ggml: new optimization interface (ggml/988)

ggml/CMakeLists.txt

@@ -228,6 +228,7 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-cann.h
     include/ggml-cuda.h
     include/ggml-kompute.h
+    include/ggml-opt.h
     include/ggml-metal.h
     include/ggml-rpc.h
     include/ggml-sycl.h

ggml/include/ggml-backend.h

@@ -86,7 +86,7 @@ extern "C" {
     GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
 
-    // "offset" refers to the offset of the tensor data for setting/getting data
+    // "offset" refers to the offset in tensor->data for setting/getting data
     GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_memset(   struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
@@ -242,14 +242,20 @@ extern "C" {
         ggml_backend_sched_reserve(sched, reserve_graph);
 
         // compute
-        graph = build_graph(sched);
+        graph = build_graph(sched); // the graph and its tensors are single-use in terms of allocation, multi-use in terms of computation
-        ggml_backend_sched_graph_compute(sched, graph);
+        for (int i = 0; i < 10; ++i) {
+            ggml_backend_sched_graph_compute(sched, graph); // on the first iteration the graph is allocated automatically
+        }
 
         // if there are graph inputs:
-        ggml_backend_sched_reset(sched);
+        graph = build_graph(sched); // get a new graph that is not allocated (the metadata for the old graph is freed once ggml_free is called)
-        ggml_backend_sched_alloc_graph(sched, graph);
+        ggml_backend_sched_reset(sched); // clear the allocation of the previous graph
-        ggml_backend_tensor_set(input_tensor, ...);
+        ggml_backend_sched_alloc_graph(sched, graph); // explicitly allocate the new graph but do not execute it
-        ggml_backend_sched_graph_compute(sched, graph);
+        ggml_backend_tensor_set(input_tensor, ...); // copy data to the newly allocated graph tensors
+        ggml_backend_sched_graph_compute(sched, graph); // execute the graph
+
+        // as an alternative to the above it is also possible to assign the inputs to a dedicated context and
+        // allocate them statically via ggml_backend_alloc_ctx_tensors
     }
     */
 
@@ -264,7 +270,7 @@ extern "C" {
     //
     typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
 
-    // Initialize a backend scheduler
+    // Initialize a backend scheduler, backends with low index are given priority over backends with high index
     GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
     GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
 
@@ -289,7 +295,9 @@ extern "C" {
     GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
     GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
 
-    // Reset all assignments and allocators - must be called before changing the node backends
+    // Reset all assignments and allocators - must be called before changing the node backends or allocating a new graph.
+    // This in effect deallocates all tensors that were previously allocated and leaves them with dangling pointers.
+    // The correct way to use this API is to discard the deallocated tensors and create new ones.
     GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
 
     // Set a callback to be called for each resulting node during graph compute

ggml/include/ggml-opt.h

@@ -0,0 +1,216 @@
+// This file contains functionality for training models using GGML.
+// It is not strictly needed vs. just vanilla GGML but it provides a more high-level interface for common needs such as datasets.
+// At the bottom of this file especially there are relatively high-level functions that are suitable use or adaptation in user code.
+//
+// Module maintainer: Johannes Gäßler (@JohannesGaessler, johannesg@5d6.de)
+
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#include <stdint.h>
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    struct ggml_opt_dataset;
+    struct ggml_opt_context;
+    struct ggml_opt_result;
+
+    typedef struct ggml_opt_dataset * ggml_opt_dataset_t;
+    typedef struct ggml_opt_context * ggml_opt_context_t;
+    typedef struct ggml_opt_result  * ggml_opt_result_t;
+
+    // ====== Loss ======
+
+    // built-in loss types, i.e. the built-in quantities minimized by the optimizer
+    // custom loss types can be defined via mean or sum which simply reduce the outputs for all datapoints to a single value
+    enum ggml_opt_loss_type {
+        GGML_OPT_LOSS_TYPE_MEAN,
+        GGML_OPT_LOSS_TYPE_SUM,
+        GGML_OPT_LOSS_TYPE_CROSS_ENTROPY,
+        GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR,
+    };
+
+    // ====== Dataset ======
+
+    GGML_API ggml_opt_dataset_t ggml_opt_dataset_init(
+            int64_t ne_datapoint, // number of elements per datapoint
+            int64_t ne_label,     // number of elements per label
+            int64_t ndata,        // total number of datapoints/labels
+            int64_t ndata_shard); // number of datapoints/labels per shard (unit at which the dataset is shuffled/copied)
+    GGML_API void ggml_opt_dataset_free(ggml_opt_dataset_t dataset);
+
+    // get underlying tensors that store the data
+    GGML_API struct ggml_tensor * ggml_opt_dataset_data  (ggml_opt_dataset_t dataset); // shape = [ne_datapoint, ndata]
+    GGML_API struct ggml_tensor * ggml_opt_dataset_labels(ggml_opt_dataset_t dataset); // shape = [nd_label,     ndata]
+
+    // shuffle idata first datapoints from dataset with RNG from opt_ctx, shuffle all datapoints if idata is negative
+    GGML_API void ggml_opt_dataset_shuffle(ggml_opt_context_t opt_ctx, ggml_opt_dataset_t dataset, int64_t idata);
+
+    // get batch at position ibatch from dataset and copy the data to data_batch and labels_batch
+    GGML_API void ggml_opt_dataset_get_batch(
+            ggml_opt_dataset_t   dataset,
+            struct ggml_tensor * data_batch,   // shape = [ne_datapoint, ndata_batch]
+            struct ggml_tensor * labels_batch, // shape = [ne_label,     ndata_batch]
+            int64_t              ibatch);
+
+    // ====== Model / Context ======
+
+    enum ggml_opt_build_type {
+        GGML_OPT_BUILD_TYPE_FORWARD,
+        GGML_OPT_BUILD_TYPE_GRAD,
+        GGML_OPT_BUILD_TYPE_OPT,
+    };
+
+    // parameters that control which optimizer is used and how said optimizer tries to find the minimal loss
+    struct ggml_opt_optimizer_params {
+        // AdamW optimizer parameters
+        struct {
+            float alpha; // learning rate
+            float beta1;
+            float beta2;
+            float eps;   // epsilon for numerical stability
+            float wd;    // weight decay for AdamW, use 0.0f to disable
+        } adamw;
+    };
+
+    // callback to calculate optimizer parameters prior to a backward pass
+    // userdata can be used to pass arbitrary data
+    typedef struct ggml_opt_optimizer_params (*ggml_opt_get_optimizer_params)(void * userdata);
+
+    // returns the default optimizer params (constant)
+    // userdata is not used
+    GGML_API struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata);
+
+    // parameters for initializing a new optimization context
+    struct ggml_opt_params {
+        ggml_backend_sched_t backend_sched; // defines which backends are used to construct the compute graphs
+
+        struct ggml_context * ctx_compute; // created in user code, holds non-static tensors
+
+        // the forward graph is defined by inputs and outputs
+        // those tensors and all tensors inbetween are not intended to be reusable between multiple optimization contexts
+        struct ggml_tensor * inputs;
+        struct ggml_tensor * outputs;
+
+        enum ggml_opt_loss_type  loss_type;
+        enum ggml_opt_build_type build_type;
+
+        int32_t opt_period; // after how many gradient accumulation steps an optimizer step should be done
+
+        ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
+        void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+    };
+
+    // get parameters for an optimization context with defaults set where possible
+    // parameters for which no sensible defaults exist are supplied as arguments to this function
+    GGML_API ggml_opt_params ggml_opt_default_params(
+            ggml_backend_sched_t      backend_sched,
+            struct ggml_context     * ctx_compute,
+            struct ggml_tensor      * inputs,
+            struct ggml_tensor      * outputs,
+            enum ggml_opt_loss_type   loss_type);
+
+    GGML_API ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params);
+    GGML_API void ggml_opt_free(ggml_opt_context_t opt_ctx);
+
+    // set gradients to zero, initilize loss, and optionally reset the optimizer
+    GGML_API void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer);
+
+    // get underlying tensors that store data
+    GGML_API struct ggml_tensor * ggml_opt_inputs(  ggml_opt_context_t opt_ctx); // forward graph input tensor
+    GGML_API struct ggml_tensor * ggml_opt_outputs( ggml_opt_context_t opt_ctx); // forward graph output tensor
+    GGML_API struct ggml_tensor * ggml_opt_labels(  ggml_opt_context_t opt_ctx); // labels to compare outputs against
+    GGML_API struct ggml_tensor * ggml_opt_loss(    ggml_opt_context_t opt_ctx); // scalar tensor that contains the loss
+    GGML_API struct ggml_tensor * ggml_opt_pred(    ggml_opt_context_t opt_ctx); // predictions made by outputs
+    GGML_API struct ggml_tensor * ggml_opt_ncorrect(ggml_opt_context_t opt_ctx); // number of matching predictions between outputs and labels
+
+    GGML_API struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node);
+
+    // ====== Optimization Result ======
+
+    GGML_API ggml_opt_result_t ggml_opt_result_init();
+    GGML_API void ggml_opt_result_free(ggml_opt_result_t result);
+    GGML_API void ggml_opt_result_reset(ggml_opt_result_t result);
+
+    // get data from result, uncertainties are optional and can be ignored by passing NULL
+    GGML_API void ggml_opt_result_ndata(   ggml_opt_result_t result, int64_t * ndata);                  // writes 1 value, number of datapoints
+    GGML_API void ggml_opt_result_loss(    ggml_opt_result_t result, double  * loss,     double * unc); // writes 1 value
+    GGML_API void ggml_opt_result_pred(    ggml_opt_result_t result, int32_t * pred);                   // writes ndata values
+    GGML_API void ggml_opt_result_accuracy(ggml_opt_result_t result, double  * accuracy, double * unc); // writes 1 value
+
+    // ====== Computation ======
+
+    // do forward pass, increment result if not NULL
+    GGML_API void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+
+    // do forward pass, increment result if not NULL, do backward pass
+    GGML_API void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+
+    // ############################################################################
+    // ## The high-level functions start here. They do not depend on any private ##
+    // ## functions or structs and can be copied to and adapted for user code.   ##
+    // ############################################################################
+
+    // ====== Intended Usage ======
+    //
+    // 1. Select the appropriate loss for your problem.
+    // 2. Create a dataset and set the data for the "data" tensor. Also set the "labels" tensor if your loss needs them.
+    //    Setting the shard size to 1 will be fine, it's the granularity with which data is shuffled/loaded (bigger values are faster).
+    // 3. Create a GGML graph for your model with no_alloc == true. Use two separate contexts for the tensors.
+    //    The first context should contain the model parameters and inputs and be allocated statically in user code.
+    //    The second context should contain all other tensors and will be (re)allocated automatically.
+    //    Due to this automated allocation the data of the second context is not defined when accessed in user code.
+    //    Note that the second dimension of the inputs/outputs are interpreted as the number of datapoints in those tensors.
+    // 4. Call ggml_opt_fit. If you need more control you can use ggml_opt_epoch instead.
+
+    // signature for a callback while evaluating opt_ctx on dataset, called after an evaluation
+    typedef void (*ggml_opt_epoch_callback)(
+            bool               train,       // true after training evaluation, false after validation evaluation
+            ggml_opt_context_t opt_ctx,
+            ggml_opt_dataset_t dataset,
+            ggml_opt_result_t  result,      // result associated with the dataset subsection
+            int64_t            ibatch,      // number of batches that have been evaluated so far
+            int64_t            ibatch_max,  // total number of batches in this dataset subsection
+            int64_t            t_start_us); // time at which the evaluation on the dataset subsection was started
+
+    // do training on front of dataset, do evaluation only on back of dataset
+    GGML_API void ggml_opt_epoch(
+            ggml_opt_context_t      opt_ctx,
+            ggml_opt_dataset_t      dataset,
+            ggml_opt_result_t       result_train,   // result to increment during training, ignored if NULL
+            ggml_opt_result_t       result_eval,    // result to increment during evaluation, ignored if NULL
+            int64_t                 idata_split,    // data index at which to split training and evaluation
+            ggml_opt_epoch_callback callback_train,
+            ggml_opt_epoch_callback callback_eval);
+
+    // callback that prints a progress bar on stderr
+    GGML_API void ggml_opt_epoch_callback_progress_bar(
+            bool               train,
+            ggml_opt_context_t opt_ctx,
+            ggml_opt_dataset_t dataset,
+            ggml_opt_result_t  result,
+            int64_t            ibatch,
+            int64_t            ibatch_max,
+            int64_t            t_start_us);
+
+    // fit model defined by inputs and outputs to dataset
+    GGML_API void ggml_opt_fit(
+            ggml_backend_sched_t            backend_sched,  // backend scheduler for constructing the compute graphs
+            ggml_context                  * ctx_compute,    // context with temporarily allocated tensors to calculate the outputs
+            ggml_tensor                   * inputs,         // input tensor with shape [ne_datapoint, ndata_batch]
+            ggml_tensor                   * outputs,        // output tensor, must have shape [ne_label, ndata_batch] if labels are used
+            ggml_opt_dataset_t              dataset,        // dataset with data and optionally also labels
+            enum ggml_opt_loss_type         loss_type,      // loss to minimize
+            ggml_opt_get_optimizer_params   get_opt_pars,   // callback to get optimizer params, userdata is pointer to epoch (of type int64_t)
+            int64_t                         nepoch,         // how many times the dataset should be iterated over
+            int64_t                         nbatch_logical, // datapoints optimizer step, must be a multiple of ndata_batch in inputs/outputs
+            float                           val_split,      // fraction of the dataset to use for validation, must be in [0.0f, 1.0f)
+            bool                            silent);        // whether or not info prints to stderr should be suppressed
+
+#ifdef  __cplusplus
+}
+#endif

ggml/include/ggml.h

@@ -602,7 +602,6 @@ extern "C" {
 
         int32_t flags;
 
-        struct ggml_tensor * grad;
         struct ggml_tensor * src[GGML_MAX_SRC];
 
         // source tensor and offset for views
@@ -615,7 +614,7 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-        // char padding[4];
+        char padding[8];
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -1985,28 +1984,20 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * grad,
-            float                 alpha,
+            struct ggml_tensor  * m,
-            float                 beta1,
+            struct ggml_tensor  * v,
-            float                 beta2,
+            struct ggml_tensor  * adamw_params); // parameters such a the learning rate
-            float                 eps,
-            float                 wd); // weight decay
 
     //
     // automatic differentiation
     //
 
     GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
-    GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate);
+    GGML_API void ggml_build_backward_expand(
-
+        struct ggml_context * ctx_static,  // context for static gradients (loss + gradient accumulation)
-    GGML_API void ggml_build_opt_adamw(
+        struct ggml_context * ctx_compute, // context for gradient computation
-            struct ggml_context * ctx,
+        struct ggml_cgraph  * cgraph,
-            struct ggml_cgraph  * gf,
+        bool                  accumulate); // whether or not gradients should be accumulated, requires static allocation of tensors in ctx_static
-            struct ggml_cgraph  * gb,
-            float                 alpha,
-            float                 beta1,
-            float                 beta2,
-            float                 eps,
-            float                 wd); // weight decay
 
     // graph allocation in a context
     GGML_API struct ggml_cgraph * ggml_new_graph       (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
@@ -2026,7 +2017,9 @@ extern "C" {
     GGML_API size_t ggml_graph_overhead(void);
     GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads);
 
-    GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);
+    GGML_API struct ggml_tensor * ggml_graph_get_tensor  (const struct ggml_cgraph * cgraph, const char * name);
+    GGML_API struct ggml_tensor * ggml_graph_get_grad    (const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
+    GGML_API struct ggml_tensor * ggml_graph_get_grad_acc(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
 
     GGML_API void                 ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
     GGML_API struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval);
@@ -2037,198 +2030,15 @@ extern "C" {
     // dump the graph into a file using the dot format
     GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename);
 
-    // build gradient checkpointing backward graph gb for gf using provided checkpoints
+    // TODO these functions were sandwiched in the old optimization interface, is there a better place for them?
-    // gb_tmp will contain original backward graph with rewritten backward process nodes,
-    // but without the second forward pass nodes.
-    GGML_API void ggml_build_backward_gradient_checkpointing(
-            struct ggml_context   * ctx,
-            struct ggml_cgraph    * gf,
-            struct ggml_cgraph    * gb,
-            struct ggml_cgraph    * gb_tmp,
-            struct ggml_tensor  * * checkpoints,
-            int                     n_checkpoints);
-    //
-    // optimization
-    //
-
-    // optimization methods
-    enum ggml_opt_type {
-        GGML_OPT_TYPE_ADAM,
-        GGML_OPT_TYPE_LBFGS,
-    };
-
-    // linesearch methods
-    enum ggml_linesearch {
-        GGML_LINESEARCH_DEFAULT = 1,
-
-        GGML_LINESEARCH_BACKTRACKING_ARMIJO       = 0,
-        GGML_LINESEARCH_BACKTRACKING_WOLFE        = 1,
-        GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE = 2,
-    };
-
-    // optimization return values
-    enum ggml_opt_result {
-        GGML_OPT_RESULT_OK = 0,
-        GGML_OPT_RESULT_DID_NOT_CONVERGE,
-        GGML_OPT_RESULT_NO_CONTEXT,
-        GGML_OPT_RESULT_INVALID_WOLFE,
-        GGML_OPT_RESULT_FAIL,
-        GGML_OPT_RESULT_CANCEL,
-
-        GGML_LINESEARCH_FAIL = -128,
-        GGML_LINESEARCH_MINIMUM_STEP,
-        GGML_LINESEARCH_MAXIMUM_STEP,
-        GGML_LINESEARCH_MAXIMUM_ITERATIONS,
-        GGML_LINESEARCH_INVALID_PARAMETERS,
-    };
-
-    typedef void (*ggml_opt_callback)(void * data, int accum_step, float * sched, bool * cancel);
     typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data);
 
     // Set callback for all future logging events.
     // If this is not called, or NULL is supplied, everything is output on stderr.
     GGML_API void ggml_log_set(ggml_log_callback log_callback, void * user_data);
 
-    // optimization parameters
-    //
-    //   see ggml.c (ggml_opt_default_params) for default values
-    //
-    struct ggml_opt_params {
-        enum ggml_opt_type type;
-
-        size_t graph_size;
-
-        int n_threads;
-
-        // delta-based convergence test
-        //
-        //   if past == 0 - disabled
-        //   if past > 0:
-        //     stop if |f(x) - f(x_past)| < delta * max(1, |f(x)|)
-        //
-        int past;
-        float delta;
-
-        // maximum number of iterations without improvement
-        //
-        //   if 0 - disabled
-        //   if > 0:
-        //     assume convergence if no cost improvement in this number of iterations
-        //
-        int max_no_improvement;
-
-        bool print_forward_graph;
-        bool print_backward_graph;
-
-        int n_gradient_accumulation;
-
-        // ADAM parameters
-        struct {
-            int n_iter;
-
-            float sched; // schedule multiplier (fixed, decay or warmup)
-            float decay; // weight decay for AdamW, use 0.0f to disable
-            int   decay_min_ndim; // minimum number of tensor dimension to apply weight decay
-            float alpha; // learning rate
-            float beta1;
-            float beta2;
-            float eps;   // epsilon for numerical stability
-            float eps_f; // epsilon for convergence test
-            float eps_g; // epsilon for convergence test
-            float gclip; // gradient clipping
-        } adam;
-
-        // LBFGS parameters
-        struct {
-            int m; // number of corrections to approximate the inv. Hessian
-            int n_iter;
-            int max_linesearch;
-
-            float eps;      // convergence tolerance
-            float ftol;     // line search tolerance
-            float wolfe;
-            float min_step;
-            float max_step;
-
-            enum ggml_linesearch linesearch;
-        } lbfgs;
-    };
-
-    struct ggml_opt_context {
-        struct ggml_context * ctx;
-        struct ggml_opt_params params;
-
-        int iter;
-        int64_t nx; // number of parameter elements
-
-        bool just_initialized;
-
-        float loss_before;
-        float loss_after;
-
-        struct {
-            struct ggml_tensor * g;  // current gradient
-            struct ggml_tensor * m;  // first moment
-            struct ggml_tensor * v;  // second moment
-            struct ggml_tensor * pf; // past function values
-            float fx_best;
-            float fx_prev;
-            int n_no_improvement;
-        } adam;
-
-        struct {
-            struct ggml_tensor * x;    // current parameters
-            struct ggml_tensor * xp;   // previous parameters
-            struct ggml_tensor * g;    // current gradient
-            struct ggml_tensor * gp;   // previous gradient
-            struct ggml_tensor * d;    // search direction
-            struct ggml_tensor * pf;   // past function values
-            struct ggml_tensor * lmal; // the L-BFGS memory alpha
-            struct ggml_tensor * lmys; // the L-BFGS memory ys
-            struct ggml_tensor * lms;  // the L-BFGS memory s
-            struct ggml_tensor * lmy;  // the L-BFGS memory y
-            float fx_best;
-            float step;
-            int j;
-            int k;
-            int end;
-            int n_no_improvement;
-        } lbfgs;
-    };
-
     GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
 
-    GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type);
-
-    // optimize the function defined by the tensor f
-    GGML_API enum ggml_opt_result ggml_opt(
-            struct ggml_context * ctx,
-            struct ggml_opt_params params,
-            struct ggml_tensor * f);
-
-    // initialize optimizer context
-    GGML_API void ggml_opt_init(
-            struct ggml_context     * ctx,
-            struct ggml_opt_context * opt,
-            struct ggml_opt_params    params,
-            int64_t                   nx);
-
-    // continue optimizing the function defined by the tensor f
-    GGML_API enum ggml_opt_result ggml_opt_resume(
-            struct ggml_context * ctx,
-            struct ggml_opt_context * opt,
-            struct ggml_tensor * f);
-
-    // continue optimizing the function defined by the tensor f
-    GGML_API enum ggml_opt_result ggml_opt_resume_g(
-            struct ggml_context * ctx,
-            struct ggml_opt_context * opt,
-            struct ggml_tensor * f,
-            struct ggml_cgraph * gf,
-            struct ggml_cgraph * gb,
-            ggml_opt_callback callback,
-            void * callback_data);
-
     //
     // quantization
     //

ggml/src/CMakeLists.txt

@@ -207,9 +207,11 @@ add_library(ggml-base
             ../include/ggml-alloc.h
             ../include/ggml-backend.h
             ../include/ggml-cpp.h
+            ../include/ggml-opt.h
             ggml.c
             ggml-alloc.c
             ggml-backend.cpp
+            ggml-opt.cpp
             ggml-threading.cpp
             ggml-threading.h
             ggml-quants.c

ggml/src/ggml-alloc.c

@@ -466,18 +466,12 @@ static bool ggml_gallocr_is_own(ggml_gallocr_t galloc, struct ggml_tensor * t) {
     return ggml_gallocr_hash_get(galloc, t)->allocated;
 }
 
-static void ggml_gallocr_set_node_offset(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id, size_t offset) {
-    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
-    hn->buffer_id = buffer_id;
-    hn->offset = offset;
-    hn->allocated = true;
-}
-
 static bool ggml_gallocr_is_allocated(ggml_gallocr_t galloc, struct ggml_tensor * t) {
     return t->data != NULL || ggml_gallocr_hash_get(galloc, t)->allocated;
 }
 
 static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) {
+    GGML_ASSERT(buffer_id >= 0);
     struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
 
     if (!ggml_gallocr_is_allocated(galloc, node) && !ggml_is_view(node)) {
@@ -816,7 +810,11 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
 }
 
 static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
-    size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
+    size_t node_size = 0;
+    if (!node->data && !node->view_src) {
+        GGML_ASSERT(talloc->buffer_id >= 0); // prevent segfault when misusing the API
+        node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
+    }
     return talloc->size_max >= node_size;
 }
 

ggml/src/ggml-backend.cpp

@@ -279,7 +279,7 @@ void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, siz
     buf->iface.get_tensor(buf, tensor, data, offset, size);
 }
 
-GGML_API void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
     if (size == 0) {

ggml/src/ggml-cpu/ggml-cpu.c

@@ -12220,7 +12220,12 @@ static void ggml_compute_forward_opt_step_adamw_f32(
     const struct ggml_tensor * src0_grad    = dst->src[1];
     const struct ggml_tensor * src0_grad_m  = dst->src[2];
     const struct ggml_tensor * src0_grad_v  = dst->src[3];
+    const struct ggml_tensor * adamw_params = dst->src[4];
+
     GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_m));
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_v));
+    GGML_ASSERT(ggml_nelements(adamw_params) == 7);
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -12237,16 +12242,14 @@ static void ggml_compute_forward_opt_step_adamw_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    /* const float   gnorm = 1.0f; */
+    const float * adamw_params_ptr = ggml_get_data_f32(adamw_params);
-    int64_t       iter;   memcpy(&iter, &dst->op_params[0], sizeof(int64_t));
+    const float alpha  = adamw_params_ptr[0];
-    const float   alpha = ggml_get_op_params_f32(dst, 2);
+    const float beta1  = adamw_params_ptr[1];
-    const float   beta1 = ggml_get_op_params_f32(dst, 3);
+    const float beta2  = adamw_params_ptr[2];
-    const float   beta2 = ggml_get_op_params_f32(dst, 4);
+    const float eps    = adamw_params_ptr[3];
-    const float   eps   = ggml_get_op_params_f32(dst, 5);
+    const float wd     = adamw_params_ptr[4];
-    const float   wd    = ggml_get_op_params_f32(dst, 6);
+    const float beta1h = adamw_params_ptr[5];
-
+    const float beta2h = adamw_params_ptr[6];
-    const float beta1h  = alpha/(1.0f - powf(beta1, iter));
-    const float beta2h  =  1.0f/(1.0f - powf(beta2, iter));
 
     for (int ir = ir0; ir < ir1; ++ir) {
         const int64_t i03 = ir/(ne02*ne01);
@@ -12270,17 +12273,9 @@ static void ggml_compute_forward_opt_step_adamw_f32(
             // The weight decay is applied independently of the Adam momenta m and v.
             // This is NOT equivalent to l2 regularization that adds w[i00]*w[i00] to the loss.
             // See: https://arxiv.org/pdf/1711.05101v3.pdf
-            w[i00] = w[i00]*(1.0f - alpha*wd) - mh/vh;
+            w[i00] = w[i00]*(1.0f - alpha*wd) - alpha*mh/vh;
         }
     }
-
-    ggml_barrier(params->threadpool);
-    if (ith != 0) {
-        return;
-    }
-
-    iter++;
-    memcpy(&dst->op_params[0], &iter, sizeof(int64_t));
 }
 
 static void ggml_compute_forward_opt_step_adamw(

ggml/src/ggml-cuda/opt-step-adamw.cu

@@ -1,11 +1,11 @@
+#include "ggml-impl.h"
 #include "opt-step-adamw.cuh"
 
 #include <cstdint>
 
 static __global__ void opt_step_adamw_f32(
-    float * __restrict__ x, const float * __restrict__ g, float * __restrict__ g_m, float * __restrict__ g_v, const int64_t k,
+    float * __restrict__ x, const float * __restrict__ g, float * __restrict__ g_m, float * __restrict__ g_v,
-    const float alpha, const float beta1, const float beta2, const float eps, const float wd,
+    const float * __restrict__ pars, const int64_t k) {
-    const float beta1h, const float beta2h) {
 
     const int64_t i = (int64_t) blockIdx.x*blockDim.x + threadIdx.x;
 
@@ -13,6 +13,14 @@ static __global__ void opt_step_adamw_f32(
         return;
     }
 
+    const float alpha  = pars[0];
+    const float beta1  = pars[1];
+    const float beta2  = pars[2];
+    const float eps    = pars[3];
+    const float wd     = pars[4];
+    const float beta1h = pars[5];
+    const float beta2h = pars[6];
+
     const float gi = g[i];
     const float gmi = g_m[i]*beta1 +    gi*(1.0f - beta1);
     const float gvi = g_v[i]*beta2 + gi*gi*(1.0f - beta2);
@@ -23,17 +31,15 @@ static __global__ void opt_step_adamw_f32(
     const float mh =       gmi*beta1h;
     const float vh = sqrtf(gvi*beta2h) + eps;
 
-    x[i] = x[i]*(1.0f - alpha*wd) - mh/vh;
+    x[i] = x[i]*(1.0f - alpha*wd) - alpha*mh/vh;
 }
 
 static void opt_step_adamw_f32_cuda(
-    float * x, const float * g, float * g_m, float * g_v, const int64_t k,
+    float * x, const float * g, float * g_m, float * g_v, const float * pars, const int64_t k, cudaStream_t stream) {
-    const float alpha, const float beta1, const float beta2, const float eps, const float wd,
-    const float beta1h, const float beta2h, cudaStream_t stream) {
 
     const dim3 block_dims(CUDA_OPT_STEP_ADAMW_BLOCK_SIZE, 1, 1);
     const dim3 block_nums((k + CUDA_OPT_STEP_ADAMW_BLOCK_SIZE - 1) / CUDA_OPT_STEP_ADAMW_BLOCK_SIZE, 1, 1);
-    opt_step_adamw_f32<<<block_nums, block_dims, 0, stream>>>(x, g, g_m, g_v, k, alpha, beta1, beta2, eps, wd, beta1h, beta2h);
+    opt_step_adamw_f32<<<block_nums, block_dims, 0, stream>>>(x, g, g_m, g_v, pars, k);
 }
 
 void ggml_cuda_opt_step_adamw(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -41,40 +47,32 @@ void ggml_cuda_opt_step_adamw(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     const ggml_tensor * src0_grad    = dst->src[1];
     const ggml_tensor * src0_grad_m  = dst->src[2];
     const ggml_tensor * src0_grad_v  = dst->src[3];
+    const ggml_tensor * adamw_params = dst->src[4];
 
     GGML_ASSERT(src0->type         == GGML_TYPE_F32);
     GGML_ASSERT(src0_grad->type    == GGML_TYPE_F32);
     GGML_ASSERT(src0_grad_m->type  == GGML_TYPE_F32);
     GGML_ASSERT(src0_grad_v->type  == GGML_TYPE_F32);
+    GGML_ASSERT(adamw_params->type == GGML_TYPE_F32);
     GGML_ASSERT(ggml_is_contiguous(src0));
     GGML_ASSERT(ggml_is_contiguous(src0_grad));
     GGML_ASSERT(ggml_is_contiguous(src0_grad_m));
     GGML_ASSERT(ggml_is_contiguous(src0_grad_v));
+    GGML_ASSERT(ggml_is_contiguous(adamw_params));
     GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
     GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_m));
     GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_v));
+    GGML_ASSERT(ggml_nelements(adamw_params) == 7);
 
     float       * src0_d         = (float       *) src0->data;
     const float * src0_grad_d    = (const float *) src0_grad->data;
     float       * src0_grad_m_d  = (float       *) src0_grad_m->data;
     float       * src0_grad_v_d  = (float       *) src0_grad_v->data;
+    const float * adamw_params_d = (const float *) adamw_params->data;
 
     cudaStream_t stream = ctx.stream();
 
     const int64_t ne = ggml_nelements(src0);
 
-    int64_t iter;  memcpy(&iter,  &dst->op_params[0], sizeof(int64_t));
+    opt_step_adamw_f32_cuda(src0_d, src0_grad_d, src0_grad_m_d, src0_grad_v_d, adamw_params_d, ne, stream);
-    float   alpha; memcpy(&alpha, &dst->op_params[2], sizeof(float));
-    float   beta1; memcpy(&beta1, &dst->op_params[3], sizeof(float));
-    float   beta2; memcpy(&beta2, &dst->op_params[4], sizeof(float));
-    float   eps;   memcpy(&eps,   &dst->op_params[5], sizeof(float));
-    float   wd;    memcpy(&wd,    &dst->op_params[6], sizeof(float));
-
-    const float beta1h  = alpha/(1.0f - powf(beta1, iter));
-    const float beta2h  =  1.0f/(1.0f - powf(beta2, iter));
-
-    opt_step_adamw_f32_cuda(src0_d, src0_grad_d, src0_grad_m_d, src0_grad_v_d, ne, alpha, beta1, beta2, eps, wd, beta1h, beta2h, stream);
-
-    iter++;
-    memcpy(&dst->op_params[0], &iter, sizeof(int64_t));
 }

ggml/src/ggml-impl.h

@@ -196,7 +196,7 @@ void ggml_hash_set_reset(struct ggml_hash_set * hash_set);
 static bool ggml_hash_contains(const struct ggml_hash_set * hash_set, struct ggml_tensor * key);
 
 // returns GGML_HASHSET_FULL if table is full, otherwise the current index of the key or where it should be inserted
-static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, struct ggml_tensor * key);
+static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, const struct ggml_tensor * key);
 
 // returns GGML_HASHSET_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
 static size_t ggml_hash_insert(struct ggml_hash_set * hash_set, struct ggml_tensor * key);
@@ -210,7 +210,7 @@ static inline size_t ggml_hash(const struct ggml_tensor * p) {
     return (size_t)(uintptr_t)p >> 4;
 }
 
-static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, struct ggml_tensor * key) {
+static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, const struct ggml_tensor * key) {
     size_t h = ggml_hash(key) % hash_set->size;
 
     // linear probing
@@ -281,13 +281,14 @@ enum ggml_cgraph_eval_order {
 };
 
 struct ggml_cgraph {
-    int size;
+    int size;    // maximum number of nodes/leafs/grads/grad_accs
-    int n_nodes;
+    int n_nodes; // number of nodes currently in use
-    int n_leafs;
+    int n_leafs; // number of leafs currently in use
 
-    struct ggml_tensor ** nodes;
+    struct ggml_tensor ** nodes;     // tensors with data that can change if the graph is evaluated
-    struct ggml_tensor ** grads;
+    struct ggml_tensor ** grads;     // the outputs of these tensors are the gradients of the nodes
-    struct ggml_tensor ** leafs;
+    struct ggml_tensor ** grad_accs; // accumulators for node gradients
+    struct ggml_tensor ** leafs;     // tensors with constant data
 
     struct ggml_hash_set visited_hash_set;
 

ggml/src/ggml-metal/ggml-metal.m

@@ -3639,6 +3639,12 @@ static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
     return ctx->all_data;
 }
 
+static void ggml_backend_metal_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    memset((char *)tensor->data + offset, value, size);
+
+    UNUSED(buffer);
+}
+
 static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     memcpy((char *)tensor->data + offset, data, size);
 
@@ -3671,7 +3677,7 @@ static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
     /* .free_buffer     = */ ggml_backend_metal_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_metal_buffer_get_base,
     /* .init_tensor     = */ NULL,
-    /* .memset_tensor   = */ NULL,
+    /* .memset_tensor   = */ ggml_backend_metal_buffer_memset_tensor,
     /* .set_tensor      = */ ggml_backend_metal_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_metal_buffer_get_tensor,
     /* .cpy_tensor      = */ ggml_backend_metal_buffer_cpy_tensor,

ggml/src/ggml-opt.cpp

@@ -0,0 +1,867 @@
+#include "ggml-opt.h"
+
+#include "ggml.h"
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <inttypes.h>
+#include <map>
+#include <random>
+#include <vector>
+
+struct ggml_opt_dataset {
+    struct ggml_context   * ctx;
+    ggml_backend_buffer_t   buf;
+    struct ggml_tensor    * data;
+    struct ggml_tensor    * labels;
+
+    int64_t ndata;
+    int64_t ndata_shard;
+    size_t  nbs_data;
+    size_t  nbs_labels;
+
+    std::vector<int64_t> permutation;
+};
+
+struct ggml_opt_context {
+    ggml_backend_sched_t    backend_sched;
+    ggml_cgraph           * allocated_graph;
+    ggml_cgraph           * allocated_graph_copy;
+    struct ggml_context   * ctx_static;
+    struct ggml_context   * ctx_static_cpu;
+    struct ggml_context   * ctx_compute;
+    struct ggml_context   * ctx_copy;
+    ggml_backend_buffer_t   buf_static;
+    ggml_backend_buffer_t   buf_static_cpu;
+    std::mt19937            rng;
+
+    struct ggml_tensor * inputs;
+    struct ggml_tensor * outputs;
+    struct ggml_tensor * labels;
+
+    struct ggml_tensor * loss;
+    struct ggml_tensor * pred;
+    struct ggml_tensor * ncorrect;
+
+    struct ggml_cgraph * gf;
+    struct ggml_cgraph * gb_grad;
+    struct ggml_cgraph * gb_opt;
+
+    int64_t iter;
+    int32_t opt_period;
+    int32_t opt_i;
+    bool    loss_per_datapoint;
+
+    ggml_opt_get_optimizer_params get_opt_pars;
+    void * get_opt_pars_ud;
+    struct ggml_tensor * adamw_params;
+};
+
+struct ggml_opt_result {
+    int64_t              ndata    = 0;
+    std::vector<float>   loss;
+    std::vector<int32_t> pred;
+    int64_t              ncorrect = 0;
+
+    bool loss_per_datapoint = false;
+    int64_t opt_period = -1;
+};
+
+// ====== Dataset ======
+
+ggml_opt_dataset_t ggml_opt_dataset_init(int64_t ne_datapoint, int64_t ne_label, int64_t ndata, int64_t ndata_shard) {
+    GGML_ASSERT(ne_datapoint >  0);
+    GGML_ASSERT(ne_label     >= 0);
+    GGML_ASSERT(ndata        >  0);
+    GGML_ASSERT(ndata_shard  >  0);
+
+    ggml_opt_dataset_t result = new ggml_opt_dataset;
+    result->ndata       = ndata;
+    result->ndata_shard = ndata_shard;
+
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ 2*ggml_tensor_overhead(),
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        result->ctx = ggml_init(params);
+    }
+
+    result->data = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_datapoint, ndata);
+    result->nbs_data = ggml_nbytes(result->data) * ndata_shard/ndata;
+
+    if (ne_label > 0) {
+        result->labels = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_label, ndata);
+        result->nbs_labels = ggml_nbytes(result->labels) * ndata_shard/ndata;
+    } else {
+        result->labels = nullptr;
+        result->nbs_labels = 0;
+    }
+
+    result->buf = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx, ggml_backend_cpu_buffer_type());
+
+    const int64_t nshards = ndata/ndata_shard;
+    result->permutation.resize(nshards);
+    for (int64_t i = 0; i < nshards; ++i) {
+        result->permutation[i] = i;
+    }
+    return result;
+}
+
+void ggml_opt_dataset_free(ggml_opt_dataset_t dataset) {
+    ggml_backend_buffer_free(dataset->buf);
+    ggml_free(dataset->ctx);
+    delete dataset;
+}
+
+struct ggml_tensor * ggml_opt_dataset_data(ggml_opt_dataset_t dataset) {
+    return dataset->data;
+}
+
+struct ggml_tensor * ggml_opt_dataset_labels(ggml_opt_dataset_t dataset) {
+    return dataset->labels;
+}
+
+void ggml_opt_dataset_shuffle(ggml_opt_context_t opt_ctx, ggml_opt_dataset_t dataset, int64_t idata) {
+    GGML_ASSERT(idata <= dataset->ndata);
+
+    if (idata < 0) {
+        std::shuffle(dataset->permutation.begin(), dataset->permutation.end(), opt_ctx->rng);
+        return;
+    }
+
+    GGML_ASSERT(idata % dataset->ndata_shard == 0);
+    const int64_t ishard_max = idata / dataset->ndata_shard;
+    std::shuffle(dataset->permutation.begin(), dataset->permutation.begin() + ishard_max, opt_ctx->rng);
+}
+
+void ggml_opt_dataset_get_batch(ggml_opt_dataset_t dataset, struct ggml_tensor * data_batch, struct ggml_tensor * labels_batch, int64_t ibatch) {
+    GGML_ASSERT(   data_batch && ggml_is_contiguous(data_batch));
+    GGML_ASSERT(!labels_batch || ggml_is_contiguous(labels_batch));
+    GGML_ASSERT((labels_batch == nullptr) == (dataset->labels == nullptr));
+
+    const size_t nb_data_batch = ggml_nbytes(data_batch);
+    GGML_ASSERT(nb_data_batch % dataset->nbs_data == 0);
+    const int64_t shards_per_batch = nb_data_batch / dataset->nbs_data;
+
+    if (labels_batch) {
+        const size_t nb_labels_batch = ggml_nbytes(labels_batch);
+        GGML_ASSERT(nb_labels_batch == shards_per_batch*dataset->nbs_labels);
+    }
+
+    GGML_ASSERT((ibatch + 1)*shards_per_batch <= int64_t(dataset->permutation.size()));
+
+    for (int64_t ishard_batch = 0; ishard_batch < shards_per_batch; ++ishard_batch) {
+        const int64_t ishard = dataset->permutation[ibatch*shards_per_batch + ishard_batch];
+
+        const char * ptr_data = (const char *) dataset->data->data + ishard*dataset->nbs_data;
+        ggml_backend_tensor_set(data_batch, ptr_data, ishard_batch*dataset->nbs_data, dataset->nbs_data);
+
+        if (!labels_batch) {
+            continue;
+        }
+
+        const char * ptr_labels = (const char *) dataset->labels->data + ishard*dataset->nbs_labels;
+        ggml_backend_tensor_set(labels_batch, ptr_labels, ishard_batch*dataset->nbs_labels, dataset->nbs_labels);
+    }
+}
+
+// ====== Model / Context ======
+
+struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata) {
+    GGML_UNUSED(userdata);
+
+    ggml_opt_optimizer_params result;
+
+    result.adamw.alpha = 0.001f;
+    result.adamw.beta1 = 0.9f;
+    result.adamw.beta2 = 0.999f;
+    result.adamw.eps   = 1e-8f;
+    result.adamw.wd    = 0.0f;
+
+    return result;
+}
+
+struct ggml_opt_params ggml_opt_default_params(
+        ggml_backend_sched_t backend_sched,
+        struct ggml_context * ctx_compute,
+        struct ggml_tensor * inputs,
+        struct ggml_tensor * outputs,
+        enum ggml_opt_loss_type loss_type) {
+    return {
+        /*backend_sched   =*/ backend_sched,
+        /*ctx_compute     =*/ ctx_compute,
+        /*inputs          =*/ inputs,
+        /*logits          =*/ outputs,
+        /*loss_type       =*/ loss_type,
+        /*build_type      =*/ GGML_OPT_BUILD_TYPE_OPT,
+        /*opt_period      =*/ 1,
+        /*get_opt_pars    =*/ ggml_opt_get_default_optimizer_params,
+        /*get_opt_pars_ud =*/ nullptr,
+    };
+}
+
+static ggml_tensor * map_tensor(std::map<ggml_tensor *, ggml_tensor *> & tensor_map, ggml_context * ctx, ggml_tensor * tensor) {
+    if (!tensor) {
+        return nullptr;
+    }
+
+    if (tensor_map.find(tensor) != tensor_map.end()) {
+        return tensor_map[tensor];
+    }
+
+    ggml_tensor * new_tensor = ggml_dup_tensor(ctx, tensor);
+    tensor_map[tensor] = new_tensor;
+
+    new_tensor->op = tensor->op;
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        new_tensor->nb[i] = tensor->nb[i];
+    }
+    new_tensor->flags = tensor->flags;
+    memcpy(new_tensor->op_params, tensor->op_params, sizeof(tensor->op_params));
+    strcpy(new_tensor->name, tensor->name);
+    new_tensor->data = tensor->data;
+    new_tensor->buffer = tensor->buffer;
+    new_tensor->extra = tensor->extra;
+    new_tensor->view_offs = tensor->view_offs;
+    new_tensor->view_src = map_tensor(tensor_map, ctx, tensor->view_src);
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        new_tensor->src[i] = map_tensor(tensor_map, ctx, tensor->src[i]);
+    }
+
+    return new_tensor;
+}
+
+static ggml_cgraph * dup_graph(ggml_context * ctx, ggml_cgraph * graph) {
+    std::map<ggml_tensor *, ggml_tensor *> tensor_map;
+
+    ggml_cgraph * new_graph = ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true);
+
+    for (int i = 0; i < graph->n_leafs; i++) {
+        ggml_build_forward_expand(new_graph, map_tensor(tensor_map, ctx, graph->leafs[i]));
+    }
+    for (int i = 0; i < graph->n_nodes; i++) {
+        ggml_build_forward_expand(new_graph, map_tensor(tensor_map, ctx, graph->nodes[i]));
+    }
+    for (int i = 0; i < graph->n_nodes; ++i) {
+        const size_t igrad_src = ggml_hash_find(&graph->visited_hash_set, graph->nodes[i]);
+        const size_t igrad_dst = ggml_hash_find(&new_graph->visited_hash_set, new_graph->nodes[i]);
+        graph->grads[igrad_dst]     = new_graph->grads[igrad_src];
+        graph->grad_accs[igrad_dst] = new_graph->grad_accs[igrad_src];
+    }
+
+    return new_graph;
+}
+
+static void ggml_opt_alloc_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph) {
+    GGML_ASSERT(graph);
+    if (opt_ctx->allocated_graph == graph) {
+        return;
+    }
+
+    ggml_backend_sched_reset(opt_ctx->backend_sched); // clear allocation of previous graph
+
+    {
+        ggml_init_params params = {
+            /*.mem_size   =*/ ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE,
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        ggml_free(opt_ctx->ctx_copy);
+        opt_ctx->ctx_copy = ggml_init(params);
+    }
+
+    opt_ctx->allocated_graph_copy = dup_graph(opt_ctx->ctx_copy, graph);
+
+    ggml_backend_sched_alloc_graph(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
+    opt_ctx->allocated_graph = graph;
+}
+
+ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
+    ggml_opt_context_t result = new struct ggml_opt_context;
+    result->backend_sched        = params.backend_sched;
+    result->allocated_graph      = nullptr;
+    result->allocated_graph_copy = nullptr;
+    result->ctx_compute          = params.ctx_compute;
+    result->ctx_copy             = nullptr;
+    result->inputs               = params.inputs;
+    result->outputs              = params.outputs;
+    result->iter                 = 1;
+    result->opt_period           = params.opt_period;
+    result->opt_i                = 0;
+    result->get_opt_pars         = params.get_opt_pars;
+    result->get_opt_pars_ud      = params.get_opt_pars_ud;
+
+    GGML_ASSERT(result->inputs->data && "the inputs must be allocated statically");
+    GGML_ASSERT(result->opt_period >= 1);
+
+    const bool accumulate = params.build_type == GGML_OPT_BUILD_TYPE_GRAD ||
+        (params.build_type == GGML_OPT_BUILD_TYPE_OPT && result->opt_period > 1);
+
+    ggml_set_input(result->inputs);
+    ggml_set_output(result->outputs);
+
+    result->gf = ggml_new_graph_custom(result->ctx_compute, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true); // Forward pass.
+    ggml_build_forward_expand(result->gf, result->outputs);
+
+    int n_param = 0;
+    for (int i = 0; i < result->gf->n_nodes; ++i) {
+        if (result->gf->nodes[i]->flags & GGML_TENSOR_FLAG_PARAM) {
+            n_param++;
+        }
+    }
+
+    {
+        // The static context is used for:
+        //   - gradients (1 tensor per param if using gradient accumulation)
+        //   - optimizer momenta (2 tensors per param)
+        //   - labels
+        //   - loss + its gradient (up to 5 tensors)
+        //   - pred
+        //   - ncorrect (2 tensors).
+        const size_t tensors_per_param = (accumulate ? 1 : 0) + (params.build_type == GGML_OPT_BUILD_TYPE_OPT ? 2 : 0);
+        const size_t size_meta = (tensors_per_param*n_param + 9) * ggml_tensor_overhead();
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ size_meta,
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        result->ctx_static = ggml_init(params);
+    }
+    {
+        // The static cpu context is used for:
+        //   - optimizer parameters (1 for the entire context)
+        const size_t size_meta = 1 * ggml_tensor_overhead();
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ size_meta,
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        result->ctx_static_cpu = ggml_init(params);
+    }
+
+
+    switch (params.loss_type) {
+        case GGML_OPT_LOSS_TYPE_MEAN: {
+            result->labels = nullptr;
+            result->loss = ggml_sum(result->ctx_static, result->outputs);
+            ggml_set_name(result->loss, "loss_sum");
+            const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs));
+            result->loss = ggml_scale(result->ctx_static, result->loss, scale);
+            ggml_set_name(result->loss, "loss_mean");
+            result->loss_per_datapoint = true;
+            break;
+        }
+        case GGML_OPT_LOSS_TYPE_SUM: {
+            result->labels = nullptr;
+            result->loss = ggml_sum(result->ctx_static, result->outputs);
+            ggml_set_name(result->loss, "loss_sum");
+            result->loss_per_datapoint = false;
+            break;
+        }
+        case GGML_OPT_LOSS_TYPE_CROSS_ENTROPY: {
+            result->labels = ggml_dup_tensor(result->ctx_static, result->outputs);
+            ggml_set_input(result->labels);
+            ggml_set_name(result->labels, "labels");
+            result->loss = ggml_cross_entropy_loss(result->ctx_static, result->outputs, result->labels);
+            ggml_set_name(result->loss, "loss_cross_entropy");
+            if (result->opt_period > 1) {
+                result->loss = ggml_scale(result->ctx_static, result->loss, 1.0f / result->opt_period);
+                ggml_set_name(result->loss, "loss_cross_entropy_scaled");
+            }
+            result->loss_per_datapoint = true;
+            break;
+        }
+        case GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR: {
+            result->labels = ggml_dup_tensor(result->ctx_static, result->outputs);
+            ggml_set_input(result->labels);
+            ggml_set_name(result->labels, "labels");
+            result->loss = ggml_sub(result->ctx_static, result->outputs, result->labels);
+            ggml_set_name(result->loss, "loss_error");
+            result->loss = ggml_sqr(result->ctx_static, result->loss);
+            ggml_set_name(result->loss, "loss_squared_error");
+            result->loss = ggml_sum(result->ctx_static, result->loss);
+            ggml_set_name(result->loss, "loss_sum_squared_error");
+            const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs));
+            result->loss = ggml_scale(result->ctx_static, result->loss, scale);
+            ggml_set_name(result->loss, "loss_mean_squared_error");
+            result->loss_per_datapoint = true;
+            break;
+        }
+    }
+    ggml_set_output(result->loss);
+    ggml_set_loss(result->loss);
+    ggml_build_forward_expand(result->gf, result->loss);
+
+    result->pred = ggml_argmax(result->ctx_static, result->outputs);
+    ggml_set_name(result->pred, "pred");
+    ggml_set_output(result->pred);
+    ggml_build_forward_expand(result->gf, result->pred);
+
+    if (result->labels) {
+        result->ncorrect = ggml_count_equal(result->ctx_static, result->pred, ggml_argmax(result->ctx_static, result->labels));
+        ggml_set_name(result->ncorrect, "ncorrect");
+        ggml_set_output(result->ncorrect);
+        ggml_build_forward_expand(result->gf, result->ncorrect);
+    } else {
+        result->ncorrect = nullptr;
+    }
+
+    if (params.build_type == GGML_OPT_BUILD_TYPE_FORWARD) {
+        result->gb_grad = nullptr;
+        result->gb_opt  = nullptr;
+
+        result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
+        result->buf_static_cpu = nullptr;
+
+        ggml_opt_alloc_graph(result, result->gf);
+
+        return result;
+    }
+
+    // gb_grad == graph backward gradients, forward pass, then backward pass to calculate gradients.
+    result->gb_grad = ggml_graph_dup(result->ctx_compute, result->gf);
+    ggml_build_backward_expand(result->ctx_static, result->ctx_compute, result->gb_grad, accumulate);
+
+    if (params.build_type == GGML_OPT_BUILD_TYPE_GRAD) {
+        result->gb_opt  = nullptr;
+
+        result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
+        result->buf_static_cpu = nullptr;
+
+        ggml_opt_alloc_graph(result, result->gb_grad);
+        ggml_graph_reset(result->gb_grad);
+
+        return result;
+    }
+
+    GGML_ASSERT(params.build_type == GGML_OPT_BUILD_TYPE_OPT);
+
+    // gb_opt == graph backward optimize, forward pass, then backward pass to calculate gradients, then optimizer step.
+    result->gb_opt = ggml_graph_dup(result->ctx_compute, result->gb_grad);
+
+    result->adamw_params = ggml_new_tensor_1d(result->ctx_static_cpu, GGML_TYPE_F32, 7);
+    ggml_set_input(result->adamw_params);
+    ggml_set_name(result->adamw_params, "adamw_params");
+
+    for (int i = result->gf->n_nodes-1; i >= 0; --i) {
+        struct ggml_tensor * node = result->gb_opt->nodes[i];
+        struct ggml_tensor * grad = ggml_graph_get_grad(result->gb_opt, node);
+
+        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
+            struct ggml_tensor * m        = ggml_dup_tensor(result->ctx_static, node);
+            struct ggml_tensor * v        = ggml_dup_tensor(result->ctx_static, node);
+            struct ggml_tensor * opt_step = ggml_opt_step_adamw(result->ctx_compute, node, grad, m, v, result->adamw_params);
+            ggml_build_forward_expand(result->gb_opt, opt_step);
+        }
+    }
+
+    result->buf_static = ggml_backend_alloc_ctx_tensors(
+        result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
+
+    result->buf_static_cpu = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx_static_cpu, ggml_backend_cpu_buffer_type());
+
+    ggml_opt_alloc_graph(result, result->gb_opt);
+    ggml_graph_reset(result->gb_opt);
+
+    return result;
+}
+
+void ggml_opt_free(ggml_opt_context_t opt_ctx) {
+    if (opt_ctx == nullptr) {
+        return;
+    }
+    ggml_backend_buffer_free(opt_ctx->buf_static);
+    ggml_backend_buffer_free(opt_ctx->buf_static_cpu);
+    ggml_free(opt_ctx->ctx_static);
+    ggml_free(opt_ctx->ctx_static_cpu);
+    delete opt_ctx;
+}
+
+void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer) {
+    if (optimizer) {
+        ggml_graph_reset(opt_ctx->gb_opt);
+        opt_ctx->iter = 1;
+    } else {
+        ggml_graph_reset(opt_ctx->gb_grad);
+    }
+}
+
+struct ggml_tensor * ggml_opt_inputs(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->inputs;
+}
+
+struct ggml_tensor * ggml_opt_outputs(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->outputs;
+}
+
+struct ggml_tensor * ggml_opt_labels(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->labels;
+}
+
+struct ggml_tensor * ggml_opt_loss(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->loss;
+}
+
+struct ggml_tensor * ggml_opt_pred(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->pred;
+}
+
+struct ggml_tensor * ggml_opt_ncorrect(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->ncorrect;
+}
+
+struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node) {
+    return ggml_graph_get_grad_acc(opt_ctx->gb_opt, node);
+}
+
+// ====== Optimization Result ======
+
+ggml_opt_result_t ggml_opt_result_init() {
+    return new ggml_opt_result;
+}
+
+void ggml_opt_result_free(ggml_opt_result_t result) {
+    delete result;
+}
+
+void ggml_opt_result_reset(ggml_opt_result_t result) {
+    result->ndata = 0;
+    result->loss.clear();
+    result->pred.clear();
+    result->ncorrect = 0;
+}
+
+void ggml_opt_result_ndata(ggml_opt_result_t result, int64_t * ndata) {
+    *ndata = result->ndata;
+}
+
+void ggml_opt_result_loss(ggml_opt_result_t result, double * loss, double * unc) {
+    const int64_t nbatches = result->loss.size(); // Number of physical batches.
+
+    if (nbatches == 0) {
+        *loss = 0.0;
+        *unc  = NAN;
+        return;
+    }
+
+    double sum         = 0.0;
+    double sum_squared = 0.0;
+
+    for (const float & loss : result->loss) {
+        // If the loss is per datapoint it was scaled by 1.0f/opt_period for each physical batch.
+        const float loss_scaled = result->loss_per_datapoint ? loss*result->opt_period : loss;
+        sum         += loss_scaled;
+        sum_squared += loss_scaled*loss_scaled;
+    }
+
+    const double mean = sum/nbatches;
+    *loss = result->loss_per_datapoint ? mean : sum;
+
+    if (!unc) {
+        return;
+    }
+
+    if (nbatches < 2) {
+        *unc = NAN;
+        return;
+    }
+
+    const double var_sum = sum_squared/nbatches - mean*mean; // variance without Bessel's correction, i.e. nbatches/(nbatches-1)
+    *unc = result->loss_per_datapoint ? sqrt(var_sum / (nbatches - 1)) : sqrt(var_sum * nbatches/(nbatches - 1));
+}
+
+void ggml_opt_result_pred(ggml_opt_result_t result, int32_t * pred) {
+    for (size_t i = 0; i < result->pred.size(); ++i) {
+        pred[i] = result->pred[i];
+    }
+}
+
+void ggml_opt_result_accuracy(ggml_opt_result_t result, double * accuracy, double * unc) {
+    *accuracy = result->ncorrect >= 0 ? double(result->ncorrect) / double(result->ndata) : NAN;
+
+    if (!unc) {
+        return;
+    }
+
+    *unc = result->ncorrect >= 0 && result->ndata >= 2 ?
+        sqrt((*accuracy) * (1.0 - (*accuracy)) / double(result->ndata - 1)) : NAN;
+}
+
+// ====== Computation ======
+
+static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph, ggml_opt_result * result) {
+    if (graph != opt_ctx->gf) {
+        struct ggml_opt_optimizer_params opt_pars = opt_ctx->get_opt_pars(opt_ctx->get_opt_pars_ud);
+
+        GGML_ASSERT(opt_pars.adamw.alpha >  0.0f);
+        GGML_ASSERT(opt_pars.adamw.beta1 >= 0.0f);
+        GGML_ASSERT(opt_pars.adamw.beta1 <= 1.0f);
+        GGML_ASSERT(opt_pars.adamw.beta2 >= 0.0f);
+        GGML_ASSERT(opt_pars.adamw.beta2 <= 1.0f);
+        GGML_ASSERT(opt_pars.adamw.eps   >= 0.0f);
+        GGML_ASSERT(opt_pars.adamw.wd    >= 0.0f);
+        GGML_ASSERT(opt_pars.adamw.wd    <= 1.0f);
+
+        // beta1, beta2 after applying warmup
+        const float beta1h = 1.0f/(1.0f - powf(opt_pars.adamw.beta1, opt_ctx->iter));
+        const float beta2h = 1.0f/(1.0f - powf(opt_pars.adamw.beta2, opt_ctx->iter));
+
+        float * adamw_par_data = ggml_get_data_f32(opt_ctx->adamw_params);
+        adamw_par_data[0] = opt_pars.adamw.alpha;
+        adamw_par_data[1] = opt_pars.adamw.beta1;
+        adamw_par_data[2] = opt_pars.adamw.beta2;
+        adamw_par_data[3] = opt_pars.adamw.eps;
+        adamw_par_data[4] = opt_pars.adamw.wd;
+        adamw_par_data[5] = beta1h;
+        adamw_par_data[6] = beta2h;
+    }
+
+    ggml_opt_alloc_graph(opt_ctx, graph);
+    ggml_backend_sched_graph_compute(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
+    opt_ctx->iter += opt_ctx->allocated_graph == opt_ctx->gb_opt;
+
+    if (!result) {
+        return;
+    }
+
+    if (result->ndata == 0) {
+        result->loss_per_datapoint = opt_ctx->loss_per_datapoint;
+        result->opt_period         = opt_ctx->opt_period;
+    } else {
+        GGML_ASSERT(result->loss_per_datapoint == opt_ctx->loss_per_datapoint);
+        GGML_ASSERT(result->opt_period         == opt_ctx->opt_period);
+    }
+
+    const int64_t ndata = opt_ctx->outputs->ne[1];
+    GGML_ASSERT(result->ndata == ndata*int64_t(result->loss.size()) && "varying batch size not supported");
+    result->ndata += ndata;
+
+    GGML_ASSERT(ggml_is_scalar(opt_ctx->loss));
+    GGML_ASSERT(opt_ctx->loss->type == GGML_TYPE_F32);
+    float loss;
+    ggml_backend_tensor_get(opt_ctx->loss, &loss, 0, ggml_nbytes(opt_ctx->loss));
+    result->loss.push_back(loss);
+
+    GGML_ASSERT(opt_ctx->pred->type == GGML_TYPE_I32);
+    std::vector<int32_t> pred(ndata);
+    ggml_backend_tensor_get(opt_ctx->pred, pred.data(), 0, ggml_nbytes(opt_ctx->pred));
+    result->pred.insert(result->pred.end(), pred.begin(), pred.end());
+
+    if (!opt_ctx->labels || result->ncorrect < 0) {
+        result->ncorrect = -1;
+        return;
+    }
+
+    GGML_ASSERT(ggml_is_scalar(opt_ctx->ncorrect));
+    GGML_ASSERT(opt_ctx->ncorrect->type == GGML_TYPE_I64);
+    int64_t ncorrect;
+    ggml_backend_tensor_get(opt_ctx->ncorrect, &ncorrect, 0, ggml_nbytes(opt_ctx->ncorrect));
+    result->ncorrect += ncorrect;
+}
+
+void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result * result) {
+    ggml_opt_eval_graph(opt_ctx, opt_ctx->gf, result);
+}
+
+void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result * result) {
+    if (opt_ctx->opt_period == 1) {
+        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_opt, result);
+        return;
+    }
+
+    const int32_t opt_i_next = (opt_ctx->opt_i + 1) % opt_ctx->opt_period;
+    if (opt_i_next == 0) {
+        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_opt, result);
+        ggml_opt_reset(opt_ctx, /*optimizer =*/ false);
+    } else {
+        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_grad, result);
+    }
+    opt_ctx->opt_i = opt_i_next;
+}
+
+// ====== High-Level Functions ======
+
+void ggml_opt_epoch(
+        ggml_opt_context_t      opt_ctx,
+        ggml_opt_dataset_t      dataset,
+        ggml_opt_result_t       result_train,
+        ggml_opt_result_t       result_eval,
+        int64_t                 idata_split,
+        ggml_opt_epoch_callback callback_train,
+        ggml_opt_epoch_callback callback_eval) {
+    struct ggml_tensor * inputs = ggml_opt_inputs(opt_ctx);
+    struct ggml_tensor * labels = ggml_opt_labels(opt_ctx);
+    struct ggml_tensor * data   = ggml_opt_dataset_data(dataset);
+    GGML_ASSERT(data->ne[0] == inputs->ne[0]);
+
+    const int64_t ndata       =   data->ne[1];
+    const int64_t ndata_batch = inputs->ne[1];
+
+    GGML_ASSERT(data->ne[1] % inputs->ne[1] == 0);
+    const int64_t nbatches = ndata/ndata_batch;
+
+    idata_split = idata_split < 0 ? ndata : idata_split;
+    GGML_ASSERT(idata_split % ndata_batch == 0);
+    const int64_t ibatch_split = idata_split / ndata_batch;
+
+    int64_t ibatch = 0;
+    int64_t t_loop_start = ggml_time_us();
+    for (; ibatch < ibatch_split; ++ibatch) {
+        ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch);
+        ggml_opt_forward_backward(opt_ctx, result_train);
+        if (callback_train) {
+            callback_train(true, opt_ctx, dataset, result_train, ibatch+1, ibatch_split, t_loop_start);
+        }
+    }
+    t_loop_start = ggml_time_us();
+    for (; ibatch < nbatches; ++ibatch) {
+        ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch);
+        ggml_opt_forward(opt_ctx, result_eval);
+        if (callback_eval) {
+            callback_eval(false, opt_ctx, dataset, result_eval, ibatch+1-ibatch_split, nbatches-ibatch_split, t_loop_start);
+        }
+    }
+}
+
+void ggml_opt_epoch_callback_progress_bar(
+        bool               train,
+        ggml_opt_context_t opt_ctx,
+        ggml_opt_dataset_t dataset,
+        ggml_opt_result_t  result,
+        int64_t            ibatch,
+        int64_t            ibatch_max,
+        int64_t            t_start_us) {
+    fprintf(stderr, "%s[", train ? "train: " : "val:   ");
+
+    constexpr int64_t bar_length = 25;
+    for (int64_t j = 0; j < bar_length; ++j) {
+        const int64_t ibatch_j = ibatch_max * j/bar_length;
+        if (ibatch_j < ibatch) {
+            fprintf(stderr, "=");
+        } else if (ibatch_max * (j - 1)/bar_length < ibatch) {
+            fprintf(stderr, ">");
+        } else {
+            fprintf(stderr, " ");
+        }
+    }
+
+    const int64_t batch_size = ggml_opt_inputs(opt_ctx)->ne[1];
+    const int64_t idata      = ibatch*batch_size;
+    const int64_t idata_max  = ibatch_max*batch_size;
+
+    double loss;
+    double loss_unc;
+    ggml_opt_result_loss(result, &loss, &loss_unc);
+
+    double accuracy;
+    double accuracy_unc;
+    ggml_opt_result_accuracy(result, &accuracy, &accuracy_unc);
+
+    const int64_t t_ibatch_us = ggml_time_us() - t_start_us;
+    int64_t t_ibatch_s = t_ibatch_us / 1000000;
+    const int64_t t_ibatch_h = t_ibatch_s / 3600;
+    t_ibatch_s -= t_ibatch_h * 3600;
+    const int64_t t_ibatch_m = t_ibatch_s / 60;
+    t_ibatch_s -= t_ibatch_m * 60;
+
+    const int64_t t_eta_us = t_ibatch_us * (ibatch_max - ibatch)/ibatch;
+    int64_t t_eta_s = t_eta_us / 1000000;
+    const int64_t t_eta_h = t_eta_s / 3600;
+    t_eta_s -= t_eta_h * 3600;
+    const int64_t t_eta_m = t_eta_s / 60;
+    t_eta_s -= t_eta_m * 60;
+
+    fprintf(stderr, "| data=%06" PRId64 "/%06" PRId64 ", loss=%.6lf+-%.6lf, accuracy=%.2lf+-%.2lf%%, "
+            "t=%02" PRId64 ":%02" PRId64 ":%02" PRId64 ", ETA=%02" PRId64 ":%02" PRId64 ":%02" PRId64 "]\r",
+            idata, idata_max, loss, loss_unc, 100.0*accuracy, 100.0*accuracy_unc,
+            t_ibatch_h, t_ibatch_m, t_ibatch_s, t_eta_h, t_eta_m, t_eta_s);
+    if (ibatch == ibatch_max) {
+        fprintf(stderr, "\n");
+    }
+    fflush(stderr);
+
+    GGML_UNUSED(dataset);
+}
+
+void ggml_opt_fit(
+        ggml_backend_sched_t            backend_sched,
+        ggml_context                  * ctx_compute,
+        ggml_tensor                   * inputs,
+        ggml_tensor                   * outputs,
+        ggml_opt_dataset_t              dataset,
+        enum ggml_opt_loss_type         loss_type,
+        ggml_opt_get_optimizer_params   get_opt_pars,
+        int64_t                         nepoch,
+        int64_t                         nbatch_logical,
+        float                           val_split,
+        bool                            silent) {
+    ggml_time_init();
+    const int64_t t_start_us = ggml_time_us();
+
+    const int64_t ndata           = ggml_opt_dataset_data(dataset)->ne[1];
+    const int64_t nbatch_physical = inputs->ne[1];
+    GGML_ASSERT(ndata          % nbatch_logical  == 0);
+    GGML_ASSERT(nbatch_logical % nbatch_physical == 0);
+
+    const int64_t opt_period       = nbatch_logical / nbatch_physical;
+    const int64_t nbatches_logical = ndata / nbatch_logical;
+
+    GGML_ASSERT(val_split >= 0.0f);
+    GGML_ASSERT(val_split <  1.0f);
+    const int64_t ibatch_split = int64_t(((1.0f - val_split) * nbatches_logical)) * opt_period; // train <-> val split index (physical)
+    const int64_t idata_split  = ibatch_split * nbatch_physical;
+
+    int64_t epoch = 1;
+
+    ggml_opt_params params = ggml_opt_default_params(backend_sched, ctx_compute, inputs, outputs, loss_type);
+    params.opt_period      = opt_period;
+    params.get_opt_pars    = get_opt_pars;
+    params.get_opt_pars_ud = &epoch;
+    ggml_opt_context_t opt_ctx = ggml_opt_init(params);
+
+    // Shuffling the data is generally useful but there is only a point if not all data is used in a single batch.
+    if (nbatch_logical < ndata) {
+        ggml_opt_dataset_shuffle(opt_ctx, dataset, -1); // Shuffle all data (train + validation).
+    }
+
+    ggml_opt_result_t result_train = ggml_opt_result_init();
+    ggml_opt_result_t result_val   = ggml_opt_result_init();
+
+    ggml_opt_epoch_callback epoch_callback = silent ? nullptr : ggml_opt_epoch_callback_progress_bar;
+
+    for (; epoch <= nepoch; ++epoch) {
+        if (nbatch_logical < idata_split) {
+            ggml_opt_dataset_shuffle(opt_ctx, dataset, idata_split);
+        }
+
+        ggml_opt_result_reset(result_train);
+        ggml_opt_result_reset(result_val);
+
+        if (!silent) {
+            fprintf(stderr, "%s: epoch %04" PRId64 "/%04" PRId64 ":\n", __func__, epoch, nepoch);
+        }
+        ggml_opt_epoch(opt_ctx, dataset, result_train, result_val, idata_split, epoch_callback, epoch_callback);
+        if (!silent) {
+            fprintf(stderr, "\n");
+        }
+    }
+
+    if (!silent) {
+        int64_t t_total_s = (ggml_time_us() - t_start_us) / 1000000;
+        const int64_t t_total_h = t_total_s / 3600;
+        t_total_s -= t_total_h * 3600;
+        const int64_t t_total_m = t_total_s / 60;
+        t_total_s -= t_total_m * 60;
+        fprintf(stderr, "%s: training took %02" PRId64 ":%02" PRId64 ":%02" PRId64 "\n", __func__, t_total_h, t_total_m, t_total_s);
+    }
+
+    ggml_opt_free(opt_ctx);
+    ggml_opt_result_free(result_train);
+    ggml_opt_result_free(result_val);
+}

tests/test-backend-ops.cpp

@@ -811,11 +811,11 @@ struct test_case {
 
         ggml_build_forward_expand(gf, out);
         ggml_graph_cpy(gf, gb);
-        ggml_build_backward_expand(ctx, gf, gb, false);
+        ggml_build_backward_expand(ctx, ctx, gb, false);
         if (expect.size() != 1 || expect[0] != 0.0f) {
             GGML_ASSERT(ggml_graph_n_nodes(gb) > ggml_graph_n_nodes(gf));
             for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-                GGML_ASSERT(!(t->flags & GGML_TENSOR_FLAG_PARAM) || t->grad->op != GGML_OP_NONE);
+                GGML_ASSERT(!(t->flags & GGML_TENSOR_FLAG_PARAM) || ggml_graph_get_grad(gb, t)->op != GGML_OP_NONE);
             }
         }
 
@@ -862,7 +862,13 @@ struct test_case {
             const char * bn = ggml_backend_name(backend);
             const int64_t ne = ggml_nelements(t);
 
-            std::vector<float> ga = tensor_to_float(t->grad);
+            std::vector<float> ga;
+            struct ggml_tensor * grad = ggml_graph_get_grad(gb, t);
+            if (grad) {
+                ga = tensor_to_float(grad);
+            } else {
+                ga.resize(ne); // default value is 0.0f
+            }
 
             for (int64_t i = 0; i < ne; ++i) { // gradient algebraic
                 // check for nans
@@ -2500,6 +2506,35 @@ struct test_sum_rows : public test_case {
     }
 };
 
+// GGML_OP_MEAN
+struct test_mean : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+
+    std::string vars() override {
+        return VARS_TO_STR2(type, ne);
+    }
+
+    test_mean(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 5, 4, 3})
+        : type(type), ne(ne) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_set_param(ctx, a);
+        ggml_set_name(a, "a");
+
+        ggml_tensor * out = ggml_mean(ctx, a);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+
+    float grad_eps() override {
+        return 0.1f * ne[0]*ne[1]*ne[2]*ne[3];
+    }
+};
+
 // GGML_OP_UPSCALE
 struct test_upscale : public test_case {
     const ggml_type type;
@@ -2834,24 +2869,14 @@ struct test_cross_entropy_loss : public test_case {
 struct test_opt_step_adamw : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
-    const float alpha;
-    const float beta1;
-    const float beta2;
-    const float eps;
-    const float wd;
 
     std::string vars() override {
-        return VARS_TO_STR7(type, ne, alpha, beta1, beta2, eps, wd);
+        return VARS_TO_STR2(type, ne);
     }
 
     test_opt_step_adamw(ggml_type type = GGML_TYPE_F32,
-            std::array<int64_t, 4> ne = {10, 5, 4, 3},
+            std::array<int64_t, 4> ne = {10, 5, 4, 3})
-            float alpha = 1e-3f,
+        : type(type), ne(ne) {}
-            float beta1 = 0.9f,
-            float beta2 = 0.999f,
-            float eps = 1e-8f,
-            float wd = 0.0f)
-        : type(type), ne(ne), alpha(alpha), beta1(beta1), beta2(beta2), eps(eps), wd(wd) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
@@ -2861,7 +2886,16 @@ struct test_opt_step_adamw : public test_case {
         ggml_tensor * grad = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
         ggml_set_name(grad, "grad");
 
-        ggml_tensor * out = ggml_opt_step_adamw(ctx, a, grad, alpha, beta1, beta2, eps, wd);
+        ggml_tensor * grad_m = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
+        ggml_set_name(grad_m, "grad_m");
+
+        ggml_tensor * grad_v = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
+        ggml_set_name(grad_v, "grad_v");
+
+        ggml_tensor * adamw_params = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 7);
+        ggml_set_name(adamw_params, "adamw_params");
+
+        ggml_tensor * out = ggml_opt_step_adamw(ctx, a, grad, grad_m, grad_v, adamw_params);
         ggml_set_name(out, "out");
 
         return out;
@@ -2869,7 +2903,7 @@ struct test_opt_step_adamw : public test_case {
 
     void initialize_tensors(ggml_context * ctx) override {
         for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-            init_tensor_uniform(t, 0.0f, 1.0f); // grad_v needs non-negative values.
+            init_tensor_uniform(t, 0.0f, 1.0f); // grad_v and adamw_params need non-negative values.
         }
     }
 
@@ -3735,6 +3769,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 
     test_cases.emplace_back(new test_sum());
     test_cases.emplace_back(new test_sum_rows());
+    test_cases.emplace_back(new test_mean());
     test_cases.emplace_back(new test_upscale());
     test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, { 512, 512, 3, 1 }, 2, true));
     test_cases.emplace_back(new test_upscale_ext());
@@ -3766,9 +3801,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     }
 
     test_cases.emplace_back(new test_cross_entropy_loss());
-    for (float wd : {0.0f, 1e-2f}) {
+    test_cases.emplace_back(new test_opt_step_adamw(GGML_TYPE_F32, {10, 5, 4, 3}));
-        test_cases.emplace_back(new test_opt_step_adamw(GGML_TYPE_F32, {10, 5, 4, 3}, 1.0f, 1e-3f, 0.9f, 0.999f, wd));
-    }
 
     // these tests are disabled to save execution time, but they can be handy for debugging
 #if 0
@@ -3938,6 +3971,8 @@ int main(int argc, char ** argv) {
         ggml_backend_free(backend);
     }
 
+    ggml_quantize_free();
+
     printf("%zu/%zu backends passed\n", n_ok, ggml_backend_dev_count());
 
     if (n_ok != ggml_backend_dev_count()) {
@@ -3945,8 +3980,6 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    ggml_quantize_free();
-
     printf("\033[1;32mOK\033[0m\n");
     return 0;
 }