diff --git a/CMakeLists.txt b/CMakeLists.txt index 38e7266dca630..09ac18b1af816 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -109,6 +109,7 @@ endif() add_executable(llama main.cpp utils.cpp + llama.cpp utils.h) add_executable(quantize diff --git a/Makefile b/Makefile index 1601079a48685..b7abaa9a029c5 100644 --- a/Makefile +++ b/Makefile @@ -188,11 +188,14 @@ ggml.o: ggml.c ggml.h utils.o: utils.cpp utils.h $(CXX) $(CXXFLAGS) -c utils.cpp -o utils.o +llama.o: llama.cpp llama.h + $(CXX) $(CXXFLAGS) -c llama.cpp -o llama.o + clean: rm -f *.o main quantize -main: main.cpp ggml.o utils.o - $(CXX) $(CXXFLAGS) main.cpp ggml.o utils.o -o main $(LDFLAGS) +main: main.cpp ggml.o utils.o llama.o + $(CXX) $(CXXFLAGS) main.cpp ggml.o utils.o llama.o -o main $(LDFLAGS) ./main -h quantize: quantize.cpp ggml.o utils.o diff --git a/llama.cpp b/llama.cpp new file mode 100644 index 0000000000000..05e37a0d6e044 --- /dev/null +++ b/llama.cpp @@ -0,0 +1,970 @@ +#include "llama.h" +#include "ggml.h" + +#include "utils.h" + +#include +#include +#include +#include +#include +#include +#include +#include + +#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) +#include +#include +#elif defined (_WIN32) +#include +#endif + +#define ANSI_COLOR_RED "\x1b[31m" +#define ANSI_COLOR_GREEN "\x1b[32m" +#define ANSI_COLOR_YELLOW "\x1b[33m" +#define ANSI_COLOR_BLUE "\x1b[34m" +#define ANSI_COLOR_MAGENTA "\x1b[35m" +#define ANSI_COLOR_CYAN "\x1b[36m" +#define ANSI_COLOR_RESET "\x1b[0m" +#define ANSI_BOLD "\x1b[1m" + +// determine number of model parts based on the dimension +static const std::map LLAMA_N_PARTS = { + { 4096, 1 }, + { 5120, 2 }, + { 6656, 4 }, + { 8192, 8 }, +}; + +// load the model's weights from a file +bool llama_model_load(const std::string & fname, llama_model & model, gpt_vocab & vocab, int n_ctx) { + fprintf(stderr, "%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str()); + + std::vector f_buf(1024*1024); + + auto fin = std::ifstream(fname, std::ios::binary); + fin.rdbuf()->pubsetbuf(f_buf.data(), f_buf.size()); + if (!fin) { + fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str()); + return false; + } + + // verify magic + { + uint32_t magic; + fin.read((char *) &magic, sizeof(magic)); + if (magic != 0x67676d6c) { + fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str()); + return false; + } + } + + int n_ff = 0; + int n_parts = 0; + + // load hparams + { + auto & hparams = model.hparams; + + fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab)); + //fin.read((char *) &hparams.n_ctx, sizeof(hparams.n_ctx)); + fin.read((char *) &hparams.n_embd, sizeof(hparams.n_embd)); + fin.read((char *) &hparams.n_mult, sizeof(hparams.n_mult)); + fin.read((char *) &hparams.n_head, sizeof(hparams.n_head)); + fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer)); + fin.read((char *) &hparams.n_rot, sizeof(hparams.n_rot)); + fin.read((char *) &hparams.f16, sizeof(hparams.f16)); + + hparams.n_ctx = n_ctx; + + n_ff = ((2*(4*hparams.n_embd)/3 + hparams.n_mult - 1)/hparams.n_mult)*hparams.n_mult; + n_parts = LLAMA_N_PARTS.at(hparams.n_embd); + + fprintf(stderr, "%s: n_vocab = %d\n", __func__, hparams.n_vocab); + fprintf(stderr, "%s: n_ctx = %d\n", __func__, hparams.n_ctx); + fprintf(stderr, "%s: n_embd = %d\n", __func__, hparams.n_embd); + fprintf(stderr, "%s: n_mult = %d\n", __func__, hparams.n_mult); + fprintf(stderr, "%s: n_head = %d\n", __func__, hparams.n_head); + fprintf(stderr, "%s: n_layer = %d\n", __func__, hparams.n_layer); + fprintf(stderr, "%s: n_rot = %d\n", __func__, hparams.n_rot); + fprintf(stderr, "%s: f16 = %d\n", __func__, hparams.f16); + fprintf(stderr, "%s: n_ff = %d\n", __func__, n_ff); + fprintf(stderr, "%s: n_parts = %d\n", __func__, n_parts); + } + + // load vocab + { + std::string word; + for (int i = 0; i < model.hparams.n_vocab; i++) { + uint32_t len; + fin.read((char *) &len, sizeof(len)); + + word.resize(len); + fin.read((char *) word.data(), len); + + vocab.token_to_id[word] = i; + vocab.id_to_token[i] = word; + + //if (i < 30000) { + // fprintf(stderr, "%s: vocab[%d] = '%s'\n", __func__, i, word.c_str()); + //} + } + } + + // for the big tensors, we have the option to store the data in 16-bit floats or quantized + // in order to save memory and also to speed up the computation + ggml_type wtype = GGML_TYPE_COUNT; + switch (model.hparams.f16) { + case 0: wtype = GGML_TYPE_F32; break; + case 1: wtype = GGML_TYPE_F16; break; + case 2: wtype = GGML_TYPE_Q4_0; break; + case 3: wtype = GGML_TYPE_Q4_1; break; + default: + { + fprintf(stderr, "%s: invalid model file '%s' (bad f16 value %d)\n", + __func__, fname.c_str(), model.hparams.f16); + return false; + } + } + + const ggml_type wtype2 = GGML_TYPE_F32; + + auto & ctx = model.ctx; + + size_t ctx_size = 0; + + { + const auto & hparams = model.hparams; + + const int n_embd = hparams.n_embd; + const int n_layer = hparams.n_layer; + const int n_ctx = hparams.n_ctx; + const int n_vocab = hparams.n_vocab; + + ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // tok_embeddings + + ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // norm + + ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // output + + ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // attention_norm + + ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wq + ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wk + ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wv + ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wo + + ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ffn_norm + + ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w1 + ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w2 + ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w3 + + ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k + ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v + + ctx_size += (5 + 10*n_layer)*256; // object overhead + + fprintf(stderr, "%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0)); + } + + // create the ggml context + { + struct ggml_init_params params = { + /*.mem_size =*/ ctx_size, + /*.mem_buffer =*/ NULL, + }; + + model.ctx = ggml_init(params); + if (!model.ctx) { + fprintf(stderr, "%s: ggml_init() failed\n", __func__); + return false; + } + } + + // prepare memory for the weights + { + const auto & hparams = model.hparams; + + const int n_embd = hparams.n_embd; + const int n_layer = hparams.n_layer; + const int n_ctx = hparams.n_ctx; + const int n_vocab = hparams.n_vocab; + + model.layers.resize(n_layer); + + model.tok_embeddings = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab); + + model.norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + model.output = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab); + + // map by name + model.tensors["tok_embeddings.weight"] = model.tok_embeddings; + + model.tensors["norm.weight"] = model.norm; + model.tensors["output.weight"] = model.output; + + for (int i = 0; i < n_layer; ++i) { + auto & layer = model.layers[i]; + + layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + + layer.wq = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); + layer.wk = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); + layer.wv = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); + layer.wo = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); + + layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + + layer.w1 = ggml_new_tensor_2d(ctx, wtype, n_embd, n_ff); + layer.w2 = ggml_new_tensor_2d(ctx, wtype, n_ff, n_embd); + layer.w3 = ggml_new_tensor_2d(ctx, wtype, n_embd, n_ff); + + // map by name + model.tensors["layers." + std::to_string(i) + ".attention_norm.weight"] = layer.attention_norm; + + model.tensors["layers." + std::to_string(i) + ".attention.wq.weight"] = layer.wq; + model.tensors["layers." + std::to_string(i) + ".attention.wk.weight"] = layer.wk; + model.tensors["layers." + std::to_string(i) + ".attention.wv.weight"] = layer.wv; + model.tensors["layers." + std::to_string(i) + ".attention.wo.weight"] = layer.wo; + + model.tensors["layers." + std::to_string(i) + ".ffn_norm.weight"] = layer.ffn_norm; + + model.tensors["layers." + std::to_string(i) + ".feed_forward.w1.weight"] = layer.w1; + model.tensors["layers." + std::to_string(i) + ".feed_forward.w2.weight"] = layer.w2; + model.tensors["layers." + std::to_string(i) + ".feed_forward.w3.weight"] = layer.w3; + } + } + + // key + value memory + { + const auto & hparams = model.hparams; + + const int n_embd = hparams.n_embd; + const int n_layer = hparams.n_layer; + const int n_ctx = hparams.n_ctx; + + const int n_mem = n_layer*n_ctx; + const int n_elements = n_embd*n_mem; + + model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements); + model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements); + + const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v); + + fprintf(stderr, "%s: memory_size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem); + } + + const size_t file_offset = fin.tellg(); + + fin.close(); + + std::vector tmp; + + for (int i = 0; i < n_parts; ++i) { + const int part_id = i; + //const int part_id = n_parts - i - 1; + + std::string fname_part = fname; + if (i > 0) { + fname_part += "." + std::to_string(i); + } + + fprintf(stderr, "%s: loading model part %d/%d from '%s'\n", __func__, i+1, n_parts, fname_part.c_str()); + + fin = std::ifstream(fname_part, std::ios::binary); + fin.rdbuf()->pubsetbuf(f_buf.data(), f_buf.size()); + fin.seekg(file_offset); + + // load weights + { + int n_tensors = 0; + size_t total_size = 0; + + fprintf(stderr, "%s: ", __func__); + + while (true) { + int32_t n_dims; + int32_t length; + int32_t ftype; + + fin.read(reinterpret_cast(&n_dims), sizeof(n_dims)); + fin.read(reinterpret_cast(&length), sizeof(length)); + fin.read(reinterpret_cast(&ftype), sizeof(ftype)); + + if (fin.eof()) { + break; + } + + int32_t nelements = 1; + int32_t ne[2] = { 1, 1 }; + for (int i = 0; i < n_dims; ++i) { + fin.read(reinterpret_cast(&ne[i]), sizeof(ne[i])); + nelements *= ne[i]; + } + + std::string name(length, 0); + fin.read(&name[0], length); + + if (model.tensors.find(name.data()) == model.tensors.end()) { + fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data()); + return false; + } + + // split_type = 0: split by columns + // split_type = 1: split by rows + int split_type = 0; + + // split_type = 0: + // regex: + // - tok_embeddings.* + // - layers.*.attention.wo.weight + // - layers.*.feed_forward.w2.weight + + // split_type = 1: + // regex: + // - output.* + // - layers.*.attention.wq.weight + // - layers.*.attention.wk.weight + // - layers.*.attention.wv.weight + // - layers.*.feed_forward.w1.weight + // - layers.*.feed_forward.w3.weight + if (name.find("tok_embeddings") != std::string::npos) { + split_type = 0; + } else if (name.find("layers") != std::string::npos) { + if (name.find("attention.wo.weight") != std::string::npos) { + split_type = 0; + } else if (name.find("feed_forward.w2.weight") != std::string::npos) { + split_type = 0; + } else { + split_type = 1; + } + } else if (name.find("output") != std::string::npos) { + split_type = 1; + } + + auto tensor = model.tensors[name.data()]; + + if (n_dims == 1) { + if (ggml_nelements(tensor) != nelements) { + fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data()); + return false; + } + } else { + if (ggml_nelements(tensor)/n_parts != nelements) { + fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data()); + return false; + } + } + + if (n_dims == 1) { + if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) { + fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n", + __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]); + return false; + } + } else { + if (split_type == 0) { + if (tensor->ne[0]/n_parts != ne[0] || tensor->ne[1] != ne[1]) { + fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n", + __func__, name.data(), tensor->ne[0]/n_parts, tensor->ne[1], ne[0], ne[1]); + return false; + } + } else { + if (tensor->ne[0] != ne[0] || tensor->ne[1]/n_parts != ne[1]) { + fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n", + __func__, name.data(), tensor->ne[0], tensor->ne[1]/n_parts, ne[0], ne[1]); + return false; + } + } + } + + if (0) { + static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", }; + fprintf(stderr, "%24s - [%5d, %5d], type = %6s, split = %d\n", name.data(), ne[0], ne[1], ftype_str[ftype], split_type); + } + + size_t bpe = 0; + + switch (ftype) { + case 0: bpe = ggml_type_size(GGML_TYPE_F32); break; + case 1: bpe = ggml_type_size(GGML_TYPE_F16); break; + case 2: bpe = ggml_type_size(GGML_TYPE_Q4_0); assert(ne[0] % 64 == 0); break; + case 3: bpe = ggml_type_size(GGML_TYPE_Q4_1); assert(ne[0] % 64 == 0); break; + default: + { + fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype); + return false; + } + }; + + if (n_dims == 1 || n_parts == 1) { + if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) { + fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n", + __func__, name.data(), ggml_nbytes(tensor), nelements*bpe); + return false; + } + + if (part_id == 0) { + fin.read(reinterpret_cast(tensor->data), ggml_nbytes(tensor)); + } else { + fin.seekg(ggml_nbytes(tensor), std::ios::cur); + } + + total_size += ggml_nbytes(tensor); + } else { + if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)/n_parts) { + fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n", + __func__, name.data(), ggml_nbytes(tensor)/n_parts, nelements*bpe); + return false; + } + + if (split_type == 0) { + const int np0 = ne[0]; + + const size_t row_size = (tensor->ne[0]/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type); + assert(row_size == tensor->nb[1]); + + for (int i1 = 0; i1 < ne[1]; ++i1) { + const size_t offset_row = i1*row_size; + const size_t offset = offset_row + ((part_id*np0)/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type); + fin.read(reinterpret_cast(tensor->data) + offset, row_size/n_parts); + } + } else { + const int np1 = ne[1]; + + const size_t row_size = (tensor->ne[0]/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type); + + for (int i1 = 0; i1 < ne[1]; ++i1) { + const size_t offset_row = (i1 + part_id*np1)*row_size; + fin.read(reinterpret_cast(tensor->data) + offset_row, row_size); + } + } + + total_size += ggml_nbytes(tensor)/n_parts; + } + + //fprintf(stderr, "%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0); + if (++n_tensors % 8 == 0) { + fprintf(stderr, "."); + fflush(stderr); + } + } + + fprintf(stderr, " done\n"); + + fprintf(stderr, "%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors); + } + + fin.close(); + } + + return true; +} + +// evaluate the transformer +// +// - model: the model +// - n_threads: number of threads to use +// - n_past: the context size so far +// - embd_inp: the embeddings of the tokens in the context +// - embd_w: the predicted logits for the next token +// +// The GPT-J model requires about 16MB of memory per input token. +// +bool llama_eval( + const llama_model & model, + const int n_threads, + const int n_past, + const std::vector & embd_inp, + std::vector & embd_w, + size_t & mem_per_token) { + const int N = embd_inp.size(); + + const auto & hparams = model.hparams; + + const int n_embd = hparams.n_embd; + const int n_layer = hparams.n_layer; + const int n_ctx = hparams.n_ctx; + const int n_head = hparams.n_head; + const int n_vocab = hparams.n_vocab; + const int n_rot = hparams.n_embd/hparams.n_head; + + const int d_key = n_embd/n_head; + + // TODO: check if this size scales with n_ctx linearly and remove constant. somehow I feel it wasn't the case + // static size_t buf_size = hparams.n_ctx*1024*1024; + static size_t buf_size = 512u*1024*1024; + static void * buf = malloc(buf_size); + + if (mem_per_token > 0 && mem_per_token*N > buf_size) { + const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead + //fprintf(stderr, "\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new); + + // reallocate + buf_size = buf_size_new; + buf = realloc(buf, buf_size); + if (buf == nullptr) { + fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size); + return false; + } + } + + struct ggml_init_params params = { + /*.mem_size =*/ buf_size, + /*.mem_buffer =*/ buf, + }; + + struct ggml_context * ctx0 = ggml_init(params); + ggml_cgraph gf = {}; + gf.n_threads = n_threads; + + struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N); + memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd)); + + struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.tok_embeddings, embd); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + struct ggml_tensor * cur; + + // norm + { + cur = ggml_rms_norm(ctx0, inpL); + + // cur = attention_norm*cur + cur = ggml_mul(ctx0, + ggml_repeat(ctx0, model.layers[il].attention_norm, cur), + cur); + } + + // self-attention + { + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + + // store key and value to memory + if (N >= 1) { + struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past)); + struct ggml_tensor * v = ggml_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past)); + + ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k)); + ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v)); + } + + // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3) + struct ggml_tensor * Q = + ggml_permute(ctx0, + ggml_rope(ctx0, + ggml_cpy(ctx0, + Qcur, + ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)), + n_past, n_rot, 0), + 0, 2, 1, 3); + + // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3) + struct ggml_tensor * K = + ggml_permute(ctx0, + ggml_rope(ctx0, + ggml_reshape_3d(ctx0, + ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd), + n_embd/n_head, n_head, n_past + N), + n_past, n_rot, 1), + 0, 2, 1, 3); + + // K * Q + struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); + + // KQ_scaled = KQ / sqrt(n_embd/n_head) + struct ggml_tensor * KQ_scaled = + ggml_scale(ctx0, + KQ, + ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head)) + ); + + // KQ_masked = mask_past(KQ_scaled) + struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past); + + // KQ = soft_max(KQ_masked) + struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); + + // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous() + struct ggml_tensor * V_trans = + ggml_permute(ctx0, + ggml_reshape_3d(ctx0, + ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd), + n_embd/n_head, n_head, n_past + N), + 1, 2, 0, 3); + + // KQV = transpose(V) * KQ_soft_max + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max); + + // KQV_merged = KQV.permute(0, 2, 1, 3) + struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); + + // cur = KQV_merged.contiguous().view(n_embd, N) + cur = ggml_cpy(ctx0, + KQV_merged, + ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N)); + + // projection (no bias) + cur = ggml_mul_mat(ctx0, + model.layers[il].wo, + cur); + } + + struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA); + + // feed-forward network + { + // norm + { + cur = ggml_rms_norm(ctx0, inpFF); + + // cur = ffn_norm*cur + cur = ggml_mul(ctx0, + ggml_repeat(ctx0, model.layers[il].ffn_norm, cur), + cur); + } + + struct ggml_tensor * tmp = ggml_mul_mat(ctx0, + model.layers[il].w3, + cur); + + + cur = ggml_mul_mat(ctx0, + model.layers[il].w1, + cur); + + // SILU activation + cur = ggml_silu(ctx0, cur); + + cur = ggml_mul(ctx0, cur, tmp); + + cur = ggml_mul_mat(ctx0, + model.layers[il].w2, + cur); + } + + cur = ggml_add(ctx0, cur, inpFF); + + // input for next layer + inpL = cur; + } + + // norm + { + inpL = ggml_rms_norm(ctx0, inpL); + + // inpL = norm*inpL + inpL = ggml_mul(ctx0, + ggml_repeat(ctx0, model.norm, inpL), + inpL); + } + + // lm_head + { + inpL = ggml_mul_mat(ctx0, model.output, inpL); + } + + // logits -> probs + //inpL = ggml_soft_max(ctx0, inpL); + + // run the computation + ggml_build_forward_expand(&gf, inpL); + ggml_graph_compute (ctx0, &gf); + + //if (n_past%100 == 0) { + // ggml_graph_print (&gf); + // ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot"); + //} + + //embd_w.resize(n_vocab*N); + //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N); + + // return result for just the last token + embd_w.resize(n_vocab); + memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab); + + if (mem_per_token == 0) { + mem_per_token = ggml_used_mem(ctx0)/N; + } + //fprintf(stderr, "used_mem = %zu\n", ggml_used_mem(ctx0)); + + ggml_free(ctx0); + + return true; +} + +static bool is_interacting = false; + +#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32) +void sigint_handler(int signo) { + printf(ANSI_COLOR_RESET); + if (signo == SIGINT) { + if (!is_interacting) { + is_interacting=true; + } else { + _exit(130); + } + } +} +#endif + +int llama_main( + gpt_params params, + gpt_vocab vocab, + llama_model model, + int64_t t_load_us, + int64_t t_main_start_us, + FILE *instream, + FILE *outstream, + FILE *errstream) { + + if (params.seed < 0) { + params.seed = time(NULL); + } + + fprintf(errstream, "%s: seed = %d\n", __func__, params.seed); + + std::mt19937 rng(params.seed); + if (params.prompt.empty()) { + params.prompt = gpt_random_prompt(rng); + } + +// params.prompt = R"(// this function checks if the number n is prime +//bool is_prime(int n) {)"; + + int n_past = 0; + + int64_t t_sample_us = 0; + int64_t t_predict_us = 0; + + std::vector logits; + + // Add a space in front of the first character to match OG llama tokenizer behavior + params.prompt.insert(0, 1, ' '); + // tokenize the prompt + std::vector embd_inp = ::llama_tokenize(vocab, params.prompt, true); + + params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size()); + + // tokenize the reverse prompt + std::vector antiprompt_inp = ::llama_tokenize(vocab, params.antiprompt, false); + + fprintf(errstream, "\n"); + fprintf(errstream, "%s: prompt: '%s'\n", __func__, params.prompt.c_str()); + fprintf(errstream, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size()); + for (int i = 0; i < (int) embd_inp.size(); i++) { + fprintf(errstream, "%6d -> '%s'\n", embd_inp[i], vocab.id_to_token.at(embd_inp[i]).c_str()); + } + fprintf(errstream, "\n"); + if (params.interactive) { +#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) + struct sigaction sigint_action; + sigint_action.sa_handler = sigint_handler; + sigemptyset (&sigint_action.sa_mask); + sigint_action.sa_flags = 0; + sigaction(SIGINT, &sigint_action, NULL); +#elif defined (_WIN32) + signal(SIGINT, sigint_handler); +#endif + + fprintf(errstream, "%s: interactive mode on.\n", __func__); + + if(antiprompt_inp.size()) { + fprintf(errstream, "%s: reverse prompt: '%s'\n", __func__, params.antiprompt.c_str()); + fprintf(errstream, "%s: number of tokens in reverse prompt = %zu\n", __func__, antiprompt_inp.size()); + for (int i = 0; i < (int) antiprompt_inp.size(); i++) { + fprintf(errstream, "%6d -> '%s'\n", antiprompt_inp[i], vocab.id_to_token.at(antiprompt_inp[i]).c_str()); + } + fprintf(errstream, "\n"); + } + } + fprintf(errstream, "sampling parameters: temp = %f, top_k = %d, top_p = %f, repeat_last_n = %i, repeat_penalty = %f\n", params.temp, params.top_k, params.top_p, params.repeat_last_n, params.repeat_penalty); + fprintf(errstream, "\n\n"); + + std::vector embd; + + // determine the required inference memory per token: + size_t mem_per_token = 0; + llama_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token); + + int last_n_size = params.repeat_last_n; + std::vector last_n_tokens(last_n_size); + std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0); + + + if (params.interactive) { + fprintf(errstream, "== Running in interactive mode. ==\n" +#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32) + " - Press Ctrl+C to interject at any time.\n" +#endif + " - Press Return to return control to LLaMa.\n" + " - If you want to submit another line, end your input in '\\'.\n"); + } + + int remaining_tokens = params.n_predict; + int input_consumed = 0; + bool input_noecho = false; + + // prompt user immediately after the starting prompt has been loaded + if (params.interactive_start) { + is_interacting = true; + } + + // set the color for the prompt which will be output initially + if (params.use_color) { + fprintf(outstream, ANSI_COLOR_YELLOW); + } + + while (remaining_tokens > 0) { + // predict + if (embd.size() > 0) { + const int64_t t_start_us = ggml_time_us(); + + if (!llama_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) { + fprintf(errstream, "Failed to predict\n"); + return 1; + } + + t_predict_us += ggml_time_us() - t_start_us; + } + + n_past += embd.size(); + embd.clear(); + + if (embd_inp.size() <= input_consumed) { + // out of user input, sample next token + const float top_k = params.top_k; + const float top_p = params.top_p; + const float temp = params.temp; + const float repeat_penalty = params.repeat_penalty; + + const int n_vocab = model.hparams.n_vocab; + + gpt_vocab::id id = 0; + + { + const int64_t t_start_sample_us = ggml_time_us(); + + id = llama_sample_top_p_top_k(vocab, logits.data() + (logits.size() - n_vocab), last_n_tokens, repeat_penalty, top_k, top_p, temp, rng); + + last_n_tokens.erase(last_n_tokens.begin()); + last_n_tokens.push_back(id); + + t_sample_us += ggml_time_us() - t_start_sample_us; + } + + // add it to the context + embd.push_back(id); + + // echo this to console + input_noecho = false; + + // decrement remaining sampling budget + --remaining_tokens; + } else { + // some user input remains from prompt or interaction, forward it to processing + while (embd_inp.size() > input_consumed) { + embd.push_back(embd_inp[input_consumed]); + last_n_tokens.erase(last_n_tokens.begin()); + last_n_tokens.push_back(embd_inp[input_consumed]); + ++input_consumed; + if (embd.size() > params.n_batch) { + break; + } + } + + // reset color to default if we there is no pending user input + if (!input_noecho && params.use_color && embd_inp.size() == input_consumed) { + fprintf(outstream, ANSI_COLOR_RESET); + } + } + + // display text + if (!input_noecho) { + for (auto id : embd) { + fprintf(outstream, "%s", vocab.id_to_token[id].c_str()); + } + fflush(outstream); + } + + // in interactive mode, and not currently processing queued inputs; + // check if we should prompt the user for more + if (params.interactive && embd_inp.size() <= input_consumed) { + // check for reverse prompt + if (antiprompt_inp.size() && std::equal(antiprompt_inp.rbegin(), antiprompt_inp.rend(), last_n_tokens.rbegin())) { + // reverse prompt found + is_interacting = true; + } + if (is_interacting) { + // currently being interactive + bool another_line=true; + while (another_line) { + fflush(outstream); + char buf[256] = {0}; + int n_read; + if(params.use_color) fprintf(outstream, ANSI_BOLD ANSI_COLOR_GREEN); + if (fscanf(instream, "%255[^\n]%n%*c", buf, &n_read) <= 0) { + // presumable empty line, consume the newline + std::ignore = fscanf(instream, "%*c"); + n_read=0; + } + if(params.use_color) fprintf(outstream, ANSI_COLOR_RESET); + + if (n_read > 0 && buf[n_read-1]=='\\') { + another_line = true; + buf[n_read-1] = '\n'; + buf[n_read] = 0; + } else { + another_line = false; + buf[n_read] = '\n'; + buf[n_read+1] = 0; + } + + std::vector line_inp = ::llama_tokenize(vocab, buf, false); + embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end()); + + remaining_tokens -= line_inp.size(); + + input_noecho = true; // do not echo this again + } + + is_interacting = false; + } + } + + // end of text token + if (embd.back() == 2) { + fprintf(errstream, " [end of text]\n"); + break; + } + } + +#if defined (_WIN32) + signal(SIGINT, SIG_DFL); +#endif + + // report timing + { + const int64_t t_main_end_us = ggml_time_us(); + + fprintf(errstream, "\n\n"); + fprintf(errstream, "%s: mem per token = %8zu bytes\n", __func__, mem_per_token); + fprintf(errstream, "%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f); + fprintf(errstream, "%s: sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f); + fprintf(errstream, "%s: predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past); + fprintf(errstream, "%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f); + } + + ggml_free(model.ctx); + + if (params.use_color) { + fprintf(outstream, ANSI_COLOR_RESET); + } + + return 0; +} diff --git a/llama.h b/llama.h new file mode 100644 index 0000000000000..7c8409d1a158e --- /dev/null +++ b/llama.h @@ -0,0 +1,71 @@ +#pragma once + +#include +#include +#include +#include + +#include "ggml.h" +#include "utils.h" + + +// default hparams (LLaMA 7B) +struct llama_hparams { + int32_t n_vocab = 32000; + int32_t n_ctx = 512; // this is provided as user input? + int32_t n_embd = 4096; + int32_t n_mult = 256; + int32_t n_head = 32; + int32_t n_layer = 32; + int32_t n_rot = 64; + int32_t f16 = 1; +}; + +struct llama_layer { + // normalization + struct ggml_tensor * attention_norm; + + // attention + struct ggml_tensor * wq; + struct ggml_tensor * wk; + struct ggml_tensor * wv; + struct ggml_tensor * wo; + + // normalization + struct ggml_tensor * ffn_norm; + + // ff + struct ggml_tensor * w1; + struct ggml_tensor * w2; + struct ggml_tensor * w3; +}; + +struct llama_model { + llama_hparams hparams; + + struct ggml_tensor * tok_embeddings; + + struct ggml_tensor * norm; + struct ggml_tensor * output; + + std::vector layers; + + // key + value memory + struct ggml_tensor * memory_k; + struct ggml_tensor * memory_v; + + // + struct ggml_context * ctx; + std::map tensors; +}; + +int llama_main( + gpt_params params, + gpt_vocab vocab, + llama_model model, + int64_t t_load_us, + int64_t t_main_start_us, + FILE *instream, + FILE *outstream, + FILE *errstream); +bool llama_model_load(const std::string & fname, llama_model & model, gpt_vocab & vocab, int n_ctx); diff --git a/main.cpp b/main.cpp index c88405b82956a..e3fc73e750a21 100644 --- a/main.cpp +++ b/main.cpp @@ -1,766 +1,6 @@ #include "ggml.h" - #include "utils.h" - -#include -#include -#include -#include -#include -#include -#include -#include - -#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) -#include -#include -#elif defined (_WIN32) -#include -#endif - -#define ANSI_COLOR_RED "\x1b[31m" -#define ANSI_COLOR_GREEN "\x1b[32m" -#define ANSI_COLOR_YELLOW "\x1b[33m" -#define ANSI_COLOR_BLUE "\x1b[34m" -#define ANSI_COLOR_MAGENTA "\x1b[35m" -#define ANSI_COLOR_CYAN "\x1b[36m" -#define ANSI_COLOR_RESET "\x1b[0m" -#define ANSI_BOLD "\x1b[1m" - -// determine number of model parts based on the dimension -static const std::map LLAMA_N_PARTS = { - { 4096, 1 }, - { 5120, 2 }, - { 6656, 4 }, - { 8192, 8 }, -}; - -// default hparams (LLaMA 7B) -struct llama_hparams { - int32_t n_vocab = 32000; - int32_t n_ctx = 512; // this is provided as user input? - int32_t n_embd = 4096; - int32_t n_mult = 256; - int32_t n_head = 32; - int32_t n_layer = 32; - int32_t n_rot = 64; - int32_t f16 = 1; -}; - -struct llama_layer { - // normalization - struct ggml_tensor * attention_norm; - - // attention - struct ggml_tensor * wq; - struct ggml_tensor * wk; - struct ggml_tensor * wv; - struct ggml_tensor * wo; - - // normalization - struct ggml_tensor * ffn_norm; - - // ff - struct ggml_tensor * w1; - struct ggml_tensor * w2; - struct ggml_tensor * w3; -}; - -struct llama_model { - llama_hparams hparams; - - struct ggml_tensor * tok_embeddings; - - struct ggml_tensor * norm; - struct ggml_tensor * output; - - std::vector layers; - - // key + value memory - struct ggml_tensor * memory_k; - struct ggml_tensor * memory_v; - - // - struct ggml_context * ctx; - std::map tensors; -}; - -// load the model's weights from a file -bool llama_model_load(const std::string & fname, llama_model & model, gpt_vocab & vocab, int n_ctx) { - fprintf(stderr, "%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str()); - - std::vector f_buf(1024*1024); - - auto fin = std::ifstream(fname, std::ios::binary); - fin.rdbuf()->pubsetbuf(f_buf.data(), f_buf.size()); - if (!fin) { - fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str()); - return false; - } - - // verify magic - { - uint32_t magic; - fin.read((char *) &magic, sizeof(magic)); - if (magic != 0x67676d6c) { - fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str()); - return false; - } - } - - int n_ff = 0; - int n_parts = 0; - - // load hparams - { - auto & hparams = model.hparams; - - fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab)); - //fin.read((char *) &hparams.n_ctx, sizeof(hparams.n_ctx)); - fin.read((char *) &hparams.n_embd, sizeof(hparams.n_embd)); - fin.read((char *) &hparams.n_mult, sizeof(hparams.n_mult)); - fin.read((char *) &hparams.n_head, sizeof(hparams.n_head)); - fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer)); - fin.read((char *) &hparams.n_rot, sizeof(hparams.n_rot)); - fin.read((char *) &hparams.f16, sizeof(hparams.f16)); - - hparams.n_ctx = n_ctx; - - n_ff = ((2*(4*hparams.n_embd)/3 + hparams.n_mult - 1)/hparams.n_mult)*hparams.n_mult; - n_parts = LLAMA_N_PARTS.at(hparams.n_embd); - - fprintf(stderr, "%s: n_vocab = %d\n", __func__, hparams.n_vocab); - fprintf(stderr, "%s: n_ctx = %d\n", __func__, hparams.n_ctx); - fprintf(stderr, "%s: n_embd = %d\n", __func__, hparams.n_embd); - fprintf(stderr, "%s: n_mult = %d\n", __func__, hparams.n_mult); - fprintf(stderr, "%s: n_head = %d\n", __func__, hparams.n_head); - fprintf(stderr, "%s: n_layer = %d\n", __func__, hparams.n_layer); - fprintf(stderr, "%s: n_rot = %d\n", __func__, hparams.n_rot); - fprintf(stderr, "%s: f16 = %d\n", __func__, hparams.f16); - fprintf(stderr, "%s: n_ff = %d\n", __func__, n_ff); - fprintf(stderr, "%s: n_parts = %d\n", __func__, n_parts); - } - - // load vocab - { - std::string word; - for (int i = 0; i < model.hparams.n_vocab; i++) { - uint32_t len; - fin.read((char *) &len, sizeof(len)); - - word.resize(len); - fin.read((char *) word.data(), len); - - vocab.token_to_id[word] = i; - vocab.id_to_token[i] = word; - - //if (i < 30000) { - // fprintf(stderr, "%s: vocab[%d] = '%s'\n", __func__, i, word.c_str()); - //} - } - } - - // for the big tensors, we have the option to store the data in 16-bit floats or quantized - // in order to save memory and also to speed up the computation - ggml_type wtype = GGML_TYPE_COUNT; - switch (model.hparams.f16) { - case 0: wtype = GGML_TYPE_F32; break; - case 1: wtype = GGML_TYPE_F16; break; - case 2: wtype = GGML_TYPE_Q4_0; break; - case 3: wtype = GGML_TYPE_Q4_1; break; - default: - { - fprintf(stderr, "%s: invalid model file '%s' (bad f16 value %d)\n", - __func__, fname.c_str(), model.hparams.f16); - return false; - } - } - - const ggml_type wtype2 = GGML_TYPE_F32; - - auto & ctx = model.ctx; - - size_t ctx_size = 0; - - { - const auto & hparams = model.hparams; - - const int n_embd = hparams.n_embd; - const int n_layer = hparams.n_layer; - const int n_ctx = hparams.n_ctx; - const int n_vocab = hparams.n_vocab; - - ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // tok_embeddings - - ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // norm - - ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // output - - ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // attention_norm - - ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wq - ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wk - ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wv - ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wo - - ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ffn_norm - - ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w1 - ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w2 - ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w3 - - ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k - ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v - - ctx_size += (5 + 10*n_layer)*256; // object overhead - - fprintf(stderr, "%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0)); - } - - // create the ggml context - { - struct ggml_init_params params = { - /*.mem_size =*/ ctx_size, - /*.mem_buffer =*/ NULL, - }; - - model.ctx = ggml_init(params); - if (!model.ctx) { - fprintf(stderr, "%s: ggml_init() failed\n", __func__); - return false; - } - } - - // prepare memory for the weights - { - const auto & hparams = model.hparams; - - const int n_embd = hparams.n_embd; - const int n_layer = hparams.n_layer; - const int n_ctx = hparams.n_ctx; - const int n_vocab = hparams.n_vocab; - - model.layers.resize(n_layer); - - model.tok_embeddings = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab); - - model.norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); - model.output = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab); - - // map by name - model.tensors["tok_embeddings.weight"] = model.tok_embeddings; - - model.tensors["norm.weight"] = model.norm; - model.tensors["output.weight"] = model.output; - - for (int i = 0; i < n_layer; ++i) { - auto & layer = model.layers[i]; - - layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); - - layer.wq = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); - layer.wk = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); - layer.wv = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); - layer.wo = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); - - layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); - - layer.w1 = ggml_new_tensor_2d(ctx, wtype, n_embd, n_ff); - layer.w2 = ggml_new_tensor_2d(ctx, wtype, n_ff, n_embd); - layer.w3 = ggml_new_tensor_2d(ctx, wtype, n_embd, n_ff); - - // map by name - model.tensors["layers." + std::to_string(i) + ".attention_norm.weight"] = layer.attention_norm; - - model.tensors["layers." + std::to_string(i) + ".attention.wq.weight"] = layer.wq; - model.tensors["layers." + std::to_string(i) + ".attention.wk.weight"] = layer.wk; - model.tensors["layers." + std::to_string(i) + ".attention.wv.weight"] = layer.wv; - model.tensors["layers." + std::to_string(i) + ".attention.wo.weight"] = layer.wo; - - model.tensors["layers." + std::to_string(i) + ".ffn_norm.weight"] = layer.ffn_norm; - - model.tensors["layers." + std::to_string(i) + ".feed_forward.w1.weight"] = layer.w1; - model.tensors["layers." + std::to_string(i) + ".feed_forward.w2.weight"] = layer.w2; - model.tensors["layers." + std::to_string(i) + ".feed_forward.w3.weight"] = layer.w3; - } - } - - // key + value memory - { - const auto & hparams = model.hparams; - - const int n_embd = hparams.n_embd; - const int n_layer = hparams.n_layer; - const int n_ctx = hparams.n_ctx; - - const int n_mem = n_layer*n_ctx; - const int n_elements = n_embd*n_mem; - - model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements); - model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements); - - const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v); - - fprintf(stderr, "%s: memory_size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem); - } - - const size_t file_offset = fin.tellg(); - - fin.close(); - - std::vector tmp; - - for (int i = 0; i < n_parts; ++i) { - const int part_id = i; - //const int part_id = n_parts - i - 1; - - std::string fname_part = fname; - if (i > 0) { - fname_part += "." + std::to_string(i); - } - - fprintf(stderr, "%s: loading model part %d/%d from '%s'\n", __func__, i+1, n_parts, fname_part.c_str()); - - fin = std::ifstream(fname_part, std::ios::binary); - fin.rdbuf()->pubsetbuf(f_buf.data(), f_buf.size()); - fin.seekg(file_offset); - - // load weights - { - int n_tensors = 0; - size_t total_size = 0; - - fprintf(stderr, "%s: ", __func__); - - while (true) { - int32_t n_dims; - int32_t length; - int32_t ftype; - - fin.read(reinterpret_cast(&n_dims), sizeof(n_dims)); - fin.read(reinterpret_cast(&length), sizeof(length)); - fin.read(reinterpret_cast(&ftype), sizeof(ftype)); - - if (fin.eof()) { - break; - } - - int32_t nelements = 1; - int32_t ne[2] = { 1, 1 }; - for (int i = 0; i < n_dims; ++i) { - fin.read(reinterpret_cast(&ne[i]), sizeof(ne[i])); - nelements *= ne[i]; - } - - std::string name(length, 0); - fin.read(&name[0], length); - - if (model.tensors.find(name.data()) == model.tensors.end()) { - fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data()); - return false; - } - - // split_type = 0: split by columns - // split_type = 1: split by rows - int split_type = 0; - - // split_type = 0: - // regex: - // - tok_embeddings.* - // - layers.*.attention.wo.weight - // - layers.*.feed_forward.w2.weight - - // split_type = 1: - // regex: - // - output.* - // - layers.*.attention.wq.weight - // - layers.*.attention.wk.weight - // - layers.*.attention.wv.weight - // - layers.*.feed_forward.w1.weight - // - layers.*.feed_forward.w3.weight - if (name.find("tok_embeddings") != std::string::npos) { - split_type = 0; - } else if (name.find("layers") != std::string::npos) { - if (name.find("attention.wo.weight") != std::string::npos) { - split_type = 0; - } else if (name.find("feed_forward.w2.weight") != std::string::npos) { - split_type = 0; - } else { - split_type = 1; - } - } else if (name.find("output") != std::string::npos) { - split_type = 1; - } - - auto tensor = model.tensors[name.data()]; - - if (n_dims == 1) { - if (ggml_nelements(tensor) != nelements) { - fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data()); - return false; - } - } else { - if (ggml_nelements(tensor)/n_parts != nelements) { - fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data()); - return false; - } - } - - if (n_dims == 1) { - if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) { - fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n", - __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]); - return false; - } - } else { - if (split_type == 0) { - if (tensor->ne[0]/n_parts != ne[0] || tensor->ne[1] != ne[1]) { - fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n", - __func__, name.data(), tensor->ne[0]/n_parts, tensor->ne[1], ne[0], ne[1]); - return false; - } - } else { - if (tensor->ne[0] != ne[0] || tensor->ne[1]/n_parts != ne[1]) { - fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n", - __func__, name.data(), tensor->ne[0], tensor->ne[1]/n_parts, ne[0], ne[1]); - return false; - } - } - } - - if (0) { - static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", }; - fprintf(stderr, "%24s - [%5d, %5d], type = %6s, split = %d\n", name.data(), ne[0], ne[1], ftype_str[ftype], split_type); - } - - size_t bpe = 0; - - switch (ftype) { - case 0: bpe = ggml_type_size(GGML_TYPE_F32); break; - case 1: bpe = ggml_type_size(GGML_TYPE_F16); break; - case 2: bpe = ggml_type_size(GGML_TYPE_Q4_0); assert(ne[0] % 64 == 0); break; - case 3: bpe = ggml_type_size(GGML_TYPE_Q4_1); assert(ne[0] % 64 == 0); break; - default: - { - fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype); - return false; - } - }; - - if (n_dims == 1 || n_parts == 1) { - if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) { - fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n", - __func__, name.data(), ggml_nbytes(tensor), nelements*bpe); - return false; - } - - if (part_id == 0) { - fin.read(reinterpret_cast(tensor->data), ggml_nbytes(tensor)); - } else { - fin.seekg(ggml_nbytes(tensor), std::ios::cur); - } - - total_size += ggml_nbytes(tensor); - } else { - if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)/n_parts) { - fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n", - __func__, name.data(), ggml_nbytes(tensor)/n_parts, nelements*bpe); - return false; - } - - if (split_type == 0) { - const int np0 = ne[0]; - - const size_t row_size = (tensor->ne[0]/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type); - assert(row_size == tensor->nb[1]); - - for (int i1 = 0; i1 < ne[1]; ++i1) { - const size_t offset_row = i1*row_size; - const size_t offset = offset_row + ((part_id*np0)/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type); - fin.read(reinterpret_cast(tensor->data) + offset, row_size/n_parts); - } - } else { - const int np1 = ne[1]; - - const size_t row_size = (tensor->ne[0]/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type); - - for (int i1 = 0; i1 < ne[1]; ++i1) { - const size_t offset_row = (i1 + part_id*np1)*row_size; - fin.read(reinterpret_cast(tensor->data) + offset_row, row_size); - } - } - - total_size += ggml_nbytes(tensor)/n_parts; - } - - //fprintf(stderr, "%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0); - if (++n_tensors % 8 == 0) { - fprintf(stderr, "."); - fflush(stderr); - } - } - - fprintf(stderr, " done\n"); - - fprintf(stderr, "%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors); - } - - fin.close(); - } - - return true; -} - -// evaluate the transformer -// -// - model: the model -// - n_threads: number of threads to use -// - n_past: the context size so far -// - embd_inp: the embeddings of the tokens in the context -// - embd_w: the predicted logits for the next token -// -// The GPT-J model requires about 16MB of memory per input token. -// -bool llama_eval( - const llama_model & model, - const int n_threads, - const int n_past, - const std::vector & embd_inp, - std::vector & embd_w, - size_t & mem_per_token) { - const int N = embd_inp.size(); - - const auto & hparams = model.hparams; - - const int n_embd = hparams.n_embd; - const int n_layer = hparams.n_layer; - const int n_ctx = hparams.n_ctx; - const int n_head = hparams.n_head; - const int n_vocab = hparams.n_vocab; - const int n_rot = hparams.n_embd/hparams.n_head; - - const int d_key = n_embd/n_head; - - // TODO: check if this size scales with n_ctx linearly and remove constant. somehow I feel it wasn't the case - // static size_t buf_size = hparams.n_ctx*1024*1024; - static size_t buf_size = 512u*1024*1024; - static void * buf = malloc(buf_size); - - if (mem_per_token > 0 && mem_per_token*N > buf_size) { - const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead - //fprintf(stderr, "\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new); - - // reallocate - buf_size = buf_size_new; - buf = realloc(buf, buf_size); - if (buf == nullptr) { - fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size); - return false; - } - } - - struct ggml_init_params params = { - /*.mem_size =*/ buf_size, - /*.mem_buffer =*/ buf, - }; - - struct ggml_context * ctx0 = ggml_init(params); - ggml_cgraph gf = {}; - gf.n_threads = n_threads; - - struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N); - memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd)); - - struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.tok_embeddings, embd); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - struct ggml_tensor * cur; - - // norm - { - cur = ggml_rms_norm(ctx0, inpL); - - // cur = attention_norm*cur - cur = ggml_mul(ctx0, - ggml_repeat(ctx0, model.layers[il].attention_norm, cur), - cur); - } - - // self-attention - { - struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); - struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); - struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); - - // store key and value to memory - if (N >= 1) { - struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past)); - struct ggml_tensor * v = ggml_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past)); - - ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v)); - } - - // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3) - struct ggml_tensor * Q = - ggml_permute(ctx0, - ggml_rope(ctx0, - ggml_cpy(ctx0, - Qcur, - ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)), - n_past, n_rot, 0), - 0, 2, 1, 3); - - // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3) - struct ggml_tensor * K = - ggml_permute(ctx0, - ggml_rope(ctx0, - ggml_reshape_3d(ctx0, - ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd), - n_embd/n_head, n_head, n_past + N), - n_past, n_rot, 1), - 0, 2, 1, 3); - - // K * Q - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - - // KQ_scaled = KQ / sqrt(n_embd/n_head) - struct ggml_tensor * KQ_scaled = - ggml_scale(ctx0, - KQ, - ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head)) - ); - - // KQ_masked = mask_past(KQ_scaled) - struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past); - - // KQ = soft_max(KQ_masked) - struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); - - // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous() - struct ggml_tensor * V_trans = - ggml_permute(ctx0, - ggml_reshape_3d(ctx0, - ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd), - n_embd/n_head, n_head, n_past + N), - 1, 2, 0, 3); - - // KQV = transpose(V) * KQ_soft_max - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max); - - // KQV_merged = KQV.permute(0, 2, 1, 3) - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - - // cur = KQV_merged.contiguous().view(n_embd, N) - cur = ggml_cpy(ctx0, - KQV_merged, - ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N)); - - // projection (no bias) - cur = ggml_mul_mat(ctx0, - model.layers[il].wo, - cur); - } - - struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA); - - // feed-forward network - { - // norm - { - cur = ggml_rms_norm(ctx0, inpFF); - - // cur = ffn_norm*cur - cur = ggml_mul(ctx0, - ggml_repeat(ctx0, model.layers[il].ffn_norm, cur), - cur); - } - - struct ggml_tensor * tmp = ggml_mul_mat(ctx0, - model.layers[il].w3, - cur); - - - cur = ggml_mul_mat(ctx0, - model.layers[il].w1, - cur); - - // SILU activation - cur = ggml_silu(ctx0, cur); - - cur = ggml_mul(ctx0, cur, tmp); - - cur = ggml_mul_mat(ctx0, - model.layers[il].w2, - cur); - } - - cur = ggml_add(ctx0, cur, inpFF); - - // input for next layer - inpL = cur; - } - - // norm - { - inpL = ggml_rms_norm(ctx0, inpL); - - // inpL = norm*inpL - inpL = ggml_mul(ctx0, - ggml_repeat(ctx0, model.norm, inpL), - inpL); - } - - // lm_head - { - inpL = ggml_mul_mat(ctx0, model.output, inpL); - } - - // logits -> probs - //inpL = ggml_soft_max(ctx0, inpL); - - // run the computation - ggml_build_forward_expand(&gf, inpL); - ggml_graph_compute (ctx0, &gf); - - //if (n_past%100 == 0) { - // ggml_graph_print (&gf); - // ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot"); - //} - - //embd_w.resize(n_vocab*N); - //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N); - - // return result for just the last token - embd_w.resize(n_vocab); - memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab); - - if (mem_per_token == 0) { - mem_per_token = ggml_used_mem(ctx0)/N; - } - //fprintf(stderr, "used_mem = %zu\n", ggml_used_mem(ctx0)); - - ggml_free(ctx0); - - return true; -} - -static bool is_interacting = false; - -#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32) -void sigint_handler(int signo) { - printf(ANSI_COLOR_RESET); - if (signo == SIGINT) { - if (!is_interacting) { - is_interacting=true; - } else { - _exit(130); - } - } -} -#endif +#include "llama.h" const char * llama_print_system_info(void) { static std::string s; @@ -793,20 +33,6 @@ int main(int argc, char ** argv) { return 1; } - if (params.seed < 0) { - params.seed = time(NULL); - } - - fprintf(stderr, "%s: seed = %d\n", __func__, params.seed); - - std::mt19937 rng(params.seed); - if (params.prompt.empty()) { - params.prompt = gpt_random_prompt(rng); - } - -// params.prompt = R"(// this function checks if the number n is prime -//bool is_prime(int n) {)"; - int64_t t_load_us = 0; gpt_vocab vocab; @@ -815,7 +41,7 @@ int main(int argc, char ** argv) { // load the model { const int64_t t_start_us = ggml_time_us(); - if (!llama_model_load(params.model, model, vocab, params.n_ctx)) { + if (!llama_model_load(params.model, model, vocab, params.n_ctx)) { fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str()); return 1; } @@ -830,234 +56,5 @@ int main(int argc, char ** argv) { params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info()); } - int n_past = 0; - - int64_t t_sample_us = 0; - int64_t t_predict_us = 0; - - std::vector logits; - - // Add a space in front of the first character to match OG llama tokenizer behavior - params.prompt.insert(0, 1, ' '); - // tokenize the prompt - std::vector embd_inp = ::llama_tokenize(vocab, params.prompt, true); - - params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size()); - - // tokenize the reverse prompt - std::vector antiprompt_inp = ::llama_tokenize(vocab, params.antiprompt, false); - - fprintf(stderr, "\n"); - fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str()); - fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size()); - for (int i = 0; i < (int) embd_inp.size(); i++) { - fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], vocab.id_to_token.at(embd_inp[i]).c_str()); - } - fprintf(stderr, "\n"); - if (params.interactive) { -#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) - struct sigaction sigint_action; - sigint_action.sa_handler = sigint_handler; - sigemptyset (&sigint_action.sa_mask); - sigint_action.sa_flags = 0; - sigaction(SIGINT, &sigint_action, NULL); -#elif defined (_WIN32) - signal(SIGINT, sigint_handler); -#endif - - fprintf(stderr, "%s: interactive mode on.\n", __func__); - - if(antiprompt_inp.size()) { - fprintf(stderr, "%s: reverse prompt: '%s'\n", __func__, params.antiprompt.c_str()); - fprintf(stderr, "%s: number of tokens in reverse prompt = %zu\n", __func__, antiprompt_inp.size()); - for (int i = 0; i < (int) antiprompt_inp.size(); i++) { - fprintf(stderr, "%6d -> '%s'\n", antiprompt_inp[i], vocab.id_to_token.at(antiprompt_inp[i]).c_str()); - } - fprintf(stderr, "\n"); - } - } - fprintf(stderr, "sampling parameters: temp = %f, top_k = %d, top_p = %f, repeat_last_n = %i, repeat_penalty = %f\n", params.temp, params.top_k, params.top_p, params.repeat_last_n, params.repeat_penalty); - fprintf(stderr, "\n\n"); - - std::vector embd; - - // determine the required inference memory per token: - size_t mem_per_token = 0; - llama_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token); - - int last_n_size = params.repeat_last_n; - std::vector last_n_tokens(last_n_size); - std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0); - - - if (params.interactive) { - fprintf(stderr, "== Running in interactive mode. ==\n" -#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32) - " - Press Ctrl+C to interject at any time.\n" -#endif - " - Press Return to return control to LLaMa.\n" - " - If you want to submit another line, end your input in '\\'.\n"); - } - - int remaining_tokens = params.n_predict; - int input_consumed = 0; - bool input_noecho = false; - - // prompt user immediately after the starting prompt has been loaded - if (params.interactive_start) { - is_interacting = true; - } - - // set the color for the prompt which will be output initially - if (params.use_color) { - printf(ANSI_COLOR_YELLOW); - } - - while (remaining_tokens > 0) { - // predict - if (embd.size() > 0) { - const int64_t t_start_us = ggml_time_us(); - - if (!llama_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) { - fprintf(stderr, "Failed to predict\n"); - return 1; - } - - t_predict_us += ggml_time_us() - t_start_us; - } - - n_past += embd.size(); - embd.clear(); - - if (embd_inp.size() <= input_consumed) { - // out of user input, sample next token - const float top_k = params.top_k; - const float top_p = params.top_p; - const float temp = params.temp; - const float repeat_penalty = params.repeat_penalty; - - const int n_vocab = model.hparams.n_vocab; - - gpt_vocab::id id = 0; - - { - const int64_t t_start_sample_us = ggml_time_us(); - - id = llama_sample_top_p_top_k(vocab, logits.data() + (logits.size() - n_vocab), last_n_tokens, repeat_penalty, top_k, top_p, temp, rng); - - last_n_tokens.erase(last_n_tokens.begin()); - last_n_tokens.push_back(id); - - t_sample_us += ggml_time_us() - t_start_sample_us; - } - - // add it to the context - embd.push_back(id); - - // echo this to console - input_noecho = false; - - // decrement remaining sampling budget - --remaining_tokens; - } else { - // some user input remains from prompt or interaction, forward it to processing - while (embd_inp.size() > input_consumed) { - embd.push_back(embd_inp[input_consumed]); - last_n_tokens.erase(last_n_tokens.begin()); - last_n_tokens.push_back(embd_inp[input_consumed]); - ++input_consumed; - if (embd.size() > params.n_batch) { - break; - } - } - - // reset color to default if we there is no pending user input - if (!input_noecho && params.use_color && embd_inp.size() == input_consumed) { - printf(ANSI_COLOR_RESET); - } - } - - // display text - if (!input_noecho) { - for (auto id : embd) { - printf("%s", vocab.id_to_token[id].c_str()); - } - fflush(stdout); - } - - // in interactive mode, and not currently processing queued inputs; - // check if we should prompt the user for more - if (params.interactive && embd_inp.size() <= input_consumed) { - // check for reverse prompt - if (antiprompt_inp.size() && std::equal(antiprompt_inp.rbegin(), antiprompt_inp.rend(), last_n_tokens.rbegin())) { - // reverse prompt found - is_interacting = true; - } - if (is_interacting) { - // currently being interactive - bool another_line=true; - while (another_line) { - fflush(stdout); - char buf[256] = {0}; - int n_read; - if(params.use_color) printf(ANSI_BOLD ANSI_COLOR_GREEN); - if (scanf("%255[^\n]%n%*c", buf, &n_read) <= 0) { - // presumable empty line, consume the newline - std::ignore = scanf("%*c"); - n_read=0; - } - if(params.use_color) printf(ANSI_COLOR_RESET); - - if (n_read > 0 && buf[n_read-1]=='\\') { - another_line = true; - buf[n_read-1] = '\n'; - buf[n_read] = 0; - } else { - another_line = false; - buf[n_read] = '\n'; - buf[n_read+1] = 0; - } - - std::vector line_inp = ::llama_tokenize(vocab, buf, false); - embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end()); - - remaining_tokens -= line_inp.size(); - - input_noecho = true; // do not echo this again - } - - is_interacting = false; - } - } - - // end of text token - if (embd.back() == 2) { - fprintf(stderr, " [end of text]\n"); - break; - } - } - -#if defined (_WIN32) - signal(SIGINT, SIG_DFL); -#endif - - // report timing - { - const int64_t t_main_end_us = ggml_time_us(); - - fprintf(stderr, "\n\n"); - fprintf(stderr, "%s: mem per token = %8zu bytes\n", __func__, mem_per_token); - fprintf(stderr, "%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f); - fprintf(stderr, "%s: sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f); - fprintf(stderr, "%s: predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past); - fprintf(stderr, "%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f); - } - - ggml_free(model.ctx); - - if (params.use_color) { - printf(ANSI_COLOR_RESET); - } - - return 0; + return llama_main(params, vocab, model, t_main_start_us, t_load_us, stdin, stdout, stderr); }