NativeServer mode, GPU/CPU classifier matrix (ROCm/SYCL/OpenVINO/Win-arm64/s390x), llama.cpp b9870 + bump automation#293

Merged

bernardladenthin merged 29 commits into

mainfrom

claude/granite-4-llama-cpp-decode-kk9xjp

Jul 4, 2026

Commits on Jul 2, 2026

Fix multi-turn tool-calling checkpoint starvation for recurrent models (Granite-4)
claude
committed

Commits on Jul 3, 2026

patch 0005: extract checkpoint conditions into named booleans
claude
committed
server: add -b/-ub/-tb/-ctk/-ctv/--jinja/--chat-template-kwargs to OpenAiServerCli
claude
committed
server: add NativeServer — run the full llama.cpp server (WebUI) in-DLL via JNI
claude
committed
server: make NativeServer the default fat-jar Main-Class (keep OpenAiCompatServer)
claude
committed
server: add ServerLauncher — one fat-jar entry, pick mode via --open-ai-compat
claude
committed
server: simplify ServerLauncher dispatch to one primitive; rename flag to --openai-compat
claude
committed
Namespace the ServerLauncher selector flag as --jllama-openai-compat
claude
committed
Upgrade llama.cpp from b9859 to b9862
claude
committed
Surface model ftype (quantization) through the Java layer and /v1/models
claude
committed
Add Linux Vulkan classifiers + Windows arm64 CPU to the build matrix
claude
committed
Emit a per-job sccache statistics table to the GitHub job summary
claude
committed
Upgrade llama.cpp from b9862 to b9864
claude
committed
Expose sse_ping_interval + audited completion params on InferenceParameters
claude
committed
CI fixes: SpotBugs suppressions, spirv-headers, clang-cl for Windows arm64
claude
committed
Fix ArchUnit violations from the new server code (layering + no-sleep)
claude
committed
Windows arm64: disable OpenMP so the clang-cl build is self-contained
claude
committed
Upgrade llama.cpp from b9864 to b9866
claude
committed
Automate llama.cpp version-bump target selection and chunking
claude
committed
Upgrade llama.cpp from b9866 to b9867
claude
committed
Add REUSE license header to the version-bump runbook doc
claude
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

NativeServer mode, GPU/CPU classifier matrix (ROCm/SYCL/OpenVINO/Win-arm64/s390x), llama.cpp b9870 + bump automation#293

NativeServer mode, GPU/CPU classifier matrix (ROCm/SYCL/OpenVINO/Win-arm64/s390x), llama.cpp b9870 + bump automation#293
bernardladenthin merged 29 commits into
mainfrom
claude/granite-4-llama-cpp-decode-kk9xjp

Commits on Jul 2, 2026

Fix multi-turn tool-calling checkpoint starvation for recurrent models (Granite-4)

Commits on Jul 3, 2026

patch 0005: extract checkpoint conditions into named booleans

server: add -b/-ub/-tb/-ctk/-ctv/--jinja/--chat-template-kwargs to OpenAiServerCli

server: add NativeServer — run the full llama.cpp server (WebUI) in-DLL via JNI

server: make NativeServer the default fat-jar Main-Class (keep OpenAiCompatServer)

server: add ServerLauncher — one fat-jar entry, pick mode via --open-ai-compat

server: simplify ServerLauncher dispatch to one primitive; rename flag to --openai-compat

Namespace the ServerLauncher selector flag as --jllama-openai-compat

Upgrade llama.cpp from b9859 to b9862

Surface model ftype (quantization) through the Java layer and /v1/models

Add Linux Vulkan classifiers + Windows arm64 CPU to the build matrix

Emit a per-job sccache statistics table to the GitHub job summary

Upgrade llama.cpp from b9862 to b9864

Expose sse_ping_interval + audited completion params on InferenceParameters

CI fixes: SpotBugs suppressions, spirv-headers, clang-cl for Windows arm64

Fix ArchUnit violations from the new server code (layering + no-sleep)

Windows arm64: disable OpenMP so the clang-cl build is self-contained

Upgrade llama.cpp from b9864 to b9866

Automate llama.cpp version-bump target selection and chunking

Upgrade llama.cpp from b9866 to b9867

Add REUSE license header to the version-bump runbook doc

Commits on Jul 4, 2026

Upgrade llama.cpp from b9867 to b9870

Add 8 GPU-backend classifiers: ROCm/HIP, SYCL, Win-arm64 OpenCL, OpenVINO

Fix vendor-toolchain installs for the 4 failing GPU classifier jobs

Fix GPU classifier build errors: HIP MSVC, SYCL/OpenVINO CRT, OpenVINO OpenCL

OpenVINO: bump to 2026.2.1 + fix OpenCL headers; raise cmake min to 3.22

Linux OpenVINO: install 2026.2.1 from archive, not the (nonexistent) 2026 apt repo

Add Linux s390x (big-endian) build with a qemu-user C++ test gate

NativeServer.main: own the server in try/finally (SonarQube S2095)