ggmlR

'GGML' Tensor Operations for Machine Learning

Provides 'R' bindings to the 'GGML' tensor library for machine learning, designed primarily for 'Vulkan' GPU acceleration with full CPU fallback. Requires 'Vulkan' 1.2+ with legacy pipeline barriers (avoids 'Synchronization2' due to 'RADV' performance issues); supports 'Push Descriptors' ('VK_KHR_push_descriptor') to eliminate descriptor pool overhead when available. 'Vulkan' support is auto-detected at build time on Linux (when 'libvulkan-dev' and 'glslc' are installed) and on Windows (when 'Vulkan' 'SDK' is installed and 'VULKAN_SDK' environment variable is set); all operations fall back to CPU transparently when no GPU is available. Supports tensors up to 5D natively (GGML_MAX_DIMS=5). Implements tensor operations, neural network layers, 'quantization', and a 'Keras'-like sequential model API for building and training networks. Includes 'AdamW' (Adam with Weight decay) and 'SGD' (Stochastic Gradient Descent) optimizers with 'MSE' (Mean Squared Error) and cross-entropy losses. Also provides a dynamic 'autograd' engine ('PyTorch'-style) with data-parallel training via 'dp_train()', broadcast arithmetic, 'f16' (half-precision) support on 'Vulkan' GPU, and a multi-head attention layer for building Transformer architectures. Supports 'ONNX' model import via built-in zero-dependency 'protobuf' parser: load 'pretrained' 'ONNX' models from 'PyTorch', 'TensorFlow', or other frameworks and run inference on 'Vulkan' GPU or CPU. Covers 50+ 'ONNX' ops including convolutions, attention primitives, normalization, quantized ops, shape operations, 'ScatterElements' (with 'Vulkan' 'atomicAdd' for GNN scatter-add), and fused custom ops (RelPosBias2D for 'BoTNet') — sufficient to run real-world models such as 'RoBERTa', 'BERT', 'GPT-NeoX', 'SqueezeNet', 'Inception v3', 'BAT-ResNeXt', 'BoTNet', and 'MNIST' out of the box. Reads 'GGUF' files natively: load 'pretrained' weights from any 'gguf'-compatible source ('llama.cpp', 'Hugging Face') with automatic weight conversion and metadata access. Uses a dedicated weight buffer architecture for zero-overhead repeated inference — weights are loaded to GPU once and never re-transferred. Serves as backend for 'LLM' (Large Language Model) inference via 'llamaR' and Stable Diffusion image generation via 'sd2R'. See <https://github.com/ggml-org/ggml> for more information about the underlying library.

Versions across snapshots

Dependencies (latest)