whisper : enable flash attention by default (#3441)

This commit is contained in:
Georgi Gerganov
2025-09-30 15:47:20 +03:00
committed by GitHub
parent 1e5ad50f8f
commit 0b3587acdd
10 changed files with 103 additions and 87 deletions

View File

@@ -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());