llama.cpp Releases

Addresses a bug related to the quantized key-value cache for dsv4 (DeepSeek V4) to ensure correct state management and consistency.

Jul 8, 2026

This release focuses on fixing unit test cases for specific operations (CONT and CPY) within the SYCL backend. It includes build artifacts for all supported platforms and architectures.

Jul 7, 2026

Adds support for SYCL operations cross_entropy_loss and cross_entropy_loss_back, expanding SYCL functionality for machine learning tasks.

Jul 7, 2026

Optimizes SYCL DMMV (Dequantize-Mul-Mat-Vec) path by setting K_QUANTS_PER_ITERATION to 1, significantly improving GPU utilization and performance on B70 hardware.

Jul 7, 2026

Enhances SYCL argsort implementation to support all unit test cases, improving robustness and test coverage of the sorting functionality.

Jul 7, 2026

Uses SYCL function calls to fix Abstract-Optimized-Toolchain (AOT) double type issues, improving type handling in SYCL implementations.

Jul 7, 2026

Improves SYCL environment variable naming by changing from 'disable' to 'enable' flags for better user experience and more intuitive configuration.

Jul 7, 2026

Note: This release appears to be from mistral.rs repository, not ggml-org/llama.cpp. It includes multiple features including Metal runtime improvements, CUDA worker splitting, distributed training enhancements, LFM 2.5 model support, and extensive performance optimizations.

Jul 7, 2026

Fixes out-of-bounds read in ngram-map during prompt shrinking operations, along with cleanup cutoff improvements for speculative decoding.

Jul 7, 2026

Vulkan fix to check src0 type in GGML_OP_SET_ROWS operation to prevent failures due to unimplemented f16 support.

Jul 7, 2026

Focuses on OpenCL flash attention decode performance optimizations with multiple kernel improvements and bug fixes for various quantization formats and model configurations.

Jul 7, 2026

Binary release with extensive platform support including macOS, Linux, Android, and Windows. The primary technical change is setting optimal default thread count for PowerPC (PPC) on Linux and AIX systems.

Jul 6, 2026

A new operation `col2im_1d` has been added to the Metal backend, supporting floating-point formats (f32, f16, bf16), mirroring functionality available in CPU and CUDA backends.

Jul 6, 2026

A significant refactoring of the CUDA backend removes the `-sm row` option and reworks the cuBLAS implementation. This includes fixes for AMD CDNA compatibility with BF16 and corrections to memory stride handling.

Jul 6, 2026

The CUDA backend for Flash Attention has been updated to validate the types of both K (Key) and V (Value) matrices, ensuring stricter type compatibility checks.

Jul 6, 2026

The CPU backend has been optimized for ARM architecture by utilizing the UE4M3 Look-Up Table (LUT) for ARM NVFP4 dot product operations, aiming to improve computational efficiency.

Jul 6, 2026

Tiled matrix multiplication has been enabled for AIX systems by reducing local stack buffer footprints to avoid segmentation faults. This change delivers approximately 2x performance improvements for specific model types on the platform.

Jul 6, 2026

A critical bug fix for the Vulkan backend addresses a 32-bit integer overflow occurring in the CEIL_DIV macro, preventing potential crashes or incorrect calculations on 32-bit platforms.

Jul 6, 2026

The build scripts have been updated to utilize the Hugging Face (HF) token when downloading UI assets, ensuring authenticated access to private or restricted resources.

Jul 6, 2026

This release enables the `-ffast-math` compiler flag for HIP (AMD GPU) builds, which optimizes floating-point arithmetic for improved performance on AMD hardware.

Jul 6, 2026