gpt4all/gpt4all-backend/starcoder.cpp

#define STARCODER_H_I_KNOW_WHAT_I_AM_DOING_WHEN_INCLUDING_THIS_FILE
#include "starcoder_impl.h"
#include "llama.h"
#include "llama-util.h"
#include "utils.h"
#include "llmodel_shared.h"

#include <cassert>
#include <cinttypes>
#include <iostream>
#include <sstream>

namespace {
const char *modelType_ = "Starcoder";
}

#define STARCODER_MAGIC 0x67676d6c

// default hparams (GPT-2 117M)
// https://huggingface.co/bigcode/gpt_bigcode-santacoder/blob/main/config.json
struct starcoder_hparams {
    int32_t n_vocab   = 49280;
    int32_t n_ctx     = 2048;
    int32_t n_embd    = 2048;
    int32_t n_head    = 16;
    int32_t n_layer   = 24;
    int32_t ftype     = 1;
};

struct starcoder_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 starcoder_model {
    starcoder_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<starcoder_layer> layers;

    // key + value memory
    llm_kv_cache kv_self;

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

    llm_buffer eval_buf;
    llm_buffer scr0_buf;
    llm_buffer scr1_buf;
    llm_buffer work_buf;
};

static bool kv_cache_init(
        const struct starcoder_hparams & hparams,
              struct llm_kv_cache & cache,
                         ggml_type   wtype,
                               int   n_ctx) {
    const int n_embd  = hparams.n_embd;
#if 0 // FIXME: This code can be made in future to support much smaller KV cache
    const int dim_head  = n_embd / hparams.n_head;
    const int dim_kv = dim_head;
#endif
    const int n_layer = hparams.n_layer;

    const int64_t n_mem      = (int64_t)n_layer*n_ctx;
    const int64_t n_elements = n_embd * n_mem;
    cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2_MiB);
    struct ggml_init_params params;
    params.mem_size   = cache.buf.size;
    params.mem_buffer = cache.buf.addr;
    params.no_alloc   = false;

    cache.ctx = ggml_init(params);
    if (!cache.ctx) {
        fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__);
        return false;
    }

    cache.k = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
    cache.v = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
    return true;
}

// load the model's weights from a file
bool starcoder_model_load(const std::string & fname, starcoder_model & model, gpt_vocab & vocab, size_t *mem_req) {
    printf("%s: loading model from '%s'\n", __func__, fname.c_str());
    if (mem_req) {
        *mem_req = 0;
    }

    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 != 0x67676d6c) {
            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;

            // if (i < 10) 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_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;

    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;

        const int head_dim = n_embd / hparams.n_head;
        const int kv_heads = hparams.n_head; // 1 if MQA else hparams.n_head
        const int kv_dim   = kv_heads * head_dim;

        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_vocab*n_embd*ggml_type_sizef(wtype);         // wte
        ctx_size +=   n_ctx*n_embd*ggml_type_sizef(GGML_TYPE_F32); // wpe
        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // lm_head

        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*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b

        ctx_size += n_layer*((n_embd + 2*kv_dim)*n_embd*ggml_type_sizef(wtype));         // c_attn_attn_w // TODO:
        ctx_size += n_layer*(       (n_embd + 2*kv_dim)*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b

        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype));           // c_attn_proj_w
        ctx_size += n_layer*(       n_embd*ggml_type_sizef(GGML_TYPE_F32));   // c_attn_proj_b

        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_F32); // memory_k
//        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v

        ctx_size += (6 + 12*n_layer)*512; // object overhead

        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
    }

    if (mem_req) {
        const int n_embd  = model.hparams.n_embd;
        const int dim_head  = n_embd / model.hparams.n_head;
        const int dim_kv = dim_head;
        const int n_layer = model.hparams.n_layer;

        const int64_t n_mem      = (int64_t)n_layer*model.hparams.n_ctx;
        const int64_t n_elements = dim_kv * n_mem;
        size_t kv_cache_size = 2u*n_elements*ggml_type_size(wtype) + 2_MiB;
        *mem_req = ctx_size + kv_cache_size;
        return false;
    }

    // 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) {
            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;

        const int head_dim = n_embd / hparams.n_head;
        const int kv_heads = hparams.n_head; // 1 if MQA else hparams.n_head
        const int kv_dim   = kv_heads * head_dim;

        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, n_embd + 2*kv_dim);
            layer.c_attn_attn_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd + 2*kv_dim);

            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); //TODO: 4*n_embd = config.n_inner
            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;
        }
    }

    // key + value memory
    {
        const auto & hparams = model.hparams;

        const int n_layer   = hparams.n_layer;
        const int n_ctx     = hparams.n_ctx;
        const int64_t n_mem = n_layer*n_ctx;

        if (!kv_cache_init(hparams, model.kv_self, GGML_TYPE_F32, model.hparams.n_ctx)) {
            fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
            ggml_free(ctx);
            return false;
        }
        const size_t memory_size = ggml_nbytes(model.kv_self.k) + ggml_nbytes(model.kv_self.v);

        printf("%s: memory_size = %8.2f MB, n_mem = %" PRId64 "\n", __func__, memory_size/1024.0/1024.0, n_mem);
    }

    // load weights
    {
        size_t total_size = 0;

        bool has_lm_head = false;

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

            auto tensor = model.tensors[name.data()];
            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(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
                return false;
            }
            if (ggml_nelements(tensor) != nelements) {
                fprintf(stderr, "%s: tensor '%s' has wrong size in model file. got %d, expected %d\n",
                        __func__, name.data(), (int) ggml_nelements(tensor), nelements);
                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)) {
                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;
            }

            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));

            // GPT-2 models share the WTE tensor as the LM head
            if (name == "model/wte" && has_lm_head == false) {
                memcpy(model.lm_head->data, tensor->data, ggml_nbytes(tensor));
            }

            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();

    model.eval_buf.resize(1280u * 1024 * 1024);
    model.scr0_buf.resize(256u * 1024 * 1024);
    model.scr1_buf.resize(256u * 1024 * 1024);

    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
//
bool starcoder_eval(
        starcoder_model & model,
        const int n_threads,
        const int n_past,
        const std::vector<gpt_vocab::id> & embd_inp,
              std::vector<float>         & 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 size_t head_dim = n_embd / n_head;

   struct ggml_init_params eval_ctx_params = {
        .mem_size = model.eval_buf.size,
        .mem_buffer = model.eval_buf.addr,
        .no_alloc = false,
    };

    struct ggml_context * ctx0 = ggml_init(eval_ctx_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));

    struct ggml_tensor * position = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
    for (int i = 0; i < N; ++i) {
        ((int32_t *) position->data)[i] = n_past + i;
    }

    // wte + wpe
    struct ggml_tensor * inpL =
        ggml_add(ctx0,
                ggml_get_rows(ctx0, model.wte, embd),
                ggml_get_rows(ctx0, model.wpe, position));

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

        ggml_set_scratch(ctx0, {0, model.scr0_buf.size, model.scr0_buf.addr, });

        // norm
        {
            // [ 768, N]
            cur = ggml_norm(ctx0, inpL);

            // cur = ln_1_g*cur + ln_1_b
            // [ 768, N]
            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));
        }

        // 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(ctx0,
                    model.layers[il].c_attn_attn_w,
                    cur);

            cur = ggml_add(ctx0,
                    ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
                    cur);
        }

        // self-attention
        {
            struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
            struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
            struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);

            // store key and value to memory
            if (N >= 1) {
                struct ggml_tensor * k = ggml_view_1d(ctx0, model.kv_self.k, N*n_embd, (ggml_element_size(model.kv_self.k)*n_embd)*(il*n_ctx + n_past));
                struct ggml_tensor * v = ggml_view_1d(ctx0, model.kv_self.v, N*n_embd, (ggml_element_size(model.kv_self.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)
            // [64, N, 12]
            struct ggml_tensor * Q =
                ggml_permute(ctx0,
                        ggml_cpy(ctx0,
                            Qcur,
                            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
                        0, 2, 1, 3);

            // 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(ctx0,
                        ggml_reshape_3d(ctx0,
                            ggml_view_1d(ctx0, model.kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.kv_self.k)*n_embd),
                            n_embd/n_head, n_head, n_past + N),
                        0, 2, 1, 3); //TODO: need to be tiled

            // 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(ctx0, K, Q); //TODO: check if it broadcasts

            // KQ_scaled = KQ / sqrt(n_embd/n_head)
            // [n_past + N, N, 12]
            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)
            // [n_past + N, N, 12]
            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);

            // KQ = soft_max(KQ_masked)
            // [n_past + N, N, 12]
            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()
            // [n_past + N, 64, 12]
            struct ggml_tensor * V_trans =
                ggml_cpy(ctx0,
                        ggml_permute(ctx0,
                            ggml_reshape_3d(ctx0,
                                ggml_view_1d(ctx0, model.kv_self.v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.kv_self.v)*n_embd),
                                n_embd/n_head, n_head, n_past + N),
                            1, 2, 0, 3),
                        ggml_new_tensor_3d(ctx0, model.kv_self.v->type, 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(ctx0, V_trans, KQ_soft_max);

            // KQV_merged = KQV.permute(0, 2, 1, 3)
            // [64, 12, N]
            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);

            // cur = KQV_merged.contiguous().view(n_embd, N)
            // [768, N]
            cur = ggml_cpy(ctx0,
                    KQV_merged,
                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
        }

        // 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(ctx0,
                    model.layers[il].c_attn_proj_w,
                    cur);

            cur = ggml_add(ctx0,
                    ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur),
                    cur);
        }

        // add the input
        cur = ggml_add(ctx0, cur, inpL);

        struct ggml_tensor * inpFF = cur;

        ggml_set_scratch(ctx0, {0, model.scr1_buf.size, model.scr1_buf.addr, });

        // feed-forward network
        {
            // norm
            {
                cur = ggml_norm(ctx0, inpFF);

                // cur = ln_2_g*cur + ln_2_b
                // [ 768, N]
                cur = ggml_add(ctx0,
                        ggml_mul(ctx0,
                            ggml_repeat(ctx0, model.layers[il].ln_2_g, cur),
                            cur),
                        ggml_repeat(ctx0, model.layers[il].ln_2_b, cur));
            }

            // 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(ctx0,
                    model.layers[il].c_mlp_fc_w,
                    cur);

            cur = ggml_add(ctx0,
                    ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
                    cur);

            // GELU activation
            // [3072, N]
            cur = ggml_gelu(ctx0, cur);

            // 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(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);
        }

        // input for next layer
        inpL = ggml_add(ctx0, cur, inpFF);
    }

    ggml_set_scratch(ctx0, {0, model.scr0_buf.size, model.scr0_buf.addr, });

    // norm
    {
        // [ 768, N]
        inpL = ggml_norm(ctx0, inpL);

        // inpL = ln_f_g*inpL + ln_f_b
        // [ 768, N]
        inpL = ggml_add(ctx0,
                ggml_mul(ctx0,
                    ggml_repeat(ctx0, model.ln_f_g, inpL),
                    inpL),
                ggml_repeat(ctx0, model.ln_f_b, inpL));
    }

    ggml_set_scratch(ctx0, { 0, 0, nullptr, });

    // inpL = WTE * inpL
    // [ 768, 50257] - model.lm_head
    // [ 768, N]     - inpL
    inpL = ggml_mul_mat(ctx0, model.lm_head, inpL);

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

    // run the computation
    ggml_build_forward_expand(&gf, inpL);
    ggml_graph_compute_g4a(model.work_buf, &gf, n_threads);

    //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 just for 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 MB\n", ggml_used_mem(ctx0)/(1024*1024));

    ggml_free(ctx0);

    return true;
}

#define MAX_RNG_STATE 64*1024
size_t starcoder_get_state_size(const starcoder_model &model) {
    const size_t s_rng_size        = sizeof(size_t);
    const size_t s_rng             = MAX_RNG_STATE;
    const size_t s_kv_size         = sizeof(size_t);
    const size_t s_kv_ntok         = sizeof(int);
    const size_t s_kv              = model.kv_self.buf.size;
    const size_t s_total = (
        + s_rng_size
        + s_rng
        + s_kv_size
        + s_kv_ntok
        + s_kv
    );
    return s_total;
}

size_t starcoder_copy_state_data(const starcoder_model &model, const std::mt19937 &rng, uint8_t *dest)
{
    uint8_t * out = dest;
    // copy rng
    {
        std::stringstream rng_ss;
        rng_ss << rng;

        const size_t rng_size = rng_ss.str().size();
        char rng_buf[MAX_RNG_STATE];

        memset(&rng_buf[0], 0, MAX_RNG_STATE);
        memcpy(&rng_buf[0], rng_ss.str().data(), rng_ss.str().size());

        memcpy(out, &rng_size,   sizeof(rng_size));   out += sizeof(rng_size);
        memcpy(out, &rng_buf[0], MAX_RNG_STATE); out += MAX_RNG_STATE;
    }

    // copy kv cache
    {
        const size_t kv_size = model.kv_self.buf.size;
        const int    kv_ntok = model.kv_self.n;

        memcpy(out, &kv_size, sizeof(kv_size)); out += sizeof(kv_size);
        memcpy(out, &kv_ntok, sizeof(kv_ntok)); out += sizeof(kv_ntok);

        if (kv_size) {
            memcpy(out, model.kv_self.buf.addr, kv_size); out += kv_size;
        }
    }

    const size_t written  = out - dest;
    assert(written == starcoder_get_state_size(model));
    fflush(stdout);
    return written;
}

size_t starcoder_set_state_data(starcoder_model *model, std::mt19937 *rng, const uint8_t *src)
{
    const uint8_t * in = src;

    // set rng
    {
        size_t rng_size;
        char   rng_buf[MAX_RNG_STATE];

        memcpy(&rng_size,   in, sizeof(rng_size));    in += sizeof(rng_size);
        memcpy(&rng_buf[0], in, MAX_RNG_STATE); in += MAX_RNG_STATE;

        std::stringstream rng_ss;
        rng_ss.str(std::string(&rng_buf[0], rng_size));
        rng_ss >> *rng;

        assert(rng_ss.fail() == false);
    }

    // set kv cache
    {
        size_t kv_size;
        int kv_ntok;

        memcpy(&kv_size, in, sizeof(kv_size)); in += sizeof(kv_size);
        memcpy(&kv_ntok, in, sizeof(kv_ntok)); in += sizeof(kv_ntok);

        if (kv_size) {
            assert(model->kv_self.buf.size == kv_size);

            void * k_data = model->kv_self.k->data; // remember data pointers
            void * v_data = model->kv_self.v->data; // because their value is stored in buf and overwritten by memcpy

            memcpy(model->kv_self.buf.addr, in, kv_size); in += kv_size;

            model->kv_self.k->data = k_data; // restore correct data pointers
            model->kv_self.v->data = v_data;

        }

        model->kv_self.n = kv_ntok;
    }

    const size_t nread    = in - src;
    assert(nread == starcoder_get_state_size(*model));
    fflush(stdout);
    return nread;
}

struct StarcoderPrivate {
    const std::string modelPath;
    bool modelLoaded;
    gpt_vocab vocab;
    starcoder_model *model = nullptr;
    int64_t n_threads = 0;
    size_t mem_per_token = 0;
    std::mt19937 rng;
};

Starcoder::Starcoder() : d_ptr(new StarcoderPrivate) {
    d_ptr->model = new starcoder_model;
    d_ptr->model->ctx = nullptr;
    d_ptr->modelLoaded = false;
}

Starcoder::~Starcoder() {
    if(d_ptr->model->ctx) {
        ggml_free(d_ptr->model->ctx);
        d_ptr->model->ctx = nullptr;
    }
    delete d_ptr->model;
}

bool Starcoder::loadModel(const std::string &modelPath)
{
    std::mt19937 rng(time(NULL));
    d_ptr->rng = rng;

    // load the model
    if (!starcoder_model_load(modelPath, *d_ptr->model, d_ptr->vocab, nullptr)) {
        std::cerr << "STARCODER ERROR: failed to load model from " <<  modelPath;
        return false;
    }

    d_ptr->n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
    d_ptr->modelLoaded = true;
    fflush(stdout);
    return true;
}

bool Starcoder::isModelLoaded() const
{
    return d_ptr->modelLoaded;
}

size_t Starcoder::requiredMem(const std::string &modelPath)
{
    starcoder_model dummy_model;
    gpt_vocab dummy_vocab;
    size_t mem_req;
    auto fin = std::ifstream(modelPath, std::ios::binary);
    starcoder_model_load(modelPath, dummy_model, dummy_vocab, &mem_req);
    return mem_req;
}

size_t Starcoder::stateSize() const
{
    return starcoder_get_state_size(*d_ptr->model);
}

size_t Starcoder::saveState(uint8_t *dest) const
{
    return starcoder_copy_state_data(*d_ptr->model, d_ptr->rng, dest);
}

size_t Starcoder::restoreState(const uint8_t *src)
{
    return starcoder_set_state_data(d_ptr->model, &d_ptr->rng, src);
}

void Starcoder::setThreadCount(int32_t n_threads)
{
    d_ptr->n_threads = n_threads;
}

int32_t Starcoder::threadCount() const
{
    return d_ptr->n_threads;
}

std::vector<LLModel::Token> Starcoder::tokenize(PromptContext &, const std::string &str) const
{
    return ::gpt_tokenize(d_ptr->vocab, str);
}

LLModel::Token Starcoder::sampleToken(PromptContext &promptCtx) const
{
    const size_t n_prev_toks = std::min((size_t) promptCtx.repeat_last_n, promptCtx.tokens.size());
    return gpt_sample_top_k_top_p(d_ptr->model->hparams.n_vocab,
        promptCtx.tokens.data() + promptCtx.tokens.size() - n_prev_toks,
        n_prev_toks,
        promptCtx.logits,
        promptCtx.top_k, promptCtx.top_p, promptCtx.temp,
        promptCtx.repeat_penalty,
        d_ptr->rng);
}

std::string Starcoder::tokenToString(Token id) const
{
    return d_ptr->vocab.id_to_token[id];
}

bool Starcoder::evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const
{
    // determine the required inference memory per token:
    static bool initialized = false;
    if (!initialized) {
        starcoder_eval(*d_ptr->model, d_ptr->n_threads, 0, { 0, 1, 2, 3 }, ctx.logits,
            d_ptr->mem_per_token);
        initialized = true;
    }

    return starcoder_eval(*d_ptr->model, d_ptr->n_threads, ctx.n_past, tokens, ctx.logits, d_ptr->mem_per_token);
}

int32_t Starcoder::contextLength() const
{
    return d_ptr->model->hparams.n_ctx;
}

const std::vector<LLModel::Token> &Starcoder::endTokens() const
{
    static const std::vector<LLModel::Token> out = { 0 };
    return out;
}

#if defined(_WIN32)
#define DLL_EXPORT __declspec(dllexport)
#else
#define DLL_EXPORT __attribute__ ((visibility ("default")))
#endif

extern "C" {
DLL_EXPORT bool is_g4a_backend_model_implementation() {
    return true;
}

DLL_EXPORT const char *get_model_type() {
    return modelType_;
}

DLL_EXPORT const char *get_build_variant() {
    return GGML_BUILD_VARIANT;
}

DLL_EXPORT bool magic_match(std::istream& f) {
    uint32_t magic = 0;
    f.read(reinterpret_cast<char*>(&magic), sizeof(magic));
    if (magic != STARCODER_MAGIC) {
         return false;
    }
    starcoder_hparams hparams;
    f.read(reinterpret_cast<char*>(&hparams), sizeof(hparams));
    if (!(hparams.n_vocab >= 49100 && hparams.n_vocab <= 49290)) {
        return false;
    }
    return true;
}

DLL_EXPORT LLModel *construct() {
    return new Starcoder;
}
}