chore: bump llama.cpp to b9596

[github-actions]

llama.cpp update

• Previous pin: b9165
• New pin: b9596
• Upstream release: https://github.com/ggml-org/llama.cpp/releases/tag/b9596
• Compare: ggml-org/llama.cpp@b9165...b9596

Upstream changelog

Release notes for b9596

Details

server: skip unused log lines on router mode (#24463)

macOS/iOS:

• macOS Apple Silicon (arm64)
• macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED
• macOS Intel (x64)
• iOS XCFramework

Linux:

• Ubuntu x64 (CPU)
• Ubuntu arm64 (CPU)
• Ubuntu s390x (CPU)
• Ubuntu x64 (Vulkan)
• Ubuntu arm64 (Vulkan)
• Ubuntu x64 (ROCm 7.2)
• Ubuntu x64 (OpenVINO)
• Ubuntu x64 (SYCL FP32) DISABLED

Android:

• Android arm64 (CPU)

Windows:

• Windows x64 (CPU)
• Windows arm64 (CPU)
• Windows x64 (CUDA 12) - CUDA 12.4 DLLs
• Windows x64 (CUDA 13) - CUDA 13.3 DLLs
• Windows x64 (Vulkan)
• Windows x64 (SYCL) DISABLED
• Windows x64 (HIP)

openEuler:

• DISABLED
• openEuler x86 (310p)
• openEuler x86 (910b, ACL Graph)
• openEuler aarch64 (310p)
• openEuler aarch64 (910b, ACL Graph)

UI:

• UI

Commit range

Commits from b9165 to b9596 (first 80)

• vulkan: optimize conv2d and implement coopmat1 support (#22620) (7799d31)
• ci : move macos jobs to the apple workflow + fix names (#23721) (5190c2e)
• ci : add [no release] keyword + fix sanitizer builds (#23728) (35a74c8)
• ci : move [no release] check to dedicated check_release job (#23734) (08bc21b)
• ci : do not allocate ccache for 3rd-party hosted runners (#23730) (0d18aaa)
• ggml-zendnn : fixed naming of matmul function (#20964) (b4c0549)
• server : fix the log message when using SSL (#23393) (7085492)
• convert: add MiniCPM5 tokenizer support (#23384) (9777256)
• docs : fix duplicated "the" in granitevision and model-conversion docs (#23767) (1d971bb)
• ci : add ccache to server builds + fix undefined sanitizer build (#23763) (0d227ec)
• vulkan: avoid preferring transfer queue on AMD UMA devices (#22455) (4d8cc0c)
• ci : remove wasm test (#23733) (b3a739c)
• ci : fix windows ccaches (#23777) (9f0e4b1)
• common : fix env names to all have LLAMA_ARG_ prefix (#23778) (6b4e4bd)
• ci : bump cuda release to 13.3 (#23749) (2d0656f)
• CUDA: restrict PDL to CTK >= 12.3 due to MSVC issues (#23742) (fda8528)
• pyproject : add conversion folder and update dependencies (#23746) (87b0a60)
• vendor : update cpp-httplib to 0.46.0 (#23650) (617255d)
• ci : move ARM jobs to self-hosted + disable kleidiai mac release (#23780) (ba4dd0b)
• vulkan: add REPEAT op support for f16 to f16. (#23298) (837bb6b)
• vulkan: use GL_NV_cooperative_matrix_decode_vector for faster matmul (#23541) (b36eefc)
• vulkan: Switch MUL_MAT_VEC to 4 K per iteration for F16/32 (#22887) (c6e4088)
• ggml-webgpu: Fix how to dispatch WG to some ops (#23750) (c40006a)
• hexagon: add support for Q4_1 in MUL_MAT and MUL_MAT_ID (#23647) (aa50b2c)
• ggml-webgpu: remove legacy constants (#23672) (f12cc6d)
• opencl: OP_GATED_DELTA_NET (#23312) (8ad8aef)
• Hexagon: OP_GATED_DELTA_NET K>1 support (#23531) (939a7dd)
• ci : refactor (#23789) (491c4d7)
• ggml: fixed Arm SVE usage bug in vec.h, vec.cpp (#22841) (e31cdaa)
• convert : add FP8 to Q8 conversion (#23250) (c522908)
• perplexity : fix format specifier in LOG_ERR (#23788) (48e7eae)
• cuda : fix KQ mask offset integer overflow in fattn MMA kernel (#23610) (09e7b76)
• docker : add ZenDNN Dockerfile (#23716) (e8d2567)
• server, ui : Add support for HTTP ETags in llama-server (#23701) (d205df6)
• vulkan: Fix memory logger unsafe iterator access (#23667) (91eb8f4)
• vulkan: fix wrong index variable in inner loop (#23665) (7c48fb8)
• chat : add Granite 4.1 chat template (#23518) (bb771cb)
• vulkan: fast path for walsh-hadamard transform (#23687) (48e7078)
• hexagon: minor refresh for HMX FA and MM (#23796) (a919001)
• server: minor tweaks to use more cpp features (#23785) (0b24686)
• CUDA: route batch>=4 quantized matmul to MMQ on AMD MFMA hardware (#23227) (bc81d47)
• mmvq Optim: add MMVQ_PARAMETERS_TURING(mmvq_parameter_table_id) for … (#23729) (d7be461)
• ggml: auto apply iGPU flag CUDA/HIP if integrated device (#23007) (30af6e2)
• test-llama-archs: fix table format [no release] (#23810) (d374e71)
• arg: Add LLAMA_ARG_API_KEY_FILE environment variable for --api-key-file (#23167) (7fb1e70)
• ci : change Vulkan builds to Release to reduce ccache (#23820) (dd15579)
• mtmd: fix gemma 4 audio rms norm eps (#23815) (d6be315)
• mtmd: n_head_kv defaults to n_head (#23782) (0b56d28)
• app : improve help output (#23805) (479a9a1)
• ci : releases use Github-hosted builds for the UI (#23823) (445b7ce)
• ui: fix audio and video modality detection (#23756) (2f6c815)
• ci : run ui publish on ubuntu-slim (#23818) (3ef2369)
• opencl: move backend info printing into its own function (#23702) (408ae2b)
• mtmd: fix gemma 4 projector pre_norm (#23822) (c8914ad)
• mtmd-debug: add color and rainbow mode (#23829) (751ebd1)
• hexagon: basic/generic op fusion support and RMS_NORM+MUL fusion (#23835) (19e92c3)
• meta : Add missing buffer set in allreduce fallback !COMPUTE clear (#23480) (33c718d)
• cuda : disables launch_fattn PDL enrollment due to compiler bug (#23825) (241cbd4)
• app : move licences to llama-app (#23824) (98e480a)
• llama: add llm_graph_input_mtp (#23643) (eef59a7)
• ngram-mod : Add missing include (#23857) (b000431)
• ggml : bump version to 0.13.1 (ggml/1523) (ea02bc3)
• sync : ggml (fe12e42)
• llama: use f16 mask for FA to save VRAM (#23764) (031ddb2)
• model : support for DeepseekV32ForCausalLM with generic DeepSeek Sparse Attention (DSA) implementation (#23346) (1f0aa2a)
• server: bump timeout to 3600s (#23842) (cb47092)
• CUDA: Check PTX version on host side to guard PDL dispatch (#23530) (6ed481e)
• mtmd: Add DeepSeekOCR 2 Support (#20975) (da3f990)
• download: add option to skip_download (#23059) (06d26df)
• ci : update macos release to use macos-26 runner (#23878) (dc71236)
• server: remove obsolete scripts (#23870) (b5f5228)
• graph : ensure DS32 kq_mask_lid is F32 (#23864) (764f1e6)
• vocab : support tokenizer for LFM2.5-8B-A1B (#23826) (2084434)
• ui: handle audio/vnd.wave as audio WAV file (#23754) (22d66b5)
• app: add llama update self updater (#23865) (5a46b46)
• server-bench : add speed-bench for speculative decoding benchmarking (#23869) (689a9a4)
• ggml-webgpu: add q4_0/q8_0 SET_ROWS (#23760) (b22da25)
• ggml-webgpu: Check earlier for WebGPU required features (#23879) (151f3a9)
• server: in SSE mode, send HTTP headers when slot starts (#23884) (0821c5f)
• llama : do not skip iGPU when only RPC devices are present (#23868) (1738129)

Web bridge review focus

Please pay extra attention to upstream changes touching:

• WebGPU, WASM, Emscripten, pthreads, or memory64 build behavior
• ggml backend APIs used by the bridge
• model loading, tokenizer, chat template, context/state persistence, or cache semantics
• CMake/build flags that can affect the generated JS/WASM artifacts

Validation

• Emscripten build passed
• Browser WebGPU/state-persistence smoke passed
• Generated bridge artifacts include wasm32 and memory64 outputs
• No stale hard-coded llama.cpp tag remains in CI/publish defaults

Automation behavior

This PR is managed from the stable branch automation/bump-llama-cpp. If another llama.cpp release appears before merge, the scheduled workflow updates this same PR instead of opening a duplicate. The workflow skips if a non-automation PR already changes llama_cpp.version.