ggmlR 0.5.1

New Features

• Export static library libggml.a for linking by dependent packages (llamaR)
• Added gguf.cpp for GGUF file format support
• Headers exported via inst/include/ for LinkingTo

CRAN Submission Fixes

• Expanded acronyms in DESCRIPTION: ‘AdamW’ (Adam with Weight decay), ‘SGD’ (Stochastic Gradient Descent), ‘MSE’ (Mean Squared Error), GPU (Graphics Processing Unit)
• Added all contributors and copyright holders to Authors@R:
  • Georgi Gerganov (GGML library author)
  • Jeffrey Quesnelle and Bowen Peng (ops.cpp contributors)
  • Mozilla Foundation (llamafile/sgemm.cpp)
• Replaced \dontrun{} with \donttest{} in all examples
• Added \value documentation to all exported functions and constants

Documentation

• Added vignette: Vulkan GPU Backend (vignette("vulkan-backend"))
• Added vignette: Multi-GPU Inference (vignette("multi-gpu"))
• Added vignette: Working with Quantized Models (vignette("quantization"))

Internal

• r_ggml_io.o moved to GGML_OBJECTS for proper symbol export
• Static library excluded from source tarball via .Rbuildignore

ggmlR 0.5.0

Major Features

• Added full optimization/training API (39 new functions)
  • ggml_opt_init(), ggml_opt_free(), ggml_opt_reset() — optimizer lifecycle
  • ggml_opt_fit() — high-level training loop
  • ggml_opt_epoch() — single epoch with R callback support
  • ggml_opt_eval(), ggml_opt_alloc() — model evaluation
  • ggml_opt_prepare_alloc() — non-static graph support
  • ggml_opt_grad_acc() — gradient accumulator access
• Dataset management
  • ggml_opt_dataset_init(), ggml_opt_dataset_free()
  • ggml_opt_dataset_data(), ggml_opt_dataset_labels()
  • ggml_opt_dataset_shuffle(), ggml_opt_dataset_get_batch()
• Training results
  • ggml_opt_result_init(), ggml_opt_result_free(), ggml_opt_result_reset()
  • ggml_opt_result_ndata(), ggml_opt_result_loss(), ggml_opt_result_accuracy()
  • ggml_opt_result_pred() — get predictions as integer vector
• Loss functions: MSE, cross-entropy, mean, sum
• Optimizers: AdamW, SGD

R Callback Support

• ggml_opt_epoch() now supports custom R callback functions
• Callbacks receive: train flag, batch index, max batches, start time, result pointer
• Built-in progress bar callback available via callback_train = TRUE

Extended Backend API (~50 new functions)

• Device management

  • ggml_backend_dev_count(), ggml_backend_dev_get(), ggml_backend_dev_by_name()
  • ggml_backend_dev_by_type() — find devices by type (CPU, GPU, etc.)
  • ggml_backend_dev_supports_op(), ggml_backend_dev_supports_buft()
  • ggml_backend_dev_memory(), ggml_backend_dev_description(), ggml_backend_dev_get_props()
  • ggml_backend_dev_init() — initialize backend from device

• Device type constants: ggml_backend_device_type_cpu/gpu/igpu/accel()

• Buffer usage constants: ggml_backend_buffer_usage_any/weights/compute()

• Backend registry

  • ggml_backend_reg_count(), ggml_backend_reg_get(), ggml_backend_reg_by_name()
  • ggml_backend_load(), ggml_backend_unload(), ggml_backend_load_all()

• Events for synchronization

  • ggml_backend_event_new(), ggml_backend_event_free()
  • ggml_backend_event_record(), ggml_backend_event_synchronize(), ggml_backend_event_wait()

• Async operations

  • ggml_backend_tensor_set_async(), ggml_backend_tensor_get_async()
  • ggml_backend_tensor_copy_async()

• Graph planning

  • ggml_backend_graph_plan_create(), ggml_backend_graph_plan_free()
  • ggml_backend_graph_plan_compute()

• Buffer management

  • ggml_backend_buffer_clear(), ggml_backend_buffer_set_usage()
  • ggml_backend_buffer_get_usage(), ggml_backend_buffer_reset(), ggml_backend_buffer_is_host()

• Direct backend initialization

  • ggml_backend_init_by_name(), ggml_backend_init_by_type(), ggml_backend_init_best()
  • ggml_backend_synchronize(), ggml_backend_get_device()

Testing

• Added 67 tests for optimization functions
• Added 60 tests for extended backend functions
• R CMD check: 0 errors, 0 warnings

ggmlR 0.4.1

• Fixed spelling notes for CRAN submission
• Updated documentation

ggmlR 0.4.0

• Added multi-GPU backend scheduler API (14 new functions)
• Added Vulkan GPU backend support (10 new functions)
• Fixed integer overflow for large tensors (>2 GB)
• Improved OpenMP handling for mixed C/C++ code

ggmlR 0.2.0

• Initial CRAN submission
• R bindings for ‘GGML’ tensor library
• Core tensor operations: creation, arithmetic, reshaping
• Neural network operations: attention, convolutions, normalization
• Activation functions: GELU, SiLU, ReLU, and variants
• Quantization support (Q4_0, Q4_1, Q8_0)
• OpenMP parallelization for CPU backend
• Computation graph API for building and executing models