whispercpp 1.3.2 → 1.3.4

data/ext/sources/examples/bench/bench.cpp

@@ -13,7 +13,7 @@ struct whisper_params {
     std::string model = "models/ggml-base.en.bin";
 
     bool use_gpu    = true;
-    bool flash_attn = false;
+    bool flash_attn = true;
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -26,11 +26,12 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
             whisper_print_usage(argc, argv, params);
             exit(0);
         }
-        else if (arg == "-t"  || arg == "--threads")    { params.n_threads  = std::stoi(argv[++i]); }
-        else if (arg == "-m"  || arg == "--model")      { params.model      = argv[++i]; }
-        else if (arg == "-w"  || arg == "--what")       { params.what       = atoi(argv[++i]); }
-        else if (arg == "-ng" || arg == "--no-gpu")     { params.use_gpu    = false; }
-        else if (arg == "-fa" || arg == "--flash-attn") { params.flash_attn = true; }
+        else if (arg == "-t"     || arg == "--threads")       { params.n_threads  = std::stoi(argv[++i]); }
+        else if (arg == "-m"     || arg == "--model")         { params.model      = argv[++i]; }
+        else if (arg == "-w"     || arg == "--what")          { params.what       = atoi(argv[++i]); }
+        else if (arg == "-ng"    || arg == "--no-gpu")        { params.use_gpu    = false; }
+        else if (arg == "-fa"    || arg == "--flash-attn")    { params.flash_attn = true; }
+        else if (arg == "-nfa"   || arg == "--no-flash-attn") { params.flash_attn = false; }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -46,15 +47,16 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     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, "  -m FNAME, --model FNAME [%-7s] model path\n",                                  params.model.c_str());
-    fprintf(stderr, "  -w N,     --what N      [%-7d] what to benchmark:\n",                          params.what);
-    fprintf(stderr, "                           %-7s  0 - whisper\n",                                 "");
-    fprintf(stderr, "                           %-7s  1 - memcpy\n",                                  "");
-    fprintf(stderr, "                           %-7s  2 - ggml_mul_mat\n",                            "");
-    fprintf(stderr, "  -ng,      --no-gpu      [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -fa,      --flash-attn  [%-7s] enable flash attention\n",                      params.flash_attn ? "true" : "false");
+    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 FNAME, --model FNAME   [%-7s] model path\n",                                  params.model.c_str());
+    fprintf(stderr, "  -w N,     --what N        [%-7d] what to benchmark:\n",                          params.what);
+    fprintf(stderr, "                             %-7s  0 - whisper\n",                                 "");
+    fprintf(stderr, "                             %-7s  1 - memcpy\n",                                  "");
+    fprintf(stderr, "                             %-7s  2 - ggml_mul_mat\n",                            "");
+    fprintf(stderr, "  -ng,      --no-gpu        [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -fa,      --flash-attn    [%-7s] enable flash attention\n",                      params.flash_attn ? "true" : "false");
+    fprintf(stderr, "  -nfa,     --no-flash-attn [%-7s] disable flash attention\n",                     params.flash_attn ? "false" : "true");
     fprintf(stderr, "\n");
 }
 
@@ -66,13 +68,12 @@ static int whisper_bench_full(const whisper_params & params) {
     cparams.use_gpu    = params.use_gpu;
     cparams.flash_attn = params.flash_attn;
 
-    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
-
     {
         fprintf(stderr, "\n");
         fprintf(stderr, "system_info: n_threads = %d / %d | %s\n", params.n_threads, std::thread::hardware_concurrency(), whisper_print_system_info());
     }
 
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
     if (ctx == nullptr) {
         fprintf(stderr, "error: failed to initialize whisper context\n");
         return 2;
@@ -100,7 +101,15 @@ static int whisper_bench_full(const whisper_params & params) {
     }
 
     // text-generation heat
-    if (int ret = whisper_decode(ctx, tokens, 1, 256, params.n_threads) != 0) {
+    for (int i = 0; i < 256; i++) {
+        if (int ret = whisper_decode(ctx, tokens, 1, i, params.n_threads) != 0) {
+            fprintf(stderr, "error: failed to decode: %d\n", ret);
+            return 4;
+        }
+    }
+
+    // batched heat
+    if (int ret = whisper_decode(ctx, tokens, 5, 0, params.n_threads) != 0) {
         fprintf(stderr, "error: failed to decode: %d\n", ret);
         return 4;
     }
@@ -156,6 +165,8 @@ static int whisper_bench_full(const whisper_params & params) {
 }
 
 int main(int argc, char ** argv) {
+    ggml_backend_load_all();
+
     whisper_params params;
 
     if (whisper_params_parse(argc, argv, params) == false) {

data/ext/sources/examples/bench.wasm/index-tmpl.html

@@ -42,6 +42,7 @@
                 <a href="../bench.wasm/">bench</a> |
                 <a href="../stream.wasm">stream</a> |
                 <a href="../command.wasm/">command</a> |
+                <a href="../wchess.wasm/">wchess</a> |
 
             <br><br>
 
@@ -191,15 +192,15 @@
 
             function loadWhisper(model) {
                 let urls = {
-                    'tiny.en': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en.bin',
-                    'base.en': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
-                    'small.en': 'https://whisper.ggerganov.com/ggml-model-whisper-small.en.bin',
-
-                    'tiny-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en-q5_1.bin',
-                    'base-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-base.en-q5_1.bin',
-                    'small-en-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-small.en-q5_1.bin',
-                    'medium-en-q5_0':'https://whisper.ggerganov.com/ggml-model-whisper-medium.en-q5_0.bin',
-                    'large-q5_0':    'https://whisper.ggerganov.com/ggml-model-whisper-large-q5_0.bin',
+                    'tiny.en': 'https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en.bin',
+                    'base.en': 'https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin',
+                    'small.en': 'https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.en.bin',
+
+                    'tiny-en-q5_1':  'https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en-q5_1.bin',
+                    'base-en-q5_1':  'https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en-q5_1.bin',
+                    'small-en-q5_1': 'https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.en-q5_1.bin',
+                    'medium-en-q5_0':'https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.en-q5_0.bin',
+                    'large-q5_0':    'https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-q5_0.bin',
                 };
 
                 let sizes = {

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

@@ -75,7 +75,7 @@ struct whisper_params {
     bool no_timestamps   = false;
     bool log_score       = false;
     bool use_gpu         = true;
-    bool flash_attn      = false;
+    bool flash_attn      = true;
     bool suppress_nst    = false;
 
     std::string language  = "en";
@@ -193,6 +193,7 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
         else if (arg == "-ls"   || arg == "--log-score")       { params.log_score       = true; }
         else if (arg == "-ng"   || arg == "--no-gpu")          { params.use_gpu         = false; }
         else if (arg == "-fa"   || arg == "--flash-attn")      { params.flash_attn      = true; }
+        else if (arg == "-nfa"  || arg == "--no-flash-attn")   { params.flash_attn      = false; }
         else if (arg == "-sns"  || arg == "--suppress-nst")    { params.suppress_nst    = true; }
         else if (                  arg == "--suppress-regex")  { params.suppress_regex  = ARGV_NEXT; }
         else if (                  arg == "--grammar")         { params.grammar         = ARGV_NEXT; }
@@ -202,7 +203,7 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
         else if (                  arg == "--vad")                         { params.vad                         = true; }
         else if (arg == "-vm"   || arg == "--vad-model")                   { params.vad_model                   = ARGV_NEXT; }
         else if (arg == "-vt"   || arg == "--vad-threshold")               { params.vad_threshold               = std::stof(ARGV_NEXT); }
-        else if (arg == "-vsd"  || arg == "--vad-min-speech-duration-ms")  { params.vad_min_speech_duration_ms  = std::stoi(ARGV_NEXT); }
+        else if (arg == "-vspd" || arg == "--vad-min-speech-duration-ms")  { params.vad_min_speech_duration_ms  = std::stoi(ARGV_NEXT); }
         else if (arg == "-vsd"  || arg == "--vad-min-silence-duration-ms") { params.vad_min_speech_duration_ms  = std::stoi(ARGV_NEXT); }
         else if (arg == "-vmsd" || arg == "--vad-max-speech-duration-s")   { params.vad_max_speech_duration_s   = std::stof(ARGV_NEXT); }
         else if (arg == "-vp"   || arg == "--vad-speech-pad-ms")           { params.vad_speech_pad_ms           = std::stoi(ARGV_NEXT); }
@@ -271,7 +272,8 @@ static void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params
     fprintf(stderr, "  -dtw MODEL --dtw MODEL         [%-7s] compute token-level timestamps\n",                 params.dtw.c_str());
     fprintf(stderr, "  -ls,       --log-score         [%-7s] log best decoder scores of tokens\n",              params.log_score?"true":"false");
     fprintf(stderr, "  -ng,       --no-gpu            [%-7s] disable GPU\n",                                    params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -fa,       --flash-attn        [%-7s] flash attention\n",                                params.flash_attn ? "true" : "false");
+    fprintf(stderr, "  -fa,       --flash-attn        [%-7s] enable flash attention\n",                         params.flash_attn ? "true" : "false");
+    fprintf(stderr, "  -nfa,      --no-flash-attn     [%-7s] disable flash attention\n",                        params.flash_attn ? "false" : "true");
     fprintf(stderr, "  -sns,      --suppress-nst      [%-7s] suppress non-speech tokens\n",                     params.suppress_nst ? "true" : "false");
     fprintf(stderr, "  --suppress-regex REGEX         [%-7s] regular expression matching tokens to suppress\n", params.suppress_regex.c_str());
     fprintf(stderr, "  --grammar GRAMMAR              [%-7s] GBNF grammar to guide decoding\n",                 params.grammar.c_str());
@@ -909,6 +911,8 @@ static void output_lrc(struct whisper_context * ctx, std::ofstream & fout, const
 static void cb_log_disable(enum ggml_log_level , const char * , void * ) { }
 
 int main(int argc, char ** argv) {
+    ggml_backend_load_all();
+
 #if defined(_WIN32)
     // Set the console output code page to UTF-8, while command line arguments
     // are still encoded in the system's code page. In this way, we can print
@@ -988,7 +992,6 @@ int main(int argc, char ** argv) {
     }
 
     // whisper init
-
     struct whisper_context_params cparams = whisper_context_default_params();
 
     cparams.use_gpu    = params.use_gpu;

data/ext/sources/examples/command/command.cpp

@@ -42,7 +42,7 @@ struct whisper_params {
     bool print_energy  = false;
     bool no_timestamps = true;
     bool use_gpu       = true;
-    bool flash_attn    = false;
+    bool flash_attn    = true;
 
     std::string language  = "en";
     std::string model     = "models/ggml-base.en.bin";
@@ -66,28 +66,29 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
             whisper_print_usage(argc, argv, params);
             exit(0);
         }
-        else if (arg == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
-        else if (arg == "-pms" || arg == "--prompt-ms")     { params.prompt_ms     = std::stoi(argv[++i]); }
-        else if (arg == "-cms" || arg == "--command-ms")    { params.command_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 == "-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 == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
-        else if (arg == "-fa"  || arg == "--flash-attn")    { params.flash_attn    = true; }
-        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
-        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
-        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
-        else if (arg == "-cmd" || arg == "--commands")      { params.commands      = argv[++i]; }
-        else if (arg == "-p"   || arg == "--prompt")        { params.prompt        = argv[++i]; }
-        else if (arg == "-ctx" || arg == "--context")       { params.context       = argv[++i]; }
-        else if (                 arg == "--grammar")       { params.grammar       = argv[++i]; }
-        else if (                 arg == "--grammar-penalty") { params.grammar_penalty = std::stof(argv[++i]); }
-        else if (                 arg == "--suppress-regex") { params.suppress_regex = argv[++i]; }
+        else if (arg == "-t"     || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
+        else if (arg == "-pms"   || arg == "--prompt-ms")     { params.prompt_ms     = std::stoi(argv[++i]); }
+        else if (arg == "-cms"   || arg == "--command-ms")    { params.command_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 == "-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 == "-ng"    || arg == "--no-gpu")        { params.use_gpu       = false; }
+        else if (arg == "-fa"    || arg == "--flash-attn")    { params.flash_attn    = true; }
+        else if (arg == "-nfa"   || arg == "--no-flash-attn") { params.flash_attn    = false; }
+        else if (arg == "-l"     || arg == "--language")      { params.language      = argv[++i]; }
+        else if (arg == "-m"     || arg == "--model")         { params.model         = argv[++i]; }
+        else if (arg == "-f"     || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else if (arg == "-cmd"   || arg == "--commands")      { params.commands      = argv[++i]; }
+        else if (arg == "-p"     || arg == "--prompt")        { params.prompt        = argv[++i]; }
+        else if (arg == "-ctx"   || arg == "--context")       { params.context       = argv[++i]; }
+        else if (                   arg == "--grammar")       { params.grammar       = argv[++i]; }
+        else if (                   arg == "--grammar-penalty") { params.grammar_penalty = std::stof(argv[++i]); }
+        else if (                   arg == "--suppress-regex") { params.suppress_regex = argv[++i]; }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -116,7 +117,8 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     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, "  -ng,        --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -fa,        --flash-attn     [%-7s] flash attention\n",                             params.flash_attn ? "true" : "false");
+    fprintf(stderr, "  -fa,        --flash-attn     [%-7s] enbale flash attention\n",                      params.flash_attn ? "true" : "false");
+    fprintf(stderr, "  -nfa,       --no-flash-attn  [%-7s] disable flash attention\n",                     params.flash_attn ? "false" : "true");
     fprintf(stderr, "  -l LANG,    --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
     fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
     fprintf(stderr, "  -f FNAME,   --file FNAME     [%-7s] text output file name\n",                       params.fname_out.c_str());
@@ -251,7 +253,7 @@ static std::vector<std::string> get_words(const std::string &txt) {
 
 // command-list mode
 // guide the transcription to match the most likely command from a provided list
-static int process_command_list(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
+static int process_command_list(struct whisper_context * ctx, audio_async &audio, const whisper_params &params, std::ofstream &fout) {
     fprintf(stderr, "\n");
     fprintf(stderr, "%s: guided mode\n", __func__);
 
@@ -444,12 +446,16 @@ static int process_command_list(struct whisper_context * ctx, audio_async &audio
 
                     const float prob = probs_id[0].first;
                     const int index = probs_id[0].second;
+                    const char * best_command = allowed_commands[index].c_str();
 
                     fprintf(stdout, "\n");
                     fprintf(stdout, "%s: detected command: %s%s%s | p = %f | t = %d ms\n", __func__,
-                            "\033[1m", allowed_commands[index].c_str(), "\033[0m", prob,
+                            "\033[1m", best_command, "\033[0m", prob,
                             (int) std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count());
                     fprintf(stdout, "\n");
+                    if (fout.is_open()) {
+                        fout << best_command << std::endl;
+                    }
                 }
             }
 
@@ -462,7 +468,7 @@ static int process_command_list(struct whisper_context * ctx, audio_async &audio
 
 // always-prompt mode
 // transcribe the voice into text after valid prompt
-static int always_prompt_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
+static int always_prompt_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params, std::ofstream & fout) {
     bool is_running = true;
     bool ask_prompt = true;
 
@@ -528,6 +534,9 @@ static int always_prompt_transcription(struct whisper_context * ctx, audio_async
 
                 if ((sim > 0.7f) && (command.size() > 0)) {
                     fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
+                    if (fout.is_open()) {
+                        fout << command << std::endl;
+                    }
                 }
 
                 fprintf(stdout, "\n");
@@ -542,7 +551,7 @@ static int always_prompt_transcription(struct whisper_context * ctx, audio_async
 
 // general-purpose mode
 // freely transcribe the voice into text
-static int process_general_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
+static int process_general_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params, std::ofstream & fout) {
     bool is_running  = true;
     bool have_prompt = false;
     bool ask_prompt  = true;
@@ -662,8 +671,10 @@ static int process_general_transcription(struct whisper_context * ctx, audio_asy
                     } else {
                         // cut the prompt from the decoded text
                         const std::string command = ::trim(txt.substr(best_len));
-
                         fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
+                        if (fout.is_open()) {
+                            fout << command << std::endl;
+                        }
                     }
 
                     fprintf(stdout, "\n");
@@ -678,6 +689,8 @@ static int process_general_transcription(struct whisper_context * ctx, audio_asy
 }
 
 int main(int argc, char ** argv) {
+    ggml_backend_load_all();
+
     whisper_params params;
 
     if (whisper_params_parse(argc, argv, params) == false) {
@@ -698,6 +711,10 @@ int main(int argc, char ** argv) {
     cparams.flash_attn = params.flash_attn;
 
     struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
+    if (ctx == nullptr) {
+        fprintf(stderr, "error: failed to initialize whisper context\n");
+        return 2;
+    }
 
     // print some info about the processing
     {
@@ -757,13 +774,22 @@ int main(int argc, char ** argv) {
         }
     }
 
+    std::ofstream fout;
+    if (params.fname_out.length() > 0) {
+        fout.open(params.fname_out);
+        if (!fout.is_open()) {
+            fprintf(stderr, "%s: failed to open output file '%s'!\n", __func__, params.fname_out.c_str());
+            return 1;
+        }
+    }
+
     if (ret_val == 0) {
         if (!params.commands.empty()) {
-            ret_val = process_command_list(ctx, audio, params);
+            ret_val = process_command_list(ctx, audio, params, fout);
         } else if (!params.prompt.empty() && params.grammar_parsed.rules.empty()) {
-            ret_val = always_prompt_transcription(ctx, audio, params);
+            ret_val = always_prompt_transcription(ctx, audio, params, fout);
         } else {
-            ret_val = process_general_transcription(ctx, audio, params);
+            ret_val = process_general_transcription(ctx, audio, params, fout);
         }
     }
 

data/ext/sources/examples/command.wasm/index-tmpl.html

@@ -42,6 +42,7 @@
                 <a href="../bench.wasm/">bench</a> |
                 <a href="../stream.wasm">stream</a> |
                 <a href="../command.wasm/">command</a> |
+                <a href="../wchess.wasm/">wchess</a> |
 
             <br><br>
 
@@ -174,11 +175,11 @@
 
             function loadWhisper(model) {
                 let urls = {
-                    'tiny.en': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en.bin',
-                    'base.en': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
+                    'tiny.en': 'https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en.bin',
+                    'base.en': 'https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin',
 
-                    'tiny-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en-q5_1.bin',
-                    'base-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-base.en-q5_1.bin',
+                    'tiny-en-q5_1':  'https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en-q5_1.bin',
+                    'base-en-q5_1':  'https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en-q5_1.bin',
                 };
 
                 let sizes = {

data/ext/sources/examples/common-ggml.cpp

@@ -72,6 +72,7 @@ bool ggml_common_quantize_0(
         case GGML_FTYPE_MOSTLY_IQ4_XS:
         case GGML_FTYPE_MOSTLY_IQ1_M:
         case GGML_FTYPE_MOSTLY_BF16:
+        case GGML_FTYPE_MOSTLY_MXFP4:
                 {
                     fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
                     return false;
@@ -211,6 +212,7 @@ bool ggml_common_quantize_0(
                 case GGML_TYPE_BF16:
                 case GGML_TYPE_TQ1_0:
                 case GGML_TYPE_TQ2_0:
+                case GGML_TYPE_MXFP4:
                 case GGML_TYPE_COUNT:
                     {
                         fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));

data/ext/sources/examples/common-whisper.cpp

@@ -112,13 +112,20 @@ bool read_audio_data(const std::string & fname, std::vector<float>& pcmf32, std:
     }
 
     if (stereo) {
-		pcmf32s.resize(2);
-		pcmf32s[0].resize(frame_count);
-		pcmf32s[1].resize(frame_count);
-		for (uint64_t i = 0; i < frame_count; i++) {
-			pcmf32s[0][i] = pcmf32[2*i];
-			pcmf32s[1][i] = pcmf32[2*i + 1];
-		}
+        std::vector<float> stereo_data = pcmf32;
+        pcmf32.resize(frame_count);
+
+        for (uint64_t i = 0; i < frame_count; i++) {
+            pcmf32[i] = (stereo_data[2*i] + stereo_data[2*i + 1]);
+        }
+
+        pcmf32s.resize(2);
+        pcmf32s[0].resize(frame_count);
+        pcmf32s[1].resize(frame_count);
+        for (uint64_t i = 0; i < frame_count; i++) {
+            pcmf32s[0][i] = stereo_data[2*i];
+            pcmf32s[1][i] = stereo_data[2*i + 1];
+        }
     }
 
     ma_decoder_uninit(&decoder);

data/ext/sources/examples/lsp/lsp.cpp

@@ -31,7 +31,7 @@ struct whisper_params {
     bool print_special = false;
     bool print_energy  = false;
     bool use_gpu       = true;
-    bool flash_attn    = false;
+    bool flash_attn    = true;
 
     std::string language  = "en";
     std::string model     = "models/ggml-base.en.bin";
@@ -62,21 +62,22 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
             whisper_print_usage(argc, argv, params);
             exit(0);
         }
-        else if (arg == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
-        else if (arg == "-pms" || arg == "--prompt-ms")     { params.prompt_ms     = std::stoi(argv[++i]); }
-        else if (arg == "-cms" || arg == "--command-ms")    { params.command_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 == "-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 == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
-        else if (arg == "-fa"  || arg == "--flash-attn")    { params.flash_attn    = true; }
-        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
-        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
+        else if (arg == "-t"      || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
+        else if (arg == "-pms"    || arg == "--prompt-ms")     { params.prompt_ms     = std::stoi(argv[++i]); }
+        else if (arg == "-cms"    || arg == "--command-ms")    { params.command_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 == "-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 == "-ng"     || arg == "--no-gpu")        { params.use_gpu       = false; }
+        else if (arg == "-fa"     || arg == "--flash-attn")    { params.flash_attn    = true; }
+        else if (arg == "-nfa"    || arg == "--no-flash-attn") { params.flash_attn    = false; }
+        else if (arg == "-l"      || arg == "--language")      { params.language      = argv[++i]; }
+        else if (arg == "-m"      || arg == "--model")         { params.model         = argv[++i]; }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -105,7 +106,8 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     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, "  -ng,        --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -fa,        --flash-attn     [%-7s] flash attention\n",                             params.flash_attn ? "true" : "false");
+    fprintf(stderr, "  -fa,        --flash-attn     [%-7s] enable flash attention\n",                      params.flash_attn ? "true" : "false");
+    fprintf(stderr, "  -nfa,       --no-flash-attn  [%-7s] disable flash attention\n",                     params.flash_attn ? "false" : "true");
     fprintf(stderr, "  -l LANG,    --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
     fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
     fprintf(stderr, "\n");
@@ -424,6 +426,8 @@ static void process_loop(struct whisper_context * ctx, audio_async &audio, const
 }
 
 int main(int argc, char ** argv) {
+    ggml_backend_load_all();
+
     whisper_params params;
     if (whisper_params_parse(argc, argv, params) == false) {
         return 1;

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

@@ -1,4 +1,5 @@
 #include "ggml.h"
+#include "ggml-backend.h"
 
 #include "common.h"
 #include "common-ggml.h"
@@ -176,6 +177,8 @@ static bool whisper_model_quantize(const std::string & fname_inp, const std::str
 }
 
 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);

data/ext/sources/examples/server/CMakeLists.txt

@@ -1,3 +1,6 @@
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+
 set(TARGET whisper-server)
 add_executable(${TARGET} server.cpp httplib.h)