diff --git a/CMakeLists.txt b/CMakeLists.txt index 065185e..56ed417 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -1,8 +1,13 @@ cmake_minimum_required (VERSION 3.0) project(turbopilot VERSION 0.1.0) +set(CMAKE_EXPORT_COMPILE_COMMANDS "on") +set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin) +set(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/lib") add_subdirectory(extern/ggml) add_subdirectory(extern/argparse) add_subdirectory(extern/spdlog) -add_subdirectory(src) \ No newline at end of file +add_subdirectory(src) + +set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin) \ No newline at end of file diff --git a/include/turbopilot/gptj.hpp b/include/turbopilot/gptj.hpp new file mode 100644 index 0000000..da56cde --- /dev/null +++ b/include/turbopilot/gptj.hpp @@ -0,0 +1,83 @@ +#ifndef __TURBOPILOT_GPTJ_H +#define __TURBOPILOT_GPTJ_H + +#include + +#include +#include + +// default hparams (GPT-J 6B) +struct gptj_hparams { + int32_t n_vocab = 50400; + int32_t n_ctx = 2048; + int32_t n_embd = 4096; + int32_t n_head = 16; + int32_t n_layer = 28; + int32_t n_rot = 64; + int32_t ftype = 1; +}; + +struct gptj_layer { + // normalization + struct ggml_tensor * ln_1_g; + struct ggml_tensor * ln_1_b; + + // attention + struct ggml_tensor * c_attn_q_proj_w; + struct ggml_tensor * c_attn_k_proj_w; + struct ggml_tensor * c_attn_v_proj_w; + + struct ggml_tensor * c_attn_proj_w; + + // ff + struct ggml_tensor * c_mlp_fc_w; + struct ggml_tensor * c_mlp_fc_b; + + struct ggml_tensor * c_mlp_proj_w; + struct ggml_tensor * c_mlp_proj_b; +}; + +struct gptj_model { + gptj_hparams hparams; + + // normalization + struct ggml_tensor * ln_f_g; + struct ggml_tensor * ln_f_b; + + struct ggml_tensor * wte; // position embedding + + struct ggml_tensor * lmh_g; // language model head + struct ggml_tensor * lmh_b; // language model bias + + std::vector layers; + + // key + value memory + struct ggml_tensor * memory_k; + struct ggml_tensor * memory_v; + + // + struct ggml_context * ctx; + std::map tensors; +}; + + + +class GPTJModel : public TurbopilotModel { + +public: + GPTJModel(ModelConfig config, std::mt19937 &rng) : TurbopilotModel(config, rng){ + this->model = new gptj_model{}; + this->vocab = new gpt_vocab{}; + } + virtual ~GPTJModel(); + bool load_model(std::string path); + virtual std::stringstream predict(std::string prompt, int max_length); + +private: + gptj_model *model = NULL; + gpt_vocab *vocab = NULL; + + +}; + +#endif \ No newline at end of file diff --git a/include/turbopilot/model.hpp b/include/turbopilot/model.hpp new file mode 100644 index 0000000..a4b8977 --- /dev/null +++ b/include/turbopilot/model.hpp @@ -0,0 +1,61 @@ +#ifndef __TURBOPILOT_MODEL_H +#define __TURBOPILOT_MODEL_H + +#include +#include +#include +#include +#include +#include + +struct gpt_vocab +{ + using id = int32_t; + using token = std::string; + + std::map token_to_id; + std::map id_to_token; + std::vector special_tokens; + + void add_special_token(const std::string &token); +}; + +std::vector gpt_tokenize(const gpt_vocab &vocab, const std::string &text); + +gpt_vocab::id gpt_sample_top_k_top_p( + const gpt_vocab &vocab, + const float *logits, + int top_k, + double top_p, + double temp, + std::mt19937 &rng); + +struct ModelConfig +{ + int n_threads = 4; + int32_t top_k = 40; + float top_p = 0.95f; + float temp = 0.80f; + float repeat_penalty = 1.10f; + int32_t seed = -1; // RNG seed + int32_t n_ctx = 512; // context size + int32_t n_batch = 512; // batch size for prompt processing (must be >=32 to use BLAS) +}; + +class TurbopilotModel +{ + +public: + TurbopilotModel(ModelConfig config, std::mt19937 &rng) : + config(config), + rng(rng) + {} + virtual bool load_model(std::string model_path) = 0; + virtual std::stringstream predict(std::string prompt, int max_length) = 0; + +protected: + ModelConfig config; + std::mt19937 &rng; +}; + +#endif //__TURBOPILOT_MODEL_H \ No newline at end of file diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt new file mode 100644 index 0000000..246f15a --- /dev/null +++ b/src/CMakeLists.txt @@ -0,0 +1,21 @@ +set(TURBOPILOT_TARGET turbopilot) + + + + +add_executable(${TURBOPILOT_TARGET} + main.cpp + gptj.cpp + common.cpp + ../include/turbopilot/model.hpp + ../include/turbopilot/gptj.hpp + ) + + +target_include_directories(${TURBOPILOT_TARGET} PRIVATE + ../include + ../extern/spdlog/include +) + + +target_link_libraries(${TURBOPILOT_TARGET} PRIVATE ggml argparse) \ No newline at end of file diff --git a/src/common.cpp b/src/common.cpp new file mode 100644 index 0000000..03faaaf --- /dev/null +++ b/src/common.cpp @@ -0,0 +1,162 @@ +#include "turbopilot/model.hpp" + +#include +#include +#include + + +void gpt_vocab::add_special_token(const std::string & token) { + special_tokens.push_back(token); +} + +void gpt_split_words(std::string str, std::vector& words) { + const std::string pattern = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)"; + const std::regex re(pattern); + std::smatch m; + + while (std::regex_search(str, m, re)) { + for (auto x : m) { + words.push_back(x); + } + str = m.suffix(); + } +} + +std::vector gpt_tokenize(const gpt_vocab & vocab, const std::string & text) { + std::vector words; + + // first split the text into words + { + std::string str = text; + + // Generate the subpattern from the special_tokens vector if it's not empty + if (!vocab.special_tokens.empty()) { + const std::regex escape(R"([\[\\\^\$\.\|\?\*\+\(\)\{\}])"); + std::string special_tokens_subpattern; + for (const auto & token : vocab.special_tokens) { + if (!special_tokens_subpattern.empty()) { + special_tokens_subpattern += "|"; + } + special_tokens_subpattern += std::regex_replace(token, escape, R"(\$&)"); + } + + std::regex re(special_tokens_subpattern); + std::smatch m; + // Split the text by special tokens. + while (std::regex_search(str, m, re)) { + // Split the substrings in-between special tokens into words. + gpt_split_words(m.prefix(), words); + // Add matched special tokens as words. + for (auto x : m) { + words.push_back(x); + } + str = m.suffix(); + } + // Remaining text without special tokens will be handled below. + } + + gpt_split_words(str, words); + } + + // find the longest token that forms each word in words: + std::vector tokens; + for (const auto & word : words) { + for (int i = 0; i < (int) word.size(); ){ + for (int j = word.size() - 1; j >= i; j--){ + auto cand = word.substr(i, j-i+1); + auto it = vocab.token_to_id.find(cand); + if (it != vocab.token_to_id.end()){ // word.substr(i, j-i+1) in vocab + tokens.push_back(it->second); + i = j + 1; + break; + } + else if (j == i){ // word.substr(i, 1) has no matching + fprintf(stderr, "%s: unknown token '%s'\n", __func__, word.substr(i, 1).data()); + i++; + } + } + } + } + + return tokens; +} + +gpt_vocab::id gpt_sample_top_k_top_p( + const gpt_vocab & vocab, + const float * logits, + int top_k, + double top_p, + double temp, + std::mt19937 & rng) { + int n_logits = vocab.id_to_token.size(); + + std::vector> logits_id; + logits_id.reserve(n_logits); + + { + const double scale = 1.0/temp; + for (int i = 0; i < n_logits; ++i) { + logits_id.push_back(std::make_pair(logits[i]*scale, i)); + } + } + + // find the top K tokens + std::partial_sort( + logits_id.begin(), + logits_id.begin() + top_k, logits_id.end(), + [](const std::pair & a, const std::pair & b) { + return a.first > b.first; + }); + + logits_id.resize(top_k); + + double maxl = -INFINITY; + for (const auto & kv : logits_id) { + maxl = std::max(maxl, kv.first); + } + + // compute probs for the top K tokens + std::vector probs; + probs.reserve(logits_id.size()); + + double sum = 0.0; + for (const auto & kv : logits_id) { + double p = exp(kv.first - maxl); + probs.push_back(p); + sum += p; + } + + // normalize the probs + for (auto & p : probs) { + p /= sum; + } + + if (top_p < 1.0f) { + double cumsum = 0.0f; + for (int i = 0; i < top_k; i++) { + cumsum += probs[i]; + if (cumsum >= top_p) { + top_k = i + 1; + probs.resize(top_k); + logits_id.resize(top_k); + break; + } + } + + cumsum = 1.0/cumsum; + for (int i = 0; i < (int) probs.size(); i++) { + probs[i] *= cumsum; + } + } + + //printf("\n"); + //for (int i = 0; i < (int) probs.size(); i++) { + // printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]); + //} + //exit(0); + + std::discrete_distribution<> dist(probs.begin(), probs.end()); + int idx = dist(rng); + + return logits_id[idx].second; +} \ No newline at end of file diff --git a/src/gptj.cpp b/src/gptj.cpp new file mode 100644 index 0000000..ba8eae7 --- /dev/null +++ b/src/gptj.cpp @@ -0,0 +1,642 @@ +#include +#include + +#include + +#include +#include + +#if defined(_MSC_VER) +#pragma warning(disable: 4244 4267) // possible loss of data +#endif + + +// 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 gptj_eval( + const gptj_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_rot; + + static size_t buf_size = 256u*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 + //printf("\n{}: 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) { + spdlog::error("{}: failed to allocate {} bytes\n", __func__, buf_size); + return false; + } + } + + struct ggml_init_params params = { + /*.mem_size =*/ buf_size, + /*.mem_buffer =*/ buf, + /*.no_alloc =*/ false, + }; + + struct ggml_context * ctx0 = ggml_init(params); + struct ggml_cgraph gf = {}; + + struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N); + memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd)); + + // wte + struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte, embd); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * cur; + + // norm + { + cur = ggml_norm(ctx0, inpL); + + // cur = ln_1_g*cur + ln_1_b + cur = ggml_add(ctx0, + ggml_mul(ctx0, + ggml_repeat(ctx0, model.layers[il].ln_1_g, cur), + cur), + ggml_repeat(ctx0, model.layers[il].ln_1_b, cur)); + } + + struct ggml_tensor * inpSA = cur; + + // self-attention + { + struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0); + struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0); + + // store key and value to memory + { + struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_v_proj_w, cur)); + + 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_2d(ctx0, model.memory_v, N, n_embd, + ( n_ctx)*ggml_element_size(model.memory_v), + (il*n_ctx)*ggml_element_size(model.memory_v)*n_embd + n_past*ggml_element_size(model.memory_v)); + + 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, + Qcur, + 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_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), + 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_inplace(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_inplace(ctx0, KQ_scaled, n_past); + + // KQ = soft_max(KQ_masked) + struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(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 = + ggml_view_3d(ctx0, model.memory_v, + n_past + N, n_embd/n_head, n_head, + n_ctx*ggml_element_size(model.memory_v), + n_ctx*ggml_element_size(model.memory_v)*n_embd/n_head, + il*n_ctx*ggml_element_size(model.memory_v)*n_embd); + + // KQV = transpose(V) * KQ_soft_max + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, 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].c_attn_proj_w, + cur); + } + + struct ggml_tensor * inpFF = cur; + + // feed-forward network + // this is independent of the self-attention result, so it could be done in parallel to the self-attention + { + // note here we pass inpSA instead of cur + cur = ggml_mul_mat(ctx0, + model.layers[il].c_mlp_fc_w, + inpSA); + + cur = ggml_add(ctx0, + ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur), + cur); + + // GELU activation + cur = ggml_gelu(ctx0, cur); + + // projection + // cur = proj_w*cur + proj_b + cur = ggml_mul_mat(ctx0, + model.layers[il].c_mlp_proj_w, + cur); + + cur = ggml_add(ctx0, + ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur), + cur); + } + + // self-attention + FF + cur = ggml_add(ctx0, cur, inpFF); + + // input for next layer + inpL = ggml_add(ctx0, cur, inpL); + } + + // norm + { + inpL = ggml_norm(ctx0, inpL); + + // inpL = ln_f_g*inpL + ln_f_b + inpL = ggml_add(ctx0, + ggml_mul(ctx0, + ggml_repeat(ctx0, model.ln_f_g, inpL), + inpL), + ggml_repeat(ctx0, model.ln_f_b, inpL)); + } + + // lm_head + { + inpL = ggml_mul_mat(ctx0, model.lmh_g, inpL); + + inpL = ggml_add(ctx0, + ggml_repeat(ctx0, model.lmh_b, inpL), + inpL); + } + + // logits -> probs + //inpL = ggml_soft_max_inplace(ctx0, inpL); + + // run the computation + ggml_build_forward_expand(&gf, inpL); + ggml_graph_compute_with_ctx(ctx0, &gf, n_threads); + + //if (n_past%100 == 0) { + // ggml_graph_print (&gf); + // ggml_graph_dump_dot(&gf, NULL, "gpt-j.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; + } + //printf("used_mem = %zu\n", ggml_used_mem(ctx0)); + + ggml_free(ctx0); + + return true; +} + + +GPTJModel::~GPTJModel(){ + ggml_free(model->ctx); + free(model); +} + +bool GPTJModel::load_model(std::string fname) { + spdlog::info("{}: loading model from '{}' - please wait ...\n", __func__, fname.c_str()); + + auto fin = std::ifstream(fname, std::ios::binary); + if (!fin) { + spdlog::error("{}: failed to open '{}'\n", __func__, fname.c_str()); + return false; + } + + // verify magic + { + uint32_t magic; + fin.read((char *) &magic, sizeof(magic)); + if (magic != GGML_FILE_MAGIC) { + spdlog::error("{}: invalid model file '{}' (bad magic)\n", __func__, fname.c_str()); + return false; + } + } + + // 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_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.ftype, sizeof(hparams.ftype)); + + const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR; + + spdlog::info("{}: n_vocab = {}\n", __func__, hparams.n_vocab); + spdlog::info("{}: n_ctx = {}\n", __func__, hparams.n_ctx); + spdlog::info("{}: n_embd = {}\n", __func__, hparams.n_embd); + spdlog::info("{}: n_head = {}\n", __func__, hparams.n_head); + spdlog::info("{}: n_layer = {}\n", __func__, hparams.n_layer); + spdlog::info("{}: n_rot = {}\n", __func__, hparams.n_rot); + spdlog::info("{}: ftype = {}\n", __func__, hparams.ftype); + spdlog::info("{}: qntvr = {}\n", __func__, qntvr); + + hparams.ftype %= GGML_QNT_VERSION_FACTOR; + } + + // load vocab + { + int32_t n_vocab = 0; + fin.read((char *) &n_vocab, sizeof(n_vocab)); + + if (n_vocab != model->hparams.n_vocab) { + spdlog::error("{}: invalid model file '{}' (bad vocab size {} != {})\n", + __func__, fname.c_str(), n_vocab, model->hparams.n_vocab); + return false; + } + + std::string word; + std::vector buf(128); + + for (int i = 0; i < n_vocab; i++) { + uint32_t len; + fin.read((char *) &len, sizeof(len)); + + buf.resize(len); + fin.read((char *) buf.data(), len); + word.assign(buf.data(), len); + + vocab->token_to_id[word] = i; + vocab->id_to_token[i] = word; + } + } + + // 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_ftype_to_ggml_type((ggml_ftype) (model->hparams.ftype)); + if (wtype == GGML_TYPE_COUNT) { + spdlog::error("{}: invalid model file '{}' (bad ftype value {})\n", + __func__, fname.c_str(), model->hparams.ftype); + return false; + } + + 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*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g + ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b + + ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // wte + + ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // lmh_g + ctx_size += n_vocab*ggml_type_sizef(GGML_TYPE_F32); // lmh_b + + ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g + ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b + + ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_q_proj_w + ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_k_proj_w + ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_v_proj_w + + ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_proj_w + + ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype)); // c_mlp_fc_w + ctx_size += n_layer*( 4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b + + ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype)); // c_mlp_proj_w + ctx_size += n_layer*( n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b + + ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F16); // memory_k + ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F16); // memory_v + + ctx_size += (5 + 10*n_layer)*512; // object overhead + + spdlog::info("{}: ggml ctx size = {} 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, + /*.no_alloc =*/ false, + }; + + model->ctx = ggml_init(params); + if (!model->ctx) { + spdlog::error("{}: 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_vocab = hparams.n_vocab; + + model->layers.resize(n_layer); + + model->wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab); + + model->ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + model->ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + + model->lmh_g = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab); + model->lmh_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_vocab); + + // map by name + model->tensors["transformer.wte.weight"] = model->wte; + + model->tensors["transformer.ln_f.weight"] = model->ln_f_g; + model->tensors["transformer.ln_f.bias"] = model->ln_f_b; + + model->tensors["lm_head.weight"] = model->lmh_g; + model->tensors["lm_head.bias"] = model->lmh_b; + + for (int i = 0; i < n_layer; ++i) { + auto & layer = model->layers[i]; + + layer.ln_1_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + layer.ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + + layer.c_attn_q_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); + layer.c_attn_k_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); + layer.c_attn_v_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); + + layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); + + layer.c_mlp_fc_w = ggml_new_tensor_2d(ctx, wtype, n_embd, 4*n_embd); + layer.c_mlp_fc_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd); + + layer.c_mlp_proj_w = ggml_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd); + layer.c_mlp_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + + // map by name + model->tensors["transformer.h." + std::to_string(i) + ".ln_1.weight"] = layer.ln_1_g; + model->tensors["transformer.h." + std::to_string(i) + ".ln_1.bias"] = layer.ln_1_b; + + model->tensors["transformer.h." + std::to_string(i) + ".attn.q_proj.weight"] = layer.c_attn_q_proj_w; + model->tensors["transformer.h." + std::to_string(i) + ".attn.k_proj.weight"] = layer.c_attn_k_proj_w; + model->tensors["transformer.h." + std::to_string(i) + ".attn.v_proj.weight"] = layer.c_attn_v_proj_w; + + model->tensors["transformer.h." + std::to_string(i) + ".attn.out_proj.weight"] = layer.c_attn_proj_w; + + model->tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.weight"] = layer.c_mlp_fc_w; + model->tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.bias"] = layer.c_mlp_fc_b; + + model->tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.weight"] = layer.c_mlp_proj_w; + model->tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.bias"] = layer.c_mlp_proj_b; + } + } + + // 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_F16, n_elements); + model->memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements); + + const size_t memory_size = ggml_nbytes(model->memory_k) + ggml_nbytes(model->memory_v); + + spdlog::info("{}: memory_size = {} MB, n_mem = {}\n", __func__, memory_size/1024.0/1024.0, n_mem); + } + + // load weights + { + int n_tensors = 0; + size_t total_size = 0; + + spdlog::info("{}: ", __func__); + + while (true) { + int32_t n_dims; + int32_t length; + int32_t ttype; + + fin.read(reinterpret_cast(&n_dims), sizeof(n_dims)); + fin.read(reinterpret_cast(&length), sizeof(length)); + fin.read(reinterpret_cast(&ttype), sizeof(ttype)); + + 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()) { + spdlog::error("{}: unknown tensor '{}' in model file\n", __func__, name.data()); + return false; + } + + auto tensor = model->tensors[name.data()]; + if (ggml_nelements(tensor) != nelements) { + spdlog::error("{}: tensor '{}' has wrong size in model file\n", __func__, name.data()); + return false; + } + + if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) { + spdlog::error("{}: tensor '{}' has wrong shape in model file: got [{}, {}], expected [{}, {}]\n", + __func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]); + return false; + } + + // for debugging + if (0) { + printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor)); + } + + const size_t bpe = ggml_type_size(ggml_type(ttype)); + + if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) { + spdlog::error("{}: tensor '{}' has wrong size in model file: got %zu, expected %zu\n", + __func__, name.data(), ggml_nbytes(tensor), nelements*bpe); + return false; + } + + fin.read(reinterpret_cast(tensor->data), ggml_nbytes(tensor)); + + //printf("%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ttype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0); + total_size += ggml_nbytes(tensor); + if (++n_tensors % 8 == 0) { + printf("."); + fflush(stdout); + } + + } + + printf("\n"); + spdlog::info(" done\n"); + + spdlog::info("{}: model size = {:06.2f} MB / num tensors = {}\n", __func__, total_size/1024.0/1024.0, n_tensors); + } + + fin.close(); + + return true; +} + +std::stringstream GPTJModel::predict(std::string prompt, int max_length) { + + std::stringstream result; + // tokenize the prompt + std::vector embd_inp = ::gpt_tokenize((*vocab), prompt); + + int n_past = 0; + + int64_t t_sample_us = 0; + int64_t t_predict_us = 0; + + int n_predict = std::min(max_length, model->hparams.n_ctx - (int) embd_inp.size()); + + spdlog::debug("{}: number of tokens in prompt = {}", __func__, embd_inp.size()); + + std::vector embd; + + // determine the required inference memory per token: + size_t mem_per_token = 0; + + std::vector logits; + + gptj_eval((*model), config.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token); + + for (int i = embd.size(); i < embd_inp.size() + n_predict; i++) { + // predict + if (embd.size() > 0) { + const int64_t t_start_us = ggml_time_us(); + + if (!gptj_eval((*model), config.n_threads, n_past, embd, logits, mem_per_token)) { + throw std::runtime_error("Failed to predict"); + } + + t_predict_us += ggml_time_us() - t_start_us; + } + + n_past += embd.size(); + embd.clear(); + + if (i >= embd_inp.size()) { + // sample next token + const int top_k = config.top_k; + const float top_p = config.top_p; + const float temp = config.temp; + + const int n_vocab = model->hparams.n_vocab; + + gpt_vocab::id id = 0; + + { + const int64_t t_start_sample_us = ggml_time_us(); + + id = gpt_sample_top_k_top_p((*vocab), logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, rng); + + t_sample_us += ggml_time_us() - t_start_sample_us; + } + + // add it to the context + embd.push_back(id); + } else { + // if here, it means we are still processing the input prompt + for (int k = i; k < embd_inp.size(); k++) { + embd.push_back(embd_inp[k]); + if (embd.size() > config.n_batch) { + break; + } + } + i += embd.size() - 1; + } + + // display text + for (auto id : embd) { + result << vocab->id_to_token[id].c_str(); + //printf("%s", vocab->id_to_token[id].c_str()); + } + fflush(stdout); + + // end of text token + if (embd.back() == 50256) { + break; + } + } + + return result; +} diff --git a/src/main.cpp b/src/main.cpp new file mode 100644 index 0000000..32b8717 --- /dev/null +++ b/src/main.cpp @@ -0,0 +1,84 @@ +#include +#include +#include + +#include + +#include +#include "turbopilot/model.hpp" + +#include "turbopilot/gptj.hpp" + +int main(int argc, char **argv) +{ + + argparse::ArgumentParser program("turbopilot"); + + program.add_argument("-f", "--model-file") + .help("Path to the model that turbopilot should serve") + .required(); + + program.add_argument("-t", "--model-type") + .help("The type of model to load. Can be codegen/gpt-j or starcoder architectures.") + .default_value("codegen"); + + program.add_argument("-p", "--port") + .help("The tcp port that turbopilot should listen on") + .default_value("18080"); + + program.add_argument("-r", "--random-seed") + .help("Set the random seed for RNG functions") + .default_value(-1) + .scan<'i', int>(); + + + try + { + program.parse_args(argc, argv); + } + catch (const std::runtime_error &err) + { + std::cerr << err.what() << std::endl; + std::cerr << program; + return 1; + } + + ggml_time_init(); + + const int64_t t_main_start_us = ggml_time_us(); + + + TurbopilotModel *model = NULL; + + auto model_type = program.get("--model-type"); + + ModelConfig config{}; + std::mt19937 rng(program.get("--random-seed")); + + if(model_type.compare("codegen") == 0) { + spdlog::info("Initializing GPT-J type model for '{}' model", model_type); + model = new GPTJModel(config, rng); + }else{ + spdlog::error("Invalid model type: {}", model_type); + } + + spdlog::info("Attempt to load model from {}", program.get("--model-type")); + int64_t t_load_us = 0; + const int64_t t_start_us = ggml_time_us(); + auto loaded = model->load_model(program.get("--model-file")); + + if(!loaded){ + spdlog::error("Failed to load model"); + return -1; + } + + t_load_us = ggml_time_us() - t_start_us; + + spdlog::info("Loaded model in {:0.2f}ms", t_load_us/1000.0f); + + auto result = model->predict("test", 100); + + spdlog::info("output: {}", result.str()); + + free(model); +} \ No newline at end of file