talk.wasm : refactoring + update README.md

examples/talk.wasm/CMakeLists.txt

@@ -6,6 +6,7 @@ set(TARGET libtalk)
 
 add_executable(${TARGET}
     emscripten.cpp
+    gpt-2.cpp
     )
 
 target_link_libraries(${TARGET} PRIVATE

examples/talk.wasm/README.md

@@ -16,7 +16,13 @@ This demo leverages 2 modern neural network models to create a high-quality voic
 
 The web page does the processing locally on your machine. The processing of these heavy neural network models in the
 browser is possible by implementing them efficiently in C/C++ and using the browser's WebAssembly SIMD capabilities for
-extra performance. For more detailed information, checkout the [current repository](https://github.com/ggerganov/whisper.cpp).
+extra performance:
+
+- The Whisper C++ implementation is here: [whisper.h](/whisper.h) / [whisper.cpp](/whisper.cpp)
+- The GPT-2 C++ implementation is here: [gpt-2.h](gpt-2.h) / [gpt-2.cpp](gpt-2.cpp)
+- Both models use a custom tensor library implemented in C: [ggml.h](/ggml.h) / [ggml.c](/ggml.c)
+- The HTML/JS layer is here: [index-tmpl.html](index-tmpl.html)
+- The Emscripten bridge between C/C++ and JS is here: [emscripten.cpp](emscripten.cpp)
 
 In order to run the models, the web page first needs to download the model data which is about ~350 MB. The model data
 is then cached in your browser's cache and can be reused in future visits without downloading it again.
@@ -33,11 +39,11 @@ In order to run this demo efficiently, you need to have the following:
 Notice that this demo is using the smallest GPT-2 model, so the generated text responses are not always very good.
 Also, the prompting strategy can likely be improved to achieve better results.
 
-The demo is quite computationally heavy - it's not usual to run these transformer models in a browser. Typically, they
-run on powerful GPU hardware. So for better experience, you do need to have a powerful computer.
+The demo is quite computationally heavy, so you need a fast CPU. It's not usual to run these transformer models in a
+browser. Typically, they run on powerful GPUs.
 
-Probably in the near future, mobile browsers will start supporting WASM SIMD. This will allow to run the demo on your
-phone or tablet. But for now this functionality is not supported on mobile devices (at least not on iPhone).
+Currently, mobile browsers do not support the Fixed-width SIMD WebAssembly capability, so you cannot run this demo
+on a phone or a tablet. Hopefully, in the near future this will become supported.
 
 ## Todo
 

examples/talk.wasm/emscripten.cpp

@@ -1,985 +1,21 @@
 #include "ggml.h"
+#include "gpt-2.h"
 #include "whisper.h"
 
 #include <emscripten.h>
 #include <emscripten/bind.h>
 
 #include <atomic>
-#include <cassert>
 #include <cmath>
-#include <cstdio>
-#include <cstring>
-#include <fstream>
-#include <map>
 #include <mutex>
 #include <string>
 #include <thread>
 #include <vector>
 #include <regex>
-#include <random>
-
-std::string to_timestamp(int64_t t) {
-    int64_t sec = t/100;
-    int64_t msec = t - sec*100;
-    int64_t min = sec/60;
-    sec = sec - min*60;
-
-    char buf[32];
-    snprintf(buf, sizeof(buf), "%02d:%02d.%03d", (int) min, (int) sec, (int) msec);
-
-    return std::string(buf);
-}
-
-/////////////////////// GPT-2 BEGIN /////////////////////////
-// TODO: move to a separate file
-
-//
-// Vocab utils
-//
-
-struct gpt_vocab {
-    using id    = int32_t;
-    using token = std::string;
-
-    std::map<token, id> token_to_id;
-    std::map<id, token> id_to_token;
-};
-
-void replace(std::string & str, const std::string & needle, const std::string & replacement) {
-    size_t pos = 0;
-    while ((pos = str.find(needle, pos)) != std::string::npos) {
-        str.replace(pos, needle.length(), replacement);
-        pos += replacement.length();
-    }
-}
-
-std::map<std::string, int32_t> json_parse(const std::string & fname) {
-    std::map<std::string, int32_t> result;
-
-    // read file into string
-    std::string json;
-    {
-        std::ifstream ifs(fname);
-        if (!ifs) {
-            fprintf(stderr, "Failed to open %s\n", fname.c_str());
-            exit(1);
-        }
-
-        json = std::string((std::istreambuf_iterator<char>(ifs)),
-                (std::istreambuf_iterator<char>()));
-    }
-
-    if (json[0] != '{') {
-        return result;
-    }
-
-    // parse json
-    {
-        bool has_key  = false;
-        bool in_token = false;
-
-        std::string str_key = "";
-        std::string str_val = "";
-
-        int n = json.size();
-        for (int i = 1; i < n; ++i) {
-            if (!in_token) {
-                if (json[i] == ' ') continue;
-                if (json[i] == '"') {
-                    in_token = true;
-                    continue;
-                }
-            } else {
-                if (json[i] == '\\' && i+1 < n) {
-                    if (has_key == false) {
-                        str_key += json[i];
-                    } else {
-                        str_val += json[i];
-                    }
-                    ++i;
-                } else if (json[i] == '"') {
-                    if (has_key == false) {
-                        has_key = true;
-                        ++i;
-                        while (json[i] == ' ') ++i;
-                        ++i; // :
-                        while (json[i] == ' ') ++i;
-                        if (json[i] != '\"') {
-                            while (json[i] != ',' && json[i] != '}') {
-                                str_val += json[i++];
-                            }
-                            has_key = false;
-                        } else {
-                            in_token = true;
-                            continue;
-                        }
-                    } else {
-                        has_key = false;
-                    }
-
-                    ::replace(str_key, "\\u0120", " " ); // \u0120 -> space
-                    ::replace(str_key, "\\u010a", "\n"); // \u010a -> new line
-                    ::replace(str_key, "\\\"",    "\""); // \\\"   -> "
-
-                    try {
-                        result[str_key] = std::stoi(str_val);
-                    } catch (...) {
-                        //fprintf(stderr, "%s: ignoring key '%s' with value '%s'\n", fname.c_str(), str_key.c_str(), str_val.c_str());
-
-                    }
-                    str_key = "";
-                    str_val = "";
-                    in_token = false;
-                    continue;
-                }
-                if (has_key == false) {
-                    str_key += json[i];
-                } else {
-                    str_val += json[i];
-                }
-            }
-        }
-    }
-
-    return result;
-}
-
-std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text) {
-    std::vector<std::string> words;
-
-    // first split the text into words
-    {
-        std::string str = text;
-        std::string pat = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
-
-        std::regex re(pat);
-        std::smatch m;
-
-        while (std::regex_search(str, m, re)) {
-            for (auto x : m) {
-                words.push_back(x);
-            }
-            str = m.suffix();
-        }
-    }
-
-    // find the longest tokens that form the words:
-    std::vector<gpt_vocab::id> tokens;
-    for (const auto & word : words) {
-        if (word.size() == 0) continue;
-
-        int i = 0;
-        int n = word.size();
-        while (i < n) {
-            int j = n;
-            while (j > i) {
-                auto it = vocab.token_to_id.find(word.substr(i, j-i));
-                if (it != vocab.token_to_id.end()) {
-                    tokens.push_back(it->second);
-                    i = j;
-                    break;
-                }
-                --j;
-            }
-            if (i == n) {
-                break;
-            }
-            if (j == i) {
-                auto sub = word.substr(i, 1);
-                if (vocab.token_to_id.find(sub) != vocab.token_to_id.end()) {
-                    tokens.push_back(vocab.token_to_id.at(sub));
-                } else {
-                    fprintf(stderr, "%s: unknown token '%s'\n", __func__, sub.data());
-                }
-                ++i;
-            }
-        }
-    }
-
-    return tokens;
-}
-
-bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab) {
-    printf("%s: loading vocab from '%s'\n", __func__, fname.c_str());
-
-    vocab.token_to_id = ::json_parse(fname);
-
-    for (const auto & kv : vocab.token_to_id) {
-        vocab.id_to_token[kv.second] = kv.first;
-    }
-
-    printf("%s: vocab size = %d\n", __func__, (int) vocab.token_to_id.size());
-
-    // print the vocabulary
-    //for (auto kv : vocab.token_to_id) {
-    //    printf("'%s' -> %d\n", kv.first.data(), kv.second);
-    //}
-
-    return true;
-}
-
-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<std::pair<double, gpt_vocab::id>> logits_id;
-    logits_id.reserve(n_logits);
-
-    for (int i = 0; i < n_logits; i++) {
-        logits_id.push_back(std::make_pair(logits[i], i));
-    }
-
-    // find the top K tokens
-    std::partial_sort(
-            logits_id.begin(),
-            logits_id.begin() + top_k, logits_id.end(),
-            [](const std::pair<double, gpt_vocab::id> & a, const std::pair<double, gpt_vocab::id> & b) {
-        return a.first > b.first;
-    });
-
-    logits_id.resize(top_k);
-
-    // normalize
-    {
-        double sum = 0.0f;
-        for (int i = 0; i < (int)logits_id.size(); i++) {
-            sum += logits_id[i].first;
-        }
-
-        sum = 1.0/sum;
-        for (int i = 0; i < (int)logits_id.size(); i++) {
-            logits_id[i].first *= sum;
-        }
-    }
-
-    if (top_p < 1.0f) {
-        {
-            double cumsum = 0.0f;
-            for (int i = 0; i < top_k; i++) {
-                cumsum += logits_id[i].first;
-                if (cumsum >= top_p) {
-                    logits_id.resize(i+1);
-                    break;
-                }
-            }
-        }
-
-        // normalize again
-        {
-            double sum = 0.0f;
-            for (int i = 0; i < (int)logits_id.size(); i++) {
-                sum += logits_id[i].first;
-            }
-
-            sum = 1.0/sum;
-            for (int i = 0; i < (int)logits_id.size(); i++) {
-                logits_id[i].first *= sum;
-            }
-        }
-    }
-
-    //printf("\n");
-    //for (int i = 0; i < (int)logits_id.size(); i++) {
-    //    printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), logits_id[i].first);
-    //}
-    //exit(0);
-
-    // sample from the obtained distribution
-    std::vector<double> probs;
-    probs.reserve(logits_id.size());
-
-    for (int i = 0; i < (int) logits_id.size(); i++) {
-        probs.push_back(logits_id[i].first);
-    }
-
-    std::discrete_distribution<> dist(probs.begin(), probs.end());
-    int idx = dist(rng);
-
-    return logits_id[idx].second;
-}
-
-// 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 f16     = 1;
-};
-
-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_trans; // transposed for efficiency
-    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
-
-    std::vector<gpt2_layer> layers;
-
-    // key + value memory
-    struct ggml_tensor * memory_k;
-    struct ggml_tensor * memory_v;
-
-    //
-    struct ggml_context * ctx;
-    std::map<std::string, struct ggml_tensor *> tensors;
-};
-
-// load the model's weights from a file
-bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab) {
-    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 != 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.f16,     sizeof(hparams.f16));
-
-        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: f16     = %d\n", __func__, hparams.f16);
-    }
-
-    // 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;
-        for (int i = 0; i < n_vocab; i++) {
-            uint32_t len;
-            fin.read((char *) &len, sizeof(len));
-
-            word.resize(len);
-            fin.read((char *) word.data(), len);
-
-            vocab.token_to_id[word] = i;
-            vocab.id_to_token[i] = word;
-        }
-    }
-
-    // for the big tensors, we have the option to store the data in 16-bit floats
-    // in order to save memory and also to speed up the computation
-    const ggml_type wtype = model.hparams.f16 ? GGML_TYPE_F16 : GGML_TYPE_F32;
-
-    auto & ctx = model.ctx;
-
-    size_t ctx_size = 0;
-
-    {
-        const auto & hparams = model.hparams;
-
-        const int n_embd  = hparams.n_embd;
-        const int n_layer = hparams.n_layer;
-        const int n_ctx   = hparams.n_ctx;
-        const int n_vocab = hparams.n_vocab;
-
-        ctx_size += n_embd*ggml_type_size(GGML_TYPE_F32); // ln_f_g
-        ctx_size += n_embd*ggml_type_size(GGML_TYPE_F32); // ln_f_b
-
-        ctx_size += n_vocab*n_embd*ggml_type_size(wtype);         // wte
-        ctx_size +=   n_ctx*n_embd*ggml_type_size(GGML_TYPE_F32); // wpe
-
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_1_g
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_1_b
-
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_2_g
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_2_b
-
-        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_size(wtype));         // c_attn_attn_w
-        ctx_size += n_layer*(       3*n_embd*ggml_type_size(GGML_TYPE_F32)); // c_attn_attn_b
-
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_size(wtype));           // c_attn_proj_w
-        ctx_size += n_layer*(       n_embd*ggml_type_size(GGML_TYPE_F32));   // c_attn_proj_b
-
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_size(wtype));         // c_mlp_fc_w
-        ctx_size += n_layer*(       4*n_embd*ggml_type_size(GGML_TYPE_F32)); // c_mlp_fc_b
-
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_size(wtype));         // c_mlp_proj_w
-        ctx_size += n_layer*(         n_embd*ggml_type_size(GGML_TYPE_F32)); // c_mlp_proj_b
-
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_size(GGML_TYPE_F32); // memory_k
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_size(GGML_TYPE_F32); // memory_v
-
-        ctx_size += (6 + 12*n_layer)*256; // object overhead
-
-        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
-    }
-
-    // create the ggml context
-    {
-        struct ggml_init_params params = {
-            .mem_size   = ctx_size,
-            .mem_buffer = NULL,
-        };
-
-        model.ctx = ggml_init(params);
-        if (!model.ctx) {
-            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
-            return false;
-        }
-    }
-
-    // prepare memory for the weights
-    {
-        const auto & hparams = model.hparams;
-
-        const int n_embd  = hparams.n_embd;
-        const int n_layer = hparams.n_layer;
-        const int n_ctx   = hparams.n_ctx;
-        const int n_vocab = hparams.n_vocab;
-
-        model.layers.resize(n_layer);
-
-        model.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);
-
-        // 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;
-
-        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,         3*n_embd, 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,         4*n_embd, n_embd);
-            layer.c_mlp_fc_b         = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
-
-            layer.c_mlp_proj_w_trans = 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_trans;
-            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_embd  = hparams.n_embd;
-        const int n_layer = hparams.n_layer;
-        const int n_ctx   = hparams.n_ctx;
-
-        const int n_mem      = n_layer*n_ctx;
-        const int n_elements = n_embd*n_mem;
-
-        model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
-        model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
-
-        const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
-
-        printf("%s: memory size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
-    }
-
-    // load weights
-    {
-        size_t total_size = 0;
-
-        while (true) {
-            int32_t n_dims;
-            int32_t length;
-            int32_t ftype;
-
-            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
-            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
-            fin.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
-
-            if (fin.eof()) {
-                break;
-            }
-
-            int32_t nelements = 1;
-            int32_t ne[2] = { 1, 1 };
-            for (int i = 0; i < n_dims; ++i) {
-                fin.read(reinterpret_cast<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 (ggml_nelements(tensor) != nelements) {
-                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
-                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.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
-                return false;
-            }
-
-            const size_t bpe = (ftype == 0) ? sizeof(float) : sizeof(ggml_fp16_t);
-
-            if (nelements*bpe != 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));
-
-            //printf("%24s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
-            total_size += ggml_nbytes(tensor);
-        }
-
-        printf("%s: model size  = %8.2f MB\n", __func__, total_size/1024.0/1024.0);
-    }
-
-    fin.close();
-
-    return true;
-}
-
-// evaluate the transformer
-//
-//   - model:     the model
-//   - n_threads: number of threads to use
-//   - n_past:    the context size so far
-//   - embd_inp:  the embeddings of the tokens in the context
-//   - embd_w:    the predicted probabilities of the next token
-//
-bool gpt2_eval(
-        const gpt2_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;
-
-    static size_t buf_size = 512u*1024*1024;
-    static void * buf = malloc(buf_size);
-
-    if (mem_per_token > 0 && mem_per_token*N > buf_size) {
-        const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
-        printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
-
-        // reallocate
-        buf_size = buf_size_new;
-        buf = realloc(buf, buf_size);
-        if (buf == nullptr) {
-            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
-            return false;
-        }
-    }
-
-    struct ggml_init_params params = {
-        .mem_size   = buf_size,
-        .mem_buffer = buf,
-    };
-
-    struct ggml_context * ctx0 = ggml_init(params);
-    struct ggml_cgraph gf = { .n_threads = n_threads };
-
-    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
-    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
-
-    struct ggml_tensor * 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;
-
-        // 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,
-                    ggml_transpose(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.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
-                struct ggml_tensor * v = ggml_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));
-
-                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
-                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
-            }
-
-            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
-            // [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.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);
-
-            // 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);
-
-            // KQ_scaled = KQ / sqrt(n_embd/n_head)
-            // [n_past + N, N, 12]
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale(ctx0,
-                        KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
-                        );
-
-            // KQ_masked = mask_past(KQ_scaled)
-            // [n_past + N, N, 12]
-            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
-
-            // KQ = soft_max(KQ_masked)
-            // [n_past + N, N, 12]
-            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
-
-            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
-            // [n_past + N, 64, 12]
-            struct ggml_tensor * V_trans =
-                ggml_permute(ctx0,
-                        ggml_reshape_3d(ctx0,
-                            ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
-                            n_embd/n_head, n_head, n_past + N),
-                        1, 2, 0, 3);
-
-            // KQV = transpose(V) * KQ_soft_max
-            // [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,
-                    ggml_transpose(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;
-
-        // 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,
-                    ggml_transpose(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_trans,
-                    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);
-    }
-
-    // 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));
-    }
-
-    // inpL = WTE * inpL
-    // [ 768, 50257] - model.wte
-    // [ 768, N]     - inpL
-    inpL = ggml_mul_mat(ctx0, model.wte, inpL);
-
-    // logits -> probs
-    inpL = ggml_soft_max(ctx0, inpL);
-
-    // run the computation
-    ggml_build_forward_expand(&gf, inpL);
-    ggml_graph_compute       (ctx0, &gf);
-
-    //if (n_past%100 == 0) {
-    //    ggml_graph_print   (&gf);
-    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
-    //}
-
-    //embd_w.resize(n_vocab*N);
-    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
-
-    // return result for just the last token
-    embd_w.resize(n_vocab);
-    memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
-
-    if (mem_per_token == 0) {
-        mem_per_token = ggml_used_mem(ctx0)/N;
-    }
-    //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
-
-    ggml_free(ctx0);
-
-    return true;
-}
-
-/////////////////////////////// GPT-2 END ////////////////////////////////
 
 constexpr int N_THREAD = 8;
 
-struct gpt2_state {
-    std::string prompt_base = R"(Hello, how are you?
-I'm fine, thanks. How are you?
-Thanks, I'm fine too. What are you doing?
-I'm just sitting here.
-It's a lovely day, isn't it?
-Yes, it is.
-Did you know that I'm a robot?
-I wasn't aware of that.
-)";
-
-    std::mt19937 rng;
-
-    gpt_vocab vocab;
-    gpt2_model model;
-
-    int32_t n_threads = std::min(N_THREAD, (int) std::thread::hardware_concurrency());
-    int32_t n_predict = 32; // new tokens to predict
-
-    // sampling parameters
-    int32_t top_k = 40;
-    float   top_p = 0.9f;
-    float   temp  = 1.0f;
-};
-
-struct gpt2_state g_gpt2;
-
-std::vector<float> g_pcmf32;
-
+struct gpt2_context * g_gpt2;
 std::vector<struct whisper_context *> g_contexts(4, nullptr);
 
 std::mutex g_mutex;
@@ -991,60 +27,18 @@ std::string g_text_to_speak = "";
 std::string g_status = "";
 std::string g_status_forced = "";
 
-std::string gpt2_gen_text(const std::string & prompt) {
-    int n_past = 0;
-
-    std::vector<float> embd_w;
-
-    // tokenize the prompt
-    std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(g_gpt2.vocab, prompt);
-
-    g_gpt2.n_predict = std::min(g_gpt2.n_predict, g_gpt2.model.hparams.n_ctx - (int) embd_inp.size());
-
-    std::vector<gpt_vocab::id> embd = embd_inp;
-
-    size_t mem_per_token = 3000000;
-
-    std::string result;
-
-    for (int i = embd.size(); i < embd_inp.size() + g_gpt2.n_predict; i++) {
-        // predict
-        if (embd.size() > 0) {
-            if (!gpt2_eval(g_gpt2.model, g_gpt2.n_threads, n_past, embd, embd_w, mem_per_token)) {
-                printf("gpt-2: failed to generate text\n");
-                return "";
-            }
-        }
-
-        n_past += embd.size();
-        embd.clear();
-
-        {
-            // sample next token
-            const int   top_k = g_gpt2.top_k;
-            const float top_p = g_gpt2.top_p;
-            const float temp  = g_gpt2.temp;
-
-            const int n_vocab = g_gpt2.model.hparams.n_vocab;
-
-            const gpt_vocab::id id = gpt_sample_top_k_top_p(g_gpt2.vocab, embd_w.data() + (embd_w.size() - n_vocab), top_k, top_p, temp, g_gpt2.rng);
-
-            // add it to the context
-            embd.push_back(id);
-        }
+std::vector<float> g_pcmf32;
 
-        result += g_gpt2.vocab.id_to_token[embd[0]];
+std::string to_timestamp(int64_t t) {
+    int64_t sec = t/100;
+    int64_t msec = t - sec*100;
+    int64_t min = sec/60;
+    sec = sec - min*60;
 
-        // end of text token
-        if (embd.back() == 50256 ||
-            g_gpt2.vocab.id_to_token[embd.back()] == "." ||
-            g_gpt2.vocab.id_to_token[embd.back()] == "!" ||
-            g_gpt2.vocab.id_to_token[embd.back()] == "?") {
-            break;
-        }
-    }
+    char buf[32];
+    snprintf(buf, sizeof(buf), "%02d:%02d.%03d", (int) min, (int) sec, (int) msec);
 
-    return result;
+    return std::string(buf);
 }
 
 void talk_set_status(const std::string & status) {
@@ -1072,26 +66,13 @@ void talk_main(size_t index) {
 
     wparams.language             = "en";
 
-    g_gpt2.rng = std::mt19937(time(NULL));
-
-    // load the model
-    {
-        const int64_t t_start_us = ggml_time_us();
-
-        if (!gpt2_model_load("gpt-2.bin", g_gpt2.model, g_gpt2.vocab)) {
-            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, "gpt-2.bin");
-            return;
-        }
-
-        const int64_t t_load_us = ggml_time_us() - t_start_us;
-
-        printf("gpt-2: model loaded in %d ms\n", (int) (t_load_us/1000));
-    }
+    g_gpt2 = gpt2_init("gpt-2.bin");
 
     printf("talk: using %d threads\n", N_THREAD);
 
     std::vector<float> pcmf32;
 
+    // whisper context
     auto & ctx = g_contexts[index];
 
     const int64_t step_samples = 2*WHISPER_SAMPLE_RATE;
@@ -1211,7 +192,7 @@ void talk_main(size_t index) {
 
             talk_set_status("'" + text_heard + "' - thinking how to respond ...");
 
-            const std::vector<gpt_vocab::id> tokens = ::gpt_tokenize(g_gpt2.vocab, text_heard);
+            const std::vector<gpt_vocab::id> tokens = gpt2_tokenize(g_gpt2, text_heard.c_str());
 
             printf("whisper: number of tokens: %d, '%s'\n", (int) tokens.size(), text_heard.c_str());
 
@@ -1220,11 +201,11 @@ void talk_main(size_t index) {
 
             {
                 std::lock_guard<std::mutex> lock(g_mutex);
-                prompt_base = g_gpt2.prompt_base;
+                prompt_base = gpt2_get_prompt(g_gpt2);
             }
 
             if (tokens.size() > 0) {
-                text_to_speak = gpt2_gen_text(prompt_base + text_heard + "\n");
+                text_to_speak = gpt2_gen_text(g_gpt2, (prompt_base + text_heard + "\n").c_str(), 32);
                 text_to_speak = std::regex_replace(text_to_speak, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
                 text_to_speak = text_to_speak.substr(0, text_to_speak.find_first_of("\n"));
 
@@ -1245,7 +226,7 @@ void talk_main(size_t index) {
                 }
                 prompt_base += text_heard + "\n" + text_to_speak + "\n";
             } else {
-                text_to_speak = gpt2_gen_text(prompt_base);
+                text_to_speak = gpt2_gen_text(g_gpt2, prompt_base.c_str(), 32);
                 text_to_speak = std::regex_replace(text_to_speak, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
                 text_to_speak = text_to_speak.substr(0, text_to_speak.find_first_of("\n"));
 
@@ -1269,13 +250,15 @@ void talk_main(size_t index) {
                 t_last = std::chrono::high_resolution_clock::now();
                 g_text_to_speak = text_to_speak;
                 g_pcmf32.clear();
-                g_gpt2.prompt_base = prompt_base;
+                gpt2_set_prompt(g_gpt2, prompt_base.c_str());
             }
 
             talk_set_status("speaking ...");
         }
     }
 
+    gpt2_free(g_gpt2);
+
     if (index < g_contexts.size()) {
         whisper_free(g_contexts[index]);
         g_contexts[index] = nullptr;
@@ -1351,7 +334,7 @@ EMSCRIPTEN_BINDINGS(talk) {
 
         {
             std::lock_guard<std::mutex> lock(g_mutex);
-            text_context = g_gpt2.prompt_base;
+            text_context = gpt2_get_prompt(g_gpt2);
         }
 
         return text_context;
@@ -1389,7 +372,7 @@ EMSCRIPTEN_BINDINGS(talk) {
     emscripten::function("set_prompt", emscripten::optional_override([](const std::string & prompt) {
         {
             std::lock_guard<std::mutex> lock(g_mutex);
-            g_gpt2.prompt_base = prompt;
+            gpt2_set_prompt(g_gpt2, prompt.c_str());
         }
     }));
 }

examples/talk.wasm/gpt-2.cpp

@@ -0,0 +1,925 @@
+#include "ggml.h"
+#include "gpt-2.h"
+
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <thread>
+#include <vector>
+#include <regex>
+#include <random>
+
+/////////////////////// GPT-2 BEGIN /////////////////////////
+
+//
+// Vocab utils
+//
+
+std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text) {
+    std::vector<std::string> words;
+
+    // first split the text into words
+    {
+        std::string str = text;
+        std::string pat = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
+
+        std::regex re(pat);
+        std::smatch m;
+
+        while (std::regex_search(str, m, re)) {
+            for (auto x : m) {
+                words.push_back(x);
+            }
+            str = m.suffix();
+        }
+    }
+
+    // find the longest tokens that form the words:
+    std::vector<gpt_vocab::id> tokens;
+    for (const auto & word : words) {
+        if (word.size() == 0) continue;
+
+        int i = 0;
+        int n = word.size();
+        while (i < n) {
+            int j = n;
+            while (j > i) {
+                auto it = vocab.token_to_id.find(word.substr(i, j-i));
+                if (it != vocab.token_to_id.end()) {
+                    tokens.push_back(it->second);
+                    i = j;
+                    break;
+                }
+                --j;
+            }
+            if (i == n) {
+                break;
+            }
+            if (j == i) {
+                auto sub = word.substr(i, 1);
+                if (vocab.token_to_id.find(sub) != vocab.token_to_id.end()) {
+                    tokens.push_back(vocab.token_to_id.at(sub));
+                } else {
+                    fprintf(stderr, "%s: unknown token '%s'\n", __func__, sub.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<std::pair<double, gpt_vocab::id>> logits_id;
+    logits_id.reserve(n_logits);
+
+    for (int i = 0; i < n_logits; i++) {
+        logits_id.push_back(std::make_pair(logits[i], i));
+    }
+
+    // find the top K tokens
+    std::partial_sort(
+            logits_id.begin(),
+            logits_id.begin() + top_k, logits_id.end(),
+            [](const std::pair<double, gpt_vocab::id> & a, const std::pair<double, gpt_vocab::id> & b) {
+        return a.first > b.first;
+    });
+
+    logits_id.resize(top_k);
+
+    // normalize
+    {
+        double sum = 0.0f;
+        for (int i = 0; i < (int)logits_id.size(); i++) {
+            sum += logits_id[i].first;
+        }
+
+        sum = 1.0/sum;
+        for (int i = 0; i < (int)logits_id.size(); i++) {
+            logits_id[i].first *= sum;
+        }
+    }
+
+    if (top_p < 1.0f) {
+        {
+            double cumsum = 0.0f;
+            for (int i = 0; i < top_k; i++) {
+                cumsum += logits_id[i].first;
+                if (cumsum >= top_p) {
+                    logits_id.resize(i+1);
+                    break;
+                }
+            }
+        }
+
+        // normalize again
+        {
+            double sum = 0.0f;
+            for (int i = 0; i < (int)logits_id.size(); i++) {
+                sum += logits_id[i].first;
+            }
+
+            sum = 1.0/sum;
+            for (int i = 0; i < (int)logits_id.size(); i++) {
+                logits_id[i].first *= sum;
+            }
+        }
+    }
+
+    //printf("\n");
+    //for (int i = 0; i < (int)logits_id.size(); i++) {
+    //    printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), logits_id[i].first);
+    //}
+    //exit(0);
+
+    // sample from the obtained distribution
+    std::vector<double> probs;
+    probs.reserve(logits_id.size());
+
+    for (int i = 0; i < (int) logits_id.size(); i++) {
+        probs.push_back(logits_id[i].first);
+    }
+
+    std::discrete_distribution<> dist(probs.begin(), probs.end());
+    int idx = dist(rng);
+
+    return logits_id[idx].second;
+}
+
+// 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 f16     = 1;
+};
+
+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_trans; // transposed for efficiency
+    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
+
+    std::vector<gpt2_layer> layers;
+
+    // key + value memory
+    struct ggml_tensor * memory_k;
+    struct ggml_tensor * memory_v;
+
+    //
+    struct ggml_context * ctx;
+    std::map<std::string, struct ggml_tensor *> tensors;
+};
+
+// load the model's weights from a file
+bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab) {
+    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 != 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.f16,     sizeof(hparams.f16));
+
+        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: f16     = %d\n", __func__, hparams.f16);
+    }
+
+    // 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;
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            fin.read((char *) &len, sizeof(len));
+
+            word.resize(len);
+            fin.read((char *) word.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // for the big tensors, we have the option to store the data in 16-bit floats
+    // in order to save memory and also to speed up the computation
+    const ggml_type wtype = model.hparams.f16 ? GGML_TYPE_F16 : GGML_TYPE_F32;
+
+    auto & ctx = model.ctx;
+
+    size_t ctx_size = 0;
+
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        ctx_size += n_embd*ggml_type_size(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_size(GGML_TYPE_F32); // ln_f_b
+
+        ctx_size += n_vocab*n_embd*ggml_type_size(wtype);         // wte
+        ctx_size +=   n_ctx*n_embd*ggml_type_size(GGML_TYPE_F32); // wpe
+
+        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_1_b
+
+        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_2_g
+        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_2_b
+
+        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_size(wtype));         // c_attn_attn_w
+        ctx_size += n_layer*(       3*n_embd*ggml_type_size(GGML_TYPE_F32)); // c_attn_attn_b
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_size(wtype));           // c_attn_proj_w
+        ctx_size += n_layer*(       n_embd*ggml_type_size(GGML_TYPE_F32));   // c_attn_proj_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_size(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_size(GGML_TYPE_F32)); // c_mlp_fc_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_size(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_size(GGML_TYPE_F32)); // c_mlp_proj_b
+
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_size(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_size(GGML_TYPE_F32); // memory_v
+
+        ctx_size += (6 + 12*n_layer)*256; // object overhead
+
+        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
+    }
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            .mem_size   = ctx_size,
+            .mem_buffer = NULL,
+        };
+
+        model.ctx = ggml_init(params);
+        if (!model.ctx) {
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+
+    // prepare memory for the weights
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        model.layers.resize(n_layer);
+
+        model.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);
+
+        // 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;
+
+        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,         3*n_embd, 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,         4*n_embd, n_embd);
+            layer.c_mlp_fc_b         = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
+
+            layer.c_mlp_proj_w_trans = 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_trans;
+            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_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+
+        const int n_mem      = n_layer*n_ctx;
+        const int n_elements = n_embd*n_mem;
+
+        model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
+        model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
+
+        const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
+
+        printf("%s: memory size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
+    }
+
+    // load weights
+    {
+        size_t total_size = 0;
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ftype;
+
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+
+            if (fin.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<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 (ggml_nelements(tensor) != nelements) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
+                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.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
+                return false;
+            }
+
+            const size_t bpe = (ftype == 0) ? sizeof(float) : sizeof(ggml_fp16_t);
+
+            if (nelements*bpe != 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));
+
+            //printf("%24s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
+            total_size += ggml_nbytes(tensor);
+        }
+
+        printf("%s: model size  = %8.2f MB\n", __func__, total_size/1024.0/1024.0);
+    }
+
+    fin.close();
+
+    return true;
+}
+
+// evaluate the transformer
+//
+//   - model:     the model
+//   - n_threads: number of threads to use
+//   - n_past:    the context size so far
+//   - embd_inp:  the embeddings of the tokens in the context
+//   - embd_w:    the predicted probabilities of the next token
+//
+bool gpt2_eval(
+        const gpt2_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;
+
+    static size_t buf_size = 512u*1024*1024;
+    static void * buf = malloc(buf_size);
+
+    if (mem_per_token > 0 && mem_per_token*N > buf_size) {
+        const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
+        printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
+
+        // reallocate
+        buf_size = buf_size_new;
+        buf = realloc(buf, buf_size);
+        if (buf == nullptr) {
+            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
+            return false;
+        }
+    }
+
+    struct ggml_init_params params = {
+        .mem_size   = buf_size,
+        .mem_buffer = buf,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph gf = { .n_threads = n_threads };
+
+    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
+
+    struct ggml_tensor * 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;
+
+        // 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,
+                    ggml_transpose(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.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));
+
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
+            // [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.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);
+
+            // 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);
+
+            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
+                        );
+
+            // KQ_masked = mask_past(KQ_scaled)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
+
+            // KQ = soft_max(KQ_masked)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
+
+            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
+            // [n_past + N, 64, 12]
+            struct ggml_tensor * V_trans =
+                ggml_permute(ctx0,
+                        ggml_reshape_3d(ctx0,
+                            ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
+                            n_embd/n_head, n_head, n_past + N),
+                        1, 2, 0, 3);
+
+            // KQV = transpose(V) * KQ_soft_max
+            // [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,
+                    ggml_transpose(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;
+
+        // 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,
+                    ggml_transpose(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_trans,
+                    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);
+    }
+
+    // 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));
+    }
+
+    // inpL = WTE * inpL
+    // [ 768, 50257] - model.wte
+    // [ 768, N]     - inpL
+    inpL = ggml_mul_mat(ctx0, model.wte, inpL);
+
+    // logits -> probs
+    inpL = ggml_soft_max(ctx0, inpL);
+
+    // run the computation
+    ggml_build_forward_expand(&gf, inpL);
+    ggml_graph_compute       (ctx0, &gf);
+
+    //if (n_past%100 == 0) {
+    //    ggml_graph_print   (&gf);
+    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
+    //}
+
+    //embd_w.resize(n_vocab*N);
+    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+
+    // return result for just the last token
+    embd_w.resize(n_vocab);
+    memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+
+    if (mem_per_token == 0) {
+        mem_per_token = ggml_used_mem(ctx0)/N;
+    }
+    //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
+
+    ggml_free(ctx0);
+
+    return true;
+}
+
+/////////////////////////////// GPT-2 END ////////////////////////////////
+
+constexpr int N_THREAD = 8;
+
+struct gpt2_context {
+    std::string prompt_base = R"(Hello, how are you?
+I'm fine, thanks. How are you?
+Thanks, I'm fine too. What are you doing?
+I'm just sitting here.
+It's a lovely day, isn't it?
+Yes, it is.
+Did you know that I'm a robot?
+I wasn't aware of that.
+)";
+
+    std::mt19937 rng;
+
+    gpt_vocab vocab;
+    gpt2_model model;
+
+    int32_t n_threads = std::min(N_THREAD, (int) std::thread::hardware_concurrency());
+
+    // sampling parameters
+    int32_t top_k = 40;
+    float   top_p = 0.9f;
+    float   temp  = 1.0f;
+};
+
+struct gpt2_context * gpt2_init(const char * path_model) {
+    gpt2_context * ctx = new gpt2_context;
+
+    ctx->rng = std::mt19937(time(NULL));
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!gpt2_model_load(path_model, ctx->model, ctx->vocab)) {
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, "gpt-2.bin");
+            return nullptr;
+        }
+
+        const int64_t t_load_us = ggml_time_us() - t_start_us;
+
+        printf("gpt-2: model loaded in %d ms\n", (int) (t_load_us/1000));
+    }
+
+    return ctx;
+}
+
+void gpt2_free(struct gpt2_context * ctx) {
+    delete ctx;
+}
+
+const char * gpt2_get_prompt(struct gpt2_context * ctx) {
+    return ctx->prompt_base.c_str();
+}
+
+void gpt2_set_prompt(struct gpt2_context * ctx, const char * prompt) {
+    ctx->prompt_base = prompt;
+}
+
+std::vector<gpt_vocab::id> gpt2_tokenize(const gpt2_context * ctx, const char * text) {
+    return ::gpt_tokenize(ctx->vocab, text);
+}
+
+std::string gpt2_gen_text(gpt2_context * ctx, const char * text, int max_tokens) {
+    int n_past = 0;
+
+    std::vector<float> embd_w;
+
+    // tokenize the prompt
+    std::vector<gpt_vocab::id> embd_inp = ::gpt2_tokenize(ctx, text);
+
+    int n_predict = std::min(max_tokens, ctx->model.hparams.n_ctx - (int) embd_inp.size());
+
+    std::vector<gpt_vocab::id> embd = embd_inp;
+
+    size_t mem_per_token = 3000000;
+
+    std::string result;
+
+    for (int i = embd.size(); i < embd_inp.size() + n_predict; i++) {
+        // predict
+        if (embd.size() > 0) {
+            if (!gpt2_eval(ctx->model, ctx->n_threads, n_past, embd, embd_w, mem_per_token)) {
+                printf("gpt-2: failed to generate text\n");
+                return "";
+            }
+        }
+
+        n_past += embd.size();
+        embd.clear();
+
+        {
+            // sample next token
+            const int   top_k = ctx->top_k;
+            const float top_p = ctx->top_p;
+            const float temp  = ctx->temp;
+
+            const int n_vocab = ctx->model.hparams.n_vocab;
+
+            const gpt_vocab::id id = gpt_sample_top_k_top_p(ctx->vocab, embd_w.data() + (embd_w.size() - n_vocab), top_k, top_p, temp, ctx->rng);
+
+            // add it to the context
+            embd.push_back(id);
+        }
+
+        result += ctx->vocab.id_to_token[embd[0]];
+
+        // end of text token
+        if (embd.back() == 50256 ||
+            ctx->vocab.id_to_token[embd.back()] == "." ||
+            ctx->vocab.id_to_token[embd.back()] == "!" ||
+            ctx->vocab.id_to_token[embd.back()] == "?") {
+            break;
+        }
+    }
+
+    return result;
+}

examples/talk.wasm/gpt-2.h

@@ -0,0 +1,27 @@
+#pragma once
+
+// TODO: Change to C-style API and move to ./examples for easy reuse.
+
+#include <vector>
+#include <map>
+#include <string>
+
+struct gpt_vocab {
+    using id    = int32_t;
+    using token = std::string;
+
+    std::map<token, id> token_to_id;
+    std::map<id, token> id_to_token;
+};
+
+struct gpt2_context;
+
+struct gpt2_context * gpt2_init(const char * path_model);
+void gpt2_free(struct gpt2_context * ctx);
+
+const char * gpt2_get_prompt(struct gpt2_context * ctx);
+void gpt2_set_prompt(struct gpt2_context * ctx, const char * prompt);
+
+std::vector<gpt_vocab::id> gpt2_tokenize(const gpt2_context * ctx, const char * text);
+
+std::string gpt2_gen_text(gpt2_context * ctx, const char * text, int max_tokens);

examples/talk.wasm/index-tmpl.html

@@ -504,7 +504,7 @@
 
             function startRecording() {
                 if (!context) {
-                    context = new AudioContext({sampleRate: 16000});
+                    context = new AudioContext({sampleRate: 16000, noiseSuppression: true});
                 }
 
                 Module.set_status("");