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
+4 -2
View File
@@ -36,7 +36,7 @@ struct whisper_params {
bool tinydiarize = false;
bool save_audio = false; // save audio to wav file
bool use_gpu = true;
bool flash_attn = false;
bool flash_attn = true;
std::string language = "en";
std::string model = "models/ggml-base.en.bin";
@@ -74,6 +74,7 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
else if (arg == "-sa" || arg == "--save-audio") { params.save_audio = 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 {
fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
@@ -111,7 +112,8 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
fprintf(stderr, " -tdrz, --tinydiarize [%-7s] enable tinydiarize (requires a tdrz model)\n", params.tinydiarize ? "true" : "false");
fprintf(stderr, " -sa, --save-audio [%-7s] save the recorded audio to a file\n", params.save_audio ? "true" : "false");
fprintf(stderr, " -ng, --no-gpu [%-7s] disable GPU inference\n", params.use_gpu ? "false" : "true");
fprintf(stderr, " -fa, --flash-attn [%-7s] flash attention during inference\n", params.flash_attn ? "true" : "false");
fprintf(stderr, " -fa, --flash-attn [%-7s] enable flash attention during inference\n", params.flash_attn ? "true" : "false");
fprintf(stderr, " -nfa, --no-flash-attn [%-7s] disable flash attention during inference\n", params.flash_attn ? "false" : "true");
fprintf(stderr, "\n");
}