ggerganov/llama.cpp

arch : add missing tensor name for Qwen 3.5

CANN: split RoPE cache init into host and device phases

fix: correct misspellings in code comments

common : simplify autoparser tagged parser rules

common : cleanup logs and modernize the progress bar

common : gpt-oss handle builtins and unsolicited tool calls testing

opencl: fix leak in Adreno q8_0 path ggml OpenCL

CI: Enable CPU and Vulkan ARM64 Release devops

webui: fix syntax highlighting lost after streaming for non-common languages examples server

CANN: Add suport for Qwen35 ops testing ggml Ascend NPU

server: respect the ignore eos flag examples python server

Fix undefined timing measurement errors in server context examples server

llama-server: translating structured generation request parameters from responses API format to completions API format examples python server

[SYCL] Enhance flash-attention performance ggml SYCL

tests: allow exporting graph ops from HF file without downloading weights testing

Add API key server support with optional arguments --api-key and --ju… examples python

common : init in params parser, add Windows UTF-8 support testing examples server

server: improve Responses API compliance and Codex CLI compatibility examples python server

ggml-cuda: fix ROCm multi-GPU illegal memory access in recurrent state restore Nvidia GPU ggml

ggml-cuda: ds_read_b128 for q4_0 and q4_1 mmq kernels Nvidia GPU ggml

contrib : clarify code origin guidelines

cpp: Adding new arch RUGPT3XL model python

Cross-backend profiler documentation Nvidia GPU Vulkan examples python ggml SYCL Apple Metal Ascend NPU OpenCL IBM zDNN Hexagon OpenVINO WebGPU

[CUDA ] Write an optimized flash_attn_stream_k_fixup kernel Nvidia GPU ggml

ggml-cpu: fix fallback for RVV kernels without zvfh ggml

examples : add llama-eval examples python

Support for DeepseekV32ForCausalLM with DeepSeek Sparse Attention (DSA) model testing Nvidia GPU python ggml

chore(docs): update list of UIs

ggml-webgpu: Add the support of `MUL_MAT_ID` documentation ggml WebGPU

common: add two-phase graceful reasoning budget termination ... testing