Skip to main content
ggml supports automatic differentiation through reverse-mode (backpropagation). Every operation that has a differentiable implementation provides both a forward function and a backward function. The backward function computes the adjoint of each input tensor given the adjoint of the output.

​
How it works

  1. Forward pass — define the function and compute its value.
  2. Backward pass — ggml automatically builds gradient nodes that propagate the loss gradient back through every operation in the graph.
  3. Read gradients — retrieve the gradient tensor for any parameter after computation.

​
Marking trainable parameters

Call ggml_set_param to mark a tensor as a trainable parameter. This sets GGML_TENSOR_FLAG_PARAM on the tensor and tells the autodiff engine to compute gradients for it. 
struct ggml_tensor * x = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
ggml_set_param(x);  // x is an input variable / trainable parameter
ggml_set_param does not allocate a gradient tensor immediately. Gradient storage is allocated when you call ggml_build_backward_expand.

​
Full example: f(x) = a·x² + b

This example is taken directly from the ggml.h header comments.

​
Define the function

struct ggml_init_params params = {
    .mem_size   = 16*1024*1024,
    .mem_buffer = NULL,
};

struct ggml_context * ctx = ggml_init(params);

struct ggml_tensor * x = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
ggml_set_param(x);  // x is the variable we differentiate with respect to

struct ggml_tensor * a  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
struct ggml_tensor * b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
struct ggml_tensor * x2 = ggml_mul(ctx, x, x);
struct ggml_tensor * f  = ggml_add(ctx, ggml_mul(ctx, a, x2), b);

​
Build forward and backward graphs

// Build the forward graph (grads=true is required for backward pass)
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, /*grads=*/true);
ggml_build_forward_expand(gf, f);

// Allocate gradient accumulator tensors and attach backward nodes.
// grad_accs is an array of ggml_tensor* with one entry per graph output node.
// The ggml-opt.h high-level API manages grad_accs automatically.
// For manual use, pass an array of NULL pointers (ggml allocates gradient tensors):
struct ggml_tensor * grad_accs[1] = { NULL };
ggml_build_backward_expand(ctx, gf, grad_accs);

​
Set values and compute

ggml_set_f32(x, 2.0f);
ggml_set_f32(a, 3.0f);
ggml_set_f32(b, 4.0f);

// Reset gradient accumulators to zero before each backward pass
ggml_graph_reset(gf);

// Run both forward and backward passes
ggml_graph_compute_with_ctx(ctx, gf, /*n_threads=*/1);

​
Read gradients

// f = a*x^2 + b  =>  df/dx = 2*a*x = 2*3*2 = 12
struct ggml_tensor * grad_x = ggml_graph_get_grad(gf, x);
printf("df/dx = %f\n", ggml_get_f32_1d(grad_x, 0));  // 12.0

​
ggml_build_backward_expand

void ggml_build_backward_expand(
    struct ggml_context *  ctx,
    struct ggml_cgraph  *  cgraph,
    struct ggml_tensor  ** grad_accs);
  • ctx — the context used to allocate gradient tensors
  • cgraph — a forward graph previously built with ggml_build_forward_expand; must have been created with grads = true
  • grad_accs — array of ggml_tensor * with one entry per output node in the forward graph; pass NULL entries to have ggml allocate gradient accumulator tensors automatically
After this call, the graph contains both forward and backward nodes. Calling ggml_graph_compute will execute them in the correct order.

​
Accessing gradients

// Gradient tensor for a node in the forward graph
struct ggml_tensor * ggml_graph_get_grad(
    const struct ggml_cgraph * cgraph,
    const struct ggml_tensor * node);

// Gradient accumulator (for gradient accumulation across multiple batches)
struct ggml_tensor * ggml_graph_get_grad_acc(
    const struct ggml_cgraph * cgraph,
    const struct ggml_tensor * node);
ggml_graph_get_grad returns NULL for tensors that are not reachable by any parameter in the graph (i.e., tensors where no gradient flows).

​
Gradient accumulation

ggml supports accumulating gradients across multiple forward/backward passes before applying an optimizer step — useful for simulating larger batch sizes. 
for (int step = 0; step < accumulation_steps; step++) {
    // Load next mini-batch into input tensors
    load_batch(inputs, step);

    // Do NOT reset gradients between accumulation steps;
    // call ggml_graph_reset only before the first step in an accumulation window.
    if (step == 0) {
        ggml_graph_reset(gf);
    }

    ggml_graph_compute_with_ctx(ctx, gf, n_threads);
}

// Now apply optimizer using the accumulated gradients
struct ggml_tensor * grad = ggml_graph_get_grad_acc(gf, param);
ggml_graph_reset zeroes gradient accumulators and sets the loss gradient seed to 1.0. Call it once at the start of each accumulation window.

​
Loss tensors

Mark the final scalar output as a loss to signal the optimizer: 
struct ggml_tensor * loss = ggml_cross_entropy_loss(ctx, logits, labels);
ggml_set_loss(loss);  // sets GGML_TENSOR_FLAG_LOSS
Multiple loss tensors sum together. The backward pass seeds these tensors with gradient 1.0 automatically.

​
High-level training API

For training workloads, ggml-opt.h provides a higher-level interface that manages forward/backward graph construction, gradient accumulation, and optimizer steps: 
#include "ggml-opt.h"

ggml_opt_fit(
    backend_sched,
    ctx_compute,
    inputs,
    outputs,
    dataset,
    GGML_OPT_LOSS_TYPE_CROSS_ENTROPY,
    GGML_OPT_OPTIMIZER_TYPE_ADAMW,
    get_opt_pars,
    /*nepoch=*/10,
    /*nbatch_logical=*/256,
    /*val_split=*/0.1f,
    /*silent=*/false);
See ggml-opt.h for the full API.