llama.cpp Releases
609 releases of ggml-org/llama.cpp
- b10107
Fixed a crash on Windows with the Hexagon backend when the op_poll option is enabled.
Jul 24, 2026
- b10106
Fixes an issue preventing external compilation of the q1_0 Multi-Quantization Matrix (MMQ) kernels.
Jul 24, 2026
- b10105
Refactors mmap, mlock, and directio arguments into a single unified `load-mode` argument, requiring a change in command-line usage.
Jul 24, 2026
- b10103
Adds f16 type support to the leaky relu activation function in the Metal backend.
Jul 24, 2026
- b10099
Improves NVFP4 W4A4 activation quantization on CUDA with fused kernels and optimized memory access patterns.
Jul 23, 2026
- b10098
Updates Hexagon activation operations with optimized microkernels and improved memory handling for non-contiguous data.
Jul 23, 2026
- b10094
Automatically infers the speculative type from draft repo sidecars (mtp/dflash/eagle3) when no explicit `--spec-type` is provided.
Jul 23, 2026
- b10093
Fixes the DeepSeek4 chat template to explicitly follow reference behavior and adds support for the DeepSeekv4 `drop_reasoning` flag.
Jul 23, 2026
- b10092
Enables PowerPC backend variants on AIX and fixes a missing semicolon in the PowerPC Q0 matmul implementation.
Jul 23, 2026
- b10091
This release fixes a bug related to SYCL package shared library lookup. It includes standard binaries for macOS, Linux, Windows, Android, and openEuler.
Jul 22, 2026
- b10090
Adds a depthwise convolution 2D (CONV_2D_DW) kernel to the WebGPU backend, ported from the Vulkan backend implementation.
Jul 22, 2026
- b10089
Significantly optimizes the CUDA backend by implementing support for k-quants, i-quants, and mxfp4 in the GET_ROWS operation, enabling efficient device-side embedding lookups.
Jul 22, 2026
- b10088
A maintenance release focused on fixing a tensor operation issue in the llama-arch module for DeepSeek4 APE, along with disabling specific backend builds.
Jul 22, 2026
- b10087
Adds support for new hardware devices, specifically the Laguna XS.2 and Laguna M.1.
Jul 22, 2026
- b10085
Fixes Qwen3-VL vision position embedding interpolation by switching from align_corners=False to align_corners=True, aligning with the reference implementation and fixing grounding coordinate scaling issues.
Jul 22, 2026
- b10084
Adds a safety check for tensor types when reusing Hexagon (DSP) descriptors to prevent errors or crashes associated with mismatched tensor configurations.
Jul 22, 2026
- b10083
Adds a sqrt_softplus function to the top-k Mixture of Experts (MoE) implementation for dsv4, enhancing activation function capabilities.
Jul 22, 2026
- b10082
Adds a warning that triggers once when a specific weight type has no available KleidiAI kernel, improving debuggability and user awareness of hardware acceleration limits.
Jul 22, 2026
- b10081
Fixes a bug where draft repositories were not resolving to their requested sidecar models when using the --spec-type draft-dflash flag, ensuring speculative decoding works as expected.
Jul 22, 2026
- b10080
Updates server error handling to return HTTP 400 instead of 500 for validation errors involving the X-Conversation-Id header, and fixes the UI to properly filter the backend_sampling placeholder.
Jul 22, 2026