master
ggml: backend-agnostic tensor parallelism (experimental) (#19378)
ggml-cpu: Optimized x86 and generic cpu q1_0 dot (follow up) (#21636)
ggml-cuda: flush legacy pool on OOM and retry (#22155)
hexagon: optimize HMX matmul operations (#21071)
metal: Implement ROLL op (#21946)
ggml-cuda: native bf16 flash attention for vec kernel (#20525)
opencl: refactor q8_0 set_tensor and mul_mat host side dispatch for Adreno (#21938)
rpc : refactor the RPC transport (#21998)
[SYCL] Fix reorder MMVQ assert on unaligned vocab sizes (#22035)
vulkan: Support F16 OP_FILL (#22177)
ggml-webgpu: updated matrix-vector multiplication (#21738)
ggml : remove ggml-ext.h (#21869)
hexagon: enable offloading to Hexagon on Windows on Snapdragon (#19150)
Tensor-parallel: Fix delayed AllReduce on Gemma-4 MoE (#22129)
ggml : add OpenVINO backend (#15307)
ggml: add graph_reused (#21764)
ggml: add Q1_0 1-bit quantization support (CPU) (#21273)
fix: free ctx_copy in ggml_opt_free to plug per-training-session leak (#21592)
ggml : build backends as libraries (#10256)
remove CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS (#10797)
ggml : Print backtrace on uncaught C++ exceptions (ggml/1232)
llama: fix llama-model-saver (#20503)