ggerganov/llama.cpp

mtmd : rename mtmd_get_audio_bitrate to mtmd_get_audio_sample_rate examples

opencl: add `set`, i32 for `cpy` ggml OpenCL

[WebGPU] Fix wait logic for inflight jobs devops ggml

hexagon: add llama-completion windows runner script script

opencl: add q6_K gemm and gemv kernels for Adreno ggml OpenCL

server: Add OpenRouter-compatible reasoning API examples server

Hybrid model cache: add `--checkpoint-every-nb` examples server

llama : add attention weights extraction API [EXPERIMENTAL] examples python

vulkan: Fix data races in coopmat1 mul_mat(_id) Vulkan ggml

CUDA: Add BF16 path to CUBLAS and increase precision of FP16 path Nvidia GPU ggml

fix: correct EXAONE3 FFN_DOWN tensor mapping prefix python

fix: speculative decoding broken on hybrid SSM/MoE (Qwen3.5 MoE)

vendor : update cpp-httplib to 0.36.0 script python

kleidiai : support for concurrent sme and neon kernel execution documentation ggml

cli: add /think command to toggle reasoning examples

ggml-webgpu: Add the support of `GGML_OP_CONCAT` documentation ggml

cli: Don't clear system prompt when using '/clear' examples

webui: Improvements for Models Selector UI examples server

cmake: fix ARM feature detection hang on platforms without SVE/SME ggml

llama: parallel model loading across GPU contexts

ggml : add NVFP4 quantization type support for metal testing python ggml Apple Metal

vulkan: Fix ErrorOutOfHostMemory on Intel GPU when loading large models with --no-mmap Vulkan ggml

server: fix infinite retry loop when KV cache is full examples server

fix(docs): correct typos found during code review documentation model script testing Nvidia GPU Vulkan examples python server ggml SYCL Apple Metal Ascend NPU OpenCL jinja parser

contributing: limit open PRs for new contributors to 1

cann: support flash attention for head dim not multiple of 16 ggml Ascend NPU

vulkan: add UMA zero-copy async transfers and fix event_record deferred memcpy handling testing Vulkan ggml

vulkan: add sparse OOM fallback for large UMA allocations and chunked staging fallback testing Vulkan ggml

feat: add --cache-only flag to skip model re-download

server: add Qwen3-Reranker instruction support examples python server