ggerganov/llama.cpp

rpc : read cached tensor files with fread instead of std::ifstream ggml

vulkan: use shared-memory reduction for dequant mul_mat_vec on Imagination PowerVR Vulkan ggml

CUDA/HIP: derive MMVQ nwarps from kernel launch bounds ggml CUDA

model, mtmd: fix gemma4 vision handling model mtmd

args: remove mmap/mlock/dio flags from arg parser documentation server

fix: speculative: zero MTP carrier at sequence start

ggml-cpu(s390x): fix q5_1 uninitialized v_acc ggml merge ready

respect -fitc from llama-bench instead of using the benchmark context size

memory : avoid allocating V cache for indexer (it's not used) in Qwen3.8-Flash-Next (qwen4exp)

ggml-cpu : add OS XSAVE checks to x86 CPU feature scoring. ggml

fix(tool-call): capture non-string args raw to fix array<object> parsing documentation model build testing Vulkan devops server ggml SYCL Apple Metal Ascend NPU OpenCL CUDA OpenVINO conversion

llama: refactor lazy mode auto, fix iGPU regression examples

tests : use 1 thread for data initialization testing

convert: fix LoRA conversion crash for Qwen3.5 V-head reorder conversion

src : add n_expert_used_max function

common : make build info output stream configurable examples

server : encode images on the GPU by borrowing the idle compute buffers server mtmd

api : mark `llama_context_params.ctx_other` as `const`

ROCm: resolve TOP_K kernels ggml CUDA

fix edge case for repetition count threshold testing

graph : keep the backend sampling subgraph static across ubatches also helps fix CI issue

HIP : add shared-memory tiled transpose fast path for F32 concat on RDNA3.5 testing ggml CUDA draft

server : fix context checkpoint eviction on prompts shorter than checkpoint_min_step server

metal : skip the empty half of the mul_mm_id token tile, load iq2/iq3 codebooks as uint32 testing ggml Apple Metal

compilation: detect and enable AVX-VNNI compilation on MSVC ggml

common : add defer_loading to tool definitions documentation testing server

add gt-asr-1.7b server mtmd conversion

fix the msvc clang compile test backend ops mismatch ggml

imatrix : parallelize with OpenMP examples

Fixed GBNF grammar generation for empty object schemas testing examples