llama.cpp Releases
609 releases of ggml-org/llama.cpp
- b10017
Adjusts SYCL memory allocation thresholds to handle larger models running on devices with limited VRAM, improving stability for systems with memory overcommitment.
Jul 15, 2026
- b10016
Introduces Flash Attention support for the SYCL backend using the XMX engine and oneDNN graph API, significantly improving prefill speeds on Battlemage architecture.
Jul 15, 2026
- b10015
A minor maintenance update fixing an OpenCL compatibility issue and releasing standard binaries for all supported platforms.
Jul 15, 2026
- b10012
A maintenance release addressing a Hexagon signal enum narrowing issue and providing standard binary builds across macOS, Linux, Android, Windows, and openEuler platforms.
Jul 14, 2026
- b10011
Refactored prompt cache state ownership in the server by moving prompt state data to `server_prompt_cache` and adding handling for batched slot clearing.
Jul 14, 2026
- b10010
Added `--cors-*` options to the server to allow configuration of CORS headers, including a special 'localhost' value, with accompanying tests.
Jul 14, 2026
- b10007
Fixed a bug in the OpenCL backend where it failed to initialize on devices without the `cl_khr_integer_dot_product` extension by preventing the use of unavailable dot product kernels.
Jul 14, 2026
- b10005
Fixed sequence management issues in the DeepseekV4 implementation, specifically correcting `seq_rm` and implementing proper `seq_cp` and update context logic.
Jul 14, 2026
- b10004
Added full support for f16 as SET_ROWS source in the Vulkan and CPU backends, improving compatibility and fixing denormalized number handling on Intel GPUs.
Jul 14, 2026
- b10003
Refactored the tokenize tool to use shared argument parsing (`common_params_parse`) for consistency, exposing new model sourcing and tokenization flags.
Jul 14, 2026
- b10002
Adds a set of utility functions to check the contiguity of inner tensor dimensions and releases standard binaries for supported platforms.
Jul 14, 2026
- b10001
Fixes a critical segfault in the `test-export-graph-ops` test when called without arguments and releases standard binaries for supported platforms.
Jul 14, 2026
- b10000
Uniformizes the im2col destination type casting logic across all convolution operations to improve consistency and fixes casting logic to allow any kernel type.
Jul 14, 2026
- b9999
Adds support for SME2 (ARMv9.2) f32 kernels with dynamic scheduling and releases standard binaries for multiple platforms including macOS, Linux, Android, Windows, and openEuler.
Jul 14, 2026
- b9996
A fix to ensure log messages are flushed before the application exits when `usage()` is called, preventing potential message loss due to asynchronous logging from other threads.
Jul 14, 2026
- b9995
Optimized the SYCL backend for Intel's Battlemage architecture by setting the vector thread count to 256, which is preferred by these chips, while keeping the default at 128 for other devices like ARC Alchemist.
Jul 14, 2026
- b9994
Added support for the Q2_0 quantization format on the Metal backend (macOS/iOS). This allows for smaller model sizes and faster inference on Apple Silicon devices.
Jul 14, 2026
- b9993
Added support for the Hy3 (Hunyuan 3) model architecture, a MoE decoder stack with specific routing mechanisms. It includes support for MTP (Multi-Token Prediction) speculative decoding for this specific model.
Jul 13, 2026
- b9992
The Multi-MatMul-Q (MMQ) kernel configuration for CUDA was refactored to improve code organization and maintainability. This includes fixing the Blackwell configuration and removing legacy code to streamline the kernel management.
Jul 13, 2026
- b9990
Added support for the Minimax2 EAGLE3 speculative decoding model, including a fix for a null pointer issue. This allows users to leverage faster inference speeds through speculative decoding with this specific model architecture.
Jul 13, 2026