + llama + llama-langchain4j together).
- release profile (GPG + Central Publishing) moved to the parent pom so the reactor
  deploy signs every module incl. the parent .pom; removed from the two child poms.
- clang-format.yml, sonarqube.yml, REUSE.toml, .gitignore re-rooted to llama/.
  codeql.yml left as a root reactor compile (now also covers llama-langchain4j).

Verified locally: full reactor builds; `mvn -pl llama -am install` (the langchain4j
jobs' install step) installs parent+core; `mvn -N -P release validate` parses the
moved release profile; `mvn -f llama-langchain4j verify` is green (7 mapping tests,
4 self-skip, main+sources+javadoc jars). Native path re-rooting is validated by CI.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rt1paYztGJ2AKUuBuAGDXE
---
 .github/build.bat                  |   5 +
 .github/build.sh                   |   8 +
 .github/workflows/clang-format.yml |   7 +-
 .github/workflows/publish.yml      | 273 +++++++++++++----------------
 .github/workflows/sonarqube.yml    |   2 +-
 .gitignore                         |  26 +--
 REUSE.toml                         |   8 +-
 llama-langchain4j/pom.xml          |  50 +-----
 llama/pom.xml                      |  43 +----
 pom.xml                            |  49 ++++++
 10 files changed, 221 insertions(+), 250 deletions(-)

diff --git a/.github/build.bat b/.github/build.bat
index d48f0957..a9df77f5 100755
--- a/.github/build.bat
+++ b/.github/build.bat
@@ -6,6 +6,11 @@ REM SPDX-License-Identifier: MIT
 @echo off
 setlocal enabledelayedexpansion
 
+REM The core project (CMakeLists.txt + src\) lives in the `llama\` module of the Maven
+REM reactor. Re-root here once (relative to this script's own location) so cmake
+REM configures the module regardless of the caller's CWD.
+cd /d "%~dp0..\llama" || exit /b 1
+
 REM ---------------------------------------------------------------------------
 REM Optional shared compiler cache: sccache fronting Depot Cache (WebDAV).
 REM Mirrors build.sh's sccache_can_wrap_compiler() probe. Because sccache *is*
diff --git a/.github/build.sh b/.github/build.sh
index 480e2009..7a47ab65 100755
--- a/.github/build.sh
+++ b/.github/build.sh
@@ -5,6 +5,14 @@
 #
 # SPDX-License-Identifier: MIT
 
+# The core project (CMakeLists.txt + src/) lives in the `llama/` module of the Maven
+# reactor. Re-root here once — every native build delegates to this script (incl. the
+# build_cuda_linux.sh / build_opencl_*.sh wrappers that `exec .github/build.sh`) — so
+# cmake configures the module regardless of the caller's CWD. Anchored to this script's
+# own location (via BASH_SOURCE), so it works from the repo root and inside the dockcross
+# container whose workdir is the mounted repo root.
+cd "$(dirname "${BASH_SOURCE[0]}")/../llama" || exit 1
+
 mkdir -p build
 
 # Build parallelism. Defaults to all cores; RAM-limited CI runners (notably GitHub's
diff --git a/.github/workflows/clang-format.yml b/.github/workflows/clang-format.yml
index fee925c6..623b6db5 100644
--- a/.github/workflows/clang-format.yml
+++ b/.github/workflows/clang-format.yml
@@ -31,8 +31,9 @@ jobs:
       - name: Check C++ formatting
         run: |
           clang-format --version
-          # All hand-written C++ sources; the generated JNI header (src/main/cpp/jllama.h,
-          # produced by `javac -h`) is intentionally excluded.
-          files=$(find src/main/cpp src/test/cpp -type f \( -name '*.cpp' -o -name '*.hpp' \) | sort)
+          # All hand-written C++ sources; the generated JNI header (llama/src/main/cpp/jllama.h,
+          # produced by `javac -h`) is intentionally excluded. Sources live in the `llama/`
+          # reactor module; clang-format finds llama/.clang-format by walking up from each file.
+          files=$(find llama/src/main/cpp llama/src/test/cpp -type f \( -name '*.cpp' -o -name '*.hpp' \) | sort)
           echo "Checking:"; echo "$files"
           clang-format --dry-run --Werror $files
diff --git a/.github/workflows/publish.yml b/.github/workflows/publish.yml
index b4ec2ad5..0d4eb6ff 100644
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@@ -50,7 +50,7 @@ env:
   # at src/test/resources/images/test-image.jpg (see the README in that
   # directory for licensing). No download step is needed; CI just points
   # mvn test at the committed path.
-  VISION_IMAGE_PATH: "src/test/resources/images/test-image.jpg"
+  VISION_IMAGE_PATH: "llama/src/test/resources/images/test-image.jpg"
 permissions:
   contents: read
 jobs:
@@ -133,25 +133,26 @@ jobs:
           java-version: '21'
           distribution: temurin
       - name: Spotless check (fail fast on format violations)
-        run: mvn -B --no-transfer-progress spotless:check
+        run: mvn -B --no-transfer-progress -f llama/pom.xml spotless:check
       - name: SpotBugs check (fail fast on static-analysis findings)
-        run: mvn -B --no-transfer-progress -DskipTests -Denforcer.skip=true compile spotbugs:check
+        run: mvn -B --no-transfer-progress -f llama/pom.xml -DskipTests -Denforcer.skip=true compile spotbugs:check
       - name: Print internal package dependency graph (jdeps, informational)
         continue-on-error: true
         run: |
-          mvn -B --no-transfer-progress -DskipTests -Denforcer.skip=true compile
+          mvn -B --no-transfer-progress -f llama/pom.xml -DskipTests -Denforcer.skip=true compile
           echo "=== internal package dependency graph (jdeps, bytecode) ==="
-          jdeps -verbose:package target/classes | grep 'net.ladenthin.llama' || true
+          jdeps -verbose:package llama/target/classes | grep 'net.ladenthin.llama' || true
 
   # ---------------------------------------------------------------------------
   # Sibling module `llama-langchain4j` (LangChain4j adapters). Pure Java, no native
   # code and no per-classifier matrix: it compiles against the core's stable Java API
   # (identical across every classifier) and the backend is a runtime choice for the
-  # consumer. This job installs the core Java jar, guards that the module version is in
-  # lockstep with the core, then builds + tests the module (Java 17; langchain4j 1.x
-  # baseline). It runs its mapping unit tests; the model-backed integration test
-  # self-skips without a GGUF. `verify` also builds the javadoc/sources jars so a
-  # release-time javadoc break is caught here in PR CI.
+  # consumer. This job installs the parent + core into the local repo, then builds + tests
+  # the module (Java 17; langchain4j 1.x baseline). It runs its mapping unit tests; the
+  # model-backed integration test self-skips without a GGUF. `verify` also builds the
+  # javadoc/sources jars so a release-time javadoc break is caught here in PR CI. Version
+  # lockstep is now guaranteed by construction (both modules inherit the parent's version),
+  # so the old lockstep guard is gone.
   # ---------------------------------------------------------------------------
 
   test-java-llama-langchain4j:
@@ -164,19 +165,9 @@ jobs:
         with:
           java-version: ${{ env.JAVA_VERSION }}
           distribution: temurin
-      - name: Version lockstep guard (module version must equal core version)
-        shell: bash
-        run: |
-          CORE=$(mvn -q -DforceStdout help:evaluate -Dexpression=project.version | tail -n1)
-          MOD=$(mvn -q -DforceStdout -f llama-langchain4j/pom.xml help:evaluate -Dexpression=project.version | tail -n1)
-          echo "core=$CORE module=$MOD"
-          if [ "$CORE" != "$MOD" ]; then
-            echo "::error::Version drift: core ($CORE) != llama-langchain4j ($MOD). Keep llama-langchain4j/pom.xml <version> in lockstep with the root pom.xml."
-            exit 1
-          fi
-      - name: Install core net.ladenthin:llama jar (Java only)
+      - name: Install parent + core net.ladenthin:llama into the local repo (Java only)
         run: >
-          mvn -B --no-transfer-progress -DskipTests -Denforcer.skip=true
+          mvn -B --no-transfer-progress -pl llama -am -DskipTests -Denforcer.skip=true
           -Dspotless.check.skip=true -Dspotbugs.skip=true
           -Dmaven.javadoc.skip=true -Dmaven.source.skip=true -Dgpg.skip=true install
       - name: Build and test llama-langchain4j
@@ -202,7 +193,7 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: Linux-x86_64-libraries
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
       - name: Restore shared GGUF model cache (populated by download-models; no re-download)
         uses: actions/cache@v6
         with:
@@ -212,9 +203,9 @@ jobs:
         with:
           distribution: 'temurin'
           java-version: ${{ env.JAVA_VERSION }}
-      - name: Install core net.ladenthin:llama jar (bundles the downloaded native library)
+      - name: Install parent + core net.ladenthin:llama (bundles the downloaded native library)
         run: >
-          mvn -B --no-transfer-progress -DskipTests -Denforcer.skip=true
+          mvn -B --no-transfer-progress -pl llama -am -DskipTests -Denforcer.skip=true
           -Dspotless.check.skip=true -Dspotbugs.skip=true
           -Dmaven.javadoc.skip=true -Dmaven.source.skip=true -Dgpg.skip=true install
       - name: Run llama-langchain4j model-backed integration tests (reused cached models)
@@ -242,7 +233,7 @@ jobs:
         id: tag
         shell: bash
         run: |
-          TAG=$(grep -oE 'GIT_TAG[[:space:]]+b[0-9]+' CMakeLists.txt | grep -oE 'b[0-9]+' | head -1)
+          TAG=$(grep -oE 'GIT_TAG[[:space:]]+b[0-9]+' llama/CMakeLists.txt | grep -oE 'b[0-9]+' | head -1)
           if [ -z "$TAG" ]; then
             echo "could not resolve llama.cpp GIT_TAG (b<nnnn>) from CMakeLists.txt" >&2
             exit 1
@@ -284,14 +275,14 @@ jobs:
             done )
           # llama-ui-embed is a self-contained C++17 host tool (no npm) — build + run it.
           g++ -O2 -std=c++17 -o llama-ui-embed embed.cpp
-          mkdir -p "$GITHUB_WORKSPACE/webui-generated"
+          mkdir -p "$GITHUB_WORKSPACE/llama/webui-generated"
           ./llama-ui-embed \
-            "$GITHUB_WORKSPACE/webui-generated/ui.cpp" \
-            "$GITHUB_WORKSPACE/webui-generated/ui.h" \
+            "$GITHUB_WORKSPACE/llama/webui-generated/ui.cpp" \
+            "$GITHUB_WORKSPACE/llama/webui-generated/ui.h" \
             dist
           echo "=== generated WebUI assets ==="
-          ls -la "$GITHUB_WORKSPACE/webui-generated"
-          if grep -q LLAMA_UI_HAS_ASSETS "$GITHUB_WORKSPACE/webui-generated/ui.h"; then
+          ls -la "$GITHUB_WORKSPACE/llama/webui-generated"
+          if grep -q LLAMA_UI_HAS_ASSETS "$GITHUB_WORKSPACE/llama/webui-generated/ui.h"; then
             echo "LLAMA_UI_HAS_ASSETS: present (real WebUI embedded)"
           else
             echo "ERROR: embed produced an empty asset table" >&2
@@ -301,7 +292,7 @@ jobs:
         uses: actions/upload-artifact@v7
         with:
           name: webui-generated
-          path: ${{ github.workspace }}/webui-generated/
+          path: ${{ github.workspace }}/llama/webui-generated/
           retention-days: 1
           if-no-files-found: error
 
@@ -332,7 +323,7 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: webui-generated
-          path: ${{ github.workspace }}/webui-generated/
+          path: ${{ github.workspace }}/llama/webui-generated/
       - name: Display CPU Info
         shell: bash
         run: |
@@ -349,7 +340,7 @@ jobs:
         uses: actions/upload-artifact@v7
         with:
           name: linux-libraries-cuda
-          path: ${{ github.workspace }}/src/main/resources_linux_cuda/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_linux_cuda/net/ladenthin/llama/
 
   crosscompile-linux-x86_64:
     name: Cross-Compile manylinux2014 x86_64
@@ -372,7 +363,7 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: webui-generated
-          path: ${{ github.workspace }}/webui-generated/
+          path: ${{ github.workspace }}/llama/webui-generated/
       - name: Display CPU Info
         shell: bash
         run: |
@@ -389,7 +380,7 @@ jobs:
         uses: actions/upload-artifact@v7
         with:
           name: Linux-x86_64-libraries
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
 
   crosscompile-linux-aarch64:
     name: Build and Test Linux aarch64
@@ -414,7 +405,7 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: webui-generated
-          path: ${{ github.workspace }}/webui-generated/
+          path: ${{ github.workspace }}/llama/webui-generated/
       - uses: actions/setup-java@v5
         with:
           distribution: 'temurin'
@@ -436,15 +427,15 @@ jobs:
       - name: Build libraries
         shell: bash
         run: |
-          mvn --no-transfer-progress compile
+          mvn --no-transfer-progress -f llama/pom.xml compile
           .github/build.sh "-DOS_NAME=Linux -DOS_ARCH=aarch64 -DGGML_NATIVE=OFF -DBUILD_TESTING=ON"
       - name: Run C++ unit tests
-        run: ctest --test-dir build --output-on-failure
+        run: ctest --test-dir llama/build --output-on-failure
       - name: Upload artifacts
         uses: actions/upload-artifact@v7
         with:
           name: Linux-aarch64-libraries
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
 
   crosscompile-android-aarch64:
     name: Cross-Compile Android aarch64
@@ -464,7 +455,7 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: webui-generated
-          path: ${{ github.workspace }}/webui-generated/
+          path: ${{ github.workspace }}/llama/webui-generated/
       - name: Display CPU Info
         shell: bash
         run: |
@@ -481,7 +472,7 @@ jobs:
         uses: actions/upload-artifact@v7
         with:
           name: Linux-Android-aarch64-libraries
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
 
   crosscompile-android-aarch64-opencl:
     name: Cross-Compile Android aarch64 (OpenCL/Adreno)
@@ -502,7 +493,7 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: webui-generated
-          path: ${{ github.workspace }}/webui-generated/
+          path: ${{ github.workspace }}/llama/webui-generated/
       - name: Build libraries
         shell: bash
         run: |
@@ -511,7 +502,7 @@ jobs:
         uses: actions/upload-artifact@v7
         with:
           name: android-libraries-opencl
-          path: ${{ github.workspace }}/src/main/resources_android_opencl/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_android_opencl/net/ladenthin/llama/
 
   # ---------------------------------------------------------------------------
   # Native build jobs — produce release artifacts + run C++ unit tests
@@ -532,7 +523,7 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: webui-generated
-          path: ${{ github.workspace }}/webui-generated/
+          path: ${{ github.workspace }}/llama/webui-generated/
       - uses: actions/setup-java@v5
         with:
           distribution: 'temurin'
@@ -552,15 +543,15 @@ jobs:
       - name: Build libraries
         shell: bash
         run: |
-          mvn --no-transfer-progress compile
+          mvn --no-transfer-progress -f llama/pom.xml compile
           .github/build.sh -DLLAMA_METAL=OFF -DGGML_NATIVE=OFF -DBUILD_TESTING=ON
       - name: Run C++ unit tests
-        run: ctest --test-dir build --output-on-failure
+        run: ctest --test-dir llama/build --output-on-failure
       - name: Upload artifacts
         uses: actions/upload-artifact@v7
         with:
           name: macos-15-libraries
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
 
   build-macos-arm64-metal:
     name: Build and Test macOS 14 arm64 (Metal)
@@ -577,7 +568,7 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: webui-generated
-          path: ${{ github.workspace }}/webui-generated/
+          path: ${{ github.workspace }}/llama/webui-generated/
       - uses: actions/setup-java@v5
         with:
           distribution: 'temurin'
@@ -597,15 +588,15 @@ jobs:
       - name: Build libraries
         shell: bash
         run: |
-          mvn --no-transfer-progress compile
+          mvn --no-transfer-progress -f llama/pom.xml compile
           .github/build.sh -DLLAMA_METAL_EMBED_LIBRARY=ON -DBUILD_TESTING=ON
       - name: Run C++ unit tests
-        run: ctest --test-dir build --output-on-failure
+        run: ctest --test-dir llama/build --output-on-failure
       - name: Upload artifacts
         uses: actions/upload-artifact@v7
         with:
           name: macos-14-libraries
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
 
   build-windows-x86_64-msvc:
     name: Build and Test Windows 2025 x86_64 (MSVC / VS 2026, classifier)
@@ -617,7 +608,7 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: webui-generated
-          path: ${{ github.workspace }}/webui-generated/
+          path: ${{ github.workspace }}/llama/webui-generated/
       - name: Display CPU Info
         shell: pwsh
         run: |
@@ -634,12 +625,12 @@ jobs:
         run: |
           .github\build.bat -G "Visual Studio 18 2026" -A "x64" -DOS_NAME=Windows -DOS_ARCH=x86_64 -DBUILD_TESTING=ON
       - name: Run C++ unit tests
-        run: ctest --test-dir build --output-on-failure
+        run: ctest --test-dir llama/build --output-on-failure
       - name: Upload artifacts
         uses: actions/upload-artifact@v7
         with:
           name: Windows-x86_64-msvc
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
 
   build-windows-x86-msvc:
     name: Build and Test Windows 2025 x86 (MSVC / VS 2026, classifier)
@@ -651,7 +642,7 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: webui-generated
-          path: ${{ github.workspace }}/webui-generated/
+          path: ${{ github.workspace }}/llama/webui-generated/
       - name: Display CPU Info
         shell: pwsh
         run: |
@@ -668,12 +659,12 @@ jobs:
         run: |
           .github\build.bat -G "Visual Studio 18 2026" -A "Win32" -DOS_NAME=Windows -DOS_ARCH=x86 -DBUILD_TESTING=ON
       - name: Run C++ unit tests
-        run: ctest --test-dir build --output-on-failure
+        run: ctest --test-dir llama/build --output-on-failure
       - name: Upload artifacts
         uses: actions/upload-artifact@v7
         with:
           name: Windows-x86-msvc
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
 
   # ---------------------------------------------------------------------------
   # Windows Ninja Multi-Config + sccache — the DEFAULT Windows CPU natives.
@@ -703,7 +694,7 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: webui-generated
-          path: ${{ github.workspace }}/webui-generated/
+          path: ${{ github.workspace }}/llama/webui-generated/
       - name: Set up MSVC developer environment (x64)
         uses: ilammy/msvc-dev-cmd@v1
         with:
@@ -736,12 +727,12 @@ jobs:
         run: |
           .github\build.bat -G "Ninja Multi-Config" -DOS_NAME=Windows -DOS_ARCH=x86_64 -DBUILD_TESTING=ON
       - name: Run C++ unit tests
-        run: ctest --test-dir build --output-on-failure
+        run: ctest --test-dir llama/build --output-on-failure
       - name: Upload artifacts
         uses: actions/upload-artifact@v7
         with:
           name: Windows-x86_64-libraries
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
 
   build-windows-x86:
     name: Build and Test Windows 2025 x86 (Ninja Multi-Config + sccache, default)
@@ -757,7 +748,7 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: webui-generated
-          path: ${{ github.workspace }}/webui-generated/
+          path: ${{ github.workspace }}/llama/webui-generated/
       - name: Set up MSVC developer environment (x86)
         uses: ilammy/msvc-dev-cmd@v1
         with:
@@ -790,12 +781,12 @@ jobs:
         run: |
           .github\build.bat -G "Ninja Multi-Config" -DOS_NAME=Windows -DOS_ARCH=x86 -DBUILD_TESTING=ON
       - name: Run C++ unit tests
-        run: ctest --test-dir build --output-on-failure
+        run: ctest --test-dir llama/build --output-on-failure
       - name: Upload artifacts
         uses: actions/upload-artifact@v7
         with:
           name: Windows-x86-libraries
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
 
   # ---------------------------------------------------------------------------
   # Windows GPU classifiers (x86_64 only) — CUDA, Vulkan, OpenCL.
@@ -826,7 +817,7 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: webui-generated
-          path: ${{ github.workspace }}/webui-generated/
+          path: ${{ github.workspace }}/llama/webui-generated/
       - name: Set up MSVC developer environment (x64)
         uses: ilammy/msvc-dev-cmd@v1
         with:
@@ -864,7 +855,7 @@ jobs:
         uses: actions/upload-artifact@v7
         with:
           name: Windows-x86_64-cuda
-          path: ${{ github.workspace }}/src/main/resources_windows_cuda/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_cuda/net/ladenthin/llama/
           if-no-files-found: error
 
   build-windows-x86_64-vulkan:
@@ -881,7 +872,7 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: webui-generated
-          path: ${{ github.workspace }}/webui-generated/
+          path: ${{ github.workspace }}/llama/webui-generated/
       - name: Set up MSVC developer environment (x64)
         uses: ilammy/msvc-dev-cmd@v1
         with:
@@ -913,7 +904,7 @@ jobs:
         uses: actions/upload-artifact@v7
         with:
           name: Windows-x86_64-vulkan
-          path: ${{ github.workspace }}/src/main/resources_windows_vulkan/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_vulkan/net/ladenthin/llama/
           if-no-files-found: error
 
   build-windows-x86_64-opencl:
@@ -930,7 +921,7 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: webui-generated
-          path: ${{ github.workspace }}/webui-generated/
+          path: ${{ github.workspace }}/llama/webui-generated/
       - name: Set up MSVC developer environment (x64)
         uses: ilammy/msvc-dev-cmd@v1
         with:
@@ -957,7 +948,7 @@ jobs:
         uses: actions/upload-artifact@v7
         with:
           name: Windows-x86_64-opencl
-          path: ${{ github.workspace }}/src/main/resources_windows_opencl/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_opencl/net/ladenthin/llama/
           if-no-files-found: error
 
   # ---------------------------------------------------------------------------
@@ -983,10 +974,10 @@ jobs:
           cat /proc/cpuinfo
       - name: Build libraries
         run: |
-          mvn -q --no-transfer-progress compile
+          mvn -q --no-transfer-progress -f llama/pom.xml compile
           .github/build.sh -DBUILD_TESTING=ON
       - name: Run C++ unit tests
-        run: ctest --test-dir build --output-on-failure
+        run: ctest --test-dir llama/build --output-on-failure
 
   build-macos-arm64-metal-15:
     name: Build and Test macOS 15 arm64 (Metal)
@@ -1003,7 +994,7 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: webui-generated
-          path: ${{ github.workspace }}/webui-generated/
+          path: ${{ github.workspace }}/llama/webui-generated/
       - uses: actions/setup-java@v5
         with:
           distribution: 'temurin'
@@ -1023,15 +1014,15 @@ jobs:
       - name: Build libraries
         shell: bash
         run: |
-          mvn --no-transfer-progress compile
+          mvn --no-transfer-progress -f llama/pom.xml compile
           .github/build.sh -DLLAMA_METAL_EMBED_LIBRARY=ON -DGGML_NATIVE=OFF -DBUILD_TESTING=ON
       - name: Run C++ unit tests
-        run: ctest --test-dir build --output-on-failure
+        run: ctest --test-dir llama/build --output-on-failure
       - name: Upload artifacts
         uses: actions/upload-artifact@v7
         with:
           name: macos-15-metal-libraries
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
 
   # ---------------------------------------------------------------------------
   # Java test jobs — download release artifact, run mvn test
@@ -1054,7 +1045,7 @@ jobs:
       - uses: actions/download-artifact@v8
         with:
           name: Linux-x86_64-libraries
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
       # GGUF models are downloaded + cached ONCE by the upstream `download-models` job
       # (this job `needs:` it), so here we only RESTORE the shared cache — no per-job
       # download logic. GGUF is platform-independent, so ubuntu + macOS + Windows share
@@ -1079,7 +1070,7 @@ jobs:
           echo "${{ github.workspace }}/core.%e.%p" | sudo tee /proc/sys/kernel/core_pattern
       - name: Run tests
         run: |
-          mvn -e --no-transfer-progress -P jcstress test \
+          mvn -e --no-transfer-progress -f llama/pom.xml -P jcstress test \
             -Dnet.ladenthin.llama.tool.model=models/${TOOL_MODEL_NAME} \
             -Dnet.ladenthin.llama.nomic.path=models/${NOMIC_EMBED_MODEL_NAME} \
             -Dnet.ladenthin.llama.vision.model=models/${VISION_MODEL_NAME} \
@@ -1091,15 +1082,15 @@ jobs:
         if: success()
         with:
           name: jacoco-report
-          path: target/site/jacoco/jacoco.xml
+          path: llama/target/site/jacoco/jacoco.xml
           if-no-files-found: ignore
       - name: Run PIT mutation tests
-        run: mvn --batch-mode --no-transfer-progress test-compile org.pitest:pitest-maven:mutationCoverage
+        run: mvn --batch-mode --no-transfer-progress -f llama/pom.xml test-compile org.pitest:pitest-maven:mutationCoverage
       - name: Extract PIT survivors
         if: always()
         run: |
           echo "=== PIT Survived Mutations ==="
-          for html_file in $(find target/pit-reports -name "*.html" -type f 2>/dev/null | sort); do
+          for html_file in $(find llama/target/pit-reports -name "*.html" -type f 2>/dev/null | sort); do
             if grep -q "SURVIVED" "$html_file"; then
               echo "Found survivors in $html_file:"
               grep -B 2 -A 3 "SURVIVED" "$html_file"
@@ -1108,7 +1099,7 @@ jobs:
           done
       - uses: actions/upload-artifact@v7
         if: always()
-        with: { name: pit-reports, path: target/pit-reports/ }
+        with: { name: pit-reports, path: llama/target/pit-reports/ }
       - name: Memory after tests
         if: always()
         run: free -h
@@ -1146,13 +1137,13 @@ jobs:
         # -Dtest matches simple class names, not package globs; the default suite is
         # excluded from the vmlens package via pom.xml managed surefire <excludes>).
         run: >-
-          mvn --batch-mode --no-transfer-progress -Pvmlens test
+          mvn --batch-mode --no-transfer-progress -f llama/pom.xml -Pvmlens test
           -Dtest=VmlensInterleavingSmokeTest,SessionStateInterleavingTest -DfailIfNoTests=false
       - uses: actions/upload-artifact@v7
         if: always()
         with:
           name: vmlens-report
-          path: target/vmlens-report/
+          path: llama/target/vmlens-report/
           if-no-files-found: ignore
 
   test-java-macos-arm64-metal:
@@ -1172,7 +1163,7 @@ jobs:
       - uses: actions/download-artifact@v8
         with:
           name: macos-14-libraries
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
       # GGUF models are downloaded + cached ONCE by the upstream `download-models` job
       # (this job `needs:` it), so here we only RESTORE the shared cache — no per-job
       # download logic. GGUF is platform-independent, so ubuntu + macOS + Windows share
@@ -1195,7 +1186,7 @@ jobs:
         run: ulimit -c unlimited
       - name: Run tests
         run: |
-          mvn -e --no-transfer-progress -Dnet.ladenthin.llama.test.ngl=0 test \
+          mvn -e --no-transfer-progress -f llama/pom.xml -Dnet.ladenthin.llama.test.ngl=0 test \
             -Dnet.ladenthin.llama.tool.model=models/${TOOL_MODEL_NAME} \
             -Dnet.ladenthin.llama.nomic.path=models/${NOMIC_EMBED_MODEL_NAME} \
             -Dnet.ladenthin.llama.vision.model=models/${VISION_MODEL_NAME} \
@@ -1236,7 +1227,7 @@ jobs:
       - uses: actions/download-artifact@v8
         with:
           name: macos-15-libraries
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
       # GGUF models are downloaded + cached ONCE by the upstream `download-models` job
       # (this job `needs:` it), so here we only RESTORE the shared cache — no per-job
       # download logic. GGUF is platform-independent, so ubuntu + macOS + Windows share
@@ -1259,7 +1250,7 @@ jobs:
         run: ulimit -c unlimited
       - name: Run tests
         run: |
-          mvn -e --no-transfer-progress test \
+          mvn -e --no-transfer-progress -f llama/pom.xml test \
             -Dnet.ladenthin.llama.tool.model=models/${TOOL_MODEL_NAME} \
             -Dnet.ladenthin.llama.nomic.path=models/${NOMIC_EMBED_MODEL_NAME} \
             -Dnet.ladenthin.llama.vision.model=models/${VISION_MODEL_NAME} \
@@ -1300,7 +1291,7 @@ jobs:
       - uses: actions/download-artifact@v8
         with:
           name: macos-15-metal-libraries
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
       # GGUF models are downloaded + cached ONCE by the upstream `download-models` job
       # (this job `needs:` it), so here we only RESTORE the shared cache — no per-job
       # download logic. GGUF is platform-independent, so ubuntu + macOS + Windows share
@@ -1323,7 +1314,7 @@ jobs:
         run: ulimit -c unlimited
       - name: Run tests
         run: |
-          mvn -e --no-transfer-progress test \
+          mvn -e --no-transfer-progress -f llama/pom.xml test \
             -Dnet.ladenthin.llama.tool.model=models/${TOOL_MODEL_NAME} \
             -Dnet.ladenthin.llama.nomic.path=models/${NOMIC_EMBED_MODEL_NAME} \
             -Dnet.ladenthin.llama.vision.model=models/${VISION_MODEL_NAME} \
@@ -1367,7 +1358,7 @@ jobs:
       - uses: actions/download-artifact@v8
         with:
           name: Windows-x86_64-libraries
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
       # GGUF models are downloaded + cached ONCE by the upstream `download-models` job
       # (this job `needs:` it), so here we only RESTORE the shared cache — no per-job
       # download logic. validate-models is kept as an integrity guard so a partial/absent
@@ -1405,7 +1396,7 @@ jobs:
           Get-ItemProperty -Path $key | Format-List
       - name: Run tests
         run: |
-          mvn -e --no-transfer-progress test `
+          mvn -e --no-transfer-progress -f llama/pom.xml test `
             "-Dnet.ladenthin.llama.tool.model=models/$env:TOOL_MODEL_NAME" `
             "-Dnet.ladenthin.llama.nomic.path=models/$env:NOMIC_EMBED_MODEL_NAME" `
             "-Dnet.ladenthin.llama.vision.model=models/$env:VISION_MODEL_NAME" `
@@ -1429,7 +1420,7 @@ jobs:
             ${{ github.workspace }}\target\surefire-reports\*.dumpstream
             ${{ github.workspace }}\target\surefire-reports\*.txt
             ${{ github.workspace }}\target\surefire-reports\TEST-*.xml
-            ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/**/*
+            ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/**/*
           if-no-files-found: warn
 
   # Java/inference validation of the MSVC-built x86_64 DLL (the analogue of
@@ -1456,7 +1447,7 @@ jobs:
       - uses: actions/download-artifact@v8
         with:
           name: Windows-x86_64-msvc
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
       # GGUF models are downloaded + cached ONCE by the upstream `download-models` job
       # (this job `needs:` it), so here we only RESTORE the shared cache — no per-job
       # download logic. validate-models is kept as an integrity guard so a partial/absent
@@ -1494,7 +1485,7 @@ jobs:
           Get-ItemProperty -Path $key | Format-List
       - name: Run tests
         run: |
-          mvn -e --no-transfer-progress test `
+          mvn -e --no-transfer-progress -f llama/pom.xml test `
             "-Dnet.ladenthin.llama.tool.model=models/$env:TOOL_MODEL_NAME" `
             "-Dnet.ladenthin.llama.nomic.path=models/$env:NOMIC_EMBED_MODEL_NAME" `
             "-Dnet.ladenthin.llama.vision.model=models/$env:VISION_MODEL_NAME" `
@@ -1518,7 +1509,7 @@ jobs:
             ${{ github.workspace }}\target\surefire-reports\*.dumpstream
             ${{ github.workspace }}\target\surefire-reports\*.txt
             ${{ github.workspace }}\target\surefire-reports\TEST-*.xml
-            ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/**/*
+            ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/**/*
           if-no-files-found: warn
 
   # ---------------------------------------------------------------------------
@@ -1554,38 +1545,38 @@ jobs:
         with:
           pattern: "*-libraries"
           merge-multiple: true
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: linux-libraries-cuda
-          path: ${{ github.workspace }}/src/main/resources_linux_cuda/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_linux_cuda/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: android-libraries-opencl
-          path: ${{ github.workspace }}/src/main/resources_android_opencl/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_android_opencl/net/ladenthin/llama/
       # MSVC-built Windows natives -> `msvc-windows` classifier tree. The default JAR
       # now ships the Ninja `*-libraries` natives merged above (default flip).
       - uses: actions/download-artifact@v8
         with:
           name: Windows-x86_64-msvc
-          path: ${{ github.workspace }}/src/main/resources_windows_msvc/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_msvc/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: Windows-x86-msvc
-          path: ${{ github.workspace }}/src/main/resources_windows_msvc/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_msvc/net/ladenthin/llama/
       # Windows GPU classifiers (x86_64 only) -> one tree each.
       - uses: actions/download-artifact@v8
         with:
           name: Windows-x86_64-cuda
-          path: ${{ github.workspace }}/src/main/resources_windows_cuda/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_cuda/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: Windows-x86_64-vulkan
-          path: ${{ github.workspace }}/src/main/resources_windows_vulkan/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_vulkan/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: Windows-x86_64-opencl
-          path: ${{ github.workspace }}/src/main/resources_windows_opencl/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_opencl/net/ladenthin/llama/
       - uses: actions/setup-java@v5
         with:
           distribution: 'temurin'
@@ -1604,7 +1595,7 @@ jobs:
         uses: actions/upload-artifact@v7
         with:
           name: llama-jars
-          path: target/*.jar
+          path: llama/target/*.jar
 
   report:
     name: Report
@@ -1668,35 +1659,35 @@ jobs:
         with:
           pattern: "*-libraries"
           merge-multiple: true
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: linux-libraries-cuda
-          path: ${{ github.workspace }}/src/main/resources_linux_cuda/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_linux_cuda/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: android-libraries-opencl
-          path: ${{ github.workspace }}/src/main/resources_android_opencl/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_android_opencl/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: Windows-x86_64-msvc
-          path: ${{ github.workspace }}/src/main/resources_windows_msvc/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_msvc/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: Windows-x86-msvc
-          path: ${{ github.workspace }}/src/main/resources_windows_msvc/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_msvc/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: Windows-x86_64-cuda
-          path: ${{ github.workspace }}/src/main/resources_windows_cuda/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_cuda/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: Windows-x86_64-vulkan
-          path: ${{ github.workspace }}/src/main/resources_windows_vulkan/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_vulkan/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: Windows-x86_64-opencl
-          path: ${{ github.workspace }}/src/main/resources_windows_opencl/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_opencl/net/ladenthin/llama/
       - name: Set up Maven Central Repository
         uses: actions/setup-java@v5
         with:
@@ -1716,27 +1707,21 @@ jobs:
             *-SNAPSHOT) echo "OK: -SNAPSHOT version, continuing snapshot deploy." ;;
             *) echo "::error::Refusing to publish non-SNAPSHOT version '$VERSION' from the snapshot job. Snapshot publishing requires a -SNAPSHOT version; releases go through the v* tag path."; exit 1 ;;
           esac
-      - name: Publish snapshot
+      # One reactor deploy publishes all three artifacts at the same version:
+      # net.ladenthin:llama-parent (the pom), :llama (the core jar + classifiers), and
+      # :llama-langchain4j. The `release` profile (GPG + Central Publishing) is inherited
+      # from the parent, so every module — including the parent pom — is signed.
+      - name: Publish snapshot (reactor - parent + llama + llama-langchain4j)
         run: mvn --batch-mode --no-transfer-progress -P release,cuda,opencl-android,windows-msvc,cuda-windows,vulkan-windows,opencl-windows -Dmaven.test.skip=true deploy
         env:
           MAVEN_USERNAME: ${{ secrets.CENTRAL_USERNAME }}
           MAVEN_PASSWORD: ${{ secrets.CENTRAL_TOKEN }}
           MAVEN_GPG_PASSPHRASE: ${{ secrets.GPG_PASSPHRASE }}
-      # Deploy the sibling llama-langchain4j at the same version. The core `deploy`
-      # above ran through `install`, so net.ladenthin:llama is in the local repo for
-      # the module to resolve. Standalone module (not in the reactor), so it is a
-      # separate deploy invocation.
-      - name: Publish snapshot (llama-langchain4j)
-        run: mvn --batch-mode --no-transfer-progress -f llama-langchain4j/pom.xml -P release -Dmaven.test.skip=true deploy
-        env:
-          MAVEN_USERNAME: ${{ secrets.CENTRAL_USERNAME }}
-          MAVEN_PASSWORD: ${{ secrets.CENTRAL_TOKEN }}
-          MAVEN_GPG_PASSPHRASE: ${{ secrets.GPG_PASSPHRASE }}
       - name: Collect signed artifacts
         run: |
           mkdir -p signed-snapshot-assets
-          cp target/*.jar signed-snapshot-assets/ 2>/dev/null || true
-          cp target/*.jar.asc signed-snapshot-assets/ 2>/dev/null || true
+          cp llama/target/*.jar signed-snapshot-assets/ 2>/dev/null || true
+          cp llama/target/*.jar.asc signed-snapshot-assets/ 2>/dev/null || true
           cp llama-langchain4j/target/*.jar signed-snapshot-assets/ 2>/dev/null || true
           cp llama-langchain4j/target/*.jar.asc signed-snapshot-assets/ 2>/dev/null || true
       - uses: actions/upload-artifact@v7
@@ -1784,35 +1769,35 @@ jobs:
         with:
           pattern: "*-libraries"
           merge-multiple: true
-          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: linux-libraries-cuda
-          path: ${{ github.workspace }}/src/main/resources_linux_cuda/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_linux_cuda/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: android-libraries-opencl
-          path: ${{ github.workspace }}/src/main/resources_android_opencl/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_android_opencl/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: Windows-x86_64-msvc
-          path: ${{ github.workspace }}/src/main/resources_windows_msvc/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_msvc/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: Windows-x86-msvc
-          path: ${{ github.workspace }}/src/main/resources_windows_msvc/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_msvc/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: Windows-x86_64-cuda
-          path: ${{ github.workspace }}/src/main/resources_windows_cuda/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_cuda/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: Windows-x86_64-vulkan
-          path: ${{ github.workspace }}/src/main/resources_windows_vulkan/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_vulkan/net/ladenthin/llama/
       - uses: actions/download-artifact@v8
         with:
           name: Windows-x86_64-opencl
-          path: ${{ github.workspace }}/src/main/resources_windows_opencl/net/ladenthin/llama/
+          path: ${{ github.workspace }}/llama/src/main/resources_windows_opencl/net/ladenthin/llama/
       - name: Set up Maven Central Repository
         uses: actions/setup-java@v5
         with:
@@ -1823,27 +1808,21 @@ jobs:
           server-password: MAVEN_PASSWORD
           gpg-private-key: ${{ secrets.GPG_PRIVATE_KEY }}
           gpg-passphrase: MAVEN_GPG_PASSPHRASE
-      - name: Publish release
+      # One reactor deploy publishes all three artifacts at the same version:
+      # net.ladenthin:llama-parent (the pom), :llama (the core jar + classifiers), and
+      # :llama-langchain4j. The `release` profile (GPG + Central Publishing) is inherited
+      # from the parent, so every module — including the parent pom — is signed.
+      - name: Publish release (reactor - parent + llama + llama-langchain4j)
         run: mvn --batch-mode --no-transfer-progress -P release,cuda,opencl-android,windows-msvc,cuda-windows,vulkan-windows,opencl-windows -Dmaven.test.skip=true deploy
         env:
           MAVEN_USERNAME: ${{ secrets.CENTRAL_USERNAME }}
           MAVEN_PASSWORD: ${{ secrets.CENTRAL_TOKEN }}
           MAVEN_GPG_PASSPHRASE: ${{ secrets.GPG_PASSPHRASE }}
-      # Deploy the sibling llama-langchain4j at the same version. The core `deploy`
-      # above ran through `install`, so net.ladenthin:llama is in the local repo for
-      # the module to resolve. Standalone module (not in the reactor), so it is a
-      # separate deploy invocation.
-      - name: Publish release (llama-langchain4j)
-        run: mvn --batch-mode --no-transfer-progress -f llama-langchain4j/pom.xml -P release -Dmaven.test.skip=true deploy
-        env:
-          MAVEN_USERNAME: ${{ secrets.CENTRAL_USERNAME }}
-          MAVEN_PASSWORD: ${{ secrets.CENTRAL_TOKEN }}
-          MAVEN_GPG_PASSPHRASE: ${{ secrets.GPG_PASSPHRASE }}
       - name: Collect signed artifacts
         run: |
           mkdir -p signed-release-assets
-          cp target/*.jar signed-release-assets/ 2>/dev/null || true
-          cp target/*.jar.asc signed-release-assets/ 2>/dev/null || true
+          cp llama/target/*.jar signed-release-assets/ 2>/dev/null || true
+          cp llama/target/*.jar.asc signed-release-assets/ 2>/dev/null || true
           cp llama-langchain4j/target/*.jar signed-release-assets/ 2>/dev/null || true
           cp llama-langchain4j/target/*.jar.asc signed-release-assets/ 2>/dev/null || true
       - uses: actions/upload-artifact@v7
diff --git a/.github/workflows/sonarqube.yml b/.github/workflows/sonarqube.yml
index b6a4808b..2ee66a0a 100644
--- a/.github/workflows/sonarqube.yml
+++ b/.github/workflows/sonarqube.yml
@@ -37,4 +37,4 @@ jobs:
       - name: Build and analyze
         env:
           SONAR_TOKEN: ${{ secrets.SONAR_TOKEN }}
-        run: mvn -B --no-transfer-progress verify org.sonarsource.scanner.maven:sonar-maven-plugin:sonar -Dsonar.projectKey=bernardladenthin_java-llama.cpp
+        run: mvn -B --no-transfer-progress -f llama/pom.xml verify org.sonarsource.scanner.maven:sonar-maven-plugin:sonar -Dsonar.projectKey=bernardladenthin_java-llama.cpp
diff --git a/.gitignore b/.gitignore
index be02ca4b..ba0a69ed 100644
--- a/.gitignore
+++ b/.gitignore
@@ -38,28 +38,28 @@ hs_err_pid*
 replay_pid*
 
 models/*.gguf
-src/main/cpp/net_ladenthin_llama_*.h
-src/main/resources_cuda_linux/
+llama/src/main/cpp/net_ladenthin_llama_*.h
+llama/src/main/resources_cuda_linux/
 # Per-classifier native trees, staged by CI before the matching Maven profile runs,
 # never committed (same policy as the default-tree native libs below).
-src/main/resources_windows_msvc/
-src/main/resources_windows_cuda/
-src/main/resources_windows_vulkan/
-src/main/resources_windows_opencl/
-src/main/resources/**/*.so
-src/main/resources/**/*.dylib
-src/main/resources/**/*.dll
-src/main/resources/**/*.metal
-src/test/resources/**/*.gbnf
+llama/src/main/resources_windows_msvc/
+llama/src/main/resources_windows_cuda/
+llama/src/main/resources_windows_vulkan/
+llama/src/main/resources_windows_opencl/
+llama/src/main/resources/**/*.so
+llama/src/main/resources/**/*.dylib
+llama/src/main/resources/**/*.dll
+llama/src/main/resources/**/*.metal
+llama/src/test/resources/**/*.gbnf
 
 # Generated WebUI assets (ui.cpp/ui.h) produced once by the build-webui CI job and
 # downloaded into every native build; embedded into libjllama, never committed
 # (this repo commits no build outputs — same policy as the native libs above).
-/webui-generated/
+/llama/webui-generated/
 
 **/*.etag
 **/*.lastModified
-src/main/cpp/llama.cpp/
+llama/src/main/cpp/llama.cpp/
 
 # jcstress / jqwik test outputs (generated in repo root)
 /.jqwik-database
diff --git a/REUSE.toml b/REUSE.toml
index df4c5422..55720a91 100644
--- a/REUSE.toml
+++ b/REUSE.toml
@@ -19,7 +19,7 @@ path = [
     "docs/history/49be664_open_issues_comments.md",
     "docs/history/CHAT_INTEGRATION_SUMMARY.md",
     "docs/history/REFACTORING.md",
-    "src/test/resources/images/README.md",
+    "llama/src/test/resources/images/README.md",
     ".github/PULL_REQUEST_TEMPLATE.md",
     ".github/ISSUE_TEMPLATE/bug_report.md",
     ".github/ISSUE_TEMPLATE/feature_request.md",
@@ -43,7 +43,7 @@ SPDX-License-Identifier = "MIT"
 
 # CMakeLists.txt (CMake format not recognized by reuse annotate)
 [[annotations]]
-path = "CMakeLists.txt"
+path = "llama/CMakeLists.txt"
 SPDX-FileCopyrightText = [
     "2023-2025 Konstantin Herud",
     "2026 Bernard Ladenthin <bernard.ladenthin@gmail.com>",
@@ -79,13 +79,13 @@ SPDX-License-Identifier = "MIT"
 # diff). Glob covers every current and future patch dropped into patches/ — see CLAUDE.md
 # "Local llama.cpp source patches".
 [[annotations]]
-path = "patches/**"
+path = "llama/patches/**"
 SPDX-FileCopyrightText = "2026 Bernard Ladenthin <bernard.ladenthin@gmail.com>"
 SPDX-License-Identifier = "MIT"
 
 # Test image (binary, cannot carry inline SPDX)
 [[annotations]]
-path = "src/test/resources/images/test-image.jpg"
+path = "llama/src/test/resources/images/test-image.jpg"
 SPDX-FileCopyrightText = "2026 Bernard Ladenthin <bernard.ladenthin@gmail.com>"
 SPDX-License-Identifier = "MIT"
 
diff --git a/llama-langchain4j/pom.xml b/llama-langchain4j/pom.xml
index 45ca9944..dabee5b2 100644
--- a/llama-langchain4j/pom.xml
+++ b/llama-langchain4j/pom.xml
@@ -66,8 +66,6 @@ SPDX-License-Identifier: MIT
 		<surefire.version>3.5.6</surefire.version>
 		<source.plugin.version>3.4.0</source.plugin.version>
 		<javadoc.plugin.version>3.12.0</javadoc.plugin.version>
-		<gpg.plugin.version>3.2.8</gpg.plugin.version>
-		<central.plugin.version>0.11.0</central.plugin.version>
 	</properties>
 
 	<dependencies>
@@ -152,47 +150,9 @@ SPDX-License-Identifier: MIT
 		</plugins>
 	</build>
 
-	<profiles>
-		<!-- Mirrors the core pom.xml `release` profile: GPG-sign the artifacts and route the
-		     deploy through the Central Publishing plugin. Activated by the publish-snapshot /
-		     publish-release CI jobs (`-P release ... deploy`). -->
-		<profile>
-			<id>release</id>
-			<build>
-				<plugins>
-					<plugin>
-						<groupId>org.apache.maven.plugins</groupId>
-						<artifactId>maven-gpg-plugin</artifactId>
-						<version>${gpg.plugin.version}</version>
-						<executions>
-							<execution>
-								<id>sign-artifacts</id>
-								<phase>verify</phase>
-								<goals><goal>sign</goal></goals>
-								<configuration>
-									<keyname>${gpg.keyname}</keyname>
-									<gpgArguments>
-										<arg>--pinentry-mode</arg>
-										<arg>loopback</arg>
-									</gpgArguments>
-								</configuration>
-							</execution>
-						</executions>
-					</plugin>
-
-					<plugin>
-						<groupId>org.sonatype.central</groupId>
-						<artifactId>central-publishing-maven-plugin</artifactId>
-						<version>${central.plugin.version}</version>
-						<extensions>true</extensions>
-						<configuration>
-							<publishingServerId>central</publishingServerId>
-							<autoPublish>true</autoPublish>
-							<waitUntil>published</waitUntil>
-						</configuration>
-					</plugin>
-				</plugins>
-			</build>
-		</profile>
-	</profiles>
+	<!--
+		No `release` profile here: GPG signing + Central Publishing are inherited from the
+		parent pom (../pom.xml), so a single reactor `mvn -P release deploy` signs and
+		publishes this module alongside the core.
+	-->
 </project>
diff --git a/llama/pom.xml b/llama/pom.xml
index 4291fcaf..73b91a09 100644
--- a/llama/pom.xml
+++ b/llama/pom.xml
@@ -718,43 +718,12 @@ SPDX-License-Identifier: MIT
 	</build>
 
 	<profiles>
-		<profile>
-			<id>release</id>
-			<build>
-				<plugins>
-					<plugin>
-						<groupId>org.apache.maven.plugins</groupId>
-						<artifactId>maven-gpg-plugin</artifactId>
-						<executions>
-							<execution>
-								<id>sign-artifacts</id>
-								<phase>verify</phase>
-								<goals><goal>sign</goal></goals>
-								<configuration>
-									<keyname>${gpg.keyname}</keyname>
-									<gpgArguments>
-										<arg>--pinentry-mode</arg>
-										<arg>loopback</arg>
-									</gpgArguments>
-								</configuration>
-							</execution>
-						</executions>
-					</plugin>
-
-					<plugin>
-						<groupId>org.sonatype.central</groupId>
-						<artifactId>central-publishing-maven-plugin</artifactId>
-						<extensions>true</extensions>
-						<configuration>
-							<publishingServerId>central</publishingServerId>
-							<autoPublish>true</autoPublish>
-							<waitUntil>published</waitUntil>
-						</configuration>
-					</plugin>
-				</plugins>
-			</build>
-		</profile>
-
+		<!--
+			The `release` profile (GPG signing + Central Publishing) now lives in the parent
+			pom (../pom.xml) and is inherited, so a single reactor `mvn -P release deploy`
+			signs and publishes every module. The per-classifier native profiles below stay
+			here — they are core-specific.
+		-->
 		<profile>
 			<id>cuda</id>
 			<build>
diff --git a/pom.xml b/pom.xml
index b136d4d8..c73ba70d 100644
--- a/pom.xml
+++ b/pom.xml
@@ -59,4 +59,53 @@ SPDX-License-Identifier: MIT
 		<module>llama</module>
 		<module>llama-langchain4j</module>
 	</modules>
+
+	<profiles>
+		<!--
+			Reactor-wide release profile. Inherited by every module (llama, llama-langchain4j)
+			AND applied to this parent pom itself, so a single `mvn -P release deploy` GPG-signs
+			all artifacts (including the parent .pom, which Central requires) and routes the whole
+			reactor through the Central Publishing plugin as one deployment. Activated only by the
+			publish-snapshot / publish-release CI jobs. Plugin versions are pinned here explicitly
+			(the parent has no pluginManagement to inherit them from).
+		-->
+		<profile>
+			<id>release</id>
+			<build>
+				<plugins>
+					<plugin>
+						<groupId>org.apache.maven.plugins</groupId>
+						<artifactId>maven-gpg-plugin</artifactId>
+						<version>3.2.8</version>
+						<executions>
+							<execution>
+								<id>sign-artifacts</id>
+								<phase>verify</phase>
+								<goals><goal>sign</goal></goals>
+								<configuration>
+									<keyname>${gpg.keyname}</keyname>
+									<gpgArguments>
+										<arg>--pinentry-mode</arg>
+										<arg>loopback</arg>
+									</gpgArguments>
+								</configuration>
+							</execution>
+						</executions>
+					</plugin>
+
+					<plugin>
+						<groupId>org.sonatype.central</groupId>
+						<artifactId>central-publishing-maven-plugin</artifactId>
+						<version>0.11.0</version>
+						<extensions>true</extensions>
+						<configuration>
+							<publishingServerId>central</publishingServerId>
+							<autoPublish>true</autoPublish>
+							<waitUntil>published</waitUntil>
+						</configuration>
+					</plugin>
+				</plugins>
+			</build>
+		</profile>
+	</profiles>
 </project>

From 2d592da6ea0ee94a86ab1939f6772d712b411064 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Wed, 1 Jul 2026 20:10:39 +0000
Subject: [PATCH 03/12] docs: document the Maven reactor layout (llama/ module)
 in CLAUDE.md
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- New "Repository layout — Maven reactor" section: the root is an aggregator/parent
  POM with llama/ (core) and llama-langchain4j/ modules; core sources + build files
  now live under llama/; build.sh/build.bat cd into llama/ themselves; core Maven
  builds use -f llama/pom.xml or -pl llama -am; version bumps happen in the root pom
  only (both modules inherit). Adds the caveat that existing root-relative build
  examples in this file now run in llama/.
- Rewrote the LangChain4j section: it is now a reactor module inheriting the parent
  version (no lockstep guard), releasing via one reactor deploy (inherited release
  profile), with the install step using -pl llama -am.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rt1paYztGJ2AKUuBuAGDXE
---
 CLAUDE.md | 79 ++++++++++++++++++++++++++++++++++++-------------------
 1 file changed, 52 insertions(+), 27 deletions(-)

diff --git a/CLAUDE.md b/CLAUDE.md
index db03010f..2826fad9 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -1225,17 +1225,46 @@ keeping it clear of the JPMS module-mode javadoc trap that bit BAF. **Before rai
 javadoc source level to ≥ 9, read**
 [`../workspace/policies/jpms-module-descriptor.md`](../workspace/policies/jpms-module-descriptor.md).
 
-## LangChain4j integration (`llama-langchain4j` sibling module)
+## Repository layout — Maven reactor (`llama/` + `llama-langchain4j/`)
+
+The repo root is a thin **aggregator/parent POM** (`net.ladenthin:llama-parent`,
+`packaging=pom`) with two modules:
+
+- **`llama/`** — the native JNI core (`net.ladenthin:llama`). *All the core sources and build
+  files live here now:* `llama/src/`, `llama/CMakeLists.txt`, `llama/cmake/`, `llama/patches/`,
+  `llama/pom.xml`, `llama/spotbugs-exclude.xml`, `llama/lombok.config`, `llama/.clang-format`.
+  Its published coordinates are unchanged (`net.ladenthin:llama`), so consumers are unaffected.
+- **`llama-langchain4j/`** — the LangChain4j adapters (see below).
+
+Both modules inherit the single `<version>` from the parent, so they **ship in lockstep by
+construction** (no CI guard needed). The parent also holds the shared `release` profile (GPG +
+Central Publishing), so one reactor `mvn -P release deploy` signs and publishes all three
+artifacts (`llama-parent` pom, `llama`, `llama-langchain4j`) at the same version.
+
+**Consequences for build commands:** the core's cmake/native build runs *in `llama/`*.
+`.github/build.sh` / `build.bat` `cd` into `llama/` themselves (relative to the script), so CI
+and the dockcross containers (whose workdir stays the repo root) are unaffected. Locally, run
+core cmake builds from `llama/` (e.g. `cd llama && cmake -B build && cmake --build build`), and
+target the core with Maven via `-f llama/pom.xml` (or `-pl llama -am` from the root). A plain
+`mvn` at the root builds the whole reactor. **When a build-command example elsewhere in this
+file shows `cmake -B build` / `src/main/...` / `mvn compile` at the root, read it as running in
+`llama/`** (the paths moved; the recipes are otherwise unchanged).
+
+**Version bump:** change the `<version>` in the **root** `pom.xml` only; `llama` and
+`llama-langchain4j` inherit it. (The SNAPSHOT/`-SNAPSHOT` line and the README badge still need
+the usual manual update.)
+
+## LangChain4j integration (`llama-langchain4j` reactor module)
 
 `llama-langchain4j/` adapts a `LlamaModel` to LangChain4j's `ChatModel`,
 `StreamingChatModel`, `EmbeddingModel` and `ScoringModel` interfaces **in-process over
-JNI** (no HTTP hop). It is a **standalone sibling module**, deliberately *not* in the root
-reactor, so the native build/release pipeline is untouched.
+JNI** (no HTTP hop). It is a **reactor module** alongside the core `llama` module (see
+"Repository layout" above), so it is built, versioned and released together with the core.
 
 Why it is a **separate artifact** and not a classifier of the core: langchain4j 1.x
 requires **Java 17** (the core stays Java 8), and classifiers share the core's single POM —
 adding `langchain4j-core` there would force it (and the Java 17 floor) on every plain
-`net.ladenthin:llama` consumer. A separate `artifactId` with its own POM is the only way to
+`net.ladenthin:llama` consumer. A separate `artifactId` (its own module POM) is the only way to
 keep that dependency (and Java floor) off the core. It is pure Java with **no per-classifier
 matrix**: it compiles against the core's Java API, which is identical across every native
 classifier; the backend (CPU/CUDA/OpenCL/Vulkan) is a runtime classpath choice for the
@@ -1243,33 +1272,29 @@ consumer.
 
 Wiring:
 
-1. **`llama-langchain4j/pom.xml`** — `net.ladenthin:llama-langchain4j`, `release 17`,
-   depends on `net.ladenthin:llama:${project.version}` (so the core dep always matches the
-   module's own version) and `dev.langchain4j:langchain4j-core`. Carries its own
-   sources/javadoc/gpg + `release` profile (Central requires per-artifact signing; the module
-   has no parent to inherit them from — plugin versions are pinned in lockstep with the root
-   `pom.xml`). Java package stays `net.ladenthin.llama.langchain4j` (package name need not track
-   the artifactId).
-2. **`.github/workflows/publish.yml`** — the `test-java-llama-langchain4j` job installs the
-   core Java jar, runs a **version-lockstep guard** (module version must equal core version,
-   else the build fails — the standalone module can't inherit `${project.version}` from a
-   reactor), then `mvn -f llama-langchain4j/pom.xml verify` (7 model-free mapping unit tests
-   run; the 4 model-backed integration tests self-skip without a GGUF; `verify` also builds the
-   javadoc jar so a release-time javadoc break is caught in PR CI). The
-   `publish-snapshot`/`publish-release` jobs `needs:` this job and, after the core `deploy`
-   (which installs the core jar locally), run a second `deploy` of the module at the same
-   version. A separate **`test-java-llama-langchain4j-integration`** job runs the model-backed
-   tests (chat/streaming/embedding/scoring adapters) by **reusing** the shared GGUF cache
+1. **`llama-langchain4j/pom.xml`** — `net.ladenthin:llama-langchain4j`, `release 17`, a child of
+   `net.ladenthin:llama-parent` (so it **inherits `${project.version}`** — no hardcoded version,
+   no lockstep guard). Depends on `net.ladenthin:llama:${project.version}` and
+   `dev.langchain4j:langchain4j-core`. Builds its own sources/javadoc jars; the `release`
+   profile (GPG + Central Publishing) is **inherited from the parent**, not duplicated here.
+   Java package stays `net.ladenthin.llama.langchain4j` (package name need not track the artifactId).
+2. **`.github/workflows/publish.yml`** — the `test-java-llama-langchain4j` job installs
+   parent + core into the local repo (`mvn -pl llama -am -DskipTests install`), then
+   `mvn -f llama-langchain4j/pom.xml verify` (7 model-free mapping unit tests run; the 4
+   model-backed integration tests self-skip without a GGUF; `verify` also builds the javadoc
+   jar so a release-time javadoc break is caught in PR CI). The `publish-snapshot`/
+   `publish-release` jobs `needs:` this job; deployment is a **single reactor**
+   `mvn -P release deploy` (no separate module deploy step — the parent's inherited `release`
+   profile signs and publishes parent + llama + llama-langchain4j together at the same version).
+   A separate **`test-java-llama-langchain4j-integration`** job runs the model-backed tests
+   (chat/streaming/embedding/scoring adapters) by **reusing** the shared GGUF cache
    (`gguf-models-v1`, restore-only — no extra download) and the `Linux-x86_64-libraries` native
    artifact: it `needs: [crosscompile-linux-x86_64, download-models]` (so the cache is already
-   populated and it runs in parallel), installs the core jar with the downloaded native lib
-   bundled, and passes the already-cached chat
-   (`REASONING_MODEL_NAME`), nomic-embedding and jina-reranker model paths via the module's
+   populated and it runs in parallel), installs parent+core with the downloaded native lib
+   bundled, and passes the already-cached chat (`REASONING_MODEL_NAME`), nomic-embedding and
+   jina-reranker model paths via the module's
    `-Dnet.ladenthin.llama.langchain4j.{embedding,rerank}.model` / `net.ladenthin.llama.model.path`
    properties. It is validation-only (not a release gate); a cold cache degrades to a self-skip.
-3. **Version bumps** — when the root `pom.xml` `<version>` changes, bump
-   `llama-langchain4j/pom.xml` `<version>` to match in the same commit, or the lockstep guard
-   reds CI.
 
 **Open follow-ups** (documented in `llama-langchain4j/README.md`): tool calling
 (`ToolSpecification` ↔ jllama `ToolDefinition`), `response_format`/JSON mode, and multimodal

From 9fcf4660f6c5fd504caefa899a4eebcdab82fa8f Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Wed, 1 Jul 2026 20:18:14 +0000
Subject: [PATCH 04/12] fix(cmake): point bundled JNI headers to repo-root
 .github/include after the llama/ move
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The C++ compile failed on every native job with `fatal error: jni.h: No such
file or directory`. The bundled JNI headers live in .github/include/ at the repo
ROOT, but CMakeLists.txt now runs inside the llama/ module, so the relative
`.github/include/unix` resolved to the non-existent llama/.github/include/unix.
Anchor it to ${CMAKE_CURRENT_SOURCE_DIR}/../.github/include/{unix,windows} so it
points at the repo root regardless of build CWD. Covers Linux, macOS and Windows
(all use the bundled headers). The other CMakeLists paths use
${CMAKE_SOURCE_DIR}/… which already resolves inside llama/.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rt1paYztGJ2AKUuBuAGDXE
---
 llama/CMakeLists.txt | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/llama/CMakeLists.txt b/llama/CMakeLists.txt
index b687c1c1..51bebdfd 100644
--- a/llama/CMakeLists.txt
+++ b/llama/CMakeLists.txt
@@ -277,10 +277,13 @@ endif()
 
 # include jni.h and jni_md.h
 if(NOT DEFINED JNI_INCLUDE_DIRS)
+    # The bundled JNI headers live in .github/include/ at the REPO ROOT, one level above
+    # this module (the core moved into llama/). Anchor to the parent of the CMake source
+    # dir so the path is correct regardless of the build CWD.
     if(OS_NAME MATCHES "^Linux" OR OS_NAME STREQUAL "Mac" OR OS_NAME STREQUAL "Darwin")
-        set(JNI_INCLUDE_DIRS .github/include/unix)
+        set(JNI_INCLUDE_DIRS ${CMAKE_CURRENT_SOURCE_DIR}/../.github/include/unix)
     elseif(OS_NAME STREQUAL "Windows")
-        set(JNI_INCLUDE_DIRS .github/include/windows)
+        set(JNI_INCLUDE_DIRS ${CMAKE_CURRENT_SOURCE_DIR}/../.github/include/windows)
     # if we don't have provided headers, try to find them via Java
     else()
         find_package(Java REQUIRED)

From 71b38594faabf8dd22c909ff345397e1afd0b908 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Wed, 1 Jul 2026 20:20:18 +0000
Subject: [PATCH 05/12] style: clang-format jllama.cpp special_tokens_json
 (pinned 22.1.5)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Pre-existing formatting violation in special_tokens_json (the token-id map added
by #286 / ModelMeta) that this branch inherited unchanged through the move. The
reactor PR's clang-format check surfaced it. Reflowed with the pinned
clang-format 22.1.5; no logic change. (Implies main currently carries the same
violation — it self-resolves when this merges.)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rt1paYztGJ2AKUuBuAGDXE
---
 llama/src/main/cpp/jllama.cpp | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/llama/src/main/cpp/jllama.cpp b/llama/src/main/cpp/jllama.cpp
index c0eef843..f3871f89 100644
--- a/llama/src/main/cpp/jllama.cpp
+++ b/llama/src/main/cpp/jllama.cpp
@@ -770,9 +770,8 @@ JNIEXPORT void JNICALL Java_net_ladenthin_llama_LlamaModel_loadModelWithProgress
 // Build the special-token id map (a token is -1 / LLAMA_TOKEN_NULL when the model defines none).
 static json special_tokens_json(const llama_vocab *vocab) {
     return {
-        {"bos", llama_vocab_bos(vocab)}, {"eos", llama_vocab_eos(vocab)},
-        {"eot", llama_vocab_eot(vocab)}, {"sep", llama_vocab_sep(vocab)},
-        {"nl", llama_vocab_nl(vocab)},   {"pad", llama_vocab_pad(vocab)},
+        {"bos", llama_vocab_bos(vocab)}, {"eos", llama_vocab_eos(vocab)}, {"eot", llama_vocab_eot(vocab)},
+        {"sep", llama_vocab_sep(vocab)}, {"nl", llama_vocab_nl(vocab)},   {"pad", llama_vocab_pad(vocab)},
     };
 }
 

From 86fd2bfe0f7f7ec1da59b7f2efc79a7b7ee9f908 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Wed, 1 Jul 2026 20:26:00 +0000
Subject: [PATCH 06/12] docs: re-root the upgrade-checklist file references to
 llama/ in CLAUDE.md
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The per-section upgrade instructions (CUDA/Android/OpenCL/Windows classifier
steps and the llama.cpp version bump) name the core `pom.xml` / `CMakeLists.txt`
that a contributor would copy-paste — those now live in llama/. Re-rooted the
precise `**pom.xml**` / `**CMakeLists.txt**` callouts, the `# Edit …` lines, the
`git add …` lines, and the build-webui GIT_TAG reference to llama/. The **root**
aggregator pom.xml (the version-bump target) is deliberately left unqualified.
The loose build-command examples (cmake -B build / mvn compile) remain covered by
the "read root examples as running in llama/" note in the Repository layout section.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rt1paYztGJ2AKUuBuAGDXE
---
 CLAUDE.md | 24 ++++++++++++------------
 1 file changed, 12 insertions(+), 12 deletions(-)

diff --git a/CLAUDE.md b/CLAUDE.md
index 2826fad9..86d74121 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -16,7 +16,7 @@ To change the CUDA version, update the following **three** places:
 
 1. **`.github/build_cuda_linux.sh`** — Line 10: `sudo dnf install -y cuda-toolkit-13-2`
 2. **`.github/build_cuda_linux.sh`** — Line 12: `-DCMAKE_CUDA_COMPILER=/usr/local/cuda-13.2/bin/nvcc`
-3. **`pom.xml`** — The `<classifier>` tag in the `cuda` jar execution: `cuda13-linux-x86-64`
+3. **`llama/pom.xml`** — The `<classifier>` tag in the `cuda` jar execution: `cuda13-linux-x86-64`
 
 Also update the header comment in `build_cuda_linux.sh` and the job name in `.github/workflows/release.yaml` for clarity.
 
@@ -32,9 +32,9 @@ Example: To upgrade from 13.2 to a hypothetical 13.3:
 # Edit .github/build_cuda_linux.sh:
 #   line 10: cuda-toolkit-13-2 -> cuda-toolkit-13-3
 #   line 12: /usr/local/cuda-13.2/bin/nvcc -> /usr/local/cuda-13.3/bin/nvcc
-# Edit pom.xml classifier: cuda13-linux-x86-64 (major version only, no need to change for minor bumps)
+# Edit llama/pom.xml classifier: cuda13-linux-x86-64 (major version only, no need to change for minor bumps)
 # Edit CLAUDE.md line: Current CUDA version: **13.2** -> **13.3**
-git add .github/build_cuda_linux.sh pom.xml CLAUDE.md
+git add .github/build_cuda_linux.sh llama/pom.xml CLAUDE.md
 git commit -m "Upgrade CUDA from 13.2 to 13.3"
 ```
 
@@ -88,7 +88,7 @@ This is enforced through bionic's **weak-symbol** mechanism, *not* by bumping
 `__ANDROID_API__` or passing `-DANDROID_PLATFORM`. See "How the API gate is
 satisfied" below for why. To change anything here, update:
 
-1. **`CMakeLists.txt`** — the `add_compile_definitions(__ANDROID_UNAVAILABLE_SYMBOLS_ARE_WEAK__)`
+1. **`llama/CMakeLists.txt`** — the `add_compile_definitions(__ANDROID_UNAVAILABLE_SYMBOLS_ARE_WEAK__)`
    block and its Android-detection guard (`OS_NAME MATCHES "Android"` etc.).
 2. **`CLAUDE.md`** (this file) — the "Current Android minimum API level" line above.
 3. **`README.md`** — the minimum-API note (the `[!NOTE]` block near the Android
@@ -134,7 +134,7 @@ The default Android arm64 JAR remains CPU-only.
 
 Three places wire it together (mirrors the CUDA classifier pattern):
 
-1. **`CMakeLists.txt`** — `elseif(GGML_OPENCL)` branch routes artifacts to
+1. **`llama/CMakeLists.txt`** — `elseif(GGML_OPENCL)` branch routes artifacts to
    `src/main/resources_android_opencl/net/ladenthin/llama/${OS_NAME}/${OS_ARCH}/`.
 2. **`.github/workflows/publish.yml`** — `crosscompile-android-aarch64-opencl`
    job runs the dockcross-android-arm64 build with
@@ -142,7 +142,7 @@ Three places wire it together (mirrors the CUDA classifier pattern):
    and uploads as artifact `android-libraries-opencl`. The `package`,
    `publish-snapshot`, and `publish-release` jobs download it into
    `resources_android_opencl/` and activate the `opencl-android` Maven profile.
-3. **`pom.xml`** — the `opencl-android` profile produces a second JAR with
+3. **`llama/pom.xml`** — the `opencl-android` profile produces a second JAR with
    `<classifier>opencl-android-aarch64</classifier>` from the
    `${project.build.outputDirectory}_opencl_android` tree.
 
@@ -196,7 +196,7 @@ local / self-hosted.
 
 Wiring (mirrors the CUDA-Linux / OpenCL-Android classifier pattern):
 
-1. **`CMakeLists.txt`** — the `if(GGML_CUDA) … elseif(GGML_VULKAN) … elseif(GGML_OPENCL) … else()`
+1. **`llama/CMakeLists.txt`** — the `if(GGML_CUDA) … elseif(GGML_VULKAN) … elseif(GGML_OPENCL) … else()`
    chain is **OS-aware**: CUDA → `resources_windows_cuda` on Windows (else `resources_linux_cuda`),
    Vulkan → `resources_windows_vulkan`, OpenCL → `resources_windows_opencl` on Windows (else
    `resources_android_opencl`). The default CPU build (both generators) still emits to the canonical
@@ -225,7 +225,7 @@ Wiring (mirrors the CUDA-Linux / OpenCL-Android classifier pattern):
    The `package`, `publish-snapshot`, and `publish-release` jobs download each non-default artifact into
    its `src/main/resources_windows_{msvc,cuda,vulkan,opencl}/` tree and activate the
    `windows-msvc,cuda-windows,vulkan-windows,opencl-windows` Maven profiles.
-5. **`pom.xml`** — profiles `windows-msvc` / `cuda-windows` / `vulkan-windows` / `opencl-windows`,
+5. **`llama/pom.xml`** — profiles `windows-msvc` / `cuda-windows` / `vulkan-windows` / `opencl-windows`,
    each a separate compile pass + resource copy + classified jar (classifiers `msvc-windows` /
    `cuda13-windows-x86-64` / `vulkan-windows-x86-64` / `opencl-windows-x86-64`). Activated only in CI.
 6. **`README.md`** — the classifier table + dependency snippets in "Choosing the right classifier".
@@ -263,7 +263,7 @@ checked in (same policy as the native libs).
 Pipeline (`.github/workflows/publish.yml`):
 
 1. **`build-webui` job** (ubuntu — the *only* job that runs `npm`): resolves the
-   pinned `b<nnnn>` tag from `CMakeLists.txt`'s `GIT_TAG`, sparse-checks-out
+   pinned `b<nnnn>` tag from `llama/CMakeLists.txt`'s `GIT_TAG`, sparse-checks-out
    `ggml-org/llama.cpp@<tag>` `tools/ui`, runs the upstream Svelte build
    (`npm ci && npm run build`), gzips `dist/` into `dist/_gzip/` (LLAMA_UI_GZIP
    parity), builds the self-contained `llama-ui-embed` host tool (plain C++17, **no
@@ -472,16 +472,16 @@ re-verify the generator the same way you re-verify `patches/`.
 
 To change the llama.cpp version, update the following **three** files (and re-verify `patches/`):
 
-1. **CMakeLists.txt** — the `GIT_TAG` line for llama.cpp: `GIT_TAG        b8831`
+1. **llama/CMakeLists.txt** — the `GIT_TAG` line for llama.cpp: `GIT_TAG        b8831`
 2. **README.md** — the badge and link line with the version number
 3. **CLAUDE.md** — the "Current llama.cpp pinned version" line
 
 Example: To upgrade from b8808 to b8831:
 ```bash
-# Edit CMakeLists.txt: change GIT_TAG b8808 to b8831
+# Edit llama/CMakeLists.txt: change GIT_TAG b8808 to b8831
 # Edit README.md: change b8808 to b8831 (in both badge and link)
 # Edit CLAUDE.md: change b8808 to b8831
-git add CMakeLists.txt README.md CLAUDE.md
+git add llama/CMakeLists.txt README.md CLAUDE.md
 git commit -m "Upgrade llama.cpp from b8808 to b8831"
 git push -u origin <your-branch>
 ```

From d64a3653b0aeac69786707813ad365b98cedd19f Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Thu, 2 Jul 2026 06:59:21 +0000
Subject: [PATCH 07/12] Fix reactor CI: keep class-level @Builder, re-root
 crash-dump paths, README
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- TrainingParameters: keep class-level Lombok @Builder @Getter but satisfy the
  strict javadoc build (failOnWarnings) — replace the raw-source-invisible
  {@link #builder()} with {@code builder()} and add an explicit private all-args
  constructor so javadoc sees a real constructor (no synthetic-default warning);
  Lombok generates the builder around it. Verified: compile + javadoc:jar clean,
  TrainingParametersTest (3) green.
- publish.yml: re-root the Java test jobs' crash-dump/surefire upload globs to
  llama/ (hs_err_pid*, *.hprof, target/surefire-reports/*) now that surefire runs
  against -f llama/pom.xml; core.* stays at workspace root (absolute core_pattern).
- README: add Jlama and LangChain4j to the Similar Projects section.
- Apply clang-format 22.1.5 (train_engine.cpp) and spotless (Java) from the #289 merge.
- .gitignore: catch llama/.jqwik-database (un-anchored pattern).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rt1paYztGJ2AKUuBuAGDXE
---
 .github/workflows/publish.yml                 | 72 +++++++++----------
 .gitignore                                    |  4 +-
 README.md                                     |  8 +++
 llama/src/main/cpp/train_engine.cpp           | 11 ++-
 .../net/ladenthin/llama/LlamaTrainer.java     | 15 ++--
 .../llama/parameters/TrainingParameters.java  | 72 +++++++++++++++----
 .../llama/LlamaTrainerIntegrationTest.java    | 17 +++--
 .../parameters/TrainingParametersTest.java    | 38 +++++-----
 8 files changed, 144 insertions(+), 93 deletions(-)

diff --git a/.github/workflows/publish.yml b/.github/workflows/publish.yml
index 0d4eb6ff..4966652c 100644
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@@ -1108,13 +1108,13 @@ jobs:
         with:
           name: error-log-linux-x86_64
           path: |
-            ${{ github.workspace }}/hs_err_pid*.log
+            ${{ github.workspace }}/llama/hs_err_pid*.log
             ${{ github.workspace }}/core.*
-            ${{ github.workspace }}/*.hprof
-            ${{ github.workspace }}/target/surefire-reports/*.dump
-            ${{ github.workspace }}/target/surefire-reports/*.dumpstream
-            ${{ github.workspace }}/target/surefire-reports/*.txt
-            ${{ github.workspace }}/target/surefire-reports/TEST-*.xml
+            ${{ github.workspace }}/llama/*.hprof
+            ${{ github.workspace }}/llama/target/surefire-reports/*.dump
+            ${{ github.workspace }}/llama/target/surefire-reports/*.dumpstream
+            ${{ github.workspace }}/llama/target/surefire-reports/*.txt
+            ${{ github.workspace }}/llama/target/surefire-reports/TEST-*.xml
           if-no-files-found: warn
 
   # ---------------------------------------------------------------------------
@@ -1202,12 +1202,12 @@ jobs:
         with:
           name: error-log-macos-14-metal
           path: |
-            ${{ github.workspace }}/hs_err_pid*.log
-            ${{ github.workspace }}/*.hprof
-            ${{ github.workspace }}/target/surefire-reports/*.dump
-            ${{ github.workspace }}/target/surefire-reports/*.dumpstream
-            ${{ github.workspace }}/target/surefire-reports/*.txt
-            ${{ github.workspace }}/target/surefire-reports/TEST-*.xml
+            ${{ github.workspace }}/llama/hs_err_pid*.log
+            ${{ github.workspace }}/llama/*.hprof
+            ${{ github.workspace }}/llama/target/surefire-reports/*.dump
+            ${{ github.workspace }}/llama/target/surefire-reports/*.dumpstream
+            ${{ github.workspace }}/llama/target/surefire-reports/*.txt
+            ${{ github.workspace }}/llama/target/surefire-reports/TEST-*.xml
           if-no-files-found: warn
 
   test-java-macos-arm64-no-metal:
@@ -1266,12 +1266,12 @@ jobs:
         with:
           name: error-log-macos-15-no-metal
           path: |
-            ${{ github.workspace }}/hs_err_pid*.log
-            ${{ github.workspace }}/*.hprof
-            ${{ github.workspace }}/target/surefire-reports/*.dump
-            ${{ github.workspace }}/target/surefire-reports/*.dumpstream
-            ${{ github.workspace }}/target/surefire-reports/*.txt
-            ${{ github.workspace }}/target/surefire-reports/TEST-*.xml
+            ${{ github.workspace }}/llama/hs_err_pid*.log
+            ${{ github.workspace }}/llama/*.hprof
+            ${{ github.workspace }}/llama/target/surefire-reports/*.dump
+            ${{ github.workspace }}/llama/target/surefire-reports/*.dumpstream
+            ${{ github.workspace }}/llama/target/surefire-reports/*.txt
+            ${{ github.workspace }}/llama/target/surefire-reports/TEST-*.xml
           if-no-files-found: warn
 
   test-java-macos-arm64-metal-15:
@@ -1330,12 +1330,12 @@ jobs:
         with:
           name: error-log-macos-15-metal
           path: |
-            ${{ github.workspace }}/hs_err_pid*.log
-            ${{ github.workspace }}/*.hprof
-            ${{ github.workspace }}/target/surefire-reports/*.dump
-            ${{ github.workspace }}/target/surefire-reports/*.dumpstream
-            ${{ github.workspace }}/target/surefire-reports/*.txt
-            ${{ github.workspace }}/target/surefire-reports/TEST-*.xml
+            ${{ github.workspace }}/llama/hs_err_pid*.log
+            ${{ github.workspace }}/llama/*.hprof
+            ${{ github.workspace }}/llama/target/surefire-reports/*.dump
+            ${{ github.workspace }}/llama/target/surefire-reports/*.dumpstream
+            ${{ github.workspace }}/llama/target/surefire-reports/*.txt
+            ${{ github.workspace }}/llama/target/surefire-reports/TEST-*.xml
           if-no-files-found: warn
 
   test-java-windows-x86_64:
@@ -1413,13 +1413,13 @@ jobs:
         with:
           name: windows-output
           path: |
-            ${{ github.workspace }}\hs_err_pid*.log
-            ${{ github.workspace }}\*.hprof
+            ${{ github.workspace }}\llama\hs_err_pid*.log
+            ${{ github.workspace }}\llama\*.hprof
             ${{ github.workspace }}\dumps\*.dmp
-            ${{ github.workspace }}\target\surefire-reports\*.dump
-            ${{ github.workspace }}\target\surefire-reports\*.dumpstream
-            ${{ github.workspace }}\target\surefire-reports\*.txt
-            ${{ github.workspace }}\target\surefire-reports\TEST-*.xml
+            ${{ github.workspace }}\llama\target\surefire-reports\*.dump
+            ${{ github.workspace }}\llama\target\surefire-reports\*.dumpstream
+            ${{ github.workspace }}\llama\target\surefire-reports\*.txt
+            ${{ github.workspace }}\llama\target\surefire-reports\TEST-*.xml
             ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/**/*
           if-no-files-found: warn
 
@@ -1502,13 +1502,13 @@ jobs:
         with:
           name: windows-output-msvc
           path: |
-            ${{ github.workspace }}\hs_err_pid*.log
-            ${{ github.workspace }}\*.hprof
+            ${{ github.workspace }}\llama\hs_err_pid*.log
+            ${{ github.workspace }}\llama\*.hprof
             ${{ github.workspace }}\dumps\*.dmp
-            ${{ github.workspace }}\target\surefire-reports\*.dump
-            ${{ github.workspace }}\target\surefire-reports\*.dumpstream
-            ${{ github.workspace }}\target\surefire-reports\*.txt
-            ${{ github.workspace }}\target\surefire-reports\TEST-*.xml
+            ${{ github.workspace }}\llama\target\surefire-reports\*.dump
+            ${{ github.workspace }}\llama\target\surefire-reports\*.dumpstream
+            ${{ github.workspace }}\llama\target\surefire-reports\*.txt
+            ${{ github.workspace }}\llama\target\surefire-reports\TEST-*.xml
             ${{ github.workspace }}/llama/src/main/resources/net/ladenthin/llama/**/*
           if-no-files-found: warn
 
diff --git a/.gitignore b/.gitignore
index ba0a69ed..dfead6a8 100644
--- a/.gitignore
+++ b/.gitignore
@@ -61,8 +61,8 @@ llama/src/test/resources/**/*.gbnf
 **/*.lastModified
 llama/src/main/cpp/llama.cpp/
 
-# jcstress / jqwik test outputs (generated in repo root)
-/.jqwik-database
+# jcstress / jqwik test outputs (generated in the module dir where surefire runs, e.g. llama/)
+.jqwik-database
 
 # Local AI agent tooling (not part of the project)
 AGENTS.md
diff --git a/README.md b/README.md
index f047c43f..b6e2ca88 100644
--- a/README.md
+++ b/README.md
@@ -931,3 +931,11 @@ The system's updated C++ runtime will be used instead, resolving the crash.
 - [gptoss.java](https://github.com/mukel/gptoss.java) — GPT-OSS architecture inference.
 - [qwen35.java](https://github.com/mukel/qwen35.java) — Qwen 3.5 inference.
 - [nemotron3.java](https://github.com/mukel/nemotron3.java) — NVIDIA Nemotron-3 inference.
+
+**Pure-Java inference engines (no JNI / no llama.cpp)**
+
+- [Jlama](https://github.com/tjake/Jlama) — a full pure-Java LLM inference engine for the JVM (multiple model architectures, quantization, and distributed inference) built on the Java Vector API. A no-native alternative to the JNI approach here; different design point (pure JVM portability vs. GGUF compatibility and llama.cpp performance via JNI).
+
+**Frameworks / orchestration**
+
+- [LangChain4j](https://github.com/langchain4j/langchain4j) — LLM-application framework for Java (chat, embeddings, RAG, tool calling, agents) over a unified provider API. This project ships a first-class **in-process** integration — see the [`llama-langchain4j`](llama-langchain4j/) module — so a llama.cpp model plugs straight into LangChain4j's `ChatModel` / `StreamingChatModel` / `EmbeddingModel` / `ScoringModel` without an HTTP hop.
diff --git a/llama/src/main/cpp/train_engine.cpp b/llama/src/main/cpp/train_engine.cpp
index eade0556..e4825d2f 100644
--- a/llama/src/main/cpp/train_engine.cpp
+++ b/llama/src/main/cpp/train_engine.cpp
@@ -37,8 +37,7 @@ bool finetune(const finetune_config &cfg, std::string &err) {
         params.n_ubatch = cfg.n_ubatch;
     }
 
-    params.optimizer =
-        cfg.optimizer == 1 ? GGML_OPT_OPTIMIZER_TYPE_SGD : GGML_OPT_OPTIMIZER_TYPE_ADAMW;
+    params.optimizer = cfg.optimizer == 1 ? GGML_OPT_OPTIMIZER_TYPE_SGD : GGML_OPT_OPTIMIZER_TYPE_ADAMW;
     params.lr.lr0 = cfg.learning_rate;
     params.lr.lr_min = cfg.lr_min;
     params.lr.decay_epochs = cfg.decay_epochs;
@@ -99,8 +98,8 @@ bool finetune(const finetune_config &cfg, std::string &err) {
     ggml_opt_result_t result_eval = ggml_opt_result_init();
 
     for (params.lr.epoch = 0; params.lr.epoch < params.lr.epochs; ++params.lr.epoch) {
-        llama_opt_epoch(ctx, dataset, result_train, result_eval, idata_split,
-                        ggml_opt_epoch_callback_progress_bar, ggml_opt_epoch_callback_progress_bar);
+        llama_opt_epoch(ctx, dataset, result_train, result_eval, idata_split, ggml_opt_epoch_callback_progress_bar,
+                        ggml_opt_epoch_callback_progress_bar);
         ggml_opt_result_reset(result_train);
         ggml_opt_result_reset(result_eval);
     }
@@ -118,8 +117,8 @@ bool finetune(const finetune_config &cfg, std::string &err) {
 
 } // namespace jllama_train
 
-extern "C" JNIEXPORT jstring JNICALL
-Java_net_ladenthin_llama_LlamaTrainer_finetuneNative(JNIEnv *env, jclass, jstring jconfig) {
+extern "C" JNIEXPORT jstring JNICALL Java_net_ladenthin_llama_LlamaTrainer_finetuneNative(JNIEnv *env, jclass,
+                                                                                          jstring jconfig) {
     std::string config_json;
     if (jconfig != nullptr) {
         const char *c = env->GetStringUTFChars(jconfig, nullptr);
diff --git a/llama/src/main/java/net/ladenthin/llama/LlamaTrainer.java b/llama/src/main/java/net/ladenthin/llama/LlamaTrainer.java
index 7ba9db41..3e04553b 100644
--- a/llama/src/main/java/net/ladenthin/llama/LlamaTrainer.java
+++ b/llama/src/main/java/net/ladenthin/llama/LlamaTrainer.java
@@ -51,14 +51,13 @@ public static void finetune(TrainingParameters parameters) {
      * @throws LlamaException if the model cannot be loaded or training fails
      */
     public static void finetune(Path model, String trainingText, Path output, int epochs, float learningRate) {
-        finetune(
-                TrainingParameters.builder()
-                        .modelPath(model)
-                        .trainingText(trainingText)
-                        .outputPath(output)
-                        .epochs(epochs)
-                        .learningRate(learningRate)
-                        .build());
+        finetune(TrainingParameters.builder()
+                .modelPath(model)
+                .trainingText(trainingText)
+                .outputPath(output)
+                .epochs(epochs)
+                .learningRate(learningRate)
+                .build());
     }
 
     private static native String finetuneNative(String configJson);
diff --git a/llama/src/main/java/net/ladenthin/llama/parameters/TrainingParameters.java b/llama/src/main/java/net/ladenthin/llama/parameters/TrainingParameters.java
index 71640b12..4b5c189c 100644
--- a/llama/src/main/java/net/ladenthin/llama/parameters/TrainingParameters.java
+++ b/llama/src/main/java/net/ladenthin/llama/parameters/TrainingParameters.java
@@ -15,7 +15,7 @@
 /**
  * Immutable configuration for a {@link net.ladenthin.llama.LlamaTrainer} fine-tuning run.
  *
- * <p>Build with {@link #builder()}; only {@code modelPath} and {@code outputPath} are required, and
+ * <p>Build with {@code builder()}; only {@code modelPath} and {@code outputPath} are required, and
  * exactly one of {@code trainingText} / {@code trainingFile} should be set. All other fields default
  * to values that mirror upstream llama.cpp's fine-tuning defaults. The configuration is serialized to
  * JSON via {@link #toJson()} and parsed by the native layer, the same way {@link ModelParameters} and
@@ -38,40 +38,88 @@ public final class TrainingParameters {
     private final Path outputPath;
 
     // Number of passes over the corpus (at least 1).
-    @Builder.Default private final int epochs = 2;
+    @Builder.Default
+    private final int epochs = 2;
 
     // Learning rate at the first epoch.
-    @Builder.Default private final float learningRate = 1e-5f;
+    @Builder.Default
+    private final float learningRate = 1e-5f;
 
     // Minimum learning rate for decay, or -1 to disable decay.
-    @Builder.Default private final float lrMin = -1f;
+    @Builder.Default
+    private final float lrMin = -1f;
 
     // If > 0, decay the learning rate from learningRate to lrMin over this many epochs.
-    @Builder.Default private final float decayEpochs = -1f;
+    @Builder.Default
+    private final float decayEpochs = -1f;
 
     // Weight decay (0 disables it).
-    @Builder.Default private final float weightDecay = 0f;
+    @Builder.Default
+    private final float weightDecay = 0f;
 
     // Optimizer algorithm.
-    @Builder.Default private final Optimizer optimizer = Optimizer.ADAMW;
+    @Builder.Default
+    private final Optimizer optimizer = Optimizer.ADAMW;
 
     // Context size in tokens, or 0 to use the model's trained context.
-    @Builder.Default private final int nCtx = 0;
+    @Builder.Default
+    private final int nCtx = 0;
 
     // Layers to offload to the GPU, or -1 for automatic.
-    @Builder.Default private final int nGpuLayers = -1;
+    @Builder.Default
+    private final int nGpuLayers = -1;
 
     // Fraction of the corpus held out for validation.
-    @Builder.Default private final float valSplit = 0.05f;
+    @Builder.Default
+    private final float valSplit = 0.05f;
 
     // Logical batch size, or 0 to use the native default.
-    @Builder.Default private final int nBatch = 0;
+    @Builder.Default
+    private final int nBatch = 0;
 
     // Physical (micro) batch size, or 0 to use the native default.
-    @Builder.Default private final int nUbatch = 0;
+    @Builder.Default
+    private final int nUbatch = 0;
 
     private static final ObjectMapper MAPPER = new ObjectMapper();
 
+    // Explicit all-args constructor used by the Lombok-generated builder. Declared explicitly (rather
+    // than letting @Builder synthesize the package-private one) so Javadoc sees a real constructor and
+    // does not emit the "use of default constructor, which does not provide a comment" warning; it is
+    // private, so it is not part of the public API and is not doclint-checked.
+    private TrainingParameters(
+            Path modelPath,
+            @Nullable String trainingText,
+            @Nullable Path trainingFile,
+            Path outputPath,
+            int epochs,
+            float learningRate,
+            float lrMin,
+            float decayEpochs,
+            float weightDecay,
+            Optimizer optimizer,
+            int nCtx,
+            int nGpuLayers,
+            float valSplit,
+            int nBatch,
+            int nUbatch) {
+        this.modelPath = modelPath;
+        this.trainingText = trainingText;
+        this.trainingFile = trainingFile;
+        this.outputPath = outputPath;
+        this.epochs = epochs;
+        this.learningRate = learningRate;
+        this.lrMin = lrMin;
+        this.decayEpochs = decayEpochs;
+        this.weightDecay = weightDecay;
+        this.optimizer = optimizer;
+        this.nCtx = nCtx;
+        this.nGpuLayers = nGpuLayers;
+        this.valSplit = valSplit;
+        this.nBatch = nBatch;
+        this.nUbatch = nUbatch;
+    }
+
     /**
      * Serialize this configuration to the JSON object the native fine-tuning layer expects.
      *
diff --git a/llama/src/test/java/net/ladenthin/llama/LlamaTrainerIntegrationTest.java b/llama/src/test/java/net/ladenthin/llama/LlamaTrainerIntegrationTest.java
index c7027d22..3dea0f24 100644
--- a/llama/src/test/java/net/ladenthin/llama/LlamaTrainerIntegrationTest.java
+++ b/llama/src/test/java/net/ladenthin/llama/LlamaTrainerIntegrationTest.java
@@ -37,15 +37,14 @@ void finetuneWritesAnOutputModel(@TempDir Path tmp) throws Exception {
         }
 
         Path output = tmp.resolve("finetuned.gguf");
-        LlamaTrainer.finetune(
-                TrainingParameters.builder()
-                        .modelPath(Paths.get(modelPath))
-                        .trainingText(corpus.toString())
-                        .outputPath(output)
-                        .epochs(1)
-                        .learningRate(1e-5f)
-                        .optimizer(Optimizer.ADAMW)
-                        .build());
+        LlamaTrainer.finetune(TrainingParameters.builder()
+                .modelPath(Paths.get(modelPath))
+                .trainingText(corpus.toString())
+                .outputPath(output)
+                .epochs(1)
+                .learningRate(1e-5f)
+                .optimizer(Optimizer.ADAMW)
+                .build());
 
         assertThat(Files.exists(output), is(true));
         assertThat(Files.size(output), greaterThan(0L));
diff --git a/llama/src/test/java/net/ladenthin/llama/parameters/TrainingParametersTest.java b/llama/src/test/java/net/ladenthin/llama/parameters/TrainingParametersTest.java
index f5dfe0d3..87ea5cb0 100644
--- a/llama/src/test/java/net/ladenthin/llama/parameters/TrainingParametersTest.java
+++ b/llama/src/test/java/net/ladenthin/llama/parameters/TrainingParametersTest.java
@@ -24,12 +24,11 @@ private JsonNode json(TrainingParameters parameters) throws Exception {
 
     @Test
     void defaultsSerializeWithExpectedValues() throws Exception {
-        TrainingParameters parameters =
-                TrainingParameters.builder()
-                        .modelPath(Paths.get("base.gguf"))
-                        .trainingText("hello world")
-                        .outputPath(Paths.get("tuned.gguf"))
-                        .build();
+        TrainingParameters parameters = TrainingParameters.builder()
+                .modelPath(Paths.get("base.gguf"))
+                .trainingText("hello world")
+                .outputPath(Paths.get("tuned.gguf"))
+                .build();
 
         JsonNode node = json(parameters);
         assertThat(node.get("model_path").asText(), is("base.gguf"));
@@ -46,20 +45,19 @@ void defaultsSerializeWithExpectedValues() throws Exception {
 
     @Test
     void customValuesSerialize() throws Exception {
-        TrainingParameters parameters =
-                TrainingParameters.builder()
-                        .modelPath(Paths.get("base.gguf"))
-                        .trainingFile(Paths.get("corpus.txt"))
-                        .outputPath(Paths.get("tuned.gguf"))
-                        .epochs(5)
-                        .learningRate(3e-4f)
-                        .optimizer(Optimizer.SGD)
-                        .nCtx(512)
-                        .nGpuLayers(0)
-                        .valSplit(0.1f)
-                        .nBatch(256)
-                        .nUbatch(64)
-                        .build();
+        TrainingParameters parameters = TrainingParameters.builder()
+                .modelPath(Paths.get("base.gguf"))
+                .trainingFile(Paths.get("corpus.txt"))
+                .outputPath(Paths.get("tuned.gguf"))
+                .epochs(5)
+                .learningRate(3e-4f)
+                .optimizer(Optimizer.SGD)
+                .nCtx(512)
+                .nGpuLayers(0)
+                .valSplit(0.1f)
+                .nBatch(256)
+                .nUbatch(64)
+                .build();
 
         JsonNode node = json(parameters);
         assertThat(node.get("epochs").asInt(), is(5));

From 18ed4904bfb5b73067d01225985449ba114a2a31 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Thu, 2 Jul 2026 07:21:59 +0000
Subject: [PATCH 08/12] Make merged #289 training code pass SpotBugs + PIT
 (keep pure @Builder)

The #289 merge introduced the first Lombok-@Builder-heavy class and it had not
cleared the strict quality gates:

SpotBugs (13 findings), all fixed while keeping class-level @Builder:
- 10x MRC_METHOD_RETURNS_CONSTANT on Lombok's synthetic $default$*() methods:
  @Builder.Default emits these, and they are NOT tagged @lombok.Generated even
  with lombok.addLombokGeneratedAnnotation=true (a Lombok limitation), so
  SpotBugs cannot auto-skip them. Suppressed for the $default$* methods only.
- IMC_IMMATURE_CLASS_NO_TOSTRING: add @ToString (generated, @lombok.Generated).
- IMC_IMMATURE_CLASS_WRONG_FIELD_ORDER: move the static MAPPER above the
  instance fields.
- RCN_REDUNDANT_NULLCHECK in LlamaTrainer: the @NullMarked default treated the
  native finetuneNative() return as @NonNull; mark it @Nullable (JNI can return
  null on success), making the `error != null` guard meaningful.

PIT (was 99% < 100% threshold):
- Optimizer.getNativeValue() showed as NO_COVERAGE because its only test lived in
  TrainingParametersTest (package parameters.*), outside PIT's args.*/value.*/
  exception.*/json.* targetTests. Move it to a dedicated args.OptimizerTest so
  PIT runs it. Score back to 100% (249/249).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rt1paYztGJ2AKUuBuAGDXE
---
 llama/spotbugs-exclude.xml                    | 17 +++++++
 .../net/ladenthin/llama/LlamaTrainer.java     |  6 ++-
 .../llama/parameters/TrainingParameters.java  | 45 +++----------------
 .../ladenthin/llama/args/OptimizerTest.java   | 20 +++++++++
 .../parameters/TrainingParametersTest.java    |  6 ---
 5 files changed, 48 insertions(+), 46 deletions(-)
 create mode 100644 llama/src/test/java/net/ladenthin/llama/args/OptimizerTest.java

diff --git a/llama/spotbugs-exclude.xml b/llama/spotbugs-exclude.xml
index 6cabea32..d64bed38 100644
--- a/llama/spotbugs-exclude.xml
+++ b/llama/spotbugs-exclude.xml
@@ -605,4 +605,21 @@ SPDX-License-Identifier: MIT
         <Method name="synthesize"/>
     </Match>
 
+    <!--
+        TrainingParameters uses Lombok @Builder with @Builder.Default on its schedule/optimizer
+        fields. For each defaulted field Lombok emits a synthetic static `$default$<field>()` method
+        that simply returns the default constant, which fb-contrib flags as
+        MRC_METHOD_RETURNS_CONSTANT. Unlike Lombok's accessor/constructor output, these
+        `$default$*` methods are NOT tagged with @lombok.Generated even when
+        `lombok.addLombokGeneratedAnnotation = true` (a long-standing Lombok limitation), so
+        SpotBugs cannot skip them automatically the way it does the rest of the generated members.
+        The methods are pure generated boilerplate — returning a constant is exactly their purpose —
+        so the finding is suppressed for these synthetic methods only.
+    -->
+    <Match>
+        <Class name="net.ladenthin.llama.parameters.TrainingParameters"/>
+        <Bug pattern="MRC_METHOD_RETURNS_CONSTANT"/>
+        <Method name="~\$default\$.*"/>
+    </Match>
+
 </FindBugsFilter>
diff --git a/llama/src/main/java/net/ladenthin/llama/LlamaTrainer.java b/llama/src/main/java/net/ladenthin/llama/LlamaTrainer.java
index 3e04553b..9d751fc7 100644
--- a/llama/src/main/java/net/ladenthin/llama/LlamaTrainer.java
+++ b/llama/src/main/java/net/ladenthin/llama/LlamaTrainer.java
@@ -8,6 +8,7 @@
 import net.ladenthin.llama.exception.LlamaException;
 import net.ladenthin.llama.loader.LlamaLoader;
 import net.ladenthin.llama.parameters.TrainingParameters;
+import org.jspecify.annotations.Nullable;
 
 /**
  * In-process fine-tuning entry point, wrapping llama.cpp's ggml-opt training path
@@ -60,5 +61,8 @@ public static void finetune(Path model, String trainingText, Path output, int ep
                 .build());
     }
 
-    private static native String finetuneNative(String configJson);
+    // The native layer returns null on success and a non-empty error message on failure, so the
+    // return is genuinely @Nullable — without this the module's @NullMarked default would treat it
+    // as @NonNull and SpotBugs would flag the `error != null` guard as a redundant null check.
+    private static native @Nullable String finetuneNative(String configJson);
 }
diff --git a/llama/src/main/java/net/ladenthin/llama/parameters/TrainingParameters.java b/llama/src/main/java/net/ladenthin/llama/parameters/TrainingParameters.java
index 4b5c189c..a0cc481f 100644
--- a/llama/src/main/java/net/ladenthin/llama/parameters/TrainingParameters.java
+++ b/llama/src/main/java/net/ladenthin/llama/parameters/TrainingParameters.java
@@ -9,6 +9,7 @@
 import java.nio.file.Path;
 import lombok.Builder;
 import lombok.Getter;
+import lombok.ToString;
 import net.ladenthin.llama.args.Optimizer;
 import org.jspecify.annotations.Nullable;
 
@@ -23,8 +24,13 @@
  */
 @Builder
 @Getter
+@ToString
 public final class TrainingParameters {
 
+    // Jackson mapper for JSON serialization. Static field declared before the instance fields
+    // to satisfy fb-contrib's IMC_IMMATURE_CLASS_WRONG_FIELD_ORDER (static members come first).
+    private static final ObjectMapper MAPPER = new ObjectMapper();
+
     // Base GGUF model to fine-tune.
     private final Path modelPath;
 
@@ -81,45 +87,6 @@ public final class TrainingParameters {
     @Builder.Default
     private final int nUbatch = 0;
 
-    private static final ObjectMapper MAPPER = new ObjectMapper();
-
-    // Explicit all-args constructor used by the Lombok-generated builder. Declared explicitly (rather
-    // than letting @Builder synthesize the package-private one) so Javadoc sees a real constructor and
-    // does not emit the "use of default constructor, which does not provide a comment" warning; it is
-    // private, so it is not part of the public API and is not doclint-checked.
-    private TrainingParameters(
-            Path modelPath,
-            @Nullable String trainingText,
-            @Nullable Path trainingFile,
-            Path outputPath,
-            int epochs,
-            float learningRate,
-            float lrMin,
-            float decayEpochs,
-            float weightDecay,
-            Optimizer optimizer,
-            int nCtx,
-            int nGpuLayers,
-            float valSplit,
-            int nBatch,
-            int nUbatch) {
-        this.modelPath = modelPath;
-        this.trainingText = trainingText;
-        this.trainingFile = trainingFile;
-        this.outputPath = outputPath;
-        this.epochs = epochs;
-        this.learningRate = learningRate;
-        this.lrMin = lrMin;
-        this.decayEpochs = decayEpochs;
-        this.weightDecay = weightDecay;
-        this.optimizer = optimizer;
-        this.nCtx = nCtx;
-        this.nGpuLayers = nGpuLayers;
-        this.valSplit = valSplit;
-        this.nBatch = nBatch;
-        this.nUbatch = nUbatch;
-    }
-
     /**
      * Serialize this configuration to the JSON object the native fine-tuning layer expects.
      *
diff --git a/llama/src/test/java/net/ladenthin/llama/args/OptimizerTest.java b/llama/src/test/java/net/ladenthin/llama/args/OptimizerTest.java
new file mode 100644
index 00000000..d8e550b7
--- /dev/null
+++ b/llama/src/test/java/net/ladenthin/llama/args/OptimizerTest.java
@@ -0,0 +1,20 @@
+// SPDX-FileCopyrightText: 2026 Bernard Ladenthin <bernard.ladenthin@gmail.com>
+//
+// SPDX-License-Identifier: MIT
+
+package net.ladenthin.llama.args;
+
+import static org.hamcrest.MatcherAssert.assertThat;
+import static org.hamcrest.Matchers.is;
+
+import org.junit.jupiter.api.Test;
+
+/** Model-free tests for the {@link Optimizer} native-value mapping. */
+class OptimizerTest {
+
+    @Test
+    void nativeValuesMatchGgml() {
+        assertThat(Optimizer.ADAMW.getNativeValue(), is(0));
+        assertThat(Optimizer.SGD.getNativeValue(), is(1));
+    }
+}
diff --git a/llama/src/test/java/net/ladenthin/llama/parameters/TrainingParametersTest.java b/llama/src/test/java/net/ladenthin/llama/parameters/TrainingParametersTest.java
index 87ea5cb0..d56b53c9 100644
--- a/llama/src/test/java/net/ladenthin/llama/parameters/TrainingParametersTest.java
+++ b/llama/src/test/java/net/ladenthin/llama/parameters/TrainingParametersTest.java
@@ -70,10 +70,4 @@ void customValuesSerialize() throws Exception {
         // training_text is omitted when a corpus file is given.
         assertThat(node.has("training_text"), is(false));
     }
-
-    @Test
-    void optimizerNativeValuesMatchGgml() {
-        assertThat(Optimizer.ADAMW.getNativeValue(), is(0));
-        assertThat(Optimizer.SGD.getNativeValue(), is(1));
-    }
 }

From eddf3bebe11ab978aa17ed3dd8050bb2cbe61960 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Thu, 2 Jul 2026 07:24:31 +0000
Subject: [PATCH 09/12] javadoc: doclint=all,-missing so Lombok @Builder value
 classes pass

Javadoc runs on raw source before Lombok's processor, so it cannot see the
constructor class-level @Builder generates and emits a spurious "use of default
constructor, which does not provide a comment" (the doclint `missing` group) for
TrainingParameters. Drop only the `missing` group; reference/html/syntax/
accessibility stay fatal (failOnWarnings remains true), so a broken {@link} or
malformed HTML still fails the build. Keeps the value class pure @Builder with
no hand-written all-args constructor.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rt1paYztGJ2AKUuBuAGDXE
---
 llama/pom.xml | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)

diff --git a/llama/pom.xml b/llama/pom.xml
index 73b91a09..be7a8960 100644
--- a/llama/pom.xml
+++ b/llama/pom.xml
@@ -551,7 +551,19 @@ SPDX-License-Identifier: MIT
 					<source>1.8</source>
 					<failOnError>true</failOnError>
 					<failOnWarnings>true</failOnWarnings>
-					<doclint>all</doclint>
+					<!--
+						doclint is "all,-missing" (not plain "all") because Javadoc runs on the RAW
+						source, before Lombok's annotation processor runs, so it cannot see the
+						constructor Lombok's class-level @Builder generates. For a Lombok value class
+						like TrainingParameters it therefore reports a spurious "use of default
+						constructor, which does not provide a comment" (the `missing` group) even
+						though no public default constructor exists at compile time. Dropping only the
+						`missing` group avoids hand-writing the redundant all-args constructor that
+						@Builder already generates, while keeping the reference / html / syntax /
+						accessibility checks fatal (a broken {@link} or malformed HTML still fails the
+						build). failOnWarnings stays true so those categories remain enforced.
+					-->
+					<doclint>all,-missing</doclint>
 				</configuration>
 				<executions>
 					<execution>

From 524ce441bb8060545c6b4dbc12193169834d5f75 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Thu, 2 Jul 2026 07:30:02 +0000
Subject: [PATCH 10/12] docs: add LangChain4j integration section to main
 README
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The main README only mentioned LangChain4j in passing (Similar Projects). Add a
proper "### LangChain4j integration" section under Documentation — Maven
dependency, the borrow-the-model lifecycle, a minimal ChatModel example, and the
adapter table — mirroring how the OpenAI-compatible server is documented, and
linking to the module README for the full details.

Also fix the module README (llama-langchain4j/README.md): it still described
itself as a "sibling module, not part of the root reactor" with a standalone
install-then-build flow. Since PR #288 that is a reactor module built/versioned/
released with the core, so update the Building section to the reactor
`mvn -pl llama-langchain4j -am` flow.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rt1paYztGJ2AKUuBuAGDXE
---
 README.md                   | 40 +++++++++++++++++++++++++++++++++++++
 llama-langchain4j/README.md | 15 +++++++-------
 2 files changed, 48 insertions(+), 7 deletions(-)

diff --git a/README.md b/README.md
index b6e2ca88..6cbe38ac 100644
--- a/README.md
+++ b/README.md
@@ -704,6 +704,46 @@ tool calling depends on the model's own tool-calling quality. Pass `--api-key` (
 `OpenAiServerConfig.apiKey(...)`) to require an `Authorization: Bearer` token; the server binds to
 `127.0.0.1` by default.
 
+### LangChain4j integration
+
+A separate artifact, **`net.ladenthin:llama-langchain4j`**, adapts a `LlamaModel` to
+[LangChain4j](https://github.com/langchain4j/langchain4j)'s `ChatModel`, `StreamingChatModel`,
+`EmbeddingModel` and `ScoringModel` interfaces **in-process over JNI** — no HTTP hop, no separate
+server. It is a separate `artifactId` (not a classifier of the core) because LangChain4j 1.x
+requires **Java 17** while the core `net.ladenthin:llama` stays Java 8; keeping it separate avoids
+forcing that floor on every core consumer. It ships and versions in lockstep with the core.
+
+```xml
+<dependency>
+    <groupId>net.ladenthin</groupId>
+    <artifactId>llama-langchain4j</artifactId>
+    <version>5.0.4-SNAPSHOT</version>
+</dependency>
+```
+
+Each adapter **borrows** a `LlamaModel` you already loaded — it never loads or closes the native
+model, so you manage its lifecycle (try-with-resources), and one `LlamaModel` can back several
+adapters at once:
+
+```java
+try (LlamaModel llama = new LlamaModel(new ModelParameters().setModel("models/qwen3-0.6b.gguf"))) {
+    ChatModel chat = new JllamaChatModel(llama);
+    String reply = chat.chat("Write a haiku about lazy senior devs.");
+    System.out.println(reply);
+}
+```
+
+| Adapter | LangChain4j interface | java-llama.cpp call |
+|---------|-----------------------|---------------------|
+| `JllamaChatModel` | `ChatModel` | `LlamaModel.chat(...)` |
+| `JllamaStreamingChatModel` | `StreamingChatModel` | `LlamaModel.generateChat(...)` (token streaming) |
+| `JllamaEmbeddingModel` | `EmbeddingModel` | `LlamaModel.embed(...)` (model loaded with `enableEmbedding()`) |
+| `JllamaScoringModel` | `ScoringModel` (re-ranking) | `LlamaModel.handleRerank(...)` (model loaded with `enableReranking()`) |
+
+See [`llama-langchain4j/README.md`](llama-langchain4j/) for streaming/embedding/re-ranking
+examples and the current mapping limitations (tool calling, JSON mode, and multimodal input are
+not yet forwarded).
+
 ### Model/Inference Configuration
 
 There are two sets of parameters you can configure, `ModelParameters` and `InferenceParameters`. Both provide builder 
diff --git a/llama-langchain4j/README.md b/llama-langchain4j/README.md
index 1082d973..f4449369 100644
--- a/llama-langchain4j/README.md
+++ b/llama-langchain4j/README.md
@@ -71,16 +71,17 @@ platform the usual way (bundled in the `net.ladenthin:llama` JAR or on `java.lib
 
 ## Building
 
-This is a **sibling module**, not part of the root reactor. Install the core artifact first, then
-build here:
+This is a **reactor module** built, versioned and released together with the core (see the root
+`pom.xml` `<modules>`). To build/test just this module locally, build it with its ancestors from the
+repo root so the core is compiled/installed first:
 
 ```bash
-# from the repo root: publish the core net.ladenthin:llama jar to your local ~/.m2
-mvn -DskipTests install
+# from the repo root: build the core (-am) and this module together
+mvn -pl llama-langchain4j -am -DskipTests test
 
-# then build/test this module
-cd llama-langchain4j
-mvn test
+# or install the core first, then build here on its own
+mvn -pl llama -am -DskipTests install
+cd llama-langchain4j && mvn test
 ```
 
 The model-backed integration tests self-skip unless you point them at a GGUF. Each adapter has

From 2387f34ba6d1f43f4653abf26f60f37e89415327 Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Thu, 2 Jul 2026 07:33:24 +0000
Subject: [PATCH 11/12] docs: re-root stale README paths into llama/ after the
 reactor move

Three README references still pointed at the pre-reactor tree and now 404 as
GitHub links: the two [examples](src/test/java/examples) links and the
vision.image default path src/test/resources/images/test-image.jpg. Re-root all
three to llama/src/... to match where the sources moved. (Reported by an
automated PR review; verified the targets exist under llama/.)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rt1paYztGJ2AKUuBuAGDXE
---
 README.md | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/README.md b/README.md
index 6cbe38ac..0faa78f8 100644
--- a/README.md
+++ b/README.md
@@ -123,7 +123,7 @@ Access this library via Maven (released versions on Maven Central):
 </dependency>
 ```
 
-There are multiple [examples](src/test/java/examples).
+There are multiple [examples](llama/src/test/java/examples).
 
 ### Snapshot builds
 
@@ -318,7 +318,7 @@ Every `net.ladenthin.llama.*` system property recognised by the library, deep-sc
 | `net.ladenthin.llama.nomic.path` | unset (test self-skips) | test | `LlamaEmbeddingsTest#testNomicEmbedLoads` | Path to a Nomic embedding model (`nomic-embed-text-v1.5.f16.gguf` or a compatible BERT-family encoder). Regression test for upstream issue #98 (BERT-encoder `result_output` assertion). |
 | `net.ladenthin.llama.vision.model` | unset (test self-skips) | test | `MultimodalIntegrationTest` | Path to a vision-capable model GGUF. Any vision-capable GGUF works; CI default is `SmolVLM-500M-Instruct-Q8_0.gguf`. |
 | `net.ladenthin.llama.vision.mmproj` | unset (test self-skips) | test | `MultimodalIntegrationTest` | Matching mmproj GGUF for the vision model. |
-| `net.ladenthin.llama.vision.image` | `src/test/resources/images/test-image.jpg` (a CC-BY-4.0 / MIT-granted photo committed to the repo) | test | `MultimodalIntegrationTest` | Visual prompt image. Any png/jpeg/webp/gif works; the extension drives MIME detection. |
+| `net.ladenthin.llama.vision.image` | `llama/src/test/resources/images/test-image.jpg` (a CC-BY-4.0 / MIT-granted photo committed to the repo) | test | `MultimodalIntegrationTest` | Visual prompt image. Any png/jpeg/webp/gif works; the extension drives MIME detection. |
 | `net.ladenthin.llama.audio.model` | unset (test self-skips) | test | `AudioInputIntegrationTest` (llama.cpp discussion #13759) | Path to an audio-input model GGUF (e.g. Ultravox, Qwen2.5-Omni). |
 | `net.ladenthin.llama.audio.mmproj` | unset (test self-skips) | test | `AudioInputIntegrationTest` | Matching audio mmproj (encoder) GGUF. |
 | `net.ladenthin.llama.audio.input` | unset (test self-skips) | test | `AudioInputIntegrationTest` | `.wav`/`.mp3` audio prompt clip; the extension drives format detection. |
@@ -370,7 +370,7 @@ public class Example {
 }
 ```
 
-Also have a look at the other [examples](src/test/java/examples).
+Also have a look at the other [examples](llama/src/test/java/examples).
 
 ### Inference
 

From 56b38c0874b6b25b941b82e6a889e49cd329194e Mon Sep 17 00:00:00 2001
From: Claude <noreply@anthropic.com>
Date: Thu, 2 Jul 2026 07:37:52 +0000
Subject: [PATCH 12/12] docs: fix stale root-relative build commands in README
 after reactor move

The "Setup required" native-build steps told users to run cmake -B build at the
repo root, but CMakeLists.txt now lives in llama/ (the root is just the reactor
aggregator), so it would fail. Prefix the block with `cd llama` and update the
install-path example to /java-llama.cpp/llama/src/... . Also fix the Android
Gradle snippet to check llama/target and run mvn compile in the llama/ module.
(Reported by an automated PR review; verified against the new layout.)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rt1paYztGJ2AKUuBuAGDXE
---
 README.md | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

diff --git a/README.md b/README.md
index 0faa78f8..76450bd0 100644
--- a/README.md
+++ b/README.md
@@ -272,10 +272,12 @@ This consists of two steps: 1) Compiling the libraries and 2) putting them in th
 
 First, have a look at [llama.cpp](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md) to know which build arguments to use (e.g. for CUDA support).
 Any build option of llama.cpp works equivalently for this project.
-You then have to run the following commands in the directory of this repository (java-llama.cpp):
+You then have to run the following commands in the `llama/` module directory (the native core lives
+there; the repository root is just the Maven reactor aggregator):
 
 ```shell
-mvn compile  # don't forget this line
+cd llama       # the native core module
+mvn compile    # don't forget this line
 cmake -B build # add any other arguments for your backend, e.g. -DGGML_CUDA=ON
 cmake --build build --config Release
 ```
@@ -286,7 +288,7 @@ cmake --build build --config Release
 All compiled libraries will be put in a resources directory matching your platform, which will appear in the cmake output. For example something like:
 
 ```shell
---  Installing files to /java-llama.cpp/src/main/resources/net/ladenthin/llama/Linux/x86_64
+--  Installing files to /java-llama.cpp/llama/src/main/resources/net/ladenthin/llama/Linux/x86_64
 ```
 
 #### Library Location
@@ -871,11 +873,12 @@ git submodule add https://github.com/bernardladenthin/java-llama.cpp
 android {
     val jllamaLib = file("java-llama.cpp")
 
-    // Execute "mvn compile" if folder target/ doesn't exist at ./java-llama.cpp/
-    if (!file("$jllamaLib/target").exists()) {
+    // Execute "mvn compile" in the llama/ core module if its target/ doesn't exist
+    // (the repository root is the Maven reactor aggregator; the native core lives in llama/).
+    if (!file("$jllamaLib/llama/target").exists()) {
         exec {
             commandLine = listOf("mvn", "compile")
-            workingDir = file("java-llama.cpp/")
+            workingDir = file("java-llama.cpp/llama/")
         }
     }