chore: bump llama.cpp to b10066

[github-actions]

llama.cpp update

• Previous pin: b9915
• New pin: b10066
• Upstream release: https://github.com/ggml-org/llama.cpp/releases/tag/b10066
• Compare: ggml-org/llama.cpp@b9915...b10066

Upstream changelog

Release notes for b10066

Details

opencl: load and use kernel_gemm_moe_q6_k_f32_ns from bin kernel lib (#25797)

Website:

• https://llama.app

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)
• Ubuntu x64 (SYCL FP16)

Android:

• Android arm64 (CPU)

Windows:

• Windows x64 (CPU)
• Windows arm64 (CPU)
• Windows arm64 (OpenCL Adreno)
• Windows x64 (CUDA 12) - CUDA 12.4 DLLs
• Windows x64 (CUDA 13) - CUDA 13.3 DLLs
• Windows x64 (Vulkan)
• Windows x64 (OpenVINO)
• Windows x64 (SYCL)
• 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 b9915 to b10066 (first 80)

• ggml : fix A indexing in simd_gemm scalar tail-column path (#25390) (57b50e1)
• fix: OOB reads in UGM tokenizer (precompiled_charsmap handling) (#18750) (4a7ee31)
• metal : add set_rows with src0 f16 (#25434) (0512ef1)
• llama-eval : fix crash when answer is None in HTML dump (#25435) (da46e59)
• ui: Context usage gauge and panel (#25340) (f1161b1)
• common: auto-create prompts-log-dir at argument parsing, so all tools using the flag benefit (#25322) (f296fdf)
• llama-batch: add n_keep_tail in split_equal for recurrent models (#25278) (230ea9d)
• server-stream: follow-up on SSE Replay Buffer (#23226) (#25047) (bbebeec)
• llama: refactor fused ops (#24646) (90e0f5c)
• cuda : add support for f16->f16 GGML_OP_SET_ROWS (#25367) (ed8c261)
• Make hip quality check run on all changes (#25403) (07e012a)
• cli : move to HTTP-based implementation (#24948) (c264f65)
• hexagon: new vtcm layouts and improved pipelines for MUL_MAT, MUL_MAT_ID and FLASH_ATTN_EXT (#25425) (81ff7ab)
• vulkan: for small AMD GPUs, reduce submission threshold based on CU count (#25240) (0bbc87b)
• llama-batch: fix allowed decreasing pos in a seq (#25449) (1ee0939)
• opencl: ragged-tile MoE prefill FP16 GEMM optimization (skip padded expert tiles) (#25433) (167d057)
• vulkan: disable FA mask_opt on GCN to improve performance (#24362) (a646006)
• opencl: Q6_K GEMM/GEMV fix for ne01 of weights that are not multiples of 128. (#25464) (92366df)
• ggml-webgpu: tune subgroup split (d_split) in flash_attn_vec (#25418) (32e41fa)
• hexagon: add VISION RoPE support (#25216) (f2d1c2f)
• server : respect min-step when splitting prompt batches (#25420) (64c8b7d)
• cuda: align snake fusion matcher with the other backends (#25460) (2021515)
• ggml-hip: enable -funsafe-math-optimizations (#24668) (ccb0c34)
• metal : add CONV_2D_DW (depthwise convolution) support (#21565) (92b187c)
• llama-bench : init params.offline (#25476) (259f2e2)
• Only index by compile times + always multiply/add (#25445) (683f0c7)
• Refactor: Consistently use smart pointers in test-backend-ops (#25440) (f84a519)
• meta: add hard emphasis on agents not writing descriptions/comments (#25480) (c15c5c7)
• ggml : fix conv 2d dw (#25490) (5c3a586)
• server : move chat-template thinking probe inside the init try/catch (#24093) (82fce65)
• hexagon: tiling, tracing and optimizations for unary ops (#25474) (fb30ba9)
• cli: add --output option (#25484) (3de7dd4)
• ggml : process data in smaller chunks in CUDA ggml_top_k() and ggml_argsort() to reduce temporary buffers memory usage (#24776) (0749449)
• opencl: cluster-parallel decode FA for Adreno (#25473) (049326a)
• llama-batch: add unit test (#25471) (961e4b2)
• ggml-et: Initial ET backend (#24179) (082b326)
• llama : make all KQ masks f16 if FA is used, remove zero attention bias, remove raw_k repeats in DeepSeek V4 (#25370) (2ed3c1a)
• ui: export full message tree instead of active path only (#25501) (d46786f)
• sync : ggml (#25517) (8f114a9)
• ui: prevent tooltip from flickering open and closed on hover (#25503) (c7af942)
• cli: fix crash on wrong server base url (#25497) (1b9691b)
• server: improve tools, remove apply_diff (#25498) (c4ae9a8)
• ggml : use ggml_vqtbl1q_u8 for 32-bit compat (whisper/0) (ac17f8a)
• sync : ggml (0badc06)
• server: remove loading.html (#25500) (a935fbf)
• ui: use server modalities in non-router mode (#24874) (9f623c6)
• feat: pre-select models in the webui using alias (#25492) (07d9378)
• mtmd: deepseek-ocr v1 multi-tile (#24717) (3e706dd)
• arg: prevent duplicate spec model downloads (#25527) (22b69b6)
• hexagon: improve ARGSORT performance for small tensors (#25512) (67776ea)
• llama : make tensor-split regex patterns static (#24710) (c749cb0)
• server: accept null sampling params (#25538) (4f37f51)
• opencl: add int8 dp4 dense and MoE prefill optimization for Adreno GPUs (#25537) (1d1d9a9)
• Vulkan: route large matmuls to medium tile on Adreno (#24877) (76f2798)
• ggml : add GGML_OP_LIGHTNING_INDEXER that implements DeepSeek V3.2/V4 lightning indexer (#24231) (00f5442)
• server: refactor server_stream (#25541) (ea1f7bb)
• server: allow stream for exec_shell_command (#25526) (c92e806)
• DeepseekV4: clear cache only for seq rather than full (#25521) (13f2b28)
• cuda: Don't crash when querying memory on device with no free memory. (#25157) (3cec3bc)
• gguf : reject empty metadata keys (#24917) (d72bfa3)
• Fix conditional to display 'LLAMA_SPLIT_MODE_TENSOR not implemented for architecture' message (#24926) (e3546c7)
• server : fix image blocks in tool_result being dropped during Anthropic OpenAI conversion (#22536) (6b4dc21)
• server : evict checkpoints within min-step of each other (#25472) (0c4fa7a)
• mtmd: fix silent prompt truncation on embedded NUL (#25548) (4114ba1)
• server: Don't consider models with --no-mmproj-auto as multimodal (#25590) (8014d2c)
• vendor : update cpp-httplib to 0.50.1 (#25576) (3455882)
• server: honour per-request reasoning_budget_tokens in chat completions (#23116) (99f3dc3)
• ui: Remove recommended MCP Servers + improve MCP Servers Settings UI/UX (#25535) (38fd5c9)
• sycl: add Q2_K to DMMV reorder path (#25064) (e474bba)
• sycl: add fused top-k MoE (#25217) (efb3036)
• chat : fix reasoning leak with force-opened bare templates (#24674) (91c631b)
• gguf : add tensor shape accessor (#24405) (ad8d821)
• tests: Harmonize header use (#25616) (f4253ef)
• readme : add link to maintainer PRs (#25621) (4193ea6)
• spec: add Minimax2 eagle3 support (259ae1d)
• vulkan: Use native e2m1 and e4m3 conversions for mxfp4/nvfp4 (#25338) (e920c52)
• CUDA: refactor MMQ kernel configuration (#24127) (6eddde0)
• model: add Hy3 (hy_v3) support with MTP speculative decoding (#25395) (2969d6d)
• metal : add Q2_0 support (#25419) (14d3ba4)
• sycl: set fattn_vec_nthreads to 256 for Battlemage (#25205) (c1063ac)

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.