Using ggmlR as a Backend in Your Package

In ggmlR: 'GGML' Tensor Operations for Machine Learning

# Vignette code is executed locally (NOT_CRAN=true) but not on CRAN, where
# the CPU fallback would multi-thread and trip the "CPU time > elapsed" NOTE.
knitr::opts_chunk$set(eval = identical(Sys.getenv("NOT_CRAN"), "true"))

There are two ways another package can build on ggmlR:

1. R level — depend on the ggmlR R API (Imports: ggmlR) and call the exported ggml_* / ag_* functions. This is the simplest path and is right for packages that build, train or run models from R. See Part A below.
2. C level — link against the exported static library (libggml.a) and C headers (LinkingTo: ggmlR) to call ggml directly from your own C/C++ code. This is the path used by llamaR (LLM inference) and sd2R (Stable Diffusion). See Part B below.

The two are independent — most packages need only one. You can also combine them (use the R API for orchestration and drop to C for a custom kernel).

---

Part A — Using the ggmlR R API from your package

A1. Depend on ggmlR

In your DESCRIPTION:

Package: myPackage
...
Imports:
    ggmlR

Then import the functions you use in NAMESPACE (via roxygen @importFrom ggmlR ...), or call them fully qualified as ggmlR::ggml_fit().

#' @importFrom ggmlR ggml_model_sequential ggml_layer_dense ggml_compile
#' @importFrom ggmlR ggml_fit ggml_predict
NULL

A2. Everything the examples use is exported

The whole high-level surface is exported and therefore usable from a downstream package — the sequential, functional and autograd APIs, optimizers, schedulers, the dataloader, data-parallel training, plus the diagnostics/reproducibility helpers. For instance, the entire titanic_classification.R example (Keras-like + functional + autograd, ~30 ggmlR functions) runs unchanged from another package once ggmlR is in Imports.

my_train <- function(x, y, epochs = 50L) {
  model <- ggmlR::ggml_model_sequential() |>
    ggmlR::ggml_layer_dense(64L, activation = "relu", input_shape = ncol(x)) |>
    ggmlR::ggml_layer_dense(ncol(y), activation = "softmax") |>
    ggmlR::ggml_compile(optimizer = "adam", loss = "categorical_crossentropy")
  ggmlR::ggml_fit(model, x, y, epochs = epochs, verbose = 0L)
}

A few things worth knowing:

• S3 dispatch works across packages. ggml_compile(), ggml_fit() and ggml_predict() are S3 generics with methods registered for the sequential and functional model classes; the dispatch resolves correctly when called from your package.
• Object methods need no extra imports. Autograd objects expose methods through their environment (model$forward(), opt$step(), dl$epoch(), …). These are fields of objects returned by exported functions, so they just work — there is nothing extra to import.
• The pipe |> is base R (≥ 4.1); it is not a ggmlR dependency.
• Reproducibility & diagnostics are exported too: ggml_set_seed(), ggml_training_history() and ggml_model_backend() let your package fix seeds and inspect the loss curve / actual backend of a fitted model.
• Local helpers stay local. Example scripts define their own convenience helpers (e.g. ag_predict_colmajor(), custom autograd layers). Those are not part of the ggmlR API — copy them into your own package if you want them.

A3. tidymodels / mlr3 engines

ggmlR registers a parsnip "ggml" engine and mlr3 learners (classif.ggml, regr.ggml) in its .onLoad: if mlr3 / parsnip are already loaded they are registered immediately, otherwise a load hook registers them as soon as those packages load. A downstream package that Imports: ggmlR therefore gets these registrations with no manual call — just make sure ggmlR's namespace is loaded (any ggmlR:: call or importFrom does that). See the tidymodels and mlr3 vignettes for usage.

---

Part B — Linking ggml from your package's C/C++ code

1. What ggmlR exports

After installing ggmlR you will find:

$(R_HOME_DIR)/library/ggmlR/lib/libggml.a
$(R_HOME_DIR)/library/ggmlR/include/ggml.h
$(R_HOME_DIR)/library/ggmlR/include/ggml-backend.h
$(R_HOME_DIR)/library/ggmlR/include/ggml-alloc.h
$(R_HOME_DIR)/library/ggmlR/include/ggml-opt.h
$(R_HOME_DIR)/library/ggmlR/include/ggml-quants.h
$(R_HOME_DIR)/library/ggmlR/include/ggml-vulkan.h
$(R_HOME_DIR)/library/ggmlR/include/r_ggml_compat.h
... (full list in inst/include/)

r_ggml_compat.h redirects printf/fprintf/abort to R-safe equivalents (Rprintf, Rf_error). Always include it (or use -include r_ggml_compat.h) in your C/C++ sources.

---

2. DESCRIPTION

Add ggmlR to LinkingTo and Imports:

Package: myPackage
...
Imports:   ggmlR
LinkingTo: ggmlR

LinkingTo makes R add ggmlR/include to the compiler include path automatically.

---

3. src/Makevars

Link against the static library:

GGMLR_LIB = $(shell Rscript -e "cat(system.file('lib', package='ggmlR'))")
GGMLR_INC = $(shell Rscript -e "cat(system.file('include', package='ggmlR'))")

PKG_CPPFLAGS = -I$(GGMLR_INC) -include r_ggml_compat.h
PKG_LIBS     = $(GGMLR_LIB)/libggml.a

If ggmlR was built with Vulkan you also need to link Vulkan:

# detect Vulkan (same logic as ggmlR's own configure)
VULKAN_LIBS = $(shell pkg-config --libs vulkan 2>/dev/null)
PKG_LIBS    = $(GGMLR_LIB)/libggml.a $(VULKAN_LIBS)

---

4. configure — detect ggmlR at build time

#!/bin/sh
# configure

GGMLR_INC=$(Rscript -e "cat(system.file('include', package='ggmlR'))" 2>/dev/null)
GGMLR_LIB=$(Rscript -e "cat(system.file('lib',     package='ggmlR'))" 2>/dev/null)

if [ -z "$GGMLR_INC" ] || [ ! -f "$GGMLR_LIB/libggml.a" ]; then
  echo "ERROR: ggmlR not found. Install it first: install.packages('ggmlR')" >&2
  exit 1
fi

sed -e "s|@GGMLR_INC@|$GGMLR_INC|g" \
    -e "s|@GGMLR_LIB@|$GGMLR_LIB|g" \
    src/Makevars.in > src/Makevars

src/Makevars.in:

PKG_CPPFLAGS = -I@GGMLR_INC@ -include r_ggml_compat.h
PKG_LIBS     = @GGMLR_LIB@/libggml.a

---

5. C code — minimal example

/* src/my_model.c */
#include "ggml.h"
#include "ggml-backend.h"
#include <R.h>
#include <Rinternals.h>

SEXP R_my_inference(SEXP r_input) {
    struct ggml_init_params params = {
        .mem_size   = 256 * 1024 * 1024,  /* 256 MB */
        .mem_buffer = NULL,
        .no_alloc   = false,
    };
    struct ggml_context *ctx = ggml_init(params);

    int n = length(r_input);
    struct ggml_tensor *x = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n);
    memcpy(x->data, REAL(r_input), n * sizeof(float));

    /* ... build graph, compute ... */

    ggml_free(ctx);
    return R_NilValue;
}

Register in the .Call table (R's standard routine registration):

/* src/init.c */
#include <R.h>
#include <Rinternals.h>
#include <R_ext/Rdynload.h>

extern SEXP R_my_inference(SEXP);

static const R_CallMethodDef CallEntries[] = {
    {"R_my_inference", (DL_FUNC) &R_my_inference, 1},
    {NULL, NULL, 0}
};

void R_init_myPackage(DllInfo *dll) {
    R_registerRoutines(dll, NULL, CallEntries, NULL, NULL);
    R_useDynamicSymbols(dll, FALSE);
}

---

6. R wrapper

my_inference <- function(input) {
  .Call("R_my_inference", as.numeric(input))
}

---

7. Thread count

ggmlR manages the CPU thread count via ggmlR_get_n_threads() (exported from r_interface.c). If your package calls ggml_backend_cpu_init() directly, set the thread count to match:

#include "ggml-backend.h"
/* ggmlR_get_n_threads() is exported by ggmlR — link against libggml.a */
extern int ggmlR_get_n_threads(void);

ggml_backend_t cpu = ggml_backend_cpu_init();
ggml_backend_cpu_set_n_threads(cpu, ggmlR_get_n_threads());

This ensures your package respects the same thread limit as ggmlR (important for CRAN compliance — tests must not exceed 2 threads).

---

8. Vulkan in downstream packages

If your package needs Vulkan, the Vulkan backend is already compiled into libggml.a when ggmlR was built with Vulkan support. You only need to link -lvulkan:

VULKAN_LIBS = $(shell pkg-config --libs vulkan 2>/dev/null)
PKG_LIBS    = @GGMLR_LIB@/libggml.a $(VULKAN_LIBS)

Do not vendor ggml source again — link the pre-built libggml.a from ggmlR to avoid symbol collisions.

---

9. Real-world references

• llamaR — LLM inference (LLaMA, Mistral, …) using ggmlR as backend
• sd2R — Stable Diffusion image generation using ggmlR as backend

Both follow the pattern above: LinkingTo: ggmlR, configure script to locate headers and libggml.a, thin C wrappers, R .Call interface.

ggmlR documentation built on July 14, 2026, 1:08 a.m.