@novastera-oss/llamarn 0.4.0 → 0.4.3-beta4

package/ios/libs/llama.xcframework/macos-arm64_x86_64/llama.framework/Versions/A/Headers/ggml-metal.h

@@ -1,66 +0,0 @@
-// Note: this description is outdated
-//
-// An interface allowing to compute ggml_cgraph with Metal
-//
-// This is a fully functional interface that extends ggml with GPU support for Apple devices.
-// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, etc.)
-//
-// How it works?
-//
-// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
-// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
-// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
-//
-// You only need to make sure that all memory buffers that you used during the graph creation
-// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
-// used during the graph evaluation to determine the arguments of the compute kernels.
-//
-// Synchronization between device and host memory (for example for input and output tensors)
-// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
-//
-
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#include <stddef.h>
-#include <stdbool.h>
-
-struct ggml_tensor;
-struct ggml_cgraph;
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-//
-// backend API
-// user-code should use only these functions
-//
-
-GGML_BACKEND_API ggml_backend_t ggml_backend_metal_init(void);
-
-GGML_BACKEND_API bool ggml_backend_is_metal(ggml_backend_t backend);
-
-GGML_DEPRECATED(
-        GGML_BACKEND_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
-        "obsoleted by the new device interface - https://github.com/ggml-org/llama.cpp/pull/9713");
-
-GGML_BACKEND_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
-
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
-
-// helper to check if the device supports a specific family
-// ideally, the user code should be doing these checks
-// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
-GGML_BACKEND_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
-
-// capture all command buffers committed the next time `ggml_backend_graph_compute` is called
-GGML_BACKEND_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
-
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_metal_reg(void);
-
-#ifdef __cplusplus
-}
-#endif

package/ios/libs/llama.xcframework/macos-arm64_x86_64/llama.framework/Versions/A/Headers/ggml-opt.h

@@ -1,256 +0,0 @@
-// 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(
-            enum ggml_type type_data,    // the type for the internal data tensor
-            enum ggml_type type_label,   // the type for the internal labels tensor
-            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 int64_t              ggml_opt_dataset_ndata (ggml_opt_dataset_t dataset);
-    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);
-    GGML_API void ggml_opt_dataset_get_batch_host(
-            ggml_opt_dataset_t   dataset,
-            void               * data_batch,
-            size_t               nb_data_batch,
-            void               * labels_batch,
-            int64_t              ibatch);
-
-    // ====== Model / Context ======
-
-    enum ggml_opt_build_type {
-        GGML_OPT_BUILD_TYPE_FORWARD = 10,
-        GGML_OPT_BUILD_TYPE_GRAD    = 20,
-        GGML_OPT_BUILD_TYPE_OPT     = 30,
-    };
-
-    enum ggml_opt_optimizer_type {
-        GGML_OPT_OPTIMIZER_TYPE_ADAMW,
-        GGML_OPT_OPTIMIZER_TYPE_SGD,
-
-        GGML_OPT_OPTIMIZER_TYPE_COUNT
-    };
-
-    // parameters that control which optimizer is used and how said optimizer tries to find the minimal loss
-    struct ggml_opt_optimizer_params {
-        struct {
-            float alpha; // learning rate
-            float beta1; // first AdamW momentum
-            float beta2; // second AdamW momentum
-            float eps;   // epsilon for numerical stability
-            float wd;    // weight decay - 0.0f to disable
-        } adamw;
-        struct {
-            float alpha; // learning rate
-            float wd;    // weight decay
-        } sgd;
-    };
-
-    // 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, hard-coded values)
-    // userdata is not used
-    GGML_API struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata);
-
-    // casts userdata to ggml_opt_optimizer_params and returns it
-    GGML_API struct ggml_opt_optimizer_params ggml_opt_get_constant_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
-
-        // by default the forward graph needs to be reconstructed for each eval
-        // if ctx_compute, inputs, and outputs are set the graphs are instead allocated statically
-        struct ggml_context * ctx_compute;
-        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
-
-        // only GGML_OPT_OPTIMIZER_TYPE_ADAMW needs m, v momenta per parameter tensor
-        enum ggml_opt_optimizer_type optimizer;
-    };
-
-    // 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 struct ggml_opt_params ggml_opt_default_params(
-            ggml_backend_sched_t    backend_sched,
-            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);
-
-    GGML_API bool ggml_opt_static_graphs(ggml_opt_context_t opt_ctx); // whether the graphs are allocated_statically
-
-    // get underlying tensors that store data
-    // if not using static graphs these pointers become invalid with the next call to ggml_opt_alloc
-    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
-
-    // get the gradient accumulator for a node from the forward graph
-    GGML_API struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node);
-
-    GGML_API enum ggml_opt_optimizer_type ggml_opt_context_optimizer_type(ggml_opt_context_t); //TODO consistent naming scheme
-
-    GGML_API const char * ggml_opt_optimizer_name(enum ggml_opt_optimizer_type);
-
-    // ====== Optimization Result ======
-
-    GGML_API ggml_opt_result_t ggml_opt_result_init(void);
-    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 ======
-
-    // if not using static graphs, this function must be called prior to ggml_opt_alloc
-    GGML_API void ggml_opt_prepare_alloc(
-        ggml_opt_context_t    opt_ctx,
-        struct ggml_context * ctx_compute,
-        struct ggml_cgraph  * gf,
-        struct ggml_tensor  * inputs,
-        struct ggml_tensor  * outputs);
-
-    // allocate the next graph for evaluation, either forward or forward + backward
-    // must be called exactly once prior to calling ggml_opt_eval
-    GGML_API void ggml_opt_alloc(ggml_opt_context_t opt_ctx, bool backward);
-
-    // do forward pass, increment result if not NULL, do backward pass if allocated
-    GGML_API void ggml_opt_eval(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
-            struct ggml_context           * ctx_compute,    // context with temporarily allocated tensors to calculate the outputs
-            struct ggml_tensor            * inputs,         // input tensor with shape [ne_datapoint, ndata_batch]
-            struct 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
-            enum ggml_opt_optimizer_type    optimizer,      // sgd or adamw
-            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