llama.cpp Releases
684 releases of ggml-org/llama.cpp
- b10246
Optimized OpenCL kernel routing for large Q6_K language model heads to use flat GEMV operations, fixing performance issues with large models like Gemma-4 E2B.
Aug 3, 2026
- b10245
Fixed a bug in the Minimax M3 graph where certain input tensors were not being utilized.
Aug 3, 2026
- b10244
Refactored the MSA (Multi-Head Attention) logic for the M3 model into a dedicated memory implementation module.
Aug 3, 2026
- b10243
Optimized memory allocation by restricting indexer cache creation to specific 'full' indexer layers to reduce memory usage.
Aug 3, 2026
- b10242
Implemented a backend sampler for handling penalty sampling logic (frequency, presence, and top-k) within the CUDA backend, accompanied by extensive test coverage.
Aug 3, 2026
- b10241
Fixed critical data races in CUDA block reduction operations by adding memory barriers and implemented double-buffering for specific operations to improve performance.
Aug 3, 2026
- b10240
Added a notice regarding an upcoming change to the server's default port from 8080 to 9931.
Aug 3, 2026
- b10238
Implemented MTP support for the Qwen3-Next model, including fixes to Python type-checking and logic for computing the number of MTP layers.
Aug 3, 2026
- b10237
This release adds MTP (Mixture of Token Prediction) support for the DeepSeek V3.2 model, optimizing how MTP layers are handled during model discovery.
Aug 3, 2026
- b10236
Implemented a Lightning Indexer for DeepSeek V4 in Metal, optimized with tiled kernels and staged memory access, resulting in significant performance improvements for large context lengths.
Aug 3, 2026
- b10235
Implemented the SILU_BACK (Sigmoid Linear Unit Backward) operation for the Metal backend to support F32 precision.
Aug 2, 2026
- b10234
Added F16 (half-precision floating point) support for binary operations in the Metal backend.
Aug 2, 2026
- b10233
Fixed a potential issue in the OpenCL backend by limiting the local workgroup size for the GLU operation to prevent instability.
Aug 2, 2026
- b10232
Implemented optimized Metal kernels for DeepSeek V4 hyper-connections using SIMDgroup registers and shuffle operations.
Aug 2, 2026
- b10231
Updated the common library to support DSpark sidecar resolution, allowing it to be auto-discovered and selected similar to other speculative decoding sidecars.
Aug 2, 2026
- b10229
Fixed a critical bug in the OpenCL backend where reference counting was not incremented correctly, which could prevent profiling data from being flushed and written during cleanup.
Aug 2, 2026
- b10228
This release introduces support for DeepseekV4 MTP (Multi-Token Prediction) models and DSpark speculative decoding sidecars.
Aug 2, 2026
- b10227
A new release adding specialized parsing support for Qwen3 models, including a tagged thinking tool parser and refactoring of chat handling utilities.
Aug 2, 2026
- b10225
This release optimizes model loading by conditionally loading MiMo V2 MTP tensors only when they are actually required, improving memory efficiency for users not utilizing these specific tensors.
Aug 2, 2026
- b10226
Fixes the classification of integrated GPUs (iGPUs) in the SYCL backend to ensure correct hardware detection.
Aug 2, 2026