chore: bump llama.cpp to b9528

[github-actions]

llama.cpp update

• Previous pin: b9165
• New pin: b9528
• Upstream release: https://github.com/ggml-org/llama.cpp/releases/tag/b9528
• Compare: ggml-org/llama.cpp@b9165...b9528

Upstream changelog

Release notes for b9528

Details

ui: run npm install when package-lock.json is newer than node_modules (#24171)

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 b9528 (first 80)

• vulkan: fuse snake activation (mul, sin, sqr, mul, add) (#22855) (47c0eda)
• Update WebGPU support and add link to blog/demo (#23483) (ee7c305)
• cmake : remove STATIC from impl libraries, enable LLAMA_BUILD_APP by default (#23462) (bb28c1f)
• CUDA: fix PDL CC check for JIT compilation (#23471) (4f0e43d)
• cmake : add install() for impl libraries + fix apple builds (#23511) (bbce619)
• vocab : fix HybridDNA tokenizer (#23466) (afcda09)
• cmake : build router app only during standalone builds (#23521) (9c92e96)
• ggml-zendnn : add Q8_0 quantization support (#23414) (99d4026)
• docs: Update documentation with Granite 4.0/4.1 (#23404) (95feeab)
• SYCL: add BF16 to DMMV kernel path (~4x tg speedup on Intel Arc) (#21580) (8cc67ef)
• SYCL : gated_delta_net K>1 (#23174) (56f16f2)
• sycl : Level Zero detection in ggml_sycl_init (#23097) (bcfd198)
• SYCL: improve MoE prefill throughput (#23142) (cc9e331)
• perplexity : fix integer overflow (#23496) (ef570f6)
• server: only parse empty msg if continuing an assistant msg (#23506) (1acee6b)
• opencl: generalize Adreno MoE kernels on M (#23449) (0f3cb3f)
• vulkan: fix windows find_package of SPIRV-Headers (#23215) (95405ac)
• ggml : Check the right iface method before using the fallback 2d get (#23514) (a497476)
• model : add NVFP4 MTP scale tensors (#23563) (b0df4c0)
• requirements : bump torch to 2.11.0 (#23503) (c0c7e14)
• cmake/ui : refactor the build (#23352) (b22ff4b)
• snapdragon: update windows toolchain to use hsdk v6.6.0.0 (#23552) (cec51c7)
• hexagon: apply repl optimization in flash attn softmax as #22993 (#23455) (1c0f6db)
• opencl: batch profiling to improve speed and prevent memory leaks (#23495) (f306111)
• TP: fix entirely zero-sized slices per device (#23525) (fff63b5)
• server: add margin for draft model for fit (#23485) (83eebe9)
• cmake : fix ui build (#23592) (63248fc)
• convert : minor fixes for numpy 2.x (#23571) (5d246a7)
• ci : update build-self-hosted.yml (#23616) (549b9d8)
• ci : move most slim jobs to self-hosted runners (#23619) (28123a3)
• perplexity : fix even more integer overflows (#23623) (6d57c26)
• server: fix checkpoints creation (#22929) (e2ef8fe)
• vendor : update cpp-httplib to 0.45.1 (#23639) (9627d0f)
• ui: media attachments before text (#23467) (b964876)
• ggml : Parallelize quant LUT init (#23595) (826539c)
• ci : install host compiler on android-ndk build (#23630) (d55fb97)
• llama : document that only one on-device state can be saved per sequence (#23520) (314e729)
• ci : fix pre-tokenizer-hashes check (#23651) (062d311)
• ci : update spacemit toolchain url and enhance curl command (#23642) (5fdf07e)
• server: MTP layer kv-cache should respect draft type ctk (#23646) (6c4cbdc)
• ggml: gguf_init_from_callback and gguf_init_from_buffer (#22341) (66efd13)
• TP: fix ggml context size calculation (#22616) (ae251b5)
• ggml-alloc: fix out-of-bounds read in ggml_dyn_tallocr_remove_block (ggml/1492) (fa97041)
• ggml.h: correct ggml_silu_back arg docstring (a=dy, b=x) (ggml/1500) (b251f74)
• ggml : bump version to 0.12.1 (ggml/1508) (ce5890b)
• sync : ggml (22307b3)
• ggml : bump version to 0.13.0 (ggml/1510) (45158f4)
• sync : ggml (d161ea7)
• convert : add compressed-tensors NVFP4 support (#21095) (a4d2d4a)
• ui: fix stop/continue during an agentic loop (#23356) (5a4126a)
• CUDA: add fast walsh-hadamard transform (#23615) (c1f1e28)
• model: tag ffn_latent as MUL_MAT to fix buft probe (#23664) (328874d)
• ci : reduce PR jobs by matching backend paths (#23675) (302e2c2)
• snapdragon: bump toolchain docker to v0.7 to fix ui build issues (#23680) (4bead4e)
• metal : add apple device id (#23566) (35c9b1f)
• CUDA: missing PDL sync for FWHT, better fallback (#23690) (192d8ae)
• [WebGPU] Check batch_compute_passes before sending passes when not doing GPU profiling (#23457) (54121f7)
• ggml-webgpu: Add MMVQ path for Q4/Q8/Q2_K/Q4_K and clean up legacy MUL_MAT pipeline (#23594) (1506d39)
• model : add support for talkie-1930-13b (#22596) (c9d9829)
• tests: test-backend-ops -j to run tests in parallel (#23637) (7623de1)
• SYCL: implement ggml_sycl_pool_vmm (#22862) (581d020)
• models : Attach Mistral3 NVFP4 weight scales (#23629) (6fe90de)
• convert : support Gemma4ForCausalLM architecture (#23682) (dbe9c0c)
• ci : reduce (disable SYCL and CANN builds/releases) (#23705) (3dc7684)
• ci : move sanitizer jobs to self-hosted runners (#23713) (ef41a69)
• ci : move more CPU jobs to self-hosted runners (#23715) (678d43d)
• hexagon: add support for CONCAT op (#23648) (ef66bfa)
• ci : remove vulkan SDK dep from webgpu job (#23718) (3a3ed15)
• 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)

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.