main-sched.cpp

#include "ggml/ggml.h"
#include "ggml/ggml-alloc.h"
#include "ggml/ggml-backend.h"

#ifdef GGML_USE_CUDA
#include "ggml-cuda.h"
#endif

#ifdef GGML_USE_METAL
#include "ggml-metal.h"
#endif

#include "common.h"
#include "common-ggml.h"

#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>

#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif

#define GPT2_MAX_NODES 4096

static void ggml_log_callback_default(ggml_log_level level, const char * text, void * user_data) {
    (void) level;
    (void) user_data;
    fputs(text, stderr);
    fflush(stderr);
}

// default hparams (GPT-2 117M)
struct gpt2_hparams {
    int32_t n_vocab = 50257;
    int32_t n_ctx   = 1024;
    int32_t n_embd  = 768;
    int32_t n_head  = 12;
    int32_t n_layer = 12;
    int32_t ftype   = 1;
    float   eps     = 1e-5f;
};

struct gpt2_layer {
    // normalization
    struct ggml_tensor * ln_1_g;
    struct ggml_tensor * ln_1_b;

    struct ggml_tensor * ln_2_g;
    struct ggml_tensor * ln_2_b;

    // attention
    struct ggml_tensor * c_attn_attn_w;
    struct ggml_tensor * c_attn_attn_b;

    struct ggml_tensor * c_attn_proj_w;
    struct ggml_tensor * c_attn_proj_b;

    // mlp
    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 gpt2_model {
    gpt2_hparams hparams;

    // normalization
    struct ggml_tensor * ln_f_g;
    struct ggml_tensor * ln_f_b;

    struct ggml_tensor * wte;     // position embedding
    struct ggml_tensor * wpe;     //    token embedding
    struct ggml_tensor * lm_head; // language model head

    std::vector<gpt2_layer> layers;

    // key + value memory
    struct ggml_tensor * memory_k;
    struct ggml_tensor * memory_v;

    //
    struct ggml_context * ctx_w;

    std::vector<ggml_backend_t> backends;
    std::vector<ggml_backend_buffer_t> buffers_w;
    ggml_backend_buffer_t buffer_kv;
    ggml_backend_buffer_t buffer_input;

    std::map<std::string, struct ggml_tensor *> tensors;

    // inputs/constants
    struct ggml_tensor * embd;
    struct ggml_tensor * position;
};

void init_backends(gpt2_model & model, const gpt_params & params) {
    ggml_backend_t gpu_backend = NULL;

    // initialize the backends
#ifdef GGML_USE_CUDA
    if (params.n_gpu_layers > 0) {
        fprintf(stderr, "%s: using CUDA backend\n", __func__);
        gpu_backend = ggml_backend_cuda_init(0);
        if (!gpu_backend) {
            fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
        }
    }
#endif

#ifdef GGML_USE_METAL
    if (params.n_gpu_layers > 0) {
        fprintf(stderr, "%s: using Metal backend\n", __func__);
        ggml_backend_metal_log_set_callback(ggml_log_callback_default, nullptr);
        gpu_backend = ggml_backend_metal_init();
        if (!gpu_backend) {
            fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
        } else {
            ggml_backend_metal_set_n_cb(gpu_backend, params.n_threads);
        }
    }
#endif
    if (gpu_backend) {
        model.backends.push_back(gpu_backend);
    }

    // always add the CPU backend as a fallback
    ggml_backend_t cpu_backend = ggml_backend_cpu_init();
    ggml_backend_cpu_set_n_threads(cpu_backend, params.n_threads);
    model.backends.push_back(cpu_backend);
}

// load the model's weights from a file
bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab, const gpt_params & params) {
    printf("%s: loading model from '%s'\n", __func__, fname.c_str());

    auto fin = std::ifstream(fname, std::ios::binary);
    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 != GGML_FILE_MAGIC) {
            fprintf(stderr, "%s: invalid model file '%s' (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.ftype,   sizeof(hparams.ftype));

        const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;

        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
        printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
        printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
        printf("%s: n_head  = %d\n", __func__, hparams.n_head);
        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
        printf("%s: ftype   = %d\n", __func__, hparams.ftype);
        printf("%s: qntvr   = %d\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) {
            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
                    __func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
            return false;
        }

        std::string word;
        std::vector<char> 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) {
        fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
                __func__, fname.c_str(), model.hparams.ftype);
        return false;
    }

    auto & ctx = model.ctx_w;

    // create the ggml context
    {
        size_t n_tensors = 3 /* input */ + 2 /* kv */ + 6 + 12*model.hparams.n_layer;
        struct ggml_init_params params = {
            /*.mem_size   =*/ ggml_tensor_overhead() * n_tensors,
            /*.mem_buffer =*/ NULL,
            /*.no_alloc   =*/ true,
        };

        model.ctx_w = ggml_init(params);
        if (!model.ctx_w) {
            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
            return false;
        }
    }

    // create tensors 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.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.wte     = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
        model.wpe     = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
        model.lm_head = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);

        // map by name
        model.tensors["model/ln_f/g"] = model.ln_f_g;
        model.tensors["model/ln_f/b"] = model.ln_f_b;

        model.tensors["model/wte"]     = model.wte;
        model.tensors["model/wpe"]     = model.wpe;
        model.tensors["model/lm_head"] = model.lm_head;

        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.ln_2_g        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
            layer.ln_2_b        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);

            layer.c_attn_attn_w = ggml_new_tensor_2d(ctx, wtype,           n_embd, 3*n_embd);
            layer.c_attn_attn_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*n_embd);

            layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd, n_embd);
            layer.c_attn_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   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["model/h" + std::to_string(i) + "/ln_1/g"]        = layer.ln_1_g;
            model.tensors["model/h" + std::to_string(i) + "/ln_1/b"]        = layer.ln_1_b;

            model.tensors["model/h" + std::to_string(i) + "/ln_2/g"]        = layer.ln_2_g;
            model.tensors["model/h" + std::to_string(i) + "/ln_2/b"]        = layer.ln_2_b;

            model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/w"] = layer.c_attn_attn_w;
            model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/b"] = layer.c_attn_attn_b;

            model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/w"] = layer.c_attn_proj_w;
            model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/b"] = layer.c_attn_proj_b;

            model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/w"]    = layer.c_mlp_fc_w;
            model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/b"]    = layer.c_mlp_fc_b;

            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"]  = layer.c_mlp_proj_w;
            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"]  = layer.c_mlp_proj_b;
        }
    }

    // assign tensors to backends
    init_backends(model, params);
    ggml_backend_t backend_gpu = model.backends.front();
    ggml_backend_t backend_cpu = model.backends.back();
    std::map<std::string, ggml_backend_t> tensor_backends;
    {
        const int i_gpu_first_layer = model.hparams.n_layer - params.n_gpu_layers;
        for (auto it : model.tensors) {
            const std::string & name = it.first;
            // input tensors
            if (name == "model/wte" || name == "model/wpe") {
                if (params.n_gpu_layers > model.hparams.n_layer) {
                    tensor_backends[name] = backend_gpu;
                } else {
                    tensor_backends[name] = backend_cpu;
                }
            }
            // output tensors
            if (name == "model/ln_f/g" || name == "model/ln_f/b" || name == "model/lm_head") {
                if (params.n_gpu_layers > 0) {
                    tensor_backends[name] = backend_gpu;
                } else {
                    tensor_backends[name] = backend_cpu;
                }
            }
            // layer tensors
            if (name.substr(0, 7) == "model/h") {
                // parse layer number
                int layer = std::stoi(name.substr(7, 2));
                if (layer >= i_gpu_first_layer) {
                    tensor_backends[name] = backend_gpu;
                } else {
                    tensor_backends[name] = backend_cpu;
                }
            }
        }
    }

    // allocate buffers
    std::map<ggml_backend_t, ggml_tallocr> backend_buffers;
    for (auto backend : model.backends) {
        // compute the size of the buffer
        size_t size = 0;
        for (auto it : model.tensors) {
            if (tensor_backends[it.first] == backend) {
                size += ggml_nbytes(it.second) + 512;
            }
        }
        if (size > 0) {
            printf("%s: %8s buffer size = %8.2f MB\n", __func__, ggml_backend_name(backend), size/1024.0/1024.0);
            // allocate the buffer
            ggml_backend_buffer_t buffer = ggml_backend_alloc_buffer(backend, size);
            ggml_backend_buffer_set_usage(buffer, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
            model.buffers_w.push_back(buffer);

            // create an allocator for the buffer to allocate the tensors
            auto alloc = ggml_tallocr_new(buffer);
            backend_buffers.insert(std::make_pair(backend, std::move(alloc)));
        } else {
            model.buffers_w.push_back(NULL);
        }
    }

    // allocate 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);

        ggml_set_name(model.memory_k, "model/memory_k");
        ggml_set_name(model.memory_v, "model/memory_v");

        const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);

        printf("%s: memory size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);

        // create a backend buffer (can be in host or device memory)
        ggml_backend_t backend_kv = params.n_gpu_layers >= hparams.n_layer/2 ? backend_gpu : backend_cpu;
        printf("%s: backend_kv = %s\n", __func__, ggml_backend_name(backend_kv));
        model.buffer_kv = ggml_backend_alloc_buffer(backend_kv, memory_size + 512*2);

        // allocate the tensors into the backend buffer
        {
            ggml_tallocr alloc = ggml_tallocr_new(model.buffer_kv);

            // this updates the pointers in the tensors to point to the correct location in the buffer
            // this is necessary since the ggml_context is .no_alloc == true
            // note that the buffer can actually be a device buffer, depending on the backend
            ggml_tallocr_alloc(&alloc, model.memory_k);
            ggml_tallocr_alloc(&alloc, model.memory_v);
        }
    }

    // load weights
    {
        size_t total_size = 0;

        bool has_lm_head = false;

        std::vector<char> read_buf;

        while (true) {
            int32_t n_dims;
            int32_t length;
            int32_t ttype;

            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
            fin.read(reinterpret_cast<char *>(&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<char *>(&ne[i]), sizeof(ne[i]));
                nelements *= ne[i];
            }

            std::string name(length, 0);
            fin.read(&name[0], length);

            if (model.tensors.find(name) == model.tensors.end()) {
                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.c_str());
                return false;
            }

            auto tensor = model.tensors[name];
            ggml_set_name(tensor, name.c_str());
            if (ggml_nelements(tensor) != nelements) {
                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.c_str());
                return false;
            }

            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.c_str(), (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.c_str(), 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)) {
                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
                        __func__, name.c_str(), ggml_nbytes(tensor), nelements*bpe);
                return false;
            }

            // allocate the tensor
            ggml_backend_t backend = tensor_backends[name];
            ggml_tallocr * alloc = &backend_buffers.find(backend)->second;
            ggml_tallocr_alloc(alloc, tensor);
            //printf("%s: [%5.5s] %s\n", __func__, ggml_backend_name(backend), name.c_str());

            if (ggml_backend_is_cpu(backend)
#ifdef GGML_USE_METAL
                || ggml_backend_is_metal(backend)
#endif
                ) {
                // for the CPU and Metal backend, we can read directly into the tensor
                fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
            } else {
                // read into a temporary buffer first, then copy to device memory
                read_buf.resize(ggml_nbytes(tensor));
                fin.read(read_buf.data(), ggml_nbytes(tensor));
                ggml_backend_tensor_set(tensor, read_buf.data(), 0, ggml_nbytes(tensor));
            }

            // GPT-2 models share the WTE tensor as the LM head
            if (name == "model/wte" && has_lm_head == false) {
                ggml_tallocr * alloc_head = &backend_buffers.find(tensor_backends["model/lm_head"])->second;
                ggml_tallocr_alloc(alloc_head, model.lm_head);
                //printf("%s: [%5.5s] %s (copied)\n", __func__, ggml_backend_name(tensor_backends["model/lm_head"]), "model/lm_head");
                ggml_backend_tensor_copy(tensor, model.lm_head);
                total_size += ggml_nbytes(model.lm_head);
            }

            if (name == "model/lm_head") {
                has_lm_head = true;
            }

            total_size += ggml_nbytes(tensor);
        }
        printf("%s: model size  = %8.2f MB\n", __func__, total_size/1024.0/1024.0);
    }

    fin.close();

    // allocate input tensors
    {
        model.embd = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, model.hparams.n_ctx);
        model.position = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, model.hparams.n_ctx);

        ggml_set_name(model.embd, "in/embd");
        ggml_set_name(model.position, "in/position");

        // add input tensors to cpu backend
        size_t input_size = ggml_nbytes(model.embd) + ggml_nbytes(model.position);

        // FIXME: use cpu backend after sched impl
        ggml_backend_t backend_input = params.n_gpu_layers >= model.hparams.n_layer ? backend_gpu : backend_cpu;
        model.buffer_input = ggml_backend_alloc_buffer(backend_input, input_size + 512*3);
        printf("%s: backend_in = %s (%zu bytes)\n", __func__, ggml_backend_name(backend_input), input_size);

        // allocate the tensors into the backend buffer
        ggml_tallocr alloc = ggml_tallocr_new(model.buffer_input);
        ggml_tallocr_alloc(&alloc, model.embd);
        ggml_tallocr_alloc(&alloc, model.position);
    }

    return true;
}

// build the computation graph
struct ggml_cgraph * gpt2_graph(
        const gpt2_model & model,
        const int n_past,
        const std::vector<gpt_vocab::id> & embd_inp) {
    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;

    // since we are using ggml-alloc, this buffer only needs enough space to hold the ggml_tensor and ggml_cgraph structs, but not the tensor data
    static size_t buf_size = ggml_tensor_overhead()*GPT2_MAX_NODES + ggml_graph_overhead_custom(GPT2_MAX_NODES, false);
    static std::vector<uint8_t> buf(buf_size);

    struct ggml_init_params params = {
        /*.mem_size   =*/ buf_size,
        /*.mem_buffer =*/ buf.data(),
        /*.no_alloc   =*/ true, // the tensors will be allocated later by ggml_gallocr_alloc_graph()
    };

    struct ggml_context * ctx = ggml_init(params);

    struct ggml_cgraph  * gf = ggml_new_graph_custom(ctx, GPT2_MAX_NODES, false);

    struct ggml_tensor * embd = ggml_view_1d(ctx, model.embd, N, 0);

    // set inputs
    // TODO: move to gpt2_eval
    ggml_backend_tensor_set(model.embd, embd_inp.data(), 0, N*ggml_element_size(embd));

    struct ggml_tensor * position = ggml_view_1d(ctx, model.position, N, 0);
    for (int i = 0; i < N; ++i) {
        int32_t v = n_past + i;
        ggml_backend_tensor_set(model.position, &v, i*sizeof(int32_t), sizeof(v));
    }

    const float KQ_scale = 1.0f/sqrtf(float(model.hparams.n_embd)/model.hparams.n_head);

    // wte + wpe
    struct ggml_tensor * inpL =
        ggml_add(ctx,
                ggml_get_rows(ctx, model.wte, embd),
                ggml_get_rows(ctx, model.wpe, position));
    ggml_set_name(inpL, "inpL");
    ggml_set_name(inpL->src[0], "wte");
    ggml_set_name(inpL->src[1], "wpe");

    for (int il = 0; il < n_layer; ++il) {
        struct ggml_tensor * cur;

        // norm
        {
            // [ 768, N]
            cur = ggml_norm(ctx, inpL, hparams.eps);
            ggml_format_name(cur, "l%d.norm", il);

            // cur = ln_1_g*cur + ln_1_b
            // [ 768, N]
            cur = ggml_add(ctx,
                    ggml_mul(ctx,
                        cur,
                        model.layers[il].ln_1_g),
                    model.layers[il].ln_1_b);
            ggml_format_name(cur, "l%d.ln_1_b", il);
            ggml_format_name(cur->src[0], "l%d.ln_1_g", il);
        }

        // attn
        // [2304, 768] - model.layers[il].c_attn_attn_w
        // [2304,   1] - model.layers[il].c_attn_attn_b
        // [ 768,   N] - cur (in)
        // [2304,   N] - cur (out)
        //
        // cur = attn_w*cur + attn_b
        // [2304, N]
        {
            cur = ggml_mul_mat(ctx,
                    model.layers[il].c_attn_attn_w,
                    cur);
            ggml_format_name(cur, "l%d.attn_w", il);

            cur = ggml_add(ctx,
                    cur,
                    model.layers[il].c_attn_attn_b);
            ggml_format_name(cur, "l%d.attn_b", il);
        }

        // self-attention
        {
            struct ggml_tensor * Qcur = ggml_view_2d(ctx, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
            struct ggml_tensor * Kcur = ggml_view_2d(ctx, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
            struct ggml_tensor * Vcur = ggml_view_2d(ctx, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);

            ggml_format_name(Qcur, "l%d.Qcur", il);
            ggml_format_name(Kcur, "l%d.Kcur", il);
            ggml_format_name(Vcur, "l%d.Vcur", il);

            // store key and value to memory
            if (N >= 1) {
                struct ggml_tensor * k = ggml_view_1d(ctx, 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(ctx, 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(ctx, Kcur, k));
                ggml_build_forward_expand(gf, ggml_cpy(ctx, Vcur, v));
            }

            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
            // [64, N, 12]
            struct ggml_tensor * Q =
                ggml_permute(ctx,
                        ggml_cont_3d(ctx, Qcur, n_embd/n_head, n_head, N),
                        0, 2, 1, 3);
            ggml_format_name(Q, "l%d.Q", il);

            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
            // [64, n_past + N, 12]
            struct ggml_tensor * K =
                ggml_permute(ctx,
                        ggml_reshape_3d(ctx,
                            ggml_view_1d(ctx, 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);
            ggml_format_name(K, "l%d.K", il);

            // GG: flash attention
            //struct ggml_tensor * V =
            //    ggml_cpy(ctx0,
            //            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),
            //            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_past + N, n_embd/n_head, n_head));

            //struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, true);

            // K * Q
            // [n_past + N, N, 12]
            struct ggml_tensor * KQ = ggml_mul_mat(ctx, K, Q);
            ggml_format_name(KQ, "l%d.KQ", il);

            // KQ_scaled = KQ / sqrt(n_embd/n_head)
            // [n_past + N, N, 12]
            struct ggml_tensor * KQ_scaled = ggml_scale(ctx, KQ, KQ_scale);
            ggml_format_name(KQ_scaled, "l%d.KQ_scaled", il);

            // KQ_masked = mask_past(KQ_scaled)
            // [n_past + N, N, 12]
            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx, KQ_scaled, n_past);
            ggml_format_name(KQ_masked, "l%d.KQ_masked", il);

            // KQ = soft_max(KQ_masked)
            // [n_past + N, N, 12]
            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx, KQ_masked);
            ggml_format_name(KQ_soft_max, "l%d.KQ_soft_max", il);

            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
            // [n_past + N, 64, 12]
            struct ggml_tensor * V_trans =
                ggml_cont_3d(ctx,
                        ggml_permute(ctx,
                            ggml_reshape_3d(ctx,
                                ggml_view_1d(ctx, 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),
                        n_past + N, n_embd/n_head, n_head);

            // KQV = transpose(V) * KQ_soft_max
            // [64, N, 12]
            struct ggml_tensor * KQV = ggml_mul_mat(ctx, V_trans, KQ_soft_max);
            ggml_format_name(KQV, "l%d.KQV", il);

            // KQV_merged = KQV.permute(0, 2, 1, 3)
            // [64, 12, N]
            struct ggml_tensor * KQV_merged = ggml_permute(ctx, KQV, 0, 2, 1, 3);
            ggml_format_name(KQV_merged, "l%d.KQV_merged", il);

            // cur = KQV_merged.contiguous().view(n_embd, N)
            // [768, N]
            cur = ggml_cont_2d(ctx, KQV_merged, n_embd, N);
            ggml_format_name(cur, "l%d.KQV_merged_contiguous", il);
        }

        // projection
        // [ 768, 768] - model.layers[il].c_attn_proj_w
        // [ 768,   1] - model.layers[il].c_attn_proj_b
        // [ 768,   N] - cur (in)
        // [ 768,   N] - cur (out)
        //
        // cur = proj_w*cur + proj_b
        // [768, N]
        {
            cur = ggml_mul_mat(ctx,
                    model.layers[il].c_attn_proj_w,
                    cur);
            ggml_format_name(cur, "l%d.attn_proj_w", il);

            cur = ggml_add(ctx,
                    cur,
                    model.layers[il].c_attn_proj_b);
            ggml_format_name(cur, "l%d.attn_proj_b", il);
        }

        // add the input
        cur = ggml_add(ctx, cur, inpL);
        ggml_format_name(cur, "l%d.add", il);

        struct ggml_tensor * inpFF = cur;

        // feed-forward network
        {
            // norm
            {
                cur = ggml_norm(ctx, inpFF, hparams.eps);
                ggml_format_name(cur, "l%d.FFnorm", il);

                // cur = ln_2_g*cur + ln_2_b
                // [ 768, N]
                cur = ggml_add(ctx,
                        ggml_mul(ctx,
                            cur,
                            model.layers[il].ln_2_g),
                        model.layers[il].ln_2_b);
                ggml_format_name(cur, "l%d.ln_2_b", il);
                ggml_format_name(cur->src[0], "l%d.ln_2_g", il);
            }

            // fully connected
            // [3072, 768] - model.layers[il].c_mlp_fc_w
            // [3072,   1] - model.layers[il].c_mlp_fc_b
            // [ 768,   N] - cur (in)
            // [3072,   N] - cur (out)
            //
            // cur = fc_w*cur + fc_b
            // [3072, N]
            cur = ggml_mul_mat(ctx,
                    model.layers[il].c_mlp_fc_w,
                    cur);
            ggml_format_name(cur, "l%d.mlp_fc_w", il);

            cur = ggml_add(ctx,
                    cur,
                    model.layers[il].c_mlp_fc_b);
            ggml_format_name(cur, "l%d.mlp_fc_b", il);

            // GELU activation
            // [3072, N]
            cur = ggml_gelu(ctx, cur);
            ggml_format_name(cur, "l%d.gelu", il);

            // projection
            // [ 768, 3072] - model.layers[il].c_mlp_proj_w
            // [ 768,    1] - model.layers[il].c_mlp_proj_b
            // [3072,    N] - cur (in)
            // [ 768,    N] - cur (out)
            //
            // cur = proj_w*cur + proj_b
            // [768, N]
            cur = ggml_mul_mat(ctx,
                    model.layers[il].c_mlp_proj_w,
                    cur);
            ggml_format_name(cur, "l%d.mlp_proj_w", il);

            cur = ggml_add(ctx,
                    cur,
                    model.layers[il].c_mlp_proj_b);
            ggml_format_name(cur, "l%d.mlp_proj_b", il);
        }

        // input for next layer
        inpL = ggml_add(ctx, cur, inpFF);
        ggml_format_name(inpL, "l%d.add2", il);
    }

    // norm
    {
        // [ 768, N]
        inpL = ggml_norm(ctx, inpL, hparams.eps);
        ggml_format_name(inpL, "out_norm");

        // inpL = ln_f_g*inpL + ln_f_b
        // [ 768, N]
        inpL = ggml_add(ctx,
                ggml_mul(ctx,
                    inpL,
                    model.ln_f_g),
                model.ln_f_b);
        ggml_format_name(inpL, "out_ln_f_b");
        ggml_format_name(inpL->src[0], "out_ln_f_g");
    }

    // inpL = WTE * inpL
    // [ 768, 50257] - model.lm_head
    // [ 768, N]     - inpL
    inpL = ggml_mul_mat(ctx, model.lm_head, inpL);
    ggml_format_name(inpL, "out_lm_head");

    // logits -> probs
    //inpL = ggml_soft_max(ctx0, inpL);

    ggml_build_forward_expand(gf, inpL);

    ggml_free(ctx);

    return gf;
}

// evaluate the transformer
//
//   - model:     the model
//   - sched:     the backend scheduler
//   - 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
//
bool gpt2_eval(
        const gpt2_model & model,
        ggml_backend_sched_t sched,
        const int n_past,
        const std::vector<gpt_vocab::id> & embd_inp,
              std::vector<float>         & embd_w) {
    const int N = embd_inp.size();

    const auto & hparams = model.hparams;

    const int n_vocab = hparams.n_vocab;

    struct ggml_cgraph * gf = gpt2_graph(model, n_past, embd_inp);

    // run the computation
    ggml_backend_sched_reset(sched);
    ggml_backend_sched_graph_compute(sched, gf);

    //if (n_past%100 == 0) {
    //    ggml_graph_print   (&gf);
    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
    //}

    // in this case, the output tensor is the last one in the graph
    struct ggml_tensor * inpL = gf->nodes[gf->n_nodes - 1];

    //embd_w.resize(n_vocab*N);
    //ggml_backend_tensor_get(inpL, embd_w.data(), 0, sizeof(float)*n_vocab*N);

    // return result just for the last token
    embd_w.resize(n_vocab);
    ggml_backend_tensor_get(inpL, embd_w.data(), (n_vocab*(N-1))*sizeof(float), sizeof(float)*n_vocab);

    return true;
}

int main(int argc, char ** argv) {
    ggml_time_init();

    const int64_t t_main_start_us = ggml_time_us();

    gpt_params params;
    params.model = "models/gpt-2-117M/ggml-model.bin";

    if (gpt_params_parse(argc, argv, params) == false) {
        return 1;
    }

    if (params.seed < 0) {
        params.seed = time(NULL);
    }

    printf("%s: seed = %d\n", __func__, params.seed);

    std::mt19937 rng(params.seed);
    if (params.prompt.empty()) {
        params.prompt = gpt_random_prompt(rng);
    }

    int64_t t_load_us = 0;

    gpt_vocab vocab;
    gpt2_model model;

    // load the model
    {
        const int64_t t_start_us = ggml_time_us();

        if (!gpt2_model_load(params.model, model, vocab, params)) {
            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
            return 1;
        }

        t_load_us = ggml_time_us() - t_start_us;

        test_gpt_tokenizer(vocab, params.token_test);
    }

    // create the backend scheduler
    // the scheduler handles the allocation of the compute buffers and the scheduling of the computation between the different backends
    ggml_backend_sched_t sched;
    {
        // initialize the scheduler
        sched = ggml_backend_sched_new(model.backends.data(), NULL, model.backends.size(), GPT2_MAX_NODES, false);

        // create the worst case graph for memory usage estimation
        int n_tokens = std::min(model.hparams.n_ctx, params.n_batch);
        int n_past = model.hparams.n_ctx - n_tokens;
        struct ggml_cgraph * gf = gpt2_graph(model, n_past, std::vector<gpt_vocab::id>(n_tokens, 0));

        ggml_backend_sched_reserve(sched, gf);


        // compute the required memory
        size_t mem_size = 0;
        for (size_t i = 0; i < model.backends.size(); i++) {
            size_t size = ggml_backend_sched_get_buffer_size(sched, model.backends[i]);
            if (size > 0) {
                mem_size += size;
                printf("%s: %8s compute buffer size = %8.2f MB\n", __func__, ggml_backend_name(model.backends[i]), size/1024.0/1024.0);
                //printf("%s: %8s compute buffer size = %zu bytes\n", __func__, ggml_backend_name(model.backends[i]), size);
            }
        }

        printf("%s: total compute buffer size: %.2f MB\n", __func__, mem_size/1024.0/1024.0);
    }

    int n_past = 0;

    int64_t t_sample_us  = 0;
    int64_t t_predict_us = 0;

    std::vector<float> logits;

    // tokenize the prompt
    std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);

    params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());

    printf("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
    printf("%s: number of tokens in prompt = %zu, first 8 tokens: ", __func__, embd_inp.size());
    for (int i = 0; i < std::min(8, (int) embd_inp.size()); i++) {
        printf("%d ", embd_inp[i]);
    }
    printf("\n\n");

    // submit the input prompt token-by-token
    // this reduces the memory usage during inference, at the cost of a bit of speed at the beginning
    std::vector<gpt_vocab::id> embd;

    for (size_t i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
        // predict
        if (embd.size() > 0) {
            const int64_t t_start_us = ggml_time_us();

            if (!gpt2_eval(model, sched, n_past, embd, logits)) {
                printf("Failed to predict\n");
                return 1;
            }

            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 = params.top_k;
            const float top_p = params.top_p;
            const float temp  = params.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 (size_t k = i; k < embd_inp.size(); k++) {
                embd.push_back(embd_inp[k]);
                if (int32_t(embd.size()) >= params.n_batch) {
                    break;
                }
            }
            i += embd.size() - 1;
        }

        // display text
        for (auto id : embd) {
            printf("%s", vocab.id_to_token[id].c_str());
        }
        fflush(stdout);

        // end of text token
        if (embd.back() == 50256) {
            break;
        }
    }

    // report timing
    {
        const int64_t t_main_end_us = ggml_time_us();

        printf("\n\n");
        printf("%s:     load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
        printf("%s:   sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
        printf("%s:  predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
    }

    ggml_free(model.ctx_w);

    ggml_backend_sched_free(sched);
    ggml_backend_buffer_free(model.buffer_kv);
    for (auto buf : model.buffers_w) {
        ggml_backend_buffer_free(buf);
    }
    for (auto backend : model.backends) {
        ggml_backend_free(backend);
    }

    return 0;
}