llama.cpp b9902 adds cross-entropy loss ops, expands platform support
New release enables training loss functions directly on-device for local LLMs.
Deep Dive
The llama.cpp project released version b9902, adding SYCL support for cross_entropy_loss and cross_entropy_loss_back operations. This release includes builds for macOS (Apple Silicon, Intel), Linux (x64, ARM, s390x), Windows (x64, ARM, with CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), Android, and iOS.
Key Points
- Adds SYCL support for cross_entropy_loss and cross_entropy_loss_back operations, enabling loss computation on Intel GPUs and other SYCL hardware.
- Supports 20+ platform builds including macOS, Linux, Windows (with CUDA 12/13, Vulkan, HIP), Android, and iOS.
- Release b9902 is signed with a verified GPG key (B5690EEEBB952194) for secure distribution.
Why It Matters
Enables local fine-tuning of LLMs on consumer hardware, democratizing AI training without cloud costs.