llama.cpp Releases
609 releases of ggml-org/llama.cpp
- b9661
Implements the col2im_1d operation in the Vulkan backend, following up on the CPU implementation. This adds a new mathematical operation to the Vulkan acceleration layer.
Jun 16, 2026
- b9660
Corrects a double-escaping issue in the parsing of LFM2 tool calls, ensuring that arguments are handled correctly. Standard release binaries are provided.
Jun 15, 2026
- b9659
Fixes a token counting error in multimodal models that could lead to incorrect token management. Standard binaries are available for all supported platforms.
Jun 15, 2026
- b9658
Enhances debugging capabilities by including the full unparsed prompt in error messages. This helps developers troubleshoot parsing issues more effectively.
Jun 15, 2026
- b9656
Hardens the PEG-native parser for tool calls to better handle OpenAI-style formats and provides clearer error reporting. The release maintains standard build configurations.
Jun 15, 2026
- b9655
Addresses a grammar generator bug that surfaced during recent changes in the chat mode. The release ensures more stable interactions and includes standard platform binaries.
Jun 15, 2026
- b9654
Adds a post-decode callback for multimodal models, allowing custom processing steps after the decoding phase. Standard binaries are provided for all supported operating systems.
Jun 15, 2026
- b9653
This release expands Vulkan hardware acceleration by supporting more CONCAT operation types. It includes standard build binaries for macOS, Linux, Windows, Android, and openEuler.
Jun 15, 2026
- b9652
A quick fix for a WebAssembly build issue involving symbol collisions in the fallback mechanism. Standard release binaries are provided for supported platforms.
Jun 15, 2026
- b9651
This update optimizes the SYCL backend for K-quants by utilizing the native subgroup size for the DMMV operation to improve efficiency.
Jun 15, 2026
- b9650
This release corrects a bug in the SYCL soft_max_f32 function where the max reduction logic was incorrect.
Jun 15, 2026
- b9649
This release fixes a reorder function bug in the SYCL backend and updates the build script to ensure FP32 and FP16 support is included.
Jun 15, 2026
- b9647
This release introduces 1D pooling support in the SYCL backend and refactors existing 1D and 2D pooling code into dedicated pool.cpp/hpp files.
Jun 15, 2026
- b9646
This update refactors the SYCL backend by removing per-allocation Level Zero runtime checks and centralizing detection logic during initialization.
Jun 15, 2026
- b9645
This update adds BFloat16 (BF16) support for the Metal backend's repeat operation, enhancing performance and compatibility on Apple Silicon devices.
Jun 15, 2026
- b9644
This maintenance release resolves persistent whitespace issues in the chat functionality and grammar generation to ensure cleaner output processing.
Jun 15, 2026
- b9642
This release restricts CUDA support for the GGML_OP_REPEAT operation to F32 and F16 formats, while providing comprehensive pre-built binaries for macOS, Linux, Windows, and Android across various hardware backends.
Jun 15, 2026
- b9637
Added a dedicated chat parser for the Cohere2-MoE (North Code) model format to improve chat interaction handling.
Jun 14, 2026
- b9632
Enhanced the Jinja template engine by adding aliases for the count, d, and e filters.
Jun 14, 2026
- b9631
Fixed a bug in the CLI where preserved tokens were not being copied correctly during processing.
Jun 14, 2026