Implementation:Ggml org Ggml Ggml opt init

Overview

The optimizer configuration implementation in GGML consists of two APIs that work together to initialize gradient-based optimization:

ggml_opt_default_params -- constructs a default parameter struct for the optimizer
ggml_opt_init -- creates a fully configured optimizer context from those parameters

API Details

ggml_opt_default_params

Signature:

struct ggml_opt_params ggml_opt_default_params(
    ggml_backend_sched_t backend_sched,
    enum ggml_opt_loss_type loss_type
);

Location: src/ggml-opt.cpp:L245-260

Parameters:

Parameter	Type	Description
`backend_sched`	`ggml_backend_sched_t`	Backend scheduler handle that determines where computation is executed
`loss_type`	`enum ggml_opt_loss_type`	Loss function type: `GGML_OPT_LOSS_TYPE_CROSS_ENTROPY` (classification) or `GGML_OPT_LOSS_TYPE_MEAN` (regression)

Returns: A ggml_opt_params struct populated with the following default values:

Field	Default Value	Description
`alpha`	0.001	AdamW learning rate
`beta1`	0.9	First moment exponential decay rate
`beta2`	0.999	Second moment exponential decay rate
`eps`	1e-8	Numerical stability constant
`wd`	0.0	Weight decay coefficient

ggml_opt_init

Signature:

ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params);

Location: src/ggml-opt.cpp:L549-582

Parameters:

Parameter	Type	Description
`params`	`struct ggml_opt_params`	The optimizer parameters struct (typically obtained from `ggml_opt_default_params`)

Returns: A ggml_opt_context_t representing the fully initialized optimizer context with:

The configured optimizer (AdamW or SGD) ready for use
Built forward, backward, and optimizer computation graphs

Dependencies

Header: ggml-opt.h

Import

#include "ggml-opt.h"

Usage Example

// Obtain default AdamW parameters for cross-entropy classification
struct ggml_opt_params params = ggml_opt_default_params(backend_sched, GGML_OPT_LOSS_TYPE_CROSS_ENTROPY);

// Optionally override defaults
params.adamw.alpha = 0.0005f;
params.adamw.wd    = 0.01f;

// Initialize the optimizer context
ggml_opt_context_t opt_ctx = ggml_opt_init(params);

Source

GGML

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment