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

@@ -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; }
@@ -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());