whispercpp 1.3.6 → 1.3.7

data/ext/sources/examples/parakeet-cli/parakeet-cli.cpp

@@ -0,0 +1,243 @@
+#include "parakeet.h"
+#include "common-whisper.h"
+
+#include <cstdio>
+#include <string>
+#include <thread>
+#include <vector>
+#include <cstring>
+#include <fstream>
+
+// command-line parameters
+struct parakeet_params {
+    int32_t n_threads         = std::min(4, (int32_t) std::thread::hardware_concurrency());
+
+    bool use_gpu       = true;
+    int32_t gpu_device = 0;
+
+    bool print_segments = false;
+    bool output_txt     = false;
+    bool no_prints      = false;
+
+    std::string model       = "models/ggml-parakeet-tdt-0.6b-v3.bin";
+    std::string output_file = "";
+    std::vector<std::string> fname_inp = {};
+};
+
+static void parakeet_print_usage(int argc, char ** argv, const parakeet_params & params);
+
+static char * requires_value_error(const std::string & arg) {
+    fprintf(stderr, "error: argument %s requires value\n", arg.c_str());
+    exit(1);
+}
+
+static bool parakeet_params_parse(int argc, char ** argv, parakeet_params & params) {
+    if (const char * env_device = std::getenv("PARAKEET_ARG_DEVICE")) {
+        params.gpu_device = std::stoi(env_device);
+    }
+
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-"){
+            params.fname_inp.push_back(arg);
+            continue;
+        }
+
+        if (arg[0] != '-') {
+            params.fname_inp.push_back(arg);
+            continue;
+        }
+
+        if (arg == "-h" || arg == "--help") {
+            parakeet_print_usage(argc, argv, params);
+            exit(0);
+        }
+        #define ARGV_NEXT (((i + 1) < argc) ? argv[++i] : requires_value_error(arg))
+        else if (arg == "-t"    || arg == "--threads")         { params.n_threads         = std::stoi(ARGV_NEXT); }
+        else if (arg == "-m"    || arg == "--model")           { params.model             = ARGV_NEXT; }
+        else if (arg == "-f"    || arg == "--file")            { params.fname_inp.emplace_back(ARGV_NEXT); }
+        else if (arg == "-ng"   || arg == "--no-gpu")          { params.use_gpu           = false; }
+        else if (arg == "-dev"  || arg == "--device")          { params.gpu_device        = std::stoi(ARGV_NEXT); }
+        else if (arg == "-ps"   || arg == "--print-segments")  { params.print_segments    = true; }
+        else if (arg == "-otxt" || arg == "--output-txt")      { params.output_txt        = true; }
+        else if (arg == "-of"   || arg == "--output-file")     { params.output_file       = ARGV_NEXT; }
+        else if (arg == "-np"   || arg == "--no-prints")       { params.no_prints         = true; }
+        else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            parakeet_print_usage(argc, argv, params);
+            exit(1);
+        }
+    }
+
+    return true;
+}
+
+static void parakeet_print_usage(int /*argc*/, char ** argv, const parakeet_params & params) {
+    fprintf(stderr, "\n");
+    fprintf(stderr, "usage: %s [options] file0 file1 ...\n", argv[0]);
+    fprintf(stderr, "supported audio formats: flac, mp3, ogg, wav\n");
+    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, "  -m,     --model FILE        [%-7s] model path\n",                                  params.model.c_str());
+    fprintf(stderr, "  -f,     --file FILE         [%-7s] input audio file\n",                            "");
+    fprintf(stderr, "  -ng,    --no-gpu            [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -dev N, --device N          [%-7d] GPU device to use\n",                           params.gpu_device);
+    fprintf(stderr, "  -ps,    --print-segments    [%-7s] print segment information\n",                   params.print_segments ? "true" : "false");
+    fprintf(stderr, "  -otxt,  --output-txt        [%-7s] output result in a text file\n",                params.output_txt ? "true" : "false");
+    fprintf(stderr, "  -of,    --output-file FILE  [%-7s] output file path (without file extension)\n",   "");
+    fprintf(stderr, "  -np,    --no-prints         [%-7s] do not print anything other than the results\n", params.no_prints ? "true" : "false");
+    fprintf(stderr, "\n");
+}
+
+void token_callback(parakeet_context * ctx, parakeet_state * state, const parakeet_token_data * token_data, void * user_data) {
+    bool * is_first = (bool *) user_data;
+
+    const char * token_str = parakeet_token_to_str(ctx, token_data->id);
+    char text_buf[256];
+    parakeet_token_to_text(token_str, *is_first, text_buf, sizeof(text_buf));
+    printf("%s", text_buf);
+    fflush(stdout);
+
+    *is_first = false;
+}
+
+static void cb_log_disable(enum ggml_log_level , const char * , void * ) { }
+
+int main(int argc, char ** argv) {
+    ggml_backend_load_all();
+
+    parakeet_params params;
+
+    if (parakeet_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (params.no_prints) {
+        parakeet_log_set(cb_log_disable, NULL);
+    }
+
+    if (params.fname_inp.empty()) {
+        fprintf(stderr, "error: no input files specified\n");
+        parakeet_print_usage(argc, argv, params);
+        return 1;
+    }
+
+    struct parakeet_context_params ctx_params = parakeet_context_default_params();
+    ctx_params.use_gpu     = params.use_gpu;
+    ctx_params.gpu_device  = params.gpu_device;
+
+    if (!params.no_prints) {
+        fprintf(stderr, "Loading Parakeet model from: %s\n", params.model.c_str());
+    }
+
+
+    struct parakeet_context * pctx = parakeet_init_from_file_with_params(params.model.c_str(), ctx_params);
+    if (pctx == nullptr) {
+        fprintf(stderr, "error: failed to load Parakeet model from '%s'\n", params.model.c_str());
+        return 1;
+    }
+
+    if (!params.no_prints) {
+        fprintf(stderr, "Successfully loaded Parakeet model\n");
+        fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
+                params.n_threads, (int32_t) std::thread::hardware_concurrency(), parakeet_print_system_info());
+    }
+
+    // Process each input file
+    for (const auto & fname : params.fname_inp) {
+        if (!params.no_prints) {
+            fprintf(stderr, "\nProcessing file: %s\n", fname.c_str());
+        }
+
+        std::vector<float> pcmf32;
+        std::vector<std::vector<float>> pcmf32s;
+        if (!read_audio_data(fname.c_str(), pcmf32, pcmf32s, false)) {
+            fprintf(stderr, "error: failed to read audio file '%s'\n", fname.c_str());
+            continue;
+        }
+
+        if (pcmf32.empty()) {
+            fprintf(stderr, "error: no audio data in file '%s'\n", fname.c_str());
+            continue;
+        }
+
+        bool is_first = true;
+        struct parakeet_full_params full_params = parakeet_full_default_params(PARAKEET_SAMPLING_GREEDY);
+        full_params.n_threads           = params.n_threads;
+        full_params.new_token_callback  = token_callback;
+        full_params.new_token_callback_user_data = &is_first;
+
+        const int mel_frames = (int)(pcmf32.size() / PARAKEET_HOP_LENGTH);
+        int ret = parakeet_full(pctx, full_params, pcmf32.data(), pcmf32.size());
+
+        if (ret != 0) {
+            fprintf(stderr, "error: failed to process audio file '%s'\n", fname.c_str());
+            continue;
+        }
+
+        printf("\n");
+
+        if (params.output_txt) {
+            const std::string fname_out = (!params.output_file.empty() ? params.output_file : fname) + ".txt";
+
+            std::ofstream fout(fname_out);
+            if (fout.is_open()) {
+                const int n_segments = parakeet_full_n_segments(pctx);
+                for (int i = 0; i < n_segments; ++i) {
+                    const char * text = parakeet_full_get_segment_text(pctx, i);
+                    fout << text << "\n";
+                }
+                fout.close();
+                if (!params.no_prints) {
+                    fprintf(stderr, "Output written to: %s\n", fname_out.c_str());
+                }
+            } else {
+                fprintf(stderr, "error: failed to open '%s' for writing\n", fname_out.c_str());
+            }
+        }
+
+        if (!params.no_prints) {
+            parakeet_print_timings(pctx);
+        }
+
+        if (params.print_segments) {
+            const int n_segments = parakeet_full_n_segments(pctx);
+            fprintf(stderr, "\nSegments (%d):\n", n_segments);
+
+            for (int i = 0; i < n_segments; i++) {
+                const char * text = parakeet_full_get_segment_text(pctx, i);
+                const int64_t t0 = parakeet_full_get_segment_t0(pctx, i);
+                const int64_t t1 = parakeet_full_get_segment_t1(pctx, i);
+                const int n_tokens = parakeet_full_n_tokens(pctx, i);
+
+                fprintf(stderr, "Segment %d: [%lld -> %lld] \"%s\"\n", i, (long long)t0, (long long)t1, text);
+                fprintf(stderr, "Tokens [%d]:\n", n_tokens);
+
+                for (int j = 0; j < n_tokens; j++) {
+                    parakeet_token_data token_data = parakeet_full_get_token_data(pctx, i, j);
+                    const char * token_str = parakeet_token_to_str(pctx, token_data.id);
+
+                    fprintf(stderr, "  [%2d] id=%5d frame=%3d dur_idx=%2d dur_val=%2d p=%.4f plog=%.4f t0=%4lld t1=%4lld word_start=%s \"%s\"\n",
+                           j,
+                           token_data.id,
+                           token_data.frame_index,
+                           token_data.duration_idx,
+                           token_data.duration_value,
+                           token_data.p,
+                           token_data.plog,
+                           (long long)token_data.t0,
+                           (long long)token_data.t1,
+                           token_data.is_word_start ? "true": "false",
+                           token_str);
+                }
+            }
+        }
+    }
+
+    parakeet_free(pctx);
+
+    return 0;
+}

data/ext/sources/examples/parakeet-quantize/CMakeLists.txt

@@ -0,0 +1,7 @@
+set(TARGET parakeet-quantize)
+add_executable(${TARGET} parakeet-quantize.cpp)
+
+include(DefaultTargetOptions)
+
+target_link_libraries(${TARGET} PRIVATE common parakeet ${CMAKE_THREAD_LIBS_INIT})
+install(TARGETS ${TARGET} RUNTIME)

data/ext/sources/examples/parakeet-quantize/parakeet-quantize.cpp

@@ -0,0 +1,230 @@
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#include "common-ggml.h"
+
+#include <cassert>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <string>
+#include <vector>
+
+struct parakeet_hparams {
+    int32_t n_vocab                = 0;
+    int32_t n_audio_ctx            = 0;
+    int32_t n_audio_state          = 0;
+    int32_t n_audio_head           = 0;
+    int32_t n_audio_layer          = 0;
+    int32_t n_mels                 = 0;
+    int32_t ftype                  = 0;
+    int32_t n_fft                  = 0;
+    int32_t subsampling_factor     = 0;
+    int32_t n_subsampling_channels = 0;
+    int32_t n_conv_kernel          = 0;
+    int32_t n_pred_dim             = 0;
+    int32_t n_pred_layers          = 0;
+    int32_t n_tdt_durations        = 0;
+    int32_t n_max_tokens           = 0;
+};
+
+static bool parakeet_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
+    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;
+    }
+
+    // magic
+    {
+        uint32_t magic;
+        finp.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file (bad magic)\n", __func__);
+            return false;
+        }
+        fout.write((char *) &magic, sizeof(magic));
+    }
+
+    // hparams
+    parakeet_hparams hparams;
+    {
+        finp.read((char *) &hparams.n_vocab,                sizeof(hparams.n_vocab));
+        finp.read((char *) &hparams.n_audio_ctx,            sizeof(hparams.n_audio_ctx));
+        finp.read((char *) &hparams.n_audio_state,          sizeof(hparams.n_audio_state));
+        finp.read((char *) &hparams.n_audio_head,           sizeof(hparams.n_audio_head));
+        finp.read((char *) &hparams.n_audio_layer,          sizeof(hparams.n_audio_layer));
+        finp.read((char *) &hparams.n_mels,                 sizeof(hparams.n_mels));
+        finp.read((char *) &hparams.ftype,                  sizeof(hparams.ftype));
+        finp.read((char *) &hparams.n_fft,                  sizeof(hparams.n_fft));
+        finp.read((char *) &hparams.subsampling_factor,     sizeof(hparams.subsampling_factor));
+        finp.read((char *) &hparams.n_subsampling_channels, sizeof(hparams.n_subsampling_channels));
+        finp.read((char *) &hparams.n_conv_kernel,          sizeof(hparams.n_conv_kernel));
+        finp.read((char *) &hparams.n_pred_dim,             sizeof(hparams.n_pred_dim));
+        finp.read((char *) &hparams.n_pred_layers,          sizeof(hparams.n_pred_layers));
+        finp.read((char *) &hparams.n_tdt_durations,        sizeof(hparams.n_tdt_durations));
+        finp.read((char *) &hparams.n_max_tokens,           sizeof(hparams.n_max_tokens));
+
+        const int32_t qntvr_src = hparams.ftype / GGML_QNT_VERSION_FACTOR;
+        const int32_t ftype_dst = GGML_QNT_VERSION * GGML_QNT_VERSION_FACTOR + ftype;
+
+        fprintf(stderr, "%s: n_vocab              = %d\n",  __func__, hparams.n_vocab);
+        fprintf(stderr, "%s: n_audio_state        = %d\n",  __func__, hparams.n_audio_state);
+        fprintf(stderr, "%s: n_audio_layer        = %d\n",  __func__, hparams.n_audio_layer);
+        fprintf(stderr, "%s: n_mels               = %d\n",  __func__, hparams.n_mels);
+        fprintf(stderr, "%s: ftype (src)          = %d\n",  __func__, hparams.ftype);
+        fprintf(stderr, "%s: qntvr (src)          = %d\n",  __func__, qntvr_src);
+        fprintf(stderr, "%s: ftype (dst)          = %d\n",  __func__, ftype_dst);
+        fprintf(stderr, "%s: qntvr (dst)          = %d\n",  __func__, GGML_QNT_VERSION);
+
+        fout.write((char *) &hparams.n_vocab,                sizeof(hparams.n_vocab));
+        fout.write((char *) &hparams.n_audio_ctx,            sizeof(hparams.n_audio_ctx));
+        fout.write((char *) &hparams.n_audio_state,          sizeof(hparams.n_audio_state));
+        fout.write((char *) &hparams.n_audio_head,           sizeof(hparams.n_audio_head));
+        fout.write((char *) &hparams.n_audio_layer,          sizeof(hparams.n_audio_layer));
+        fout.write((char *) &hparams.n_mels,                 sizeof(hparams.n_mels));
+        fout.write((char *) &ftype_dst,                      sizeof(ftype_dst));
+        fout.write((char *) &hparams.n_fft,                  sizeof(hparams.n_fft));
+        fout.write((char *) &hparams.subsampling_factor,     sizeof(hparams.subsampling_factor));
+        fout.write((char *) &hparams.n_subsampling_channels, sizeof(hparams.n_subsampling_channels));
+        fout.write((char *) &hparams.n_conv_kernel,          sizeof(hparams.n_conv_kernel));
+        fout.write((char *) &hparams.n_pred_dim,             sizeof(hparams.n_pred_dim));
+        fout.write((char *) &hparams.n_pred_layers,          sizeof(hparams.n_pred_layers));
+        fout.write((char *) &hparams.n_tdt_durations,        sizeof(hparams.n_tdt_durations));
+        fout.write((char *) &hparams.n_max_tokens,           sizeof(hparams.n_max_tokens));
+    }
+
+    // mel filterbank
+    {
+        int32_t n_mel, n_fb;
+        finp.read((char *) &n_mel, sizeof(n_mel));
+        fout.write((char *) &n_mel, sizeof(n_mel));
+        finp.read((char *) &n_fb,  sizeof(n_fb));
+        fout.write((char *) &n_fb,  sizeof(n_fb));
+
+        const size_t n = (size_t) n_mel * n_fb;
+        std::vector<float> buf(n);
+        finp.read((char *) buf.data(), n * sizeof(float));
+        fout.write((char *) buf.data(), n * sizeof(float));
+    }
+
+    // window function
+    {
+        int32_t n_window;
+        finp.read((char *) &n_window, sizeof(n_window));
+        fout.write((char *) &n_window, sizeof(n_window));
+
+        std::vector<float> buf(n_window);
+        finp.read((char *) buf.data(), n_window * sizeof(float));
+        fout.write((char *) buf.data(), n_window * sizeof(float));
+    }
+
+    // TDT durations
+    {
+        std::vector<uint32_t> buf(hparams.n_tdt_durations);
+        finp.read((char *) buf.data(), hparams.n_tdt_durations * sizeof(uint32_t));
+        fout.write((char *) buf.data(), hparams.n_tdt_durations * sizeof(uint32_t));
+    }
+
+    // vocab
+    {
+        int32_t n_tokens;
+        finp.read((char *) &n_tokens, sizeof(n_tokens));
+        fout.write((char *) &n_tokens, sizeof(n_tokens));
+
+        for (int i = 0; i < n_tokens; ++i) {
+            int32_t len;
+            finp.read((char *) &len, sizeof(len));
+            fout.write((char *) &len, sizeof(len));
+
+            std::string token(len, '\0');
+            finp.read(&token[0], len);
+            fout.write(&token[0], len);
+        }
+    }
+
+    // tensors — quantize 2D weights skipping tensors that must stay F32:
+    // ggml_ssm_conv / ggml_conv2d_dw CUDA kernels require F32 weights.
+    // pos_bias_u / pos_bias_v are declared F32 in the loader.
+    const std::vector<std::string> to_quant = { ".*" };
+    std::vector<std::string> to_skip = {
+        // CUDA kernel constraints (ggml_ssm_conv / ggml_conv2d_dw require F32 weights)
+        "encoder\\.layers\\..+\\.conv\\.depthwise_conv\\.weight",
+        // Declared F32 in loader (pos_bias tensors)
+        "encoder\\.layers\\..+\\.self_attn\\.pos_bias_u",
+        "encoder\\.layers\\..+\\.self_attn\\.pos_bias_v",
+    };
+
+    // Prediction/joint tensors use n_pred_dim as their inner dimension. K-quant
+    // types (block size 256) cannot quantize 640 evenly, so keep them F32. For
+    // other types (Q8_0, Q4_0, block size 32) 640 is divisible and they can be
+    // quantized normally. The loader mirrors this logic at load time.
+    {
+        const ggml_type qtype = ggml_ftype_to_ggml_type(ftype);
+        const int32_t   blck  = ggml_blck_size(qtype);
+        if (blck > 1 && hparams.n_pred_dim % blck != 0) {
+            to_skip.push_back("decoder\\.prediction\\.embed\\.weight");
+            to_skip.push_back("decoder\\.prediction\\.dec_rnn\\.lstm\\.weight_ih_l.*");
+            to_skip.push_back("decoder\\.prediction\\.dec_rnn\\.lstm\\.weight_hh_l.*");
+            to_skip.push_back("joint\\.pred\\.weight");
+            to_skip.push_back("joint\\.joint_net\\.2\\.weight");
+        }
+    }
+
+    if (!ggml_common_quantize_0(finp, fout, ftype, to_quant, to_skip)) {
+        fprintf(stderr, "%s: failed to quantize tensors\n", __func__);
+        return false;
+    }
+
+    finp.close();
+    fout.close();
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+    ggml_backend_load_all();
+
+    if (argc != 4) {
+        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
+        ggml_print_ftypes(stderr);
+        return 1;
+    }
+
+    // initialise F16 lookup tables
+    {
+        struct ggml_init_params params = { 0, NULL, false };
+        struct ggml_context * ctx = ggml_init(params);
+        ggml_free(ctx);
+    }
+
+    const std::string fname_inp = argv[1];
+    const std::string fname_out = argv[2];
+    const ggml_ftype  ftype     = ggml_parse_ftype(argv[3]);
+
+    if (ftype == GGML_FTYPE_UNKNOWN) {
+        fprintf(stderr, "%s: invalid quantization type\n", argv[0]);
+        ggml_print_ftypes(stderr);
+        return 1;
+    }
+
+    const int64_t t_start_us = ggml_time_us();
+
+    if (!parakeet_model_quantize(fname_inp, fname_out, ftype)) {
+        fprintf(stderr, "%s: failed to quantize model from '%s'\n", argv[0], fname_inp.c_str());
+        return 1;
+    }
+
+    printf("\n%s: quantize time = %8.2f ms\n", argv[0], (ggml_time_us() - t_start_us) / 1000.0f);
+    printf("%s: output model  = %s\n",         argv[0], fname_out.c_str());
+
+    return 0;
+}