ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-06-13 04:28:07 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

whisper : use flash attention (#2152)

Browse Source 
* whisper : use flash attention in the encoder

* whisper : add kv_pad

* whisper : remove extra backend instance (huh?)

* whisper : use FA for cross-attention

* whisper : use FA for self-attention

* whisper : simplify encoder FA

* whisper : add flash_attn runtime parameter

* scripts : add bench log

* scripts : add M1 Pro bench log
This commit is contained in:
Georgi Gerganov
2024-05-15 09:38:19 +03:00
committed by GitHub
parent 9d5771ae43
commit 7094ea5e75
13 changed files with 657 additions and 172 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
examples/talk-llama/talk-llama.cpp
View File

@ -66,6 +66,7 @@ struct whisper_params {
bool no_timestamps = true;
bool verbose_prompt = false;
bool use_gpu = true;
bool flash_attn = false;
std::string person = "Georgi";
std::string bot_name = "LLaMA";
@ -105,6 +106,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
else if (arg == "-pe" || arg == "--print-energy") { params.print_energy = true; }
else if (arg == "-vp" || arg == "--verbose-prompt") { params.verbose_prompt = 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 == "-p" || arg == "--person") { params.person = argv[++i]; }
else if (arg == "-bn" || arg == "--bot-name") { params.bot_name = argv[++i]; }
else if (arg == "--session") { params.path_session = argv[++i]; }
@ -123,7 +125,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
}
}
else if (arg == "-f" || arg == "--file") { params.fname_out = argv[++i]; }
else if (arg == "-ng" || arg == "--no-gpu") { params.use_gpu = false; }
else {
fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
whisper_print_usage(argc, argv, params);
@ -154,6 +155,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
fprintf(stderr, " -pe, --print-energy [%-7s] print sound energy (for debugging)\n", params.print_energy ? "true" : "false");
fprintf(stderr, " -vp, --verbose-prompt [%-7s] print prompt at start\n", params.verbose_prompt ? "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, " -p NAME, --person NAME [%-7s] person name (for prompt selection)\n", params.person.c_str());
fprintf(stderr, " -bn NAME, --bot-name NAME [%-7s] bot name (to display)\n", params.bot_name.c_str());
fprintf(stderr, " -w TEXT, --wake-command T [%-7s] wake-up command to listen for\n", params.wake_cmd.c_str());
@ -285,7 +287,9 @@ int main(int argc, char ** argv) {
// whisper init
struct whisper_context_params cparams = whisper_context_default_params();
cparams.use_gpu = params.use_gpu;
cparams.use_gpu = params.use_gpu;
cparams.flash_attn = params.flash_attn;
struct whisper_context * ctx_wsp = whisper_init_from_file_with_params(params.model_wsp.c_str(), cparams);
if (!ctx_wsp) {
@ -316,6 +320,7 @@ int main(int argc, char ** argv) {
lcparams.n_ctx = 2048;
lcparams.seed = 1;
lcparams.n_threads = params.n_threads;
lcparams.flash_attn = params.flash_attn;
struct llama_context * ctx_llama = llama_new_context_with_model(model_llama, lcparams);