| 1 | // thread safety test |
| 2 | // - Loads a copy of the same model on each GPU, plus a copy on the CPU |
| 3 | // - Creates n_parallel (--parallel) contexts per model |
| 4 | // - Runs inference in parallel on each context |
| 5 | |
| 6 | #include <array> |
| 7 | #include <thread> |
| 8 | #include <vector> |
| 9 | #include <atomic> |
| 10 | #include "llama.h" |
| 11 | #include "arg.h" |
| 12 | #include "common.h" |
| 13 | #include "log.h" |
| 14 | #include "sampling.h" |
| 15 | |
| 16 | int main(int argc, char ** argv) { |
| 17 | common_params params; |
| 18 | |
| 19 | if (!common_params_parse(argc, argv, params, ex: LLAMA_EXAMPLE_COMMON)) { |
| 20 | return 1; |
| 21 | } |
| 22 | |
| 23 | common_init(); |
| 24 | |
| 25 | llama_backend_init(); |
| 26 | llama_numa_init(numa: params.numa); |
| 27 | |
| 28 | LOG_INF("%s\n" , common_params_get_system_info(params).c_str()); |
| 29 | |
| 30 | //llama_log_set([](ggml_log_level level, const char * text, void * /*user_data*/) { |
| 31 | // if (level == GGML_LOG_LEVEL_ERROR) { |
| 32 | // common_log_add(common_log_main(), level, "%s", text); |
| 33 | // } |
| 34 | //}, NULL); |
| 35 | |
| 36 | auto cparams = common_context_params_to_llama(params); |
| 37 | |
| 38 | // each context has a single sequence |
| 39 | cparams.n_seq_max = 1; |
| 40 | |
| 41 | int dev_count = ggml_backend_dev_count(); |
| 42 | std::vector<std::array<ggml_backend_dev_t, 2>> gpus; |
| 43 | for (int i = 0; i < dev_count; ++i) { |
| 44 | auto * dev = ggml_backend_dev_get(index: i); |
| 45 | if (dev && ggml_backend_dev_type(device: dev) == GGML_BACKEND_DEVICE_TYPE_GPU) { |
| 46 | gpus.push_back(x: {dev, nullptr}); |
| 47 | } |
| 48 | } |
| 49 | const int gpu_dev_count = (int)gpus.size(); |
| 50 | const int num_models = gpu_dev_count + 1 + 1; // GPUs + 1 CPU model + 1 layer split |
| 51 | //const int num_models = std::max(1, gpu_dev_count); |
| 52 | const int num_contexts = std::max(a: 1, b: params.n_parallel); |
| 53 | |
| 54 | std::vector<llama_model_ptr> models; |
| 55 | std::vector<std::thread> threads; |
| 56 | std::atomic<bool> failed = false; |
| 57 | |
| 58 | for (int m = 0; m < num_models; ++m) { |
| 59 | auto mparams = common_model_params_to_llama(params); |
| 60 | |
| 61 | if (m < gpu_dev_count) { |
| 62 | mparams.split_mode = LLAMA_SPLIT_MODE_NONE; |
| 63 | mparams.devices = gpus[m].data(); |
| 64 | } else if (m == gpu_dev_count) { |
| 65 | mparams.split_mode = LLAMA_SPLIT_MODE_NONE; |
| 66 | mparams.main_gpu = -1; // CPU model |
| 67 | } else { |
| 68 | mparams.split_mode = LLAMA_SPLIT_MODE_LAYER; |
| 69 | } |
| 70 | |
| 71 | llama_model * model = llama_model_load_from_file(path_model: params.model.path.c_str(), params: mparams); |
| 72 | if (model == NULL) { |
| 73 | LOG_ERR("%s: failed to load model '%s'\n" , __func__, params.model.path.c_str()); |
| 74 | return 1; |
| 75 | } |
| 76 | |
| 77 | models.emplace_back(args&: model); |
| 78 | } |
| 79 | |
| 80 | for (int m = 0; m < num_models; ++m) { |
| 81 | auto * model = models[m].get(); |
| 82 | for (int c = 0; c < num_contexts; ++c) { |
| 83 | threads.emplace_back(args: [&, m, c, model]() { |
| 84 | LOG_INF("Creating context %d/%d for model %d/%d\n" , c + 1, num_contexts, m + 1, num_models); |
| 85 | |
| 86 | llama_context_ptr ctx { llama_init_from_model(model, params: cparams) }; |
| 87 | if (ctx == NULL) { |
| 88 | LOG_ERR("failed to create context\n" ); |
| 89 | failed.store(i: true); |
| 90 | return; |
| 91 | } |
| 92 | |
| 93 | std::unique_ptr<common_sampler, decltype(&common_sampler_free)> sampler { common_sampler_init(model, params: params.sampling), common_sampler_free }; |
| 94 | if (sampler == NULL) { |
| 95 | LOG_ERR("failed to create sampler\n" ); |
| 96 | failed.store(i: true); |
| 97 | return; |
| 98 | } |
| 99 | |
| 100 | llama_batch batch = {}; |
| 101 | { |
| 102 | auto prompt = common_tokenize(ctx: ctx.get(), text: params.prompt, add_special: true); |
| 103 | if (prompt.empty()) { |
| 104 | LOG_ERR("failed to tokenize prompt\n" ); |
| 105 | failed.store(i: true); |
| 106 | return; |
| 107 | } |
| 108 | batch = llama_batch_get_one(tokens: prompt.data(), n_tokens: prompt.size()); |
| 109 | if (llama_decode(ctx: ctx.get(), batch)) { |
| 110 | LOG_ERR("failed to decode prompt\n" ); |
| 111 | failed.store(i: true); |
| 112 | return; |
| 113 | } |
| 114 | } |
| 115 | |
| 116 | const auto * vocab = llama_model_get_vocab(model); |
| 117 | std::string result = params.prompt; |
| 118 | |
| 119 | for (int i = 0; i < params.n_predict; i++) { |
| 120 | llama_token token; |
| 121 | if (batch.n_tokens > 0) { |
| 122 | token = common_sampler_sample(gsmpl: sampler.get(), ctx: ctx.get(), idx: batch.n_tokens - 1); |
| 123 | } else { |
| 124 | token = llama_vocab_bos(vocab); |
| 125 | } |
| 126 | |
| 127 | result += common_token_to_piece(ctx: ctx.get(), token); |
| 128 | |
| 129 | if (llama_vocab_is_eog(vocab, token)) { |
| 130 | break; |
| 131 | } |
| 132 | |
| 133 | batch = llama_batch_get_one(tokens: &token, n_tokens: 1); |
| 134 | |
| 135 | int ret = llama_decode(ctx: ctx.get(), batch); |
| 136 | if (ret == 1 && i > 0) { |
| 137 | LOG_INF("Context full, stopping generation.\n" ); |
| 138 | break; |
| 139 | } |
| 140 | |
| 141 | if (ret != 0) { |
| 142 | LOG_ERR("Model %d/%d, Context %d/%d: failed to decode\n" , m + 1, num_models, c + 1, num_contexts); |
| 143 | failed.store(i: true); |
| 144 | return; |
| 145 | } |
| 146 | } |
| 147 | |
| 148 | LOG_INF("Model %d/%d, Context %d/%d: %s\n\n" , m + 1, num_models, c + 1, num_contexts, result.c_str()); |
| 149 | }); |
| 150 | } |
| 151 | } |
| 152 | |
| 153 | for (auto & thread : threads) { |
| 154 | thread.join(); |
| 155 | } |
| 156 | |
| 157 | if (failed) { |
| 158 | LOG_ERR("One or more threads failed.\n" ); |
| 159 | return 1; |
| 160 | } |
| 161 | |
| 162 | LOG_INF("All threads finished without errors.\n" ); |
| 163 | return 0; |
| 164 | } |
| 165 | |