llama.cpp Releases
609 releases of ggml-org/llama.cpp
- b10048
This release fixes Transformer Parallelism (TP) support for specific model architectures including Phi3, Bert, Plamo2/3, and ChatGLM.
Jul 16, 2026
- b10047
Updates the BoringSSL vendor dependency to version 0.20260713.0 and provides pre-built binaries for macOS, Linux, Windows, and Android.
Jul 16, 2026
- b10046
Improves test coverage by implementing the `test-recurrent-state-rollback` test case and includes pre-compiled binaries for all supported platforms.
Jul 16, 2026
- b10045
Adds a server feature allowing text-only slot save and restore when using Multimodal (mtmd) models, accompanied by standard platform binaries.
Jul 16, 2026
- b10043
Adds support for CUDA virtual devices, allowing the simulation of multiple GPUs, and disables NCCL in this scenario.
Jul 16, 2026
- b10042
Activates CUDA graph support for Volta and Turing architecture GPUs to improve performance.
Jul 16, 2026
- b10041
Reduces log noise in the server by ignoring empty or missing Origin headers in the CORS handling.
Jul 16, 2026
- b10040
Fixes a bug in CUDA/HIP integration detection where the integrated flag was incorrectly set to false, preventing host buffer usage on AMD APU parts.
Jul 16, 2026
- b10038
Updates release notes to include a link to the official website.
Jul 16, 2026
- b10037
Enhances the quantization process to support manual tensor types when using the `--pure` flag.
Jul 16, 2026
- b10036
Disables specific OpenCL kernels (FA and MoE) on Adreno 850 to prevent compiler-related errors, and fixes a parsing bug for DirectX compiler versions.
Jul 16, 2026
- b10035
Optimizes the CUDA kernel for Q1_0 quantization type using the `__byte_perm` intrinsic instruction.
Jul 16, 2026
- b10034
Fixes a bug in OpenCL MoE (Mixture of Experts) kernels for specific Adreno GPU architectures that were causing corrupted output.
Jul 15, 2026
- b10032
Adds a new CUDA implementation for the lightning indexer operation using generic vector and WMMA kernels, with architectural relaxations and refactoring.
Jul 15, 2026
- b10031
Removes a mutual-exclusion check for the `--stdin` tokenization flag to align CLI behavior with completion scripts.
Jul 15, 2026
- b10025
Relaxes tensor contiguity requirements for CUDA quantized concatenation operations, allowing for more flexible memory layouts.
Jul 15, 2026
- b10021
Optimizes DeepseekV4 graph splitting logic to reduce overhead, improving performance for inference tasks on this specific model architecture.
Jul 15, 2026
- b10020
Fixes data retrieval issues in SYCL for specific quantization types, ensuring correct tensor operations.
Jul 15, 2026
- b10019
Adds support for FP16 kernel types in SYCL for depthwise convolution operations, enhancing precision handling.
Jul 15, 2026
- b10018
Implements a new `xielu` operation for the SYCL backend, expanding the library's capability for specific computational kernels.
Jul 15, 2026