Simple example

Minimal working examples of ggml computation

The examples/simple directory contains two minimal programs that each multiply two matrices using ggml. They demonstrate the two main approaches to memory and compute management:

simple-ctx

Context-based allocation. All tensors and the compute graph live in a single ggml_context. Simple to use; CPU-only.

simple-backend

Backend-based allocation. Separates graph definition from execution. Supports CPU, CUDA, Metal, and other backends.

Both programs compute A × Bᵀ for two matrices and print the result.

Context-based approach (simple-ctx.cpp)

This is the legacy API. Memory for tensors and the compute graph is allocated inside a single ggml_context using a fixed-size memory pool.

1. Calculate and allocate the memory pool

   Before creating any tensors you must calculate the total memory needed and pass it to ggml_init:

   size_t ctx_size = 0;
   ctx_size += rows_A * cols_A * ggml_type_size(GGML_TYPE_F32); // tensor a
   ctx_size += rows_B * cols_B * ggml_type_size(GGML_TYPE_F32); // tensor b
   ctx_size += 2 * ggml_tensor_overhead();  // tensor metadata
   ctx_size += ggml_graph_overhead();       // compute graph
   ctx_size += 1024;                        // general overhead
   
   struct ggml_init_params params {
       /*.mem_size   =*/ ctx_size,
       /*.mem_buffer =*/ NULL,
       /*.no_alloc   =*/ false, // allocate tensor data inside the context
   };
   
   model.ctx = ggml_init(params);
   

   Setting no_alloc = false means tensor data buffers are allocated immediately inside the memory pool.

2. Create tensors and copy data

   Allocate 2D tensors and copy the input matrices into them:

   model.a = ggml_new_tensor_2d(model.ctx, GGML_TYPE_F32, cols_A, rows_A);
   model.b = ggml_new_tensor_2d(model.ctx, GGML_TYPE_F32, cols_B, rows_B);
   
   memcpy(model.a->data, a, ggml_nbytes(model.a));
   memcpy(model.b->data, b, ggml_nbytes(model.b));
   

   Note the argument order for ggml_new_tensor_2d: dimensions are (cols, rows) because ggml uses column-major storage.

3. Build the compute graph

   Describe the computation by connecting tensors with operations. ggml_mul_mat(a, b) computes A × Bᵀ:

   struct ggml_cgraph * gf = ggml_new_graph(model.ctx);
   
   // result = a * b^T
   struct ggml_tensor * result = ggml_mul_mat(model.ctx, model.a, model.b);
   
   ggml_build_forward_expand(gf, result);
   

   ggml_build_forward_expand walks the tensor dependency tree and records all nodes needed to produce result.

4. Run the computation

   Execute the graph on the CPU:

   int n_threads = 1;
   ggml_graph_compute_with_ctx(model.ctx, gf, n_threads);
   
   // the output tensor is the last node in the graph
   struct ggml_tensor * result = ggml_graph_node(gf, -1);
   
5. Read the result and free memory

   Copy output data out of the tensor buffer, then free the context:

   std::vector<float> out_data(ggml_nelements(result));
   memcpy(out_data.data(), result->data, ggml_nbytes(result));
   
   // expected output:
   // [ 60.00 55.00 50.00 110.00
   //   90.00 54.00 54.00 126.00
   //   42.00 29.00 28.00  64.00 ]
   
   ggml_free(model.ctx);
   

Full source

#include "ggml.h"
#include "ggml-cpu.h"

#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <vector>

struct simple_model {
    struct ggml_tensor * a;
    struct ggml_tensor * b;
    struct ggml_context * ctx;
};

void load_model(simple_model & model, float * a, float * b,
                int rows_A, int cols_A, int rows_B, int cols_B) {
    size_t ctx_size = 0;
    ctx_size += rows_A * cols_A * ggml_type_size(GGML_TYPE_F32);
    ctx_size += rows_B * cols_B * ggml_type_size(GGML_TYPE_F32);
    ctx_size += 2 * ggml_tensor_overhead();
    ctx_size += ggml_graph_overhead();
    ctx_size += 1024;

    struct ggml_init_params params {
        /*.mem_size   =*/ ctx_size,
        /*.mem_buffer =*/ NULL,
        /*.no_alloc   =*/ false,
    };

    model.ctx = ggml_init(params);
    model.a = ggml_new_tensor_2d(model.ctx, GGML_TYPE_F32, cols_A, rows_A);
    model.b = ggml_new_tensor_2d(model.ctx, GGML_TYPE_F32, cols_B, rows_B);

    memcpy(model.a->data, a, ggml_nbytes(model.a));
    memcpy(model.b->data, b, ggml_nbytes(model.b));
}

struct ggml_cgraph * build_graph(const simple_model & model) {
    struct ggml_cgraph * gf = ggml_new_graph(model.ctx);
    struct ggml_tensor * result = ggml_mul_mat(model.ctx, model.a, model.b);
    ggml_build_forward_expand(gf, result);
    return gf;
}

struct ggml_tensor * compute(const simple_model & model) {
    struct ggml_cgraph * gf = build_graph(model);
    ggml_graph_compute_with_ctx(model.ctx, gf, /*n_threads=*/1);
    return ggml_graph_node(gf, -1);
}

int main(void) {
    ggml_time_init();

    const int rows_A = 4, cols_A = 2;
    float matrix_A[rows_A * cols_A] = { 2, 8, 5, 1, 4, 2, 8, 6 };

    const int rows_B = 3, cols_B = 2;
    float matrix_B[rows_B * cols_B] = { 10, 5, 9, 9, 5, 4 };

    simple_model model;
    load_model(model, matrix_A, matrix_B, rows_A, cols_A, rows_B, cols_B);

    struct ggml_tensor * result = compute(model);

    std::vector<float> out_data(ggml_nelements(result));
    memcpy(out_data.data(), result->data, ggml_nbytes(result));

    printf("mul mat (%d x %d):\n[", (int) result->ne[0], (int) result->ne[1]);
    for (int j = 0; j < result->ne[1]; j++) {
        if (j > 0) printf("\n");
        for (int i = 0; i < result->ne[0]; i++) {
            printf(" %.2f", out_data[j * result->ne[0] + i]);
        }
    }
    printf(" ]\n");

    ggml_free(model.ctx);
    return 0;
}

Backend-based approach (simple-backend.cpp)

The backend API separates graph definition from execution and works with any ggml backend — CPU, CUDA, Metal, and others. The key difference is that tensor data is allocated by the backend scheduler after the graph is built, not inside a context.

1. Initialize backends

   Load all available backends and create a scheduler that picks the best device:

   ggml_backend_load_all();
   
   model.backend     = ggml_backend_init_best();       // GPU if available, else CPU
   model.cpu_backend = ggml_backend_init_by_type(
       GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
   
   ggml_backend_t backends[2] = { model.backend, model.cpu_backend };
   model.sched = ggml_backend_sched_new(
       backends, nullptr, 2, GGML_DEFAULT_GRAPH_SIZE, false, true);
   

   The scheduler runs each graph node on the highest-priority backend that supports the operation, falling back to CPU for unsupported ops.

2. Build the compute graph with no_alloc = true

   Create a temporary context only to define the graph structure. Set no_alloc = true because the scheduler will allocate tensor data later:

   size_t buf_size = ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE
                   + ggml_graph_overhead();
   model.buf.resize(buf_size);
   
   struct ggml_init_params params0 = {
       /*.mem_size   =*/ buf_size,
       /*.mem_buffer =*/ model.buf.data(),
       /*.no_alloc   =*/ true,  // tensors are allocated later by the scheduler
   };
   
   struct ggml_context * ctx = ggml_init(params0);
   struct ggml_cgraph  * gf  = ggml_new_graph(ctx);
   
   model.a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, cols_A, rows_A);
   model.b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, cols_B, rows_B);
   
   struct ggml_tensor * result = ggml_mul_mat(ctx, model.a, model.b);
   ggml_build_forward_expand(gf, result);
   
   ggml_free(ctx); // free the context; the graph buffer outlives it
   
3. Allocate and upload tensor data

   Let the scheduler allocate backend memory, then upload the input data:

   ggml_backend_sched_reset(model.sched);
   ggml_backend_sched_alloc_graph(model.sched, gf);
   
   // upload CPU data to the backend (GPU, etc.)
   ggml_backend_tensor_set(model.a, matrix_A, 0, ggml_nbytes(model.a));
   ggml_backend_tensor_set(model.b, matrix_B, 0, ggml_nbytes(model.b));
   
4. Run the computation

   Execute the graph through the scheduler:

   ggml_backend_sched_graph_compute(model.sched, gf);
   
   struct ggml_tensor * result = ggml_graph_node(gf, -1);
   
5. Download the result and clean up

   Copy output data back to CPU memory, then free all backend resources:

   std::vector<float> out_data(ggml_nelements(result));
   ggml_backend_tensor_get(result, out_data.data(), 0, ggml_nbytes(result));
   
   // expected output:
   // [ 60.00 55.00 50.00 110.00
   //   90.00 54.00 54.00 126.00
   //   42.00 29.00 28.00  64.00 ]
   
   ggml_backend_sched_free(model.sched);
   ggml_backend_free(model.backend);
   ggml_backend_free(model.cpu_backend);
   

Full source

#include "ggml.h"
#include "ggml-backend.h"

#include <cstdio>
#include <cstring>
#include <vector>

struct simple_model {
    struct ggml_tensor * a {};
    struct ggml_tensor * b {};
    ggml_backend_t backend {};
    ggml_backend_t cpu_backend {};
    ggml_backend_sched_t sched {};
    std::vector<uint8_t> buf;
};

const int rows_A = 4, cols_A = 2;
float matrix_A[rows_A * cols_A] = { 2, 8, 5, 1, 4, 2, 8, 6 };

const int rows_B = 3, cols_B = 2;
float matrix_B[rows_B * cols_B] = { 10, 5, 9, 9, 5, 4 };

void init_model(simple_model & model) {
    ggml_backend_load_all();
    model.backend     = ggml_backend_init_best();
    model.cpu_backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
    ggml_backend_t backends[2] = { model.backend, model.cpu_backend };
    model.sched = ggml_backend_sched_new(backends, nullptr, 2,
                                         GGML_DEFAULT_GRAPH_SIZE, false, true);
}

struct ggml_cgraph * build_graph(simple_model & model) {
    size_t buf_size = ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE
                    + ggml_graph_overhead();
    model.buf.resize(buf_size);

    struct ggml_init_params params0 = {
        /*.mem_size   =*/ buf_size,
        /*.mem_buffer =*/ model.buf.data(),
        /*.no_alloc   =*/ true,
    };

    struct ggml_context * ctx = ggml_init(params0);
    struct ggml_cgraph  * gf  = ggml_new_graph(ctx);

    model.a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, cols_A, rows_A);
    model.b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, cols_B, rows_B);

    struct ggml_tensor * result = ggml_mul_mat(ctx, model.a, model.b);
    ggml_build_forward_expand(gf, result);
    ggml_free(ctx);
    return gf;
}

struct ggml_tensor * compute(simple_model & model, struct ggml_cgraph * gf) {
    ggml_backend_sched_reset(model.sched);
    ggml_backend_sched_alloc_graph(model.sched, gf);
    ggml_backend_tensor_set(model.a, matrix_A, 0, ggml_nbytes(model.a));
    ggml_backend_tensor_set(model.b, matrix_B, 0, ggml_nbytes(model.b));
    ggml_backend_sched_graph_compute(model.sched, gf);
    return ggml_graph_node(gf, -1);
}

int main(void) {
    ggml_time_init();

    simple_model model;
    init_model(model);

    struct ggml_cgraph * gf     = build_graph(model);
    struct ggml_tensor * result = compute(model, gf);

    std::vector<float> out_data(ggml_nelements(result));
    ggml_backend_tensor_get(result, out_data.data(), 0, ggml_nbytes(result));

    printf("mul mat (%d x %d):\n[", (int) result->ne[0], (int) result->ne[1]);
    for (int j = 0; j < result->ne[1]; j++) {
        if (j > 0) printf("\n");
        for (int i = 0; i < result->ne[0]; i++) {
            printf(" %.2f", out_data[j * result->ne[0] + i]);
        }
    }
    printf(" ]\n");

    ggml_backend_sched_free(model.sched);
    ggml_backend_free(model.backend);
    ggml_backend_free(model.cpu_backend);
    return 0;
}

Choosing an approach

	Context-based (simple-ctx)	Backend-based (simple-backend)
Device support	CPU only	CPU, CUDA, Metal, Vulkan, …
Memory management	Single pre-allocated pool	Scheduler allocates per-backend
Data transfer	Direct memcpy	ggml_backend_tensor_set / _get
Complexity	Lower	Higher
When to use	Prototyping, CPU-only tools	Production, GPU acceleration

For new projects targeting hardware acceleration, prefer the backend-based API. It is more verbose but works transparently across all ggml-supported devices.