|
| 1 | +#pragma once |
| 2 | + |
| 3 | +#include <chrono> |
| 4 | +#include <cstdint> |
| 5 | +#include <cstdio> |
| 6 | +#include <cstring> |
| 7 | +#include <ctime> |
| 8 | +#include <fstream> |
| 9 | +#include <memory> |
| 10 | +#include <sstream> |
| 11 | +#include <streambuf> |
| 12 | +#include <string> |
| 13 | +#include <thread> |
| 14 | +#include <vector> |
| 15 | + |
| 16 | +// header-only utilities to showcase how to directly load a model from memory |
| 17 | +#include "uint8-buff-stream-wrapper.h" |
| 18 | + |
| 19 | +namespace { |
| 20 | +bool is_split_file(const char * const model_path) { |
| 21 | + if (!model_path) { |
| 22 | + fprintf(stderr, "No model file provided\n"); |
| 23 | + exit(EXIT_FAILURE); |
| 24 | + } |
| 25 | + |
| 26 | + std::string path(model_path); |
| 27 | + return path.find("-of-") != std::string::npos; |
| 28 | +} |
| 29 | + |
| 30 | +std::vector<uint8_t> load_file_into_buffer(const char * const model_path) { |
| 31 | + std::ifstream file_stream(model_path, std::ios::binary | std::ios::ate); |
| 32 | + if (!file_stream) { |
| 33 | + fprintf(stderr, "Failed to open file %s for reading into streambuf\n", model_path); |
| 34 | + exit(EXIT_FAILURE); |
| 35 | + } |
| 36 | + |
| 37 | + const size_t file_size = file_stream.tellg(); |
| 38 | + file_stream.seekg(0, std::ios::beg); |
| 39 | + |
| 40 | + static_assert(sizeof(std::uint8_t) == sizeof(char), "uint8_t must be same size as char"); |
| 41 | + std::vector<std::uint8_t> buffer(file_size); |
| 42 | + if (!file_stream.read((char *) buffer.data(), file_size)) { |
| 43 | + fprintf(stderr, "Failed to read entire file into buffer\n"); |
| 44 | + exit(EXIT_FAILURE); |
| 45 | + } |
| 46 | + |
| 47 | + return buffer; |
| 48 | +} |
| 49 | + |
| 50 | +std::unique_ptr<std::basic_streambuf<uint8_t>> load_file_into_streambuf(const char * const model_path) { |
| 51 | + return std::make_unique<Uint8BufferStreamBuf>(load_file_into_buffer(model_path)); |
| 52 | +} |
| 53 | + |
| 54 | +struct file_entry { |
| 55 | + std::string path; |
| 56 | + std::unique_ptr<std::basic_streambuf<uint8_t>> streambuf; |
| 57 | +}; |
| 58 | + |
| 59 | +std::vector<file_entry> load_files_into_streambuf(const char * const model_path) { |
| 60 | + std::vector<file_entry> files; |
| 61 | + |
| 62 | + // Extract pattern from first file path |
| 63 | + std::string path(model_path); |
| 64 | + |
| 65 | + // Split by '-' |
| 66 | + std::vector<std::string> parts; |
| 67 | + std::stringstream ss(path); |
| 68 | + std::string item; |
| 69 | + while (std::getline(ss, item, '-')) { |
| 70 | + parts.push_back(item); |
| 71 | + } |
| 72 | + |
| 73 | + // Split the last part by '.' |
| 74 | + std::string last_part = parts.back(); |
| 75 | + parts.pop_back(); |
| 76 | + size_t dot_pos = last_part.find('.'); |
| 77 | + if (dot_pos != std::string::npos) { |
| 78 | + parts.push_back(last_part.substr(0, dot_pos)); |
| 79 | + parts.push_back(last_part.substr(dot_pos + 1)); // extension |
| 80 | + } else { |
| 81 | + parts.push_back(last_part); |
| 82 | + } |
| 83 | + |
| 84 | + // Check if we have enough parts |
| 85 | + if (parts.size() < 4) { |
| 86 | + fprintf(stderr, "Model path does not contain expected pattern\n"); |
| 87 | + exit(EXIT_FAILURE); |
| 88 | + } |
| 89 | + |
| 90 | + // Get total files from [-2] position (before the extension) |
| 91 | + int total_files = std::stoi(parts[parts.size() - 2]); |
| 92 | + |
| 93 | + // Get base path by joining all parts except -start-of-end.gguf |
| 94 | + std::string base_path; |
| 95 | + for (size_t i = 0; i < parts.size() - 4; i++) { |
| 96 | + if (i > 0) { |
| 97 | + base_path += "-"; |
| 98 | + } |
| 99 | + base_path += parts[i]; |
| 100 | + } |
| 101 | + |
| 102 | + for (int i = 1; i <= total_files; i++) { |
| 103 | + char numbered_path[1024]; |
| 104 | + snprintf(numbered_path, sizeof(numbered_path), "%s-%05d-of-%05d.gguf", base_path.c_str(), i, total_files); |
| 105 | + |
| 106 | + files.push_back({ numbered_path, load_file_into_streambuf(numbered_path) }); |
| 107 | + } |
| 108 | + |
| 109 | + return files; |
| 110 | +} |
| 111 | + |
| 112 | +file_entry load_tensor_list_file(const char * const model_path) { |
| 113 | + std::string path(model_path); |
| 114 | + |
| 115 | + // Split by '-' |
| 116 | + std::vector<std::string> parts; |
| 117 | + std::stringstream ss(path); |
| 118 | + std::string item; |
| 119 | + while (std::getline(ss, item, '-')) { |
| 120 | + parts.push_back(item); |
| 121 | + } |
| 122 | + |
| 123 | + // Split the last part by '.' |
| 124 | + std::string last_part = parts.back(); |
| 125 | + parts.pop_back(); |
| 126 | + size_t dot_pos = last_part.find('.'); |
| 127 | + if (dot_pos != std::string::npos) { |
| 128 | + parts.push_back(last_part.substr(0, dot_pos)); |
| 129 | + parts.push_back(last_part.substr(dot_pos + 1)); // extension |
| 130 | + } else { |
| 131 | + parts.push_back(last_part); |
| 132 | + } |
| 133 | + |
| 134 | + // Check if we have enough parts |
| 135 | + if (parts.size() < 4) { |
| 136 | + fprintf(stderr, "Model path does not contain expected pattern\n"); |
| 137 | + exit(EXIT_FAILURE); |
| 138 | + } |
| 139 | + |
| 140 | + // Get base path by joining all parts except -start-of-end.gguf |
| 141 | + std::string base_path; |
| 142 | + for (size_t i = 0; i < parts.size() - 4; i++) { |
| 143 | + if (i > 0) { |
| 144 | + base_path += "-"; |
| 145 | + } |
| 146 | + base_path += parts[i]; |
| 147 | + } |
| 148 | + |
| 149 | + // Construct tensor list file path |
| 150 | + std::string tensor_list_path = base_path + ".tensors.txt"; |
| 151 | + |
| 152 | + printf("Loading tensor list file: %s\n", tensor_list_path.c_str()); |
| 153 | + return { tensor_list_path, load_file_into_streambuf(tensor_list_path.c_str()) }; |
| 154 | +} |
| 155 | + |
| 156 | +llama_model * load_model_from_memory_configuration(const char * model_path, llama_model_params & model_params) { |
| 157 | + llama_model * model; |
| 158 | + std::chrono::steady_clock::time_point load_start_time; |
| 159 | + if (getenv("LLAMA_EXAMPLE_MEMORY_BUFFER")) { |
| 160 | + std::vector<uint8_t> buffer = load_file_into_buffer(model_path); |
| 161 | + fprintf(stdout, "%s: loading model from memory buffer\n", __func__); |
| 162 | + load_start_time = std::chrono::steady_clock::now(); |
| 163 | + model = llama_model_load_from_buffer(std::move(buffer), model_params); |
| 164 | + } else if (getenv("LLAMA_EXAMPLE_MEMORY_BUFFER_SPLIT")) { |
| 165 | + file_entry tensor_list_file = load_tensor_list_file(model_path); |
| 166 | + std::vector<file_entry> files = load_files_into_streambuf(model_path); |
| 167 | + fprintf(stdout, "%s: loading model from %zu file streambufs\n", __func__, files.size()); |
| 168 | + |
| 169 | + std::vector<const char *> file_paths; |
| 170 | + for (const auto & file : files) { |
| 171 | + printf("Found file %s with streambuf\n", file.path.c_str()); |
| 172 | + file_paths.push_back(file.path.c_str()); |
| 173 | + } |
| 174 | + |
| 175 | + load_start_time = std::chrono::steady_clock::now(); |
| 176 | + const char * async_load_context = "test-model-load"; |
| 177 | + std::thread fulfill_thread([&files, &tensor_list_file, &async_load_context]() { |
| 178 | + const bool success = llama_model_load_fulfill_split_future( |
| 179 | + tensor_list_file.path.c_str(), async_load_context, std::move(tensor_list_file.streambuf)); |
| 180 | + printf("Fulfilling tensor list file %s: %s\n", tensor_list_file.path.c_str(), |
| 181 | + success ? "success" : "failure"); |
| 182 | + if (!success) { |
| 183 | + exit(EXIT_FAILURE); |
| 184 | + } |
| 185 | + |
| 186 | + for (auto & file : files) { |
| 187 | + const bool success = llama_model_load_fulfill_split_future(file.path.c_str(), async_load_context, |
| 188 | + std::move(file.streambuf)); |
| 189 | + printf("Fulfilling file %s with streambuf: %s\n", file.path.c_str(), success ? "success" : "failure"); |
| 190 | + if (!success) { |
| 191 | + exit(EXIT_FAILURE); |
| 192 | + } |
| 193 | + } |
| 194 | + }); |
| 195 | + fprintf(stderr, "Loading model from splits\n"); |
| 196 | + model = llama_model_load_from_split_futures(file_paths.data(), file_paths.size(), async_load_context, |
| 197 | + tensor_list_file.path.c_str(), model_params); |
| 198 | + fulfill_thread.join(); |
| 199 | + } else if (getenv("LLAMA_EXAMPLE_FROM_FILE")) { |
| 200 | + load_start_time = std::chrono::steady_clock::now(); |
| 201 | + model = llama_model_load_from_file(model_path, model_params); |
| 202 | + } else { |
| 203 | + return nullptr; |
| 204 | + } |
| 205 | + |
| 206 | + if (model == NULL) { |
| 207 | + fprintf(stderr, "%s: error: unable to load model\n", __func__); |
| 208 | + exit(1); |
| 209 | + } |
| 210 | + std::chrono::steady_clock::time_point load_end_time = std::chrono::steady_clock::now(); |
| 211 | + std::chrono::duration<double> load_duration = load_end_time - load_start_time; |
| 212 | + fprintf(stdout, "%s: loading model took %f seconds\n", __func__, load_duration.count()); |
| 213 | + return model; |
| 214 | +} |
| 215 | + |
| 216 | +bool memory_configuration_env_is_set() { |
| 217 | + return getenv("LLAMA_EXAMPLE_MEMORY_BUFFER") || getenv("LLAMA_EXAMPLE_MEMORY_BUFFER_SPLIT") || |
| 218 | + getenv("LLAMA_EXAMPLE_FROM_FILE"); |
| 219 | +} |
| 220 | +} // namespace |
0 commit comments