Search notes:

ggml

Tensors

The library supports tensors up to 4 dimensions (#define GGML_MAX_DIMS 4).

The tensors are stored in row-major order. (See the members ne and nb of the ggml_tensor struct)

Tensors are created with ggml_new_tensor, ggml_new_tensor_1d, ggml_new_tensor_2d, ggml_new_tensor_3d or ggml_new_tensor_4d.

Each tensor operation produces a new tensor.

The library provides a forward and a backward computation function for each tensor operation.

The forward function computes the output tensor value given the input tensor values.

The backward function computes the adjoint of the input tensors given the adjoint of the output tensor.

For a detailed explanation of what this means, take a calculus class, or watch What is Automatic Differentiation?

Available tensors operations correspond to the entries in the ggml_op enum.

FP16 and FP32 are first class citizens.

Theortically, the library could be extended to support FP8 and (other?) integer data types as well.

Some possibly interesting functions …

`ggml_time_init`	To be called once when the program initializes
`ggml_init`	Allocates a memory buffer (for a tensor?) and returns a pointer to a `ggml_context`.
`ggml_new_tensor_*`	Create new tensors whose data is stored in the memory allocated with `ggml_init`.
`ggml_new_tensor_1d`	Returns a pointer to a `ggml_tensor`
`ggml_set_param`
`ggml_used_mem`	Determine amount of memory used in ???
`ggml_build_forward`, `ggml_build_forward_expand`, `ggml_build_backward`
`ggml_free`
`ggml_nelements`, `ggml_nbytes`
`ggml_graph_compute`	Requests to compute 'a function'
`ggml_blck_size`, `ggml_type_size`, `ggml_type_sizef`
`ggml_permute`
`ggml_conv_1d_1s`
`ggml_conv_1d_2s`
`ggml_quantize_q4_0`, `ggml_quantize_q4_1`
`ggml_get_f32_1d`, `ggml_set_f32_1d`
`ggml_time_ms`, `ggml_time_us`
`ggml_cycles`, `ggml_cycles_per_ms`
`ggml_print_object`, `ggml_print_objects`
`ggml_used_mem`
`ggml_add`,`ggml_mul`, …

`ggml_init_params`
`ggml_context`	A pointer to a `ggml_context` is returned by `ggml_init()`
`ggml_tensor`	A representation about a tensor stored in the momory. The struct has fields to query the tensor's size, data type and memory buffer. `src` points to the tensors from which the tensor was computed from. A pointer to a `ggml_tensor` is returned by `ggml_new_tensor_1d`
`ggml_object`
`ggml_cgraph`	Computation graph
`ggml_scratch`	Scratch buffer

`src`
`ne`	Number of elements in each dimension of the tensor
`nb`	Number of bytes (a.k.a. stride)
`data`	Pointer to the memory buffer

There are a few examples

{
    struct ggml_tensor * c = ggml_add(ctx, a, b);

    assert(c->src[0] == a);
    assert(c->src[1] == b);
}

ggml_type, ggml_op

The enum ggml_opt_type allows for the two optimization methods

Linsearch methods (enum ggml_linesearch):