[leehack]

llama.cpp update

• Previous pin: b10255
• New pin: b10276
• Upstream release: https://github.com/ggml-org/llama.cpp/releases/tag/b10276
• Compare: ggml-org/llama.cpp@b10255...b10276

Upstream changelog

Release notes for b10276

Details

Prefer npm ci over install for security (#26601)

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 b10255 to b10276 (first 80)

• sycl: parallelize the non-contiguous concat kernel (#25852) (6c8dcaa)
• ci: fix vulkan llvmpipe runs (#26533) (22dc605)
• llama : move n_vocab from llama_sampler_data to penalty_sampler (#26520) (935cad6)
• model : allow reshape of tensors during load (#26531) (1269cb1)
• convert : import bytes_to_unicode from convert_slow_tokenizer (#26217) (b06fbc9)
• vocab : validate plamo2 byte tokens (#26511) (cf06ad7)
• vulkan backend ops: implemented GATED_LINEAR_ATTN (#25601) (f26efa0)
• convert : add missing return after setting tekken vocab (#25947) (b5746d2)
• ggml : bump version to 0.18.1 (ggml/1578) (15831f5)
• sync : ggml (2e17f69)
• gguf-py: validate n_dims and guard against uint64 overflow in reader (#25401) (5788b51)
• speculative : refactor enabled configs common_speculative_init (#26510) (7bd8282)
• ci: fix pre-built binaries no longer working on macOS 15 and below (#26375) (6b5224c)
• models : fix dflash wo_a reshape on load (#26577) (1c3c967)
• mtmd: support Qwen3-TTS (note: breaking change to llama-tts binary) (#26254) (0713275)
• ui: CWD for agent (#26518) (2f56fc3)
• gguf-split: Add option to delete split parts during merge (#26538) (76c956c)
• sampler : remove "full-context windows" from history-based samplers (#26524) (a6aa6f5)
• mtmd: correcting duplicate empty audio chunks for short inputs (#26536) (474c92e)
• server: decode Windows OEM output to UTF-8 in built-in tools (#26597) (4308a4f)
• Prefer npm ci over install for security (#26601) (6ea215d)

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.

[Copilot]

Pull request overview

This PR updates the pinned upstream llama.cpp release tag used by this repository’s build/CI workflows and bridge build process.

Changes:

• Bump llama.cpp pin from b10255 to b10276.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[leehack]

The b10276 ABI blocker is fixed at 62417a8: llama_sampler_init_penalties now receives the active vocabulary size, with a static CI contract guarding the call shape.

Exact-head validation:

• GitHub CI passed: wasm32 + memory64 build, output verification, model-backed state persistence, and integrity-pinned Qwen3.5 multimodal browser smoke
• local exact b10276 wasm32/memory64 compilation passed
• local direct + worker state save/load passed
• local direct + worker Qwen3.5 multimodal inference passed with matching output
• JS type/build/syntax checks and CI reliability checks passed

No assets were published; the publish job correctly skipped on the PR event.