talk-llama : add new example + sync ggml from llama.cpp (#664)

* talk-llama : talk with LLaMA AI

* talk.llama : disable EOS token

* talk-llama : add README instructions

* ggml : fix build in debug

.gitignore

@@ -18,6 +18,7 @@ build-sanitize-thread/
 /stream
 /command
 /talk
+/talk-llama
 /bench
 
 arm_neon.h
@@ -32,3 +33,5 @@ examples/whisper.objc/whisper.objc.xcodeproj/xcuserdata/
 examples/whisper.objc/whisper.objc.xcodeproj/project.xcworkspace/xcuserdata
 
 extra/bench-gg.txt
+
+*.mlmodel*

Makefile

@@ -36,7 +36,7 @@ LDFLAGS  =
 
 # ref: https://github.com/ggerganov/whisper.cpp/issues/37
 ifneq ($(wildcard /usr/include/musl/*),)
-	CFLAGS   += -D_POSIX_SOURCE -D_GNU_SOURCE
+	CFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
 	CXXFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
 endif
 
@@ -178,7 +178,7 @@ $(info I CC:       $(CCV))
 $(info I CXX:      $(CXXV))
 $(info )
 
-default: main
+default: main bench
 
 #
 # Build library
@@ -197,7 +197,7 @@ libwhisper.so: ggml.o whisper.o
 	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o whisper.o $(LDFLAGS)
 
 clean:
-	rm -f *.o main stream command talk bench libwhisper.a libwhisper.so
+	rm -f *.o main stream command talk talk-llama bench libwhisper.a libwhisper.so
 
 #
 # Examples
@@ -212,6 +212,9 @@ main: examples/main/main.cpp $(SRC_COMMON) ggml.o whisper.o
 	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) ggml.o whisper.o -o main $(LDFLAGS)
 	./main -h
 
+bench: examples/bench/bench.cpp ggml.o whisper.o
+	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o whisper.o -o bench $(LDFLAGS)
+
 stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o
 	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o -o stream $(CC_SDL) $(LDFLAGS)
 
@@ -221,8 +224,8 @@ command: examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whi
 talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o
 	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o -o talk $(CC_SDL) $(LDFLAGS)
 
-bench: examples/bench/bench.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o whisper.o -o bench $(LDFLAGS)
+talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o
+	$(CXX) $(CXXFLAGS) examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o -o talk-llama $(CC_SDL) $(LDFLAGS)
 
 #
 # Audio samples

examples/CMakeLists.txt

@@ -63,4 +63,5 @@ else()
     add_subdirectory(command)
     add_subdirectory(bench)
     add_subdirectory(talk)
+    add_subdirectory(talk-llama)
 endif()

examples/talk-llama/.gitignore

@@ -0,0 +1,2 @@
+eleven-labs.py
+audio.mp3

examples/talk-llama/CMakeLists.txt

@@ -0,0 +1,10 @@
+if (WHISPER_SUPPORT_SDL2)
+    # talk-llama
+    set(TARGET talk-llama)
+
+    add_executable(${TARGET} talk-llama.cpp llama.cpp)
+
+    include(DefaultTargetOptions)
+
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+endif ()

examples/talk-llama/README.md

@@ -0,0 +1,32 @@
+# talk-llama
+
+Talk with an LLaMA AI in your terminal
+
+[Demo Talk](https://user-images.githubusercontent.com/1991296/228024237-848f998c-c334-46a6-bef8-3271590da83b.mp4)
+
+## Building
+
+The `talk-llama` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
+
+```bash
+# Install SDL2 on Linux
+sudo apt-get install libsdl2-dev
+
+# Install SDL2 on Mac OS
+brew install sdl2
+
+# Build the "talk-llama" executable
+make talk-llama
+
+# Run it
+./talk-llama -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/13B/ggml-model-q4_0.bin -p "Georgi" -t 8
+```
+
+- The `-mw` argument specifies the Whisper model that you would like to use. Recommended `base` or `small` for real-time experience
+- The `-ml` argument specifies the LLaMA model that you would like to use. Read the instructions in https://github.com/ggerganov/llama.cpp for information about how to obtain a `ggml` compatible LLaMA model
+
+## TTS
+
+For best experience, this example needs a TTS tool to convert the generated text responses to voice.
+You can use any TTS engine that you would like - simply edit the [speak.sh](speak.sh) script to your needs.
+By default, it is configured to use MacOS's `say`, but you can use whatever you wish.

examples/talk-llama/llama.cpp

@@ -0,0 +1,1865 @@
+#include "llama.h"
+
+#include "ggml.h"
+
+#include <cinttypes>
+#include <fstream>
+#include <random>
+#include <map>
+#include <unordered_map>
+#include <queue>
+#include <regex>
+#include <cassert>
+#include <cstring>
+
+#define LLAMA_USE_SCRATCH
+#define LLAMA_MAX_SCRATCH_BUFFERS 16
+
+#define LLAMA_ASSERT(x) \
+    do { \
+        if (!(x)) { \
+            fprintf(stderr, "LLAMA_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
+            abort(); \
+        } \
+    } while (0)
+
+
+// determine number of model parts based on the dimension
+static const std::unordered_map<int, int> LLAMA_N_PARTS = {
+    { 4096, 1 },
+    { 5120, 2 },
+    { 6656, 4 },
+    { 8192, 8 },
+};
+
+// available llama models
+enum e_model {
+    MODEL_UNKNOWN,
+    MODEL_7B,
+    MODEL_13B,
+    MODEL_30B,
+    MODEL_65B,
+};
+
+static const size_t MB = 1024*1024;
+
+// computed for n_ctx == 2048
+// TODO: dynamically determine these sizes
+//       needs modifications in ggml
+
+static const std::map<e_model, size_t> MEM_REQ_SCRATCH0 = {
+    { MODEL_7B,    512ull*MB },
+    { MODEL_13B,   512ull*MB },
+    { MODEL_30B,   512ull*MB },
+    { MODEL_65B,   512ull*MB },
+};
+
+static const std::map<e_model, size_t> MEM_REQ_SCRATCH1 = {
+    { MODEL_7B,    512ull*MB },
+    { MODEL_13B,   512ull*MB },
+    { MODEL_30B,   512ull*MB },
+    { MODEL_65B,   512ull*MB },
+};
+
+// 2*n_embd*n_ctx*n_layer*sizeof(float16)
+static const std::map<e_model, size_t> MEM_REQ_KV_SELF = {
+    { MODEL_7B,   1026ull*MB },
+    { MODEL_13B,  1608ull*MB },
+    { MODEL_30B,  3124ull*MB },
+    { MODEL_65B,  5120ull*MB },
+};
+
+// this is mostly needed for temporary mul_mat buffers to dequantize the data
+// not actually needed if BLAS is disabled
+static const std::map<e_model, size_t> MEM_REQ_EVAL = {
+    { MODEL_7B,   768ull*MB },
+    { MODEL_13B, 1024ull*MB },
+    { MODEL_30B, 1280ull*MB },
+    { MODEL_65B, 1536ull*MB },
+};
+
+// default hparams (LLaMA 7B)
+struct llama_hparams {
+    int32_t n_vocab = 32000;
+    int32_t n_ctx   = 512;   // this is provided as user input?
+    int32_t n_embd  = 4096;
+    int32_t n_mult  = 256;
+    int32_t n_head  = 32;
+    int32_t n_layer = 32;
+    int32_t n_rot   = 64;
+    int32_t f16     = 1;
+};
+
+struct llama_layer {
+    // normalization
+    struct ggml_tensor * attention_norm;
+
+    // attention
+    struct ggml_tensor * wq;
+    struct ggml_tensor * wk;
+    struct ggml_tensor * wv;
+    struct ggml_tensor * wo;
+
+    // normalization
+    struct ggml_tensor * ffn_norm;
+
+    // ff
+    struct ggml_tensor * w1;
+    struct ggml_tensor * w2;
+    struct ggml_tensor * w3;
+};
+
+struct llama_kv_cache {
+    struct ggml_tensor * k;
+    struct ggml_tensor * v;
+
+    struct ggml_context * ctx;
+
+    std::vector<uint8_t> buf;
+
+    int n; // number of tokens currently in the cache
+};
+
+struct llama_model {
+    e_model type = MODEL_UNKNOWN;
+
+    llama_hparams hparams;
+
+    struct ggml_tensor * tok_embeddings;
+
+    struct ggml_tensor * norm;
+    struct ggml_tensor * output;
+
+    std::vector<llama_layer> layers;
+
+    // context
+    struct ggml_context * ctx;
+
+    // key + value cache for the self attention
+    // TODO: move to llama_state
+    struct llama_kv_cache kv_self;
+
+    // the model memory buffer
+    std::vector<uint8_t> buf;
+
+    // tensors
+    int n_loaded;
+    std::unordered_map<std::string, struct ggml_tensor *> tensors;
+};
+
+struct llama_vocab {
+    using id    = int32_t;
+    using token = std::string;
+
+    struct token_score {
+        token tok;
+        float score;
+    };
+
+    std::unordered_map<token, id> token_to_id;
+    std::vector<token_score> id_to_token;
+};
+
+struct llama_context {
+    std::mt19937 rng;
+
+    int64_t t_load_us = 0;
+    int64_t t_start_us = 0;
+
+    int64_t t_sample_us = 0;
+    int64_t t_eval_us   = 0;
+    int64_t t_p_eval_us = 0;
+
+    int32_t n_sample = 0; // number of tokens sampled
+    int32_t n_eval   = 0; // number of eval calls
+    int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1)
+
+    llama_model model;
+    llama_vocab vocab;
+
+    size_t mem_per_token = 0;
+
+    // decode output (2-dimensional array: [n_tokens][n_vocab])
+    std::vector<float> logits;
+    bool logits_all = false;
+
+    // input embedding (1-dimensional array: [n_embd])
+    std::vector<float> embedding;
+
+    // memory buffers used to evaluate the model
+    // TODO: move in llama_state
+    std::vector<uint8_t> buf_compute;
+    std::vector<uint8_t> buf_scratch[LLAMA_MAX_SCRATCH_BUFFERS];
+
+    int    buf_last = 0;
+    size_t buf_max_size[LLAMA_MAX_SCRATCH_BUFFERS] = { 0 };
+
+    void use_buf(struct ggml_context * ctx, int i) {
+#if defined(LLAMA_USE_SCRATCH)
+        size_t last_size = 0;
+
+        if (i == -1) {
+            last_size = ggml_set_scratch(ctx, { 0, 0, nullptr, });
+        } else {
+            auto & buf = buf_scratch[i];
+            last_size = ggml_set_scratch(ctx, { 0, buf.size(), buf.data(), });
+        }
+
+        if (buf_last >= 0) {
+            buf_max_size[buf_last] = std::max(buf_max_size[buf_last], last_size);
+        }
+
+        buf_last = i;
+#else
+        (void) i;
+        (void) ctx;
+#endif
+    }
+
+    size_t get_buf_max_mem(int i) const {
+#if defined(LLAMA_USE_SCRATCH)
+        return buf_max_size[i];
+#else
+        (void) i;
+        return 0;
+#endif
+    }
+};
+
+//
+// kv cache
+//
+
+static bool kv_cache_init(
+        const struct llama_hparams & hparams,
+             struct llama_kv_cache & cache,
+                         ggml_type   wtype,
+                               int   n_ctx) {
+    const int n_embd  = hparams.n_embd;
+    const int n_layer = hparams.n_layer;
+
+    const int n_mem      = n_layer*n_ctx;
+    const int n_elements = n_embd*n_mem;
+
+    cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
+
+    struct ggml_init_params params;
+    params.mem_size   = cache.buf.size();
+    params.mem_buffer = cache.buf.data();
+
+    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;
+}
+
+static void kv_cache_free(struct llama_kv_cache & cache) {
+    if (cache.ctx) {
+        ggml_free(cache.ctx);
+        cache.ctx = nullptr;
+    }
+}
+
+struct llama_context_params llama_context_default_params() {
+    struct llama_context_params result = {
+        /*.n_ctx                       =*/ 512,
+        /*.n_parts                     =*/ -1,
+        /*.seed                        =*/ 0,
+        /*.f16_kv                      =*/ false,
+        /*.logits_all                  =*/ false,
+        /*.vocab_only                  =*/ false,
+        /*.use_mlock                   =*/ false,
+        /*.embedding                   =*/ false,
+        /*.progress_callback           =*/ nullptr,
+        /*.progress_callback_user_data =*/ nullptr,
+    };
+
+    return result;
+}
+
+//
+// model loading
+//
+
+static bool llama_model_load(
+        const std::string & fname,
+        llama_context & lctx,
+        int n_ctx,
+        int n_parts,
+        ggml_type memory_type,
+        bool vocab_only,
+        llama_progress_callback progress_callback,
+        void *progress_callback_user_data) {
+    fprintf(stderr, "%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
+
+    const int64_t t_start_us = ggml_time_us();
+
+    lctx.t_start_us = t_start_us;
+
+    std::vector<char> f_buf(1024*1024);
+
+    auto & model = lctx.model;
+    auto & vocab = lctx.vocab;
+
+    auto fin = std::ifstream(fname, std::ios::binary);
+    fin.rdbuf()->pubsetbuf(f_buf.data(), f_buf.size());
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        fin.read((char *) &magic, sizeof(magic));
+        if (magic == LLAMA_FILE_MAGIC_UNVERSIONED) {
+            fprintf(stderr, "%s: invalid model file '%s' (too old, regenerate your model files!)\n",
+                    __func__, fname.c_str());
+            return false;
+        }
+        if (magic != LLAMA_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+            return false;
+        }
+
+        uint32_t format_version;
+        fin.read((char *) &format_version, sizeof(format_version));
+
+        if (format_version != LLAMA_FILE_VERSION) {
+            fprintf(stderr, "%s: invalid model file '%s' (unsupported format version %" PRIu32 ", expected %d)\n",
+                    __func__, fname.c_str(), format_version, LLAMA_FILE_VERSION);
+            return false;
+        }
+    }
+
+    int n_ff = 0;
+
+    // load hparams
+    {
+        auto & hparams = model.hparams;
+
+        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        //fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fin.read((char *) &hparams.n_mult,  sizeof(hparams.n_mult));
+        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fin.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        fin.read((char *) &hparams.f16,     sizeof(hparams.f16));
+
+        hparams.n_ctx = n_ctx;
+
+        n_ff = ((2*(4*hparams.n_embd)/3 + hparams.n_mult - 1)/hparams.n_mult)*hparams.n_mult;
+
+        if (n_parts < 1) {
+            n_parts = LLAMA_N_PARTS.at(hparams.n_embd);
+        }
+
+        // temp warning to tell the user to use "--n_parts"
+        if (hparams.f16 == 4 && n_parts != 1) {
+            fprintf(stderr, "%s: GPTQ model detected - are you sure n_parts should be %d? we normally expect it to be 1\n", __func__, n_parts);
+            fprintf(stderr, "%s: use '--n_parts 1' if necessary\n", __func__);
+        }
+
+        if (hparams.n_layer == 32) {
+            model.type = e_model::MODEL_7B;
+        }
+
+        if (hparams.n_layer == 40) {
+            model.type = e_model::MODEL_13B;
+        }
+
+        if (hparams.n_layer == 60) {
+            model.type = e_model::MODEL_30B;
+        }
+
+        if (hparams.n_layer == 80) {
+            model.type = e_model::MODEL_65B;
+        }
+
+        fprintf(stderr, "%s: n_vocab = %d\n", __func__, hparams.n_vocab);
+        fprintf(stderr, "%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
+        fprintf(stderr, "%s: n_embd  = %d\n", __func__, hparams.n_embd);
+        fprintf(stderr, "%s: n_mult  = %d\n", __func__, hparams.n_mult);
+        fprintf(stderr, "%s: n_head  = %d\n", __func__, hparams.n_head);
+        fprintf(stderr, "%s: n_layer = %d\n", __func__, hparams.n_layer);
+        fprintf(stderr, "%s: n_rot   = %d\n", __func__, hparams.n_rot);
+        fprintf(stderr, "%s: f16     = %d\n", __func__, hparams.f16);
+        fprintf(stderr, "%s: n_ff    = %d\n", __func__, n_ff);
+        fprintf(stderr, "%s: n_parts = %d\n", __func__, n_parts);
+        fprintf(stderr, "%s: type    = %d\n", __func__, model.type);
+    }
+
+    // load vocab
+    {
+        std::string word;
+        vocab.id_to_token.resize(model.hparams.n_vocab);
+        std::vector<char> tmp(64);
+
+        for (int i = 0; i < model.hparams.n_vocab; i++) {
+            uint32_t len;
+            fin.read((char *) &len, sizeof(len));
+
+            word.resize(len);
+            if (len > 0) {
+                tmp.resize(len);
+                fin.read(tmp.data(), len);
+                word.assign(tmp.data(), len);
+            } else {
+                word.clear();
+            }
+
+            float score;
+            fin.read((char *) &score, sizeof(score));
+
+            vocab.token_to_id[word] = i;
+
+            auto &tok_score = vocab.id_to_token[i];
+            tok_score.tok = word;
+            tok_score.score = score;
+        }
+    }
+
+    if (vocab_only) {
+        return true;
+    }
+
+    // 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
+    // wtype is for per-layer weights, while vtype is for other weights
+    ggml_type wtype, vtype;
+    switch (model.hparams.f16) {
+        case 0: wtype = vtype = GGML_TYPE_F32;  break;
+        case 1: wtype = vtype = GGML_TYPE_F16;  break;
+        case 2: wtype = vtype = GGML_TYPE_Q4_0; break;
+        case 3: wtype = vtype = GGML_TYPE_Q4_1; break;
+        case 4: wtype = GGML_TYPE_Q4_1; vtype = GGML_TYPE_F16; break;
+        default:
+                {
+                    fprintf(stderr, "%s: invalid model file '%s' (bad f16 value %d)\n",
+                            __func__, fname.c_str(), model.hparams.f16);
+                    return false;
+                }
+    }
+
+    auto & ctx = model.ctx;
+
+    size_t ctx_size = 0;
+
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(vtype); // tok_embeddings
+
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // norm
+
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(vtype); // output
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // attention_norm
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wq
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wk
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wv
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wo
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ffn_norm
+
+        ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w1
+        ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w2
+        ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w3
+
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(memory_type); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(memory_type); // memory_v
+
+        ctx_size += (5 + 10*n_layer)*256; // object overhead
+
+        fprintf(stderr, "%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
+    }
+
+    // print memory requirements
+    {
+        const size_t scale = memory_type == GGML_TYPE_F32 ? 2 : 1;
+
+        // this is the total memory required to run the inference
+        const size_t mem_required =
+            ctx_size +
+            MEM_REQ_SCRATCH0.at(model.type) +
+            MEM_REQ_SCRATCH1.at(model.type) +
+            MEM_REQ_EVAL.at    (model.type);
+
+        // this is the memory required by one llama_state
+        const size_t mem_required_state =
+            scale*MEM_REQ_KV_SELF.at(model.type);
+
+        fprintf(stderr, "%s: mem required  = %7.2f MB (+ %7.2f MB per state)\n", __func__,
+                mem_required / 1024.0 / 1024.0, mem_required_state / 1024.0 / 1024.0);
+    }
+
+    // create the ggml context
+    {
+        lctx.model.buf.resize(ctx_size);
+
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ lctx.model.buf.size(),
+            /*.mem_buffer =*/ lctx.model.buf.data(),
+        };
+
+        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_vocab = hparams.n_vocab;
+
+        model.layers.resize(n_layer);
+
+        model.tok_embeddings = ggml_new_tensor_2d(ctx, vtype, n_embd, n_vocab);
+
+        model.norm   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+        model.output = ggml_new_tensor_2d(ctx, vtype,         n_embd, n_vocab);
+
+        // map by name
+        model.tensors["tok_embeddings.weight"] = model.tok_embeddings;
+
+        model.tensors["norm.weight"]   = model.norm;
+        model.tensors["output.weight"] = model.output;
+
+        for (int i = 0; i < n_layer; ++i) {
+            auto & layer = model.layers[i];
+
+            layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+            layer.wq = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
+            layer.wk = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
+            layer.wv = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
+            layer.wo = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
+
+            layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+            layer.w1 = ggml_new_tensor_2d(ctx, wtype, n_embd,   n_ff);
+            layer.w2 = ggml_new_tensor_2d(ctx, wtype,   n_ff, n_embd);
+            layer.w3 = ggml_new_tensor_2d(ctx, wtype, n_embd,   n_ff);
+
+            // map by name
+            model.tensors["layers." + std::to_string(i) + ".attention_norm.weight"] = layer.attention_norm;
+
+            model.tensors["layers." + std::to_string(i) + ".attention.wq.weight"] = layer.wq;
+            model.tensors["layers." + std::to_string(i) + ".attention.wk.weight"] = layer.wk;
+            model.tensors["layers." + std::to_string(i) + ".attention.wv.weight"] = layer.wv;
+            model.tensors["layers." + std::to_string(i) + ".attention.wo.weight"] = layer.wo;
+
+            model.tensors["layers." + std::to_string(i) + ".ffn_norm.weight"] = layer.ffn_norm;
+
+            model.tensors["layers." + std::to_string(i) + ".feed_forward.w1.weight"] = layer.w1;
+            model.tensors["layers." + std::to_string(i) + ".feed_forward.w2.weight"] = layer.w2;
+            model.tensors["layers." + std::to_string(i) + ".feed_forward.w3.weight"] = layer.w3;
+        }
+    }
+
+    const size_t file_offset = fin.tellg();
+
+    fin.close();
+
+    std::vector<uint8_t> tmp;
+
+    if (progress_callback) {
+        progress_callback(0.0, progress_callback_user_data);
+    }
+
+    for (int i = 0; i < n_parts; ++i) {
+        const int part_id = i;
+        //const int part_id = n_parts - i - 1;
+
+        std::string fname_part = fname;
+        if (i > 0) {
+            fname_part += "." + std::to_string(i);
+        }
+
+        fprintf(stderr, "%s: loading model part %d/%d from '%s'\n", __func__, i+1, n_parts, fname_part.c_str());
+
+        fin = std::ifstream(fname_part, std::ios::binary);
+        fin.rdbuf()->pubsetbuf(f_buf.data(), f_buf.size());
+
+        fin.seekg(0, fin.end);
+        const size_t file_size = fin.tellg();
+
+        fin.seekg(file_offset);
+
+        // load weights
+        {
+            size_t total_size = 0;
+
+            model.n_loaded = 0;
+
+            fprintf(stderr, "%s: ", __func__);
+
+            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;
+                }
+
+                // split_type = 0: split by columns
+                // split_type = 1: split by rows
+                int split_type = 0;
+
+                // split_type = 0:
+                // regex:
+                //   - tok_embeddings.*
+                //   - layers.*.attention.wo.weight
+                //   - layers.*.feed_forward.w2.weight
+
+                // split_type = 1:
+                // regex:
+                //   - output.*
+                //   - layers.*.attention.wq.weight
+                //   - layers.*.attention.wk.weight
+                //   - layers.*.attention.wv.weight
+                //   - layers.*.feed_forward.w1.weight
+                //   - layers.*.feed_forward.w3.weight
+                if (name.find("tok_embeddings") != std::string::npos) {
+                    split_type = 0;
+                } else if (name.find("layers") != std::string::npos) {
+                    if (name.find("attention.wo.weight") != std::string::npos) {
+                        split_type = 0;
+                    } else if (name.find("feed_forward.w2.weight") != std::string::npos) {
+                        split_type = 0;
+                    } else {
+                        split_type = 1;
+                    }
+                } else if (name.find("output") != std::string::npos) {
+                    split_type = 1;
+                }
+
+                auto tensor = model.tensors[name.data()];
+
+                if (n_dims == 1) {
+                    if (ggml_nelements(tensor) != nelements) {
+                        fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
+                        return false;
+                    }
+                } else {
+                    if (ggml_nelements(tensor)/n_parts != nelements) {
+                        fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
+                        return false;
+                    }
+                }
+
+                if (n_dims == 1) {
+                    if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
+                        fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
+                                __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
+                        return false;
+                    }
+                } else {
+                    if (split_type == 0) {
+                        if (tensor->ne[0]/n_parts != ne[0] || tensor->ne[1] != ne[1]) {
+                            fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
+                                    __func__, name.data(), tensor->ne[0]/n_parts, tensor->ne[1], ne[0], ne[1]);
+                            return false;
+                        }
+                    } else {
+                        if (tensor->ne[0] != ne[0] || tensor->ne[1]/n_parts != ne[1]) {
+                            fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
+                                    __func__, name.data(), tensor->ne[0], tensor->ne[1]/n_parts, ne[0], ne[1]);
+                            return false;
+                        }
+                    }
+                }
+
+                if (0) {
+                    static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
+                    fprintf(stderr, "%24s - [%5d, %5d], type = %6s, split = %d\n", name.data(), ne[0], ne[1], ftype_str[ftype], split_type);
+                }
+
+                size_t bpe = 0;
+
+                switch (ftype) {
+                    case 0: bpe = ggml_type_size(GGML_TYPE_F32);  break;
+                    case 1: bpe = ggml_type_size(GGML_TYPE_F16);  break;
+                    case 2: bpe = ggml_type_size(GGML_TYPE_Q4_0); assert(ne[0] % 64 == 0); break;
+                    case 3: bpe = ggml_type_size(GGML_TYPE_Q4_1); assert(ne[0] % 64 == 0); break;
+                    default:
+                            {
+                                fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
+                                return false;
+                            }
+                };
+
+                if (n_dims == 1 || n_parts == 1) {
+                    if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+                        fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+                                __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
+                        return false;
+                    }
+
+                    if (part_id == 0) {
+                        fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
+                    } else {
+                        fin.seekg(ggml_nbytes(tensor), std::ios::cur);
+                    }
+
+                    total_size += ggml_nbytes(tensor);
+                } else {
+                    if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)/n_parts) {
+                        fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+                                __func__, name.data(), ggml_nbytes(tensor)/n_parts, nelements*bpe);
+                        return false;
+                    }
+
+                    if (split_type == 0) {
+                        const int np0 = ne[0];
+
+                        const size_t row_size = (tensor->ne[0]/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type);
+                        assert(row_size == tensor->nb[1]);
+
+                        for (int i1 = 0; i1 < ne[1]; ++i1) {
+                            const size_t offset_row = i1*row_size;
+                            const size_t offset = offset_row + ((part_id*np0)/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type);
+                            fin.read(reinterpret_cast<char *>(tensor->data) + offset, row_size/n_parts);
+                        }
+                    } else {
+                        const int np1 = ne[1];
+
+                        const size_t row_size = (tensor->ne[0]/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type);
+
+                        for (int i1 = 0; i1 < ne[1]; ++i1) {
+                            const size_t offset_row = (i1 + part_id*np1)*row_size;
+                            fin.read(reinterpret_cast<char *>(tensor->data) + offset_row, row_size);
+                        }
+                    }
+
+                    total_size += ggml_nbytes(tensor)/n_parts;
+                }
+
+                //fprintf(stderr, "%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
+                model.n_loaded++;
+
+                // progress
+                if (progress_callback) {
+                    double current_file_progress = double(size_t(fin.tellg()) - file_offset) / double(file_size - file_offset);
+                    double current_progress = (double(i) + current_file_progress) / double(n_parts);
+                    progress_callback(current_progress, progress_callback_user_data);
+                }
+                if (model.n_loaded % 8 == 0) {
+                    fprintf(stderr, ".");
+                    fflush(stderr);
+                }
+            }
+
+            fprintf(stderr, " done\n");
+
+            fprintf(stderr, "%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, model.n_loaded);
+            if (model.n_loaded == 0) {
+                fprintf(stderr, "%s: WARN no tensors loaded from model file - assuming empty model for testing\n", __func__);
+            } else if (model.n_loaded != (int) model.tensors.size()) {
+                fprintf(stderr, "%s: ERROR not all tensors loaded from model file - expected %zu, got %d\n", __func__, model.tensors.size(), model.n_loaded);
+                return false;
+            }
+        }
+
+        fin.close();
+    }
+
+    lctx.t_load_us = ggml_time_us() - t_start_us;
+
+    if (progress_callback) {
+        progress_callback(1.0, progress_callback_user_data);
+    }
+
+    return true;
+}
+
+// evaluate the transformer
+//
+//   - lctx:      llama context
+//   - tokens:    new batch of tokens to process
+//   - n_past:    the context size so far
+//   - n_threads: number of threads to use
+//
+static bool llama_eval_internal(
+        llama_context & lctx,
+    const llama_token * tokens,
+            const int   n_tokens,
+            const int   n_past,
+            const int   n_threads) {
+    const int64_t t_start_us = ggml_time_us();
+
+    const int N = n_tokens;
+
+    const auto & model   = lctx.model;
+    const auto & hparams = model.hparams;
+
+    auto & kv_self = model.kv_self;
+
+    LLAMA_ASSERT(!!kv_self.ctx);
+
+    const int n_embd  = hparams.n_embd;
+    const int n_layer = hparams.n_layer;
+    const int n_ctx   = hparams.n_ctx;
+    const int n_head  = hparams.n_head;
+    const int n_vocab = hparams.n_vocab;
+    const int n_rot   = hparams.n_embd/hparams.n_head;
+
+    auto & mem_per_token = lctx.mem_per_token;
+    auto & buf_compute   = lctx.buf_compute;
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_compute.size(),
+        /*.mem_buffer =*/ buf_compute.data(),
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+
+    // for big prompts, if BLAS is enabled, it is better to use only one thread
+    // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance
+    ggml_cgraph gf = {};
+    gf.n_threads = N > 255 && ggml_cpu_has_blas() ? 1 : n_threads;
+
+    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    memcpy(embd->data, tokens, N*ggml_element_size(embd));
+
+    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.tok_embeddings, embd);
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * inpSA = inpL;
+
+        struct ggml_tensor * cur;
+
+        lctx.use_buf(ctx0, 0);
+
+        // norm
+        {
+            cur = ggml_rms_norm(ctx0, inpL);
+
+            // cur = attention_norm*cur
+            cur = ggml_mul(ctx0,
+                        ggml_repeat(ctx0, model.layers[il].attention_norm, cur),
+                        cur);
+        }
+
+        // self-attention
+        {
+            struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+            struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+            struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+
+            // store key and value to memory
+            if (N >= 1) {
+                struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_1d(ctx0, kv_self.v, N*n_embd, (ggml_element_size(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)
+            struct ggml_tensor * Q =
+                ggml_permute(ctx0,
+                        ggml_rope(ctx0,
+                            ggml_cpy(ctx0,
+                                Qcur,
+                                ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
+                            n_past, n_rot, 0),
+                        0, 2, 1, 3);
+
+            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
+            struct ggml_tensor * K =
+                ggml_permute(ctx0,
+                        ggml_rope(ctx0,
+                            ggml_reshape_3d(ctx0,
+                                ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.k)*n_embd),
+                                n_embd/n_head, n_head, n_past + N),
+                            n_past, n_rot, 1),
+                        0, 2, 1, 3);
+
+            // K * Q
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+
+            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head)));
+
+            // KQ_masked = mask_past(KQ_scaled)
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
+
+            // KQ = soft_max(KQ_masked)
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
+
+            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
+            struct ggml_tensor * V_trans =
+                ggml_cpy(ctx0,
+                    ggml_permute(ctx0,
+                            ggml_reshape_3d(ctx0,
+                                ggml_view_1d(ctx0, kv_self.v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.v)*n_embd),
+                                n_embd/n_head, n_head, n_past + N),
+                            1, 2, 0, 3),
+                    ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_embd/n_head, n_head));
+
+            // KQV = transpose(V) * KQ_soft_max
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            // cur = KQV_merged.contiguous().view(n_embd, N)
+            cur = ggml_cpy(ctx0,
+                    KQV_merged,
+                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+
+            // projection (no bias)
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].wo,
+                    cur);
+        }
+
+        lctx.use_buf(ctx0, 1);
+
+        struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA);
+
+        // feed-forward network
+        {
+            // norm
+            {
+                cur = ggml_rms_norm(ctx0, inpFF);
+
+                // cur = ffn_norm*cur
+                cur = ggml_mul(ctx0,
+                        ggml_repeat(ctx0, model.layers[il].ffn_norm, cur),
+                        cur);
+            }
+
+            struct ggml_tensor * tmp = ggml_mul_mat(ctx0,
+                    model.layers[il].w3,
+                    cur);
+
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].w1,
+                    cur);
+
+            // SILU activation
+            cur = ggml_silu(ctx0, cur);
+
+            cur = ggml_mul(ctx0, cur, tmp);
+
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].w2,
+                    cur);
+        }
+
+        cur = ggml_add(ctx0, cur, inpFF);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    lctx.use_buf(ctx0, 0);
+
+    // used at the end to optionally extract the embeddings
+    struct ggml_tensor * embeddings = NULL;
+
+    // norm
+    {
+
+        inpL = ggml_rms_norm(ctx0, inpL);
+
+        // inpL = norm*inpL
+        inpL = ggml_mul(ctx0,
+                    ggml_repeat(ctx0, model.norm, inpL),
+                    inpL);
+
+        embeddings = inpL;
+    }
+
+    // lm_head
+    inpL = ggml_mul_mat(ctx0, model.output, inpL);
+
+    lctx.use_buf(ctx0, -1);
+
+    // 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);
+
+    // extract logits
+    {
+        auto & logits_out = lctx.logits;
+
+        if (lctx.logits_all) {
+            logits_out.resize(n_vocab * N);
+            memcpy(logits_out.data(), (float *) ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+        } else {
+            // return result for just the last token
+            logits_out.resize(n_vocab);
+            memcpy(logits_out.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+        }
+    }
+
+    // extract embeddings
+    if (lctx.embedding.size()) {
+        auto & embedding_out = lctx.embedding;
+
+        embedding_out.resize(n_embd);
+        memcpy(embedding_out.data(), (float *) ggml_get_data(embeddings) + (n_embd*(N - 1)), sizeof(float)*n_embd);
+    }
+
+    if (mem_per_token == 0) {
+        mem_per_token = ggml_used_mem(ctx0)/N;
+    }
+
+#if 0
+    printf("\n%s: used_mem = %.3f MB, scratch -- %.3f MB %.3f MB\n", __func__,
+            ggml_used_mem(ctx0)/1024.0/1024.0,
+            lctx.get_buf_max_mem(0)/1024.0/1024.0,
+            lctx.get_buf_max_mem(1)/1024.0/1024.0);
+#endif
+
+    ggml_free(ctx0);
+
+    // measure the performance only for the single-token evals
+    if (N == 1) {
+        lctx.t_eval_us += ggml_time_us() - t_start_us;
+        lctx.n_eval++;
+    }
+    else if (N > 1) {
+        lctx.t_p_eval_us += ggml_time_us() - t_start_us;
+        lctx.n_p_eval += N;
+    }
+
+    return true;
+}
+
+//
+// tokenizer
+//
+
+static size_t utf8_len(char src) {
+    const size_t lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
+    uint8_t highbits = static_cast<uint8_t>(src) >> 4;
+    return lookup[highbits];
+}
+
+struct llama_sp_symbol {
+    using index = int;
+    index prev;
+    index next;
+    const char * text;
+    size_t n;
+};
+
+struct llama_sp_bigram {
+    struct comparator {
+        bool operator()(llama_sp_bigram & l, llama_sp_bigram & r) {
+            return (l.score < r.score) || (l.score == r.score && l.left > r.left);
+        }
+    };
+    using queue_storage = std::vector<llama_sp_bigram>;
+    using queue = std::priority_queue<llama_sp_bigram, queue_storage, comparator>;
+    llama_sp_symbol::index left;
+    llama_sp_symbol::index right;
+    float score;
+    size_t size;
+};
+
+// original implementation:
+// https://github.com/ggerganov/llama.cpp/commit/074bea2eb1f1349a0118239c4152914aecaa1be4
+struct llama_tokenizer {
+    llama_tokenizer(const llama_vocab & vocab): vocab_(vocab) {}
+
+    void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
+        // split string into utf8 chars
+        int index = 0;
+        size_t offs = 0;
+        while (offs < text.size()) {
+            llama_sp_symbol sym;
+            size_t char_len = std::min(text.size() - offs, utf8_len(text[offs]));
+            sym.text = text.c_str() + offs;
+            sym.n = char_len;
+            offs += char_len;
+            sym.prev = index - 1;
+            sym.next = offs == text.size() ? -1 : index + 1;
+            index++;
+            symbols_.emplace_back(std::move(sym));
+        }
+
+        // seed the work queue with all possible 2-character tokens.
+        for (size_t i = 1; i < symbols_.size(); ++i) {
+            try_add_bigram(i - 1, i);
+        }
+
+        // keep substituting the highest frequency pairs for as long as we can.
+        while (!work_queue_.empty()) {
+            auto bigram = work_queue_.top();
+            work_queue_.pop();
+
+            auto & left_sym = symbols_[bigram.left];
+            auto & right_sym = symbols_[bigram.right];
+
+            // if one of the symbols already got merged, skip it.
+            if (left_sym.n == 0 || right_sym.n == 0 ||
+                left_sym.n + right_sym.n != bigram.size) {
+                continue;
+            }
+
+            // merge the right sym into the left one
+            left_sym.n += right_sym.n;
+            right_sym.n = 0;
+
+            //printf("left = '%*s' size = %zu\n", (int) left_sym.n, left_sym.text, bigram.size);
+
+            // remove the right sym from the chain
+            left_sym.next = right_sym.next;
+            if (right_sym.next >= 0) {
+                symbols_[right_sym.next].prev = bigram.left;
+            }
+
+            // find more substitutions
+            try_add_bigram(left_sym.prev, bigram.left);
+            try_add_bigram(bigram.left, left_sym.next);
+        }
+
+        for (int i = 0; i != -1; i = symbols_[i].next) {
+            auto & symbol = symbols_[i];
+            auto token = vocab_.token_to_id.find(std::string(symbol.text, symbol.n));
+
+            if (token == vocab_.token_to_id.end()) {
+                // output any symbols that did not form tokens as bytes.
+                for (int j = 0; j < (int) symbol.n; ++j) {
+                    llama_vocab::id token_id = static_cast<uint8_t>(symbol.text[j]) + 3;
+                    output.push_back(token_id);
+                }
+            } else {
+                output.push_back((*token).second);
+            }
+        }
+    }
+
+private:
+    void try_add_bigram(int left, int right) {
+        if (left == -1 || right == -1) {
+            return;
+        }
+
+        const std::string text = std::string(symbols_[left].text, symbols_[left].n + symbols_[right].n);
+        auto token = vocab_.token_to_id.find(text);
+
+        if (token == vocab_.token_to_id.end()) {
+            return;
+        }
+
+        if (static_cast<size_t>((*token).second) >= vocab_.id_to_token.size()) {
+            return;
+        }
+
+        const auto &tok_score = vocab_.id_to_token[(*token).second];
+
+        llama_sp_bigram bigram;
+        bigram.left = left;
+        bigram.right = right;
+        bigram.score = tok_score.score;
+        bigram.size = text.size();
+        work_queue_.push(bigram);
+    }
+
+    const llama_vocab & vocab_;
+    std::vector<llama_sp_symbol> symbols_;
+    llama_sp_bigram::queue work_queue_;
+};
+
+static std::vector<llama_vocab::id> llama_tokenize(const llama_vocab & vocab, const std::string & text, bool bos) {
+    llama_tokenizer tokenizer(vocab);
+    std::vector<llama_vocab::id> output;
+
+    if (text.size() == 0) {
+        return output;
+    }
+
+    if (bos) {
+        output.push_back(1);
+    }
+
+    tokenizer.tokenize(text, output);
+    return output;
+}
+
+//
+// sampling
+//
+
+static void sample_top_k(std::vector<std::pair<double, llama_vocab::id>> & logits_id, int top_k) {
+    // find the top k tokens
+    std::partial_sort(
+            logits_id.begin(),
+            logits_id.begin() + top_k, logits_id.end(),
+            [](const std::pair<double, llama_vocab::id> & a, const std::pair<double, llama_vocab::id> & b) {
+        return a.first > b.first;
+    });
+
+    logits_id.resize(top_k);
+}
+
+static llama_vocab::id llama_sample_top_p_top_k(
+        llama_context & lctx,
+        const std::vector<llama_vocab::id> & last_n_tokens,
+        int top_k,
+        double top_p,
+        double temp,
+        double repeat_penalty) {
+    auto & rng = lctx.rng;
+
+    const int n_logits = lctx.model.hparams.n_vocab;
+
+    const auto & logits = lctx.logits;
+    const auto * plogits = logits.data() + logits.size() - n_logits;
+
+    std::vector<std::pair<double, llama_vocab::id>> logits_id;
+    logits_id.reserve(n_logits);
+
+    {
+        const double scale = 1.0/temp;
+        for (int i = 0; i < n_logits; ++i) {
+            // repetition penalty from ctrl paper (https://arxiv.org/abs/1909.05858)
+            // credit https://github.com/facebookresearch/llama/compare/main...shawwn:llama:main
+            if (std::find(last_n_tokens.begin(), last_n_tokens.end(), i) != last_n_tokens.end()) {
+                // if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
+                if (plogits[i] < 0.0) {
+                    logits_id.push_back(std::make_pair(plogits[i]*scale*repeat_penalty, i));
+                } else {
+                    logits_id.push_back(std::make_pair(plogits[i]*scale/repeat_penalty, i));
+                }
+            } else {
+                logits_id.push_back(std::make_pair(plogits[i]*scale, i));
+            }
+        }
+    }
+
+    sample_top_k(logits_id, top_k);
+
+    double maxl = -std::numeric_limits<double>::infinity();
+    for (const auto & kv : logits_id) {
+        maxl = std::max(maxl, kv.first);
+    }
+
+    // compute probs for the top k tokens
+    std::vector<double> probs;
+    probs.reserve(logits_id.size());
+
+    double sum = 0.0;
+    for (const auto & kv : logits_id) {
+        double p = exp(kv.first - maxl);
+        probs.push_back(p);
+        sum += p;
+    }
+
+    // normalize the probs
+    for (auto & p : probs) {
+        p /= sum;
+    }
+
+    if (top_p < 1.0f) {
+        double cumsum = 0.0f;
+        for (int i = 0; i < (int) probs.size(); i++) {
+            cumsum += probs[i];
+            if (cumsum >= top_p) {
+                probs.resize(i + 1);
+                logits_id.resize(i + 1);
+                break;
+            }
+        }
+
+        cumsum = 1.0/cumsum;
+        for (int i = 0; i < (int) probs.size(); i++) {
+            probs[i] *= cumsum;
+        }
+    }
+
+    //printf("\n");
+    //for (int i = 0; i < (int) 10; i++) {
+    //    printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
+    //}
+    //printf("\n\n");
+    //exit(0);
+
+    std::discrete_distribution<> dist(probs.begin(), probs.end());
+    int idx = dist(rng);
+
+    return logits_id[idx].second;
+}
+
+//
+// quantization
+//
+
+// TODO: reuse code from the llama_model_load() somehow
+bool llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, int itype, int qk) {
+    ggml_type type = GGML_TYPE_Q4_1;
+
+    switch (itype) {
+        case 2: type = GGML_TYPE_Q4_0; break;
+        case 3: type = GGML_TYPE_Q4_1; break;
+        default: fprintf(stderr, "%s: invalid quantization type %d\n", __func__, itype); return 1;
+    };
+
+    if (type != GGML_TYPE_Q4_0 && type != GGML_TYPE_Q4_1) {
+        fprintf(stderr, "%s: invalid quantization type %d\n", __func__, type);
+        return false;
+    }
+
+    llama_vocab vocab;
+
+    printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
+
+    auto finp = std::ifstream(fname_inp, std::ios::binary);
+    if (!finp) {
+        fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    auto fout = std::ofstream(fname_out, std::ios::binary);
+    if (!fout) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        finp.read((char *) &magic, sizeof(magic));
+        if (magic == LLAMA_FILE_MAGIC_UNVERSIONED) {
+            fprintf(stderr, "%s: invalid model file '%s' (too old, regenerate your model files!)\n",
+                    __func__, fname_inp.c_str());
+            return false;
+        }
+        if (magic != LLAMA_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
+            return false;
+        }
+
+        fout.write((char *) &magic, sizeof(magic));
+
+        uint32_t format_version;
+        finp.read((char *) &format_version, sizeof(format_version));
+
+        if (format_version != LLAMA_FILE_VERSION) {
+            fprintf(stderr, "%s: invalid model file '%s' (unsupported format version %" PRIu32 ", expected %d)\n",
+                    __func__, fname_inp.c_str(), format_version, LLAMA_FILE_VERSION);
+            return false;
+        }
+
+        fout.write((char *) &format_version, sizeof(format_version));
+    }
+
+    llama_hparams hparams;
+
+    // load hparams
+    {
+        finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        //finp.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        finp.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        finp.read((char *) &hparams.n_mult,  sizeof(hparams.n_mult));
+        finp.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        finp.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        finp.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_mult  = %d\n", __func__, hparams.n_mult);
+        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);
+
+        fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        //fout.write((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fout.write((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fout.write((char *) &hparams.n_mult,  sizeof(hparams.n_mult));
+        fout.write((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fout.write((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        fout.write((char *) &itype,           sizeof(hparams.f16));
+    }
+
+    // load vocab
+    {
+        const int32_t n_vocab = hparams.n_vocab;
+
+        if (n_vocab != hparams.n_vocab) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+                    __func__, fname_inp.c_str(), n_vocab, hparams.n_vocab);
+            return false;
+        }
+
+        std::string word;
+        vocab.id_to_token.resize(n_vocab);
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            finp.read ((char *) &len, sizeof(len));
+            fout.write((char *) &len, sizeof(len));
+
+            word.resize(len);
+            finp.read ((char *) word.data(), len);
+            fout.write((char *) word.data(), len);
+
+            float score;
+            finp.read ((char *) &score, sizeof(score));
+            fout.write((char *) &score, sizeof(score));
+
+            vocab.token_to_id[word] = i;
+
+            auto &tok_score = vocab.id_to_token[i];
+            tok_score.tok = word;
+            tok_score.score = score;
+        }
+    }
+
+    // load weights
+    {
+        size_t total_size_org = 0;
+        size_t total_size_new = 0;
+
+        std::vector<float> work;
+
+        std::vector<uint8_t>     data_u8;
+        std::vector<ggml_fp16_t> data_f16;
+        std::vector<float>       data_f32;
+
+        std::vector<int64_t> hist_all(1 << 4, 0);
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ftype;
+
+            finp.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            finp.read(reinterpret_cast<char *>(&length), sizeof(length));
+            finp.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+
+            if (finp.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                finp.read (reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            finp.read (&name[0], length);
+
+            {
+                static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
+                printf("%48s - [%5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ftype_str[ftype]);
+            }
+
+            // regexes of tensor names to be quantized
+            const std::vector<std::string> k_names = {
+                ".*weight",
+            };
+
+            bool quantize = false;
+            for (const auto & s : k_names) {
+                if (std::regex_match(name, std::regex(s))) {
+                    quantize = true;
+                    break;
+                }
+            }
+
+            // quantize only 2D tensors
+            quantize &= (n_dims == 2);
+
+            if (quantize) {
+                if (ftype != 0 && ftype != 1) {
+                    fprintf(stderr, "%s: unsupported ftype %d for integer quantization\n", __func__, ftype);
+                    return false;
+                }
+
+                if (ftype == 1) {
+                    data_f16.resize(nelements);
+                    finp.read(reinterpret_cast<char *>(data_f16.data()), nelements * sizeof(ggml_fp16_t));
+                    data_f32.resize(nelements);
+                    for (int i = 0; i < nelements; ++i) {
+                        data_f32[i] = ggml_fp16_to_fp32(data_f16[i]);
+                    }
+                } else {
+                    data_f32.resize(nelements);
+                    finp.read(reinterpret_cast<char *>(data_f32.data()), nelements * sizeof(float));
+                }
+
+                ftype = itype;
+            } else {
+                const int bpe = (ftype == 0) ? sizeof(float) : sizeof(uint16_t);
+
+                data_u8.resize(nelements*bpe);
+                finp.read(reinterpret_cast<char *>(data_u8.data()), nelements * bpe);
+            }
+
+            fout.write(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fout.write(reinterpret_cast<char *>(&length), sizeof(length));
+            fout.write(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+            for (int i = 0; i < n_dims; ++i) {
+                fout.write(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+            }
+            fout.write(&name[0], length);
+
+            if (quantize) {
+                printf("quantizing .. ");
+                work.resize(nelements); // for quantization
+
+                size_t cur_size = 0;
+                std::vector<int64_t> hist_cur(1 << 4, 0);
+
+                switch (type) {
+                    case GGML_TYPE_Q4_0:
+                        {
+                            cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], qk, hist_cur.data());
+                        } break;
+                    case GGML_TYPE_Q4_1:
+                        {
+                            cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], qk, hist_cur.data());
+                        } break;
+                    default:
+                        {
+                            fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, type);
+                            return false;
+                        }
+                }
+
+                fout.write(reinterpret_cast<char *>(work.data()), cur_size);
+                total_size_new += cur_size;
+
+                printf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
+                for (int i = 0; i < (int) hist_cur.size(); ++i) {
+                    hist_all[i] += hist_cur[i];
+                }
+
+                for (int i = 0; i < (int) hist_cur.size(); ++i) {
+                    printf("%5.3f ", hist_cur[i] / (float)nelements);
+                }
+                printf("\n");
+            } else {
+                printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
+                fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
+                total_size_new += data_u8.size();
+            }
+
+            total_size_org += nelements * sizeof(float);
+        }
+
+        printf("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
+        printf("%s: quant size  = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0);
+
+        {
+            int64_t sum_all = 0;
+            for (int i = 0; i < (int) hist_all.size(); ++i) {
+                sum_all += hist_all[i];
+            }
+
+            printf("%s: hist: ", __func__);
+            for (int i = 0; i < (int) hist_all.size(); ++i) {
+                printf("%5.3f ", hist_all[i] / (float)sum_all);
+            }
+            printf("\n");
+        }
+    }
+
+    finp.close();
+    fout.close();
+
+    return true;
+}
+
+//
+// interface implementation
+//
+
+struct llama_context * llama_init_from_file(
+                             const char * path_model,
+            struct llama_context_params   params) {
+    ggml_time_init();
+
+    llama_context * ctx = new llama_context;
+
+    if (params.seed <= 0) {
+        params.seed = time(NULL);
+    }
+
+    ctx->rng = std::mt19937(params.seed);
+    ctx->logits_all = params.logits_all;
+
+    ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
+
+    if (!llama_model_load(path_model, *ctx, params.n_ctx, params.n_parts, memory_type,
+                          params.vocab_only, params.progress_callback,
+                          params.progress_callback_user_data)) {
+        fprintf(stderr, "%s: failed to load model\n", __func__);
+        llama_free(ctx);
+        return nullptr;
+    }
+
+    if (params.use_mlock) {
+        char *err;
+        if (!ggml_mlock(ctx->model.ctx, &err)) {
+            fprintf(stderr, "%s\n", err);
+            free(err);
+            llama_free(ctx);
+            return nullptr;
+        }
+    }
+
+    // reserve memory for context buffers
+    {
+        if (!kv_cache_init(ctx->model.hparams, ctx->model.kv_self, memory_type, ctx->model.hparams.n_ctx)) {
+            fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
+            llama_free(ctx);
+            return nullptr;
+        }
+
+        {
+            const size_t memory_size = ggml_nbytes(ctx->model.kv_self.k) + ggml_nbytes(ctx->model.kv_self.v);
+            fprintf(stderr, "%s: kv self size  = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
+        }
+
+        const auto & hparams = ctx->model.hparams;
+
+        // resized during inference
+        if (params.logits_all) {
+            ctx->logits.reserve(hparams.n_ctx*hparams.n_vocab);
+        } else {
+            ctx->logits.reserve(hparams.n_ctx);
+        }
+
+        if (params.embedding){
+            ctx->embedding.resize(hparams.n_embd);
+        }
+
+        ctx->buf_compute.resize(MEM_REQ_EVAL.at(ctx->model.type));
+
+        ctx->buf_scratch[0].resize(MEM_REQ_SCRATCH0.at(ctx->model.type));
+        ctx->buf_scratch[1].resize(MEM_REQ_SCRATCH1.at(ctx->model.type));
+    }
+
+    return ctx;
+}
+
+void llama_free(struct llama_context * ctx) {
+    kv_cache_free(ctx->model.kv_self);
+
+    if (ctx->model.ctx) {
+        ggml_free(ctx->model.ctx);
+    }
+
+    delete ctx;
+}
+
+int llama_model_quantize(
+        const char * fname_inp,
+        const char * fname_out,
+               int   itype,
+               int   qk) {
+    if (!llama_model_quantize_internal(fname_inp, fname_out, itype, qk)) {
+        fprintf(stderr, "%s: failed to quantize\n", __func__);
+        return 1;
+    }
+
+    return 0;
+}
+
+int llama_eval(
+        struct llama_context * ctx,
+           const llama_token * tokens,
+                         int   n_tokens,
+                         int   n_past,
+                         int   n_threads) {
+    if (!llama_eval_internal(*ctx, tokens, n_tokens, n_past, n_threads)) {
+        fprintf(stderr, "%s: failed to eval\n", __func__);
+        return 1;
+    }
+
+    return 0;
+}
+
+int llama_tokenize(
+        struct llama_context * ctx,
+                  const char * text,
+                 llama_token * tokens,
+                         int   n_max_tokens,
+                        bool   add_bos) {
+    auto res = llama_tokenize(ctx->vocab, text, add_bos);
+
+    if (n_max_tokens < (int) res.size()) {
+        fprintf(stderr, "%s: too many tokens\n", __func__);
+        return -((int) res.size());
+    }
+
+    for (size_t i = 0; i < res.size(); i++) {
+        tokens[i] = res[i];
+    }
+
+    return res.size();
+}
+
+int llama_n_vocab(struct llama_context * ctx) {
+    return ctx->vocab.id_to_token.size();
+}
+
+int llama_n_ctx(struct llama_context * ctx) {
+    return ctx->model.hparams.n_ctx;
+}
+
+int llama_n_embd(struct llama_context * ctx) {
+    return ctx->model.hparams.n_embd;
+}
+
+float * llama_get_logits(struct llama_context * ctx) {
+    return ctx->logits.data();
+}
+
+float * llama_get_embeddings(struct llama_context * ctx) {
+    return ctx->embedding.data();
+}
+
+const char * llama_token_to_str(struct llama_context * ctx, llama_token token) {
+    if (token >= llama_n_vocab(ctx)) {
+        return nullptr;
+    }
+
+    return ctx->vocab.id_to_token[token].tok.c_str();
+}
+
+llama_token llama_token_bos() {
+    return 1;
+}
+
+llama_token llama_token_eos() {
+    return 2;
+}
+
+llama_token llama_sample_top_p_top_k(
+          llama_context * ctx,
+      const llama_token * last_n_tokens_data,
+                    int   last_n_tokens_size,
+                    int   top_k,
+                 double   top_p,
+                 double   temp,
+                 double   repeat_penalty) {
+    const int64_t t_start_sample_us = ggml_time_us();
+
+    llama_token result = 0;
+
+    // TODO: avoid this ...
+    const auto last_n_tokens = std::vector<llama_token>(last_n_tokens_data, last_n_tokens_data + last_n_tokens_size);
+
+    result = llama_sample_top_p_top_k(
+            *ctx,
+            last_n_tokens,
+            top_k,
+            top_p,
+            temp,
+            repeat_penalty);
+
+    ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
+    ctx->n_sample++;
+
+    return result;
+}
+
+
+void llama_print_timings(struct llama_context * ctx) {
+    const int64_t t_end_us = ggml_time_us();
+
+    const int32_t n_sample = std::max(1, ctx->n_sample);
+    const int32_t n_eval   = std::max(1, ctx->n_eval);
+    const int32_t n_p_eval = std::max(1, ctx->n_p_eval);
+
+    fprintf(stderr, "\n");
+    fprintf(stderr, "%s:        load time = %8.2f ms\n", __func__, ctx->t_load_us / 1000.0f);
+    fprintf(stderr, "%s:      sample time = %8.2f ms / %5d runs   (%8.2f ms per run)\n",   __func__, 1e-3f * ctx->t_sample_us, n_sample, 1e-3f * ctx->t_sample_us / n_sample);
+    fprintf(stderr, "%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token)\n", __func__, 1e-3f * ctx->t_p_eval_us, n_p_eval, 1e-3f * ctx->t_p_eval_us / n_p_eval);
+    fprintf(stderr, "%s:        eval time = %8.2f ms / %5d runs   (%8.2f ms per run)\n",   __func__, 1e-3f * ctx->t_eval_us,   n_eval,   1e-3f * ctx->t_eval_us   / n_eval);
+    fprintf(stderr, "%s:       total time = %8.2f ms\n", __func__, (t_end_us - ctx->t_start_us)/1000.0f);
+}
+
+void llama_reset_timings(struct llama_context * ctx) {
+    ctx->t_start_us = ggml_time_us();
+
+    ctx->t_sample_us = ctx->n_sample = 0;
+    ctx->t_eval_us   = ctx->n_eval   = 0;
+    ctx->t_p_eval_us = ctx->n_p_eval = 0;
+}
+
+const char * llama_print_system_info(void) {
+    static std::string s;
+
+    s  = "";
+    s += "AVX = "       + std::to_string(ggml_cpu_has_avx())       + " | ";
+    s += "AVX2 = "      + std::to_string(ggml_cpu_has_avx2())      + " | ";
+    s += "AVX512 = "    + std::to_string(ggml_cpu_has_avx512())    + " | ";
+    s += "FMA = "       + std::to_string(ggml_cpu_has_fma())       + " | ";
+    s += "NEON = "      + std::to_string(ggml_cpu_has_neon())      + " | ";
+    s += "ARM_FMA = "   + std::to_string(ggml_cpu_has_arm_fma())   + " | ";
+    s += "F16C = "      + std::to_string(ggml_cpu_has_f16c())      + " | ";
+    s += "FP16_VA = "   + std::to_string(ggml_cpu_has_fp16_va())   + " | ";
+    s += "WASM_SIMD = " + std::to_string(ggml_cpu_has_wasm_simd()) + " | ";
+    s += "BLAS = "      + std::to_string(ggml_cpu_has_blas())      + " | ";
+    s += "SSE3 = "      + std::to_string(ggml_cpu_has_sse3())      + " | ";
+    s += "VSX = "       + std::to_string(ggml_cpu_has_vsx())       + " | ";
+
+    return s.c_str();
+}

examples/talk-llama/llama.h

@@ -0,0 +1,153 @@
+#ifndef LLAMA_H
+#define LLAMA_H
+
+#include <stddef.h>
+#include <stdint.h>
+#include <stdbool.h>
+
+#ifdef LLAMA_SHARED
+#    ifdef _WIN32
+#        ifdef LLAMA_BUILD
+#            define LLAMA_API __declspec(dllexport)
+#        else
+#            define LLAMA_API __declspec(dllimport)
+#        endif
+#    else
+#        define LLAMA_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define LLAMA_API
+#endif
+
+#define LLAMA_FILE_VERSION 1
+#define LLAMA_FILE_MAGIC 0x67676d66 // 'ggmf' in hex
+#define LLAMA_FILE_MAGIC_UNVERSIONED 0x67676d6c // pre-versioned files
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+    //
+    // C interface
+    //
+    // TODO: show sample usage
+    //
+
+    struct llama_context;
+
+    typedef int llama_token;
+
+    typedef struct llama_token_data {
+        llama_token id;  // token id
+
+        float p;     // probability of the token
+        float plog;  // log probability of the token
+
+    } llama_token_data;
+
+    typedef void (*llama_progress_callback)(double progress, void *ctx);
+
+    struct llama_context_params {
+        int n_ctx;   // text context
+        int n_parts; // -1 for default
+        int seed;    // RNG seed, 0 for random
+
+        bool f16_kv;     // use fp16 for KV cache
+        bool logits_all; // the llama_eval() call computes all logits, not just the last one
+        bool vocab_only; // only load the vocabulary, no weights
+        bool use_mlock;  // force system to keep model in RAM
+        bool embedding;  // embedding mode only
+
+        // called with a progress value between 0 and 1, pass NULL to disable
+        llama_progress_callback progress_callback;
+        // context pointer passed to the progress callback
+        void * progress_callback_user_data;
+    };
+
+    LLAMA_API struct llama_context_params llama_context_default_params();
+
+    // Various functions for loading a ggml llama model.
+    // Allocate (almost) all memory needed for the model.
+    // Return NULL on failure
+    LLAMA_API struct llama_context * llama_init_from_file(
+                             const char * path_model,
+            struct llama_context_params   params);
+
+    // Frees all allocated memory
+    LLAMA_API void llama_free(struct llama_context * ctx);
+
+    // TODO: not great API - very likely to change
+    // Returns 0 on success
+    LLAMA_API int llama_model_quantize(
+            const char * fname_inp,
+            const char * fname_out,
+                   int   itype,
+                   int   qk);
+
+    // Run the llama inference to obtain the logits and probabilities for the next token.
+    // tokens + n_tokens is the provided batch of new tokens to process
+    // n_past is the number of tokens to use from previous eval calls
+    // Returns 0 on success
+    LLAMA_API int llama_eval(
+            struct llama_context * ctx,
+               const llama_token * tokens,
+                             int   n_tokens,
+                             int   n_past,
+                             int   n_threads);
+
+    // Convert the provided text into tokens.
+    // The tokens pointer must be large enough to hold the resulting tokens.
+    // Returns the number of tokens on success, no more than n_max_tokens
+    // Returns a negative number on failure - the number of tokens that would have been returned
+    // TODO: not sure if correct
+    LLAMA_API int llama_tokenize(
+            struct llama_context * ctx,
+                      const char * text,
+                     llama_token * tokens,
+                             int   n_max_tokens,
+                            bool   add_bos);
+
+    LLAMA_API int llama_n_vocab(struct llama_context * ctx);
+    LLAMA_API int llama_n_ctx  (struct llama_context * ctx);
+    LLAMA_API int llama_n_embd (struct llama_context * ctx);
+
+    // Token logits obtained from the last call to llama_eval()
+    // The logits for the last token are stored in the last row
+    // Can be mutated in order to change the probabilities of the next token
+    // Rows: n_tokens
+    // Cols: n_vocab
+    LLAMA_API float * llama_get_logits(struct llama_context * ctx);
+
+    // Get the embeddings for the input
+    // shape: [n_embd] (1-dimensional)
+    LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
+
+    // Token Id -> String. Uses the vocabulary in the provided context
+    LLAMA_API const char * llama_token_to_str(struct llama_context * ctx, llama_token token);
+
+    // Special tokens
+    LLAMA_API llama_token llama_token_bos();
+    LLAMA_API llama_token llama_token_eos();
+
+    // TODO: improve the last_n_tokens interface ?
+    LLAMA_API llama_token llama_sample_top_p_top_k(
+       struct llama_context * ctx,
+          const llama_token * last_n_tokens_data,
+                        int   last_n_tokens_size,
+                        int   top_k,
+                     double   top_p,
+                     double   temp,
+                     double   repeat_penalty);
+
+    // Performance information
+    LLAMA_API void llama_print_timings(struct llama_context * ctx);
+    LLAMA_API void llama_reset_timings(struct llama_context * ctx);
+
+    // Print system information
+    LLAMA_API const char * llama_print_system_info(void);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

examples/talk-llama/speak.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+
+# Usage:
+#  speak.sh <voice_id> <text-to-speak>
+
+# espeak
+# Mac OS: brew install espeak
+# Linux: apt-get install espeak
+#
+#espeak -v en-us+m$1 -s 225 -p 50 -a 200 -g 5 -k 5 "$2"
+
+# for Mac
+say "$2"
+
+# Eleven Labs
+#
+#wd=$(dirname $0)
+#script=$wd/eleven-labs.py
+#python3 $script $1 "$2" >/dev/null 2>&1
+#ffplay -autoexit -nodisp -loglevel quiet -hide_banner -i ./audio.mp3 >/dev/null 2>&1

examples/talk-llama/talk-llama.cpp

@@ -0,0 +1,529 @@
+// Talk with AI
+//
+
+#include "common.h"
+#include "common-sdl.h"
+#include "whisper.h"
+#include "llama.h"
+
+#include <cassert>
+#include <cstdio>
+#include <fstream>
+#include <regex>
+#include <string>
+#include <thread>
+#include <vector>
+#include <regex>
+
+std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
+    // initialize to prompt numer of chars, since n_tokens <= n_prompt_chars
+    std::vector<llama_token> res(text.size() + (int)add_bos);
+    int n = llama_tokenize(ctx, text.c_str(), res.data(), res.size(), add_bos);
+    assert(n >= 0);
+    res.resize(n);
+
+    return res;
+}
+
+// command-line parameters
+struct whisper_params {
+    int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t voice_ms   = 10000;
+    int32_t capture_id = -1;
+    int32_t max_tokens = 32;
+    int32_t audio_ctx  = 0;
+
+    float vad_thold    = 0.6f;
+    float freq_thold   = 100.0f;
+
+    bool speed_up      = false;
+    bool translate     = false;
+    bool print_special = false;
+    bool print_energy  = false;
+    bool no_timestamps = true;
+
+    std::string person      = "Georgi";
+    std::string language    = "en";
+    std::string model_wsp   = "models/ggml-base.en.bin";
+    std::string model_llama = "models/ggml-llama-7B.bin";
+    std::string speak       = "./examples/talk/speak.sh";
+    std::string fname_out;
+};
+
+void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
+
+bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-h" || arg == "--help") {
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+        else if (arg == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
+        else if (arg == "-vms" || arg == "--voice-ms")      { params.voice_ms      = std::stoi(argv[++i]); }
+        else if (arg == "-c"   || arg == "--capture")       { params.capture_id    = std::stoi(argv[++i]); }
+        else if (arg == "-mt"  || arg == "--max-tokens")    { params.max_tokens    = std::stoi(argv[++i]); }
+        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
+        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
+        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
+        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
+        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
+        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
+        else if (arg == "-p"   || arg == "--person")        { params.person        = argv[++i]; }
+        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
+        else if (arg == "-mw"  || arg == "--model-whisper") { params.model_wsp     = argv[++i]; }
+        else if (arg == "-ml"  || arg == "--model-llama")   { params.model_llama   = argv[++i]; }
+        else if (arg == "-s"   || arg == "--speak")         { params.speak         = argv[++i]; }
+        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
+    fprintf(stderr, "\n");
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h,       --help          [default] show this help message and exit\n");
+    fprintf(stderr, "  -t N,     --threads N     [%-7d] number of threads to use during computation\n", params.n_threads);
+    fprintf(stderr, "  -vms N,   --voice-ms N    [%-7d] voice duration in milliseconds\n",              params.voice_ms);
+    fprintf(stderr, "  -c ID,    --capture ID    [%-7d] capture device ID\n",                           params.capture_id);
+    fprintf(stderr, "  -mt N,    --max-tokens N  [%-7d] maximum number of tokens per audio chunk\n",    params.max_tokens);
+    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
+    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
+    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
+    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
+    fprintf(stderr, "  -pe,      --print-energy  [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
+    fprintf(stderr, "  -p NAME,  --person NAME   [%-7s] person name (for prompt selection)\n",          params.person.c_str());
+    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                             params.language.c_str());
+    fprintf(stderr, "  -mw FILE, --model-whisper [%-7s] whisper model file\n",                          params.model_wsp.c_str());
+    fprintf(stderr, "  -mg FILE, --model-llama   [%-7s] llama model file\n",                            params.model_llama.c_str());
+    fprintf(stderr, "  -s FILE,  --speak TEXT    [%-7s] command for TTS\n",                             params.speak.c_str());
+    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                       params.fname_out.c_str());
+    fprintf(stderr, "\n");
+}
+
+std::string transcribe(
+        whisper_context * ctx,
+        const whisper_params & params,
+        const std::vector<float> & pcmf32,
+        const std::string prompt_text,
+        float & prob,
+        int64_t & t_ms) {
+    const auto t_start = std::chrono::high_resolution_clock::now();
+
+    prob = 0.0f;
+    t_ms = 0;
+
+    std::vector<whisper_token> prompt_tokens;
+
+    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+
+    prompt_tokens.resize(1024);
+    prompt_tokens.resize(whisper_tokenize(ctx, prompt_text.c_str(), prompt_tokens.data(), prompt_tokens.size()));
+
+    wparams.print_progress   = false;
+    wparams.print_special    = params.print_special;
+    wparams.print_realtime   = false;
+    wparams.print_timestamps = !params.no_timestamps;
+    wparams.translate        = params.translate;
+    wparams.no_context       = true;
+    wparams.single_segment   = true;
+    wparams.max_tokens       = params.max_tokens;
+    wparams.language         = params.language.c_str();
+    wparams.n_threads        = params.n_threads;
+
+    wparams.prompt_tokens    = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
+    wparams.prompt_n_tokens  = prompt_tokens.empty() ? 0       : prompt_tokens.size();
+
+    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;
+
+    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
+        return "";
+    }
+
+    int prob_n = 0;
+    std::string result;
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+
+        result += text;
+
+        const int n_tokens = whisper_full_n_tokens(ctx, i);
+        for (int j = 0; j < n_tokens; ++j) {
+            const auto token = whisper_full_get_token_data(ctx, i, j);
+
+            prob += token.p;
+            ++prob_n;
+        }
+    }
+
+    if (prob_n > 0) {
+        prob /= prob_n;
+    }
+
+    const auto t_end = std::chrono::high_resolution_clock::now();
+    t_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();
+
+    return result;
+}
+
+const std::string k_prompt_whisper = R"(A conversation with a person called {1}.)";
+
+// need to have leading ' '
+const std::string k_prompt_llama = R"( Text transcript of a never ending dialog, where {0} interacts with an AI assistant named {1}.
+{1} is helpful, kind, honest, friendly, good at writing and never fails to answer {0}’s requests immediately and with details and precision.
+There are no annotations like (30 seconds passed...) or (to himself), just what {0} and {1} say aloud to each other.
+The transcript only includes text, it does not include markup like HTML and Markdown.
+{1} responds with short and concise answers.
+
+{0}{4} Hello, {1}!
+{1}{4} Hello {0}! How may I help you today?
+{0}{4} What time is it?
+{1}{4} It is {2} o'clock.
+{0}{4} What year is it?
+{1}{4} We are in {3}.
+{0}{4} What is a cat?
+{1}{4} A cat is a domestic species of small carnivorous mammal. It is the only domesticated species in the family Felidae.
+{0}{4} Name a color.
+{1}{4} Blue
+{0}{4})";
+
+int main(int argc, char ** argv) {
+    whisper_params params;
+
+    if (whisper_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (whisper_lang_id(params.language.c_str()) == -1) {
+        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
+        whisper_print_usage(argc, argv, params);
+        exit(0);
+    }
+
+    // whisper init
+
+    struct whisper_context * ctx_wsp = whisper_init_from_file(params.model_wsp.c_str());
+
+    // llama init
+
+    auto lparams = llama_context_default_params();
+
+    // tune these to your liking
+    lparams.n_ctx      = 512;
+    lparams.seed       = 1;
+    lparams.f16_kv     = true;
+
+    struct llama_context * ctx_llama = llama_init_from_file(params.model_llama.c_str(), lparams);
+
+    // print some info about the processing
+    {
+        fprintf(stderr, "\n");
+
+        if (!whisper_is_multilingual(ctx_wsp)) {
+            if (params.language != "en" || params.translate) {
+                params.language = "en";
+                params.translate = false;
+                fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
+            }
+        }
+        fprintf(stderr, "%s: processing, %d threads, lang = %s, task = %s, timestamps = %d ...\n",
+                __func__,
+                params.n_threads,
+                params.language.c_str(),
+                params.translate ? "translate" : "transcribe",
+                params.no_timestamps ? 0 : 1);
+
+        fprintf(stderr, "\n");
+    }
+
+
+    // init audio
+
+    audio_async audio(30*1000);
+    if (!audio.init(params.capture_id, WHISPER_SAMPLE_RATE)) {
+        fprintf(stderr, "%s: audio.init() failed!\n", __func__);
+        return 1;
+    }
+
+    audio.resume();
+
+    int n_iter = 0;
+
+    bool is_running  = true;
+    bool force_speak = false;
+
+    float prob0 = 0.0f;
+
+    const std::string chat_symb = ":";
+    const std::string bot_name  = "LLaMA";
+
+    std::vector<float> pcmf32_cur;
+    std::vector<float> pcmf32_prompt;
+
+    const std::string prompt_whisper = ::replace(k_prompt_whisper, "{1}", bot_name);
+
+    // construct the initial prompt for LLaMA inference
+    std::string prompt_llama = k_prompt_llama;
+
+    prompt_llama = ::replace(prompt_llama, "{0}", params.person);
+    prompt_llama = ::replace(prompt_llama, "{1}", bot_name);
+
+    {
+        // get time string
+        std::string time_str;
+        {
+            time_t t = time(0);
+            struct tm * now = localtime(&t);
+            char buf[128];
+            strftime(buf, sizeof(buf), "%H:%M", now);
+            time_str = buf;
+        }
+        prompt_llama = ::replace(prompt_llama, "{2}", time_str);
+    }
+
+    {
+        // get year string
+        std::string year_str;
+        {
+            time_t t = time(0);
+            struct tm * now = localtime(&t);
+            char buf[128];
+            strftime(buf, sizeof(buf), "%Y", now);
+            year_str = buf;
+        }
+        prompt_llama = ::replace(prompt_llama, "{3}", year_str);
+    }
+
+    prompt_llama = ::replace(prompt_llama, "{4}", chat_symb);
+
+    // evaluate the initial prompt
+
+    auto embd_inp = ::llama_tokenize(ctx_llama, prompt_llama, true);
+
+    printf("\n");
+    printf("%s : initializing - please wait ...\n", __func__);
+
+    if (llama_eval(ctx_llama, embd_inp.data(), embd_inp.size(), 0, params.n_threads)) {
+        fprintf(stderr, "%s : failed to eval\n", __func__);
+        return 1;
+    }
+
+    //fprintf(stdout, "\n");
+    //fprintf(stdout, "%s", prompt_llama.c_str());
+    //fflush(stdout);
+
+    printf("%s : done! start speaking in the microphone\n", __func__);
+    printf("\n");
+    printf("%s%s", params.person.c_str(), chat_symb.c_str());
+    fflush(stdout);
+
+    // clear audio buffer
+    audio.clear();
+
+    // text inference variables
+    const int voice_id = 2;
+    const int n_keep   = embd_inp.size();
+    const int n_ctx    = llama_n_ctx(ctx_llama);
+
+    int n_past = n_keep;
+    int n_prev = 64; // TODO arg
+
+    std::vector<llama_token> embd;
+
+    // reverse prompts for detecting when it's time to stop speaking
+    std::vector<std::string> antiprompts = {
+        params.person + chat_symb,
+    };
+
+    // main loop
+    while (is_running) {
+        // handle Ctrl + C
+        is_running = sdl_poll_events();
+
+        if (!is_running) {
+            break;
+        }
+
+        // delay
+        std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+        int64_t t_ms = 0;
+
+        {
+            audio.get(2000, pcmf32_cur);
+
+            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1250, params.vad_thold, params.freq_thold, params.print_energy) || force_speak) {
+                //fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
+
+                audio.get(params.voice_ms, pcmf32_cur);
+
+                std::string text_heard;
+
+                if (!force_speak) {
+                    text_heard = ::trim(::transcribe(ctx_wsp, params, pcmf32_cur, prompt_whisper, prob0, t_ms));
+                }
+
+                // remove text between brackets using regex
+                {
+                    std::regex re("\\[.*?\\]");
+                    text_heard = std::regex_replace(text_heard, re, "");
+                }
+
+                // remove text between brackets using regex
+                {
+                    std::regex re("\\(.*?\\)");
+                    text_heard = std::regex_replace(text_heard, re, "");
+                }
+
+                // remove all characters, except for letters, numbers, punctuation and ':', '\'', '-', ' '
+                text_heard = std::regex_replace(text_heard, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
+
+                // take first line
+                text_heard = text_heard.substr(0, text_heard.find_first_of('\n'));
+
+                // remove leading and trailing whitespace
+                text_heard = std::regex_replace(text_heard, std::regex("^\\s+"), "");
+                text_heard = std::regex_replace(text_heard, std::regex("\\s+$"), "");
+
+                const std::vector<llama_token> tokens = llama_tokenize(ctx_llama, text_heard.c_str(), false);
+
+                if (text_heard.empty() || tokens.empty() || force_speak) {
+                    //fprintf(stdout, "%s: Heard nothing, skipping ...\n", __func__);
+                    audio.clear();
+
+                    continue;
+                }
+
+                force_speak = false;
+
+                text_heard.insert(0, 1, ' ');
+                text_heard += "\n" + bot_name + chat_symb;
+                fprintf(stdout, "%s%s%s", "\033[1m", text_heard.c_str(), "\033[0m");
+                fflush(stdout);
+
+                embd = ::llama_tokenize(ctx_llama, text_heard, false);
+
+                // text inference
+                bool done = false;
+                std::string text_to_speak;
+                while (true) {
+                    // predict
+                    if (embd.size() > 0) {
+                        if (n_past + (int) embd.size() > n_ctx) {
+                            n_past = n_keep;
+
+                            // insert n_left/2 tokens at the start of embd from last_n_tokens
+                            embd.insert(embd.begin(), embd_inp.begin() + embd_inp.size() - n_prev, embd_inp.end());
+
+                            //printf("\n---\n");
+                            //printf("resetting: '");
+                            //for (int i = 0; i < (int) embd.size(); i++) {
+                            //    printf("%s", llama_token_to_str(ctx_llama, embd[i]));
+                            //}
+                            //printf("'\n");
+                            //printf("\n---\n");
+                        }
+
+                        if (llama_eval(ctx_llama, embd.data(), embd.size(), n_past, params.n_threads)) {
+                            fprintf(stderr, "%s : failed to eval\n", __func__);
+                            return 1;
+                        }
+                    }
+
+                    //printf("n_iter = %d, n_past = %d, n_ctx = %d, n_keep = %d, n_prev = %d, embd.size() = %d\n", n_iter, n_past, n_ctx, n_keep, n_prev, (int) embd.size());
+
+                    embd_inp.insert(embd_inp.end(), embd.begin(), embd.end());
+                    n_past += embd.size();
+                    embd.clear();
+
+                    if (done) break;
+
+                    {
+                        // out of user input, sample next token
+                        const float top_k          = 5;
+                        const float top_p          = 0.80f;
+                        const float temp           = 0.30f;
+                        const float repeat_penalty = 1.1764f;
+
+                        const int repeat_last_n    = 256;
+
+                        llama_token id = 0;
+
+                        {
+                            auto logits = llama_get_logits(ctx_llama);
+                            logits[llama_token_eos()] = 0;
+
+                            id = llama_sample_top_p_top_k(ctx_llama,
+                                    embd_inp.data() + std::max(0, n_past - repeat_last_n),
+                                    repeat_last_n, top_k, top_p, temp, repeat_penalty);
+                        }
+
+                        if (id != llama_token_eos()) {
+                            // add it to the context
+                            embd.push_back(id);
+
+                            text_to_speak += llama_token_to_str(ctx_llama, id);
+
+                            printf("%s", llama_token_to_str(ctx_llama, id));
+                        }
+                    }
+
+                    {
+                        std::string last_output;
+                        for (int i = embd_inp.size() - 16; i < (int) embd_inp.size(); i++) {
+                            last_output += llama_token_to_str(ctx_llama, embd_inp[i]);
+                        }
+                        last_output += llama_token_to_str(ctx_llama, embd[0]);
+
+                        for (std::string & antiprompt : antiprompts) {
+                            if (last_output.find(antiprompt.c_str(), last_output.length() - antiprompt.length(), antiprompt.length()) != std::string::npos) {
+                                done = true;
+                                text_to_speak = ::replace(text_to_speak, antiprompt, "");
+                                fflush(stdout);
+                                break;
+                            }
+                        }
+                    }
+
+                    is_running = sdl_poll_events();
+
+                    if (!is_running) {
+                        break;
+                    }
+                }
+
+                text_to_speak = ::replace(text_to_speak, "\"", "");
+                system((params.speak + " " + std::to_string(voice_id) + " \"" + text_to_speak + "\"").c_str());
+
+                audio.clear();
+
+                ++n_iter;
+            }
+        }
+    }
+
+    audio.pause();
+
+    whisper_print_timings(ctx_wsp);
+    whisper_free(ctx_wsp);
+
+    llama_print_timings(ctx_llama);
+    llama_free(ctx_llama);
+
+    return 0;
+}

examples/talk/speak.sh

@@ -7,7 +7,10 @@
 # Mac OS: brew install espeak
 # Linux: apt-get install espeak
 #
-espeak -v en-us+m$1 -s 175 -p 50 -a 200 -g 5 -k 5 "$2"
+#espeak -v en-us+m$1 -s 175 -p 50 -a 200 -g 5 -k 5 "$2"
+
+# Mac OS "say" command
+say "$2"
 
 # Eleven Labs
 #

ggml.h

@@ -198,6 +198,8 @@ struct ggml_object;
 struct ggml_context;
 
 enum ggml_type {
+    GGML_TYPE_Q4_0,
+    GGML_TYPE_Q4_1,
     GGML_TYPE_I8,
     GGML_TYPE_I16,
     GGML_TYPE_I32,
@@ -226,7 +228,9 @@ enum ggml_op {
     GGML_OP_STEP,
     GGML_OP_RELU,
     GGML_OP_GELU,
+    GGML_OP_SILU,
     GGML_OP_NORM, // normalize
+    GGML_OP_RMS_NORM,
 
     GGML_OP_MUL_MAT,
 
@@ -326,7 +330,10 @@ void ggml_print_objects(const struct ggml_context * ctx);
 int    ggml_nelements(const struct ggml_tensor * tensor);
 size_t ggml_nbytes   (const struct ggml_tensor * tensor);
 
-size_t ggml_type_size   (enum ggml_type type);
+int    ggml_blck_size (enum ggml_type type);
+size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
+float  ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float
+
 size_t ggml_element_size(const struct ggml_tensor * tensor);
 
 struct ggml_context * ggml_init(struct ggml_init_params params);
@@ -336,6 +343,9 @@ size_t ggml_used_mem(const struct ggml_context * ctx);
 
 size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch);
 
+bool ggml_mlock_supported(void);
+bool ggml_mlock(struct ggml_context * ctx, char ** err_p);
+
 struct ggml_tensor * ggml_new_tensor(
         struct ggml_context * ctx,
         enum   ggml_type type,
@@ -466,12 +476,20 @@ struct ggml_tensor * ggml_gelu(
         struct ggml_context * ctx,
         struct ggml_tensor  * a);
 
+struct ggml_tensor * ggml_silu(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
 // normalize along rows
 // TODO: eps is hardcoded to 1e-5 for now
 struct ggml_tensor * ggml_norm(
         struct ggml_context * ctx,
         struct ggml_tensor  * a);
 
+struct ggml_tensor * ggml_rms_norm(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
 // A: m rows, n columns
 // B: p rows, n columns (i.e. we transpose it internally)
 // result is m columns, p rows
@@ -726,6 +744,13 @@ enum ggml_opt_result ggml_opt(
         struct ggml_opt_params params,
         struct ggml_tensor * f);
 
+//
+// quantization
+//
+
+size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int qk, int64_t * hist);
+size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int qk, int64_t * hist);
+
 //
 // system info
 //

whisper.cpp

@@ -636,6 +636,8 @@ struct whisper_context {
     whisper_model model;
     whisper_vocab vocab;
     whisper_state * state = nullptr;
+
+    std::string path_model; // populated by whisper_init_from_file()
 };
 
 template<typename T>
@@ -1597,7 +1599,7 @@ static bool whisper_encode_internal(
                         ggml_repeat(ctx0, layer.mlp_ln_w, cur),
                         cur),
                     ggml_repeat(ctx0, layer.mlp_ln_b, cur));
-    }
+            }
 
 #ifdef WHISPER_USE_FLASH_FF
             wstate.use_buf(ctx0, 0);
@@ -1637,7 +1639,7 @@ static bool whisper_encode_internal(
                 ggml_repeat(ctx0, layer.mlp_1_b, cur),
                 cur);
 #endif
-}
+        }
 
         wstate.use_buf(ctx0, 3);
 
@@ -1841,8 +1843,6 @@ static bool whisper_decode_internal(
 
         // self-attention
         {
-            wstate.use_buf(ctx0, 1);
-
             struct ggml_tensor * Qcur = ggml_mul_mat(ctx0,
                     layer.attn_q_w,
                     cur);
@@ -1904,8 +1904,6 @@ static bool whisper_decode_internal(
             // K * Q
             struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
 
-            wstate.use_buf(ctx0, 0);
-
             //struct ggml_tensor * KQ_scaled =
             //    ggml_scale(ctx0,
             //            KQ,
@@ -1914,20 +1912,16 @@ static bool whisper_decode_internal(
 
             struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ, n_past);
 
-            wstate.use_buf(ctx0, 1);
-
             struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
 
-            wstate.use_buf(ctx0, 0);
-
             struct ggml_tensor * V_trans =
-                ggml_permute(ctx0,
-                        ggml_reshape_3d(ctx0,
-                            ggml_view_1d(ctx0, kv_self.v, (n_past + N)*n_state, il*n_ctx*ggml_element_size(kv_self.v)*n_state),
-                            n_state/n_head, n_head, n_past + N),
-                        1, 2, 0, 3);
-
-            wstate.use_buf(ctx0, 1);
+                ggml_cpy(ctx0,
+                        ggml_permute(ctx0,
+                            ggml_reshape_3d(ctx0,
+                                ggml_view_1d(ctx0, kv_self.v, (n_past + N)*n_state, il*n_ctx*ggml_element_size(kv_self.v)*n_state),
+                                n_state/n_head, n_head, n_past + N),
+                            1, 2, 0, 3),
+                        ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_state/n_head, n_head));
 
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
 
@@ -1964,8 +1958,6 @@ static bool whisper_decode_internal(
 
             cur = ggml_norm(ctx0, inpCA); // note: we use inpCA here
 
-            wstate.use_buf(ctx0, 1);
-
             // cur = ln_0_w*cur + ln_0_b
             cur = ggml_add(ctx0,
                     ggml_mul(ctx0,
@@ -1976,8 +1968,6 @@ static bool whisper_decode_internal(
 
         // cross-attention
         {
-            wstate.use_buf(ctx0, 0);
-
             struct ggml_tensor * Qcur = ggml_mul_mat(ctx0,
                     layer.cross_attn_q_w,
                     cur);
@@ -2001,12 +1991,13 @@ static bool whisper_decode_internal(
                         ggml_view_1d(ctx0, wstate.kv_cross.v, M*n_state, il*M*ggml_element_size(wstate.kv_cross.v)*n_state),
                         n_state/n_head, n_head, M);
 
-            struct ggml_tensor * V_trans = ggml_permute(ctx0, Vcross, 1, 2, 0, 3);
+            struct ggml_tensor * V_trans =
+                ggml_cpy(ctx0,
+                        ggml_permute(ctx0, Vcross, 1, 2, 0, 3),
+                        ggml_new_tensor_3d(ctx0, Vcross->type, M, n_state/n_head, n_head));
 
             // ------
 
-            wstate.use_buf(ctx0, 1);
-
             struct ggml_tensor * Q =
                 ggml_permute(ctx0,
                         ggml_cpy(ctx0,
@@ -2016,8 +2007,6 @@ static bool whisper_decode_internal(
 
             struct ggml_tensor * K = ggml_permute(ctx0, Kcross, 0, 2, 1, 3);
 
-            wstate.use_buf(ctx0, 0);
-
             // K * Q
             struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
 
@@ -2030,16 +2019,10 @@ static bool whisper_decode_internal(
             // no masking for cross-attention
             //struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
 
-            wstate.use_buf(ctx0, 1);
-
             struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ);
 
-            wstate.use_buf(ctx0, 0);
-
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
 
-            wstate.use_buf(ctx0, 1);
-
             struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
 
             // cur = KQV_merged.contiguous().view(n_state, N)
@@ -2482,7 +2465,6 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
 
     const size_t scale = ctx->model.hparams.f16 ? 1 : 2;
 
-
     if (!kv_cache_init(ctx->model.hparams, scale * MEM_REQ_KV_SELF.at(ctx->model.type), state->decoders[0].kv_self, ctx->wtype, ctx->model.hparams.n_text_ctx)) {
         fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
         return nullptr;
@@ -2503,7 +2485,6 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
         fprintf(stderr, "%s: kv cross size = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
     }
 
-
     state->logits.reserve(ctx->vocab.n_vocab * ctx->model.hparams.n_text_ctx);
 
     state->logits_id.reserve(ctx->model.hparams.n_vocab);
@@ -2554,7 +2535,13 @@ struct whisper_context * whisper_init_from_file_no_state(const char * path_model
         fin->close();
     };
 
-    return whisper_init_no_state(&loader);
+    auto ctx = whisper_init_no_state(&loader);
+
+    if (ctx) {
+        ctx->path_model = path_model;
+    }
+
+    return ctx;
 }
 
 struct whisper_context * whisper_init_from_buffer_no_state(void * buffer, size_t buffer_size) {