Felipe97/llama-cpp-compiled

1#pragma once2 3// this is a staging header for new llama.cpp API4// breaking changes and C++ are allowed. everything here should be considered WIP5// try as much as possible to not include this header in the rest of the codebase6 7#include "llama.h"8 9#include <cstdint>10#include <map>11 12// Reserve a new compute graph. It is valid until the next call to llama_graph_reserve.13LLAMA_API struct ggml_cgraph * llama_graph_reserve(14        struct llama_context * ctx,15        uint32_t n_tokens,16        uint32_t n_seqs,17        uint32_t n_outputs);18 19// Get the default ggml_type for a given ftype.20LLAMA_API ggml_type llama_ftype_get_default_type(llama_ftype ftype);21 22struct quantize_state_impl;23 24LLAMA_API quantize_state_impl * llama_quant_init(25        const llama_model * model,26        const llama_model_quantize_params * params);27 28LLAMA_API void llama_quant_free(quantize_state_impl * qs);29 30// Descriptor for constructing a mock model for quantization testing.31struct llama_quant_model_desc {32    const char * architecture;33    uint32_t n_embd;34    uint32_t n_ff;35    uint32_t n_layer;36    uint32_t n_head;37    uint32_t n_head_kv;38    uint32_t n_expert;39    uint32_t n_embd_head_k;40    uint32_t n_embd_head_v;41};42 43// Create a mock model from a metadata descriptor (for testing).44// The returned model must be freed with llama_model_free().45LLAMA_API llama_model * llama_quant_model_from_metadata(const llama_quant_model_desc * desc);46 47// Returns true if this tensor should be quantized (based on name, dims, params).48LLAMA_API bool llama_quant_tensor_allows_quantization(49        const quantize_state_impl * qs,50        const ggml_tensor * tensor);51 52// Compute quantization type assignments for a list of tensors.53// All tensors should be quantizable (use llama_quant_tensor_allows_quantization to filter).54// result_types: caller-allocated array of n_tensors elements, filled with assigned types.55LLAMA_API void llama_quant_compute_types(56        quantize_state_impl * qs,57        llama_ftype ftype,58        ggml_tensor ** tensors,59        ggml_type * result_types,60        size_t n_tensors);61 62//63// device memory querying64//65 66// "memory" as in physical memory for a buffer type, in bytes67struct llama_memory_breakdown_data {68    size_t model   = 0; // memory allocated for the model69    size_t context = 0; // memory allocated for the context70    size_t compute = 0; // memory allocated for temporary compute buffers71 72    size_t total() const {73        return model + context + compute;74    }75};76 77struct llama_device_memory_data {78    int64_t total;79    int64_t free;80    llama_memory_breakdown_data mb;81};82 83// TODO: convert to C-style data structure84using llama_memory_breakdown = std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data>;85 86LLAMA_API int32_t llama_model_n_expert (const struct llama_model * model);87LLAMA_API int32_t llama_model_n_devices(const struct llama_model * model);88 89LLAMA_API ggml_backend_dev_t llama_model_get_device(const struct llama_model * model, int i);90 91LLAMA_API llama_memory_breakdown llama_get_memory_breakdown(const struct llama_context * ctx);92 93// Set whether the context outputs nextn embeddings or not94// If masked == true,  output the embeddings only for the tokens with batch.logits != 095// If masked == false, output the embeddings for all tokens in the batch regardless of batch.logits96LLAMA_API void llama_set_embeddings_nextn(struct llama_context * ctx, bool value, bool masked);97 98// Select which appended NextN block the DECODER_MTP graph runs (offset past99// the trunk: il = n_layer() + offset). Used by the speculative NextN driver to100// chain multiple trained NextN heads. Default 0 (first head).101LLAMA_API void llama_set_nextn_layer_offset(struct llama_context * ctx, int32_t offset);102 103// mirrors:104// LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);105LLAMA_API float * llama_get_embeddings_nextn(struct llama_context * ctx);106 107// LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i);108LLAMA_API float * llama_get_embeddings_nextn_ith(struct llama_context * ctx, int32_t i);109 110// Set whether the context outputs the input embeddings of a specific layer111LLAMA_API void llama_set_embeddings_layer_inp(struct llama_context * ctx, uint32_t lid, bool value);112 113// mirrors:114// LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);115LLAMA_API float * llama_get_embeddings_layer_inp(struct llama_context * ctx, uint32_t lid);116 117LLAMA_API llama_context * llama_get_ctx_other(struct llama_context * ctx);118 119//120// model/context data extraction121//122 123LLAMA_API int32_t llama_model_dflash_selector_top_k(const struct llama_model * model);124 125// returns pointer to the target-model layer indices126LLAMA_API const int32_t * llama_model_target_layer_ids  (const struct llama_model * model);127// returns the number of extracted layers from target model128LLAMA_API uint32_t        llama_model_target_layer_ids_n(const struct llama_model * model);129 130// retrieves the whole token embedding matrix in F32 format (n_embd * n_vocab)131// returns total number of elements or 0 on error132// if out is nullptr, returns the number of tokens without writing to out133// caller must allocate enough memory for out before calling134LLAMA_API uint32_t llama_model_get_tok_embd(const struct llama_model * model, float * out);135