ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-06-20 23:55:04 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

ggml : sync (ggml-alloc, GPU, eps, etc.) (#1220)

Browse Source 
* ggml : sync (ggml-alloc, GPU, eps, etc.)

* ggml : fix build

* wasm : fix build
This commit is contained in:
Georgi Gerganov
2023-09-05 13:54:40 +03:00
committed by GitHub
parent 6780c98e19
commit 59a3d0cb57
17 changed files with 11254 additions and 4095 deletions
Download Patch File Download Diff File Expand all files Collapse all files

40
whisper.cpp
View File

@ -441,6 +441,7 @@ struct whisper_hparams {
int32_t n_text_layer = 4;
int32_t n_mels = 80;
int32_t ftype = 1;
float eps = 1e-5f;
};
// audio encoding layer
@ -1578,7 +1579,7 @@ static bool whisper_encode_internal(
{
wstate.use_buf(ctx0, 0);
cur = ggml_norm(ctx0, inpL);
cur = ggml_norm(ctx0, inpL, hparams.eps);
// cur = ln_0_w*cur + ln_0_b
cur = ggml_add(ctx0,
@ -1725,7 +1726,7 @@ static bool whisper_encode_internal(
{
wstate.use_buf(ctx0, 0);
cur = ggml_norm(ctx0, inpFF);
cur = ggml_norm(ctx0, inpFF, hparams.eps);
wstate.use_buf(ctx0, 1);
@ -1788,7 +1789,7 @@ static bool whisper_encode_internal(
{
wstate.use_buf(ctx0, 0);
cur = ggml_norm(ctx0, cur);
cur = ggml_norm(ctx0, cur, hparams.eps);
wstate.use_buf(ctx0, 1);
@ -1805,10 +1806,9 @@ static bool whisper_encode_internal(
// run the computation
{
struct ggml_cgraph gf = {};
gf.n_threads = n_threads;
ggml_build_forward_expand(&gf, cur);
ggml_graph_compute(ctx0, &gf);
ggml_build_forward_expand (&gf, cur);
ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
//ggml_graph_print(&gf);
}
@ -1851,12 +1851,11 @@ static bool whisper_encode_internal(
// pre-compute cross-attention memory
{
struct ggml_cgraph gf = {};
gf.n_threads = n_threads;
// TODO: hack to disconnect the encoded features from the previous graph
cur->op = GGML_OP_NONE;
cur->src0 = nullptr;
cur->src1 = nullptr;
cur->src[0] = nullptr;
cur->src[1] = nullptr;
for (int il = 0; il < model.hparams.n_text_layer; ++il) {
auto& layer = model.layers_decoder[il];
@ -1894,7 +1893,7 @@ static bool whisper_encode_internal(
ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcross, v));
}
ggml_graph_compute(ctx0, &gf);
ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
//ggml_graph_print(&gf);
}
@ -1965,7 +1964,6 @@ static bool whisper_decode_internal(
struct ggml_context * ctx0 = ggml_init(params);
struct ggml_cgraph gf = {};
gf.n_threads = n_threads;
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
memcpy(embd->data, tokens, N*ggml_element_size(embd));
@ -1992,7 +1990,7 @@ static bool whisper_decode_internal(
{
wstate.use_buf(ctx0, 0);
cur = ggml_norm(ctx0, inpL);
cur = ggml_norm(ctx0, inpL, hparams.eps);
// cur = ln_0_w*cur + ln_0_b
cur = ggml_add(ctx0,
@ -2119,7 +2117,7 @@ static bool whisper_decode_internal(
{
wstate.use_buf(ctx0, 0);
cur = ggml_norm(ctx0, inpCA); // note: we use inpCA here
cur = ggml_norm(ctx0, inpCA, hparams.eps); // note: we use inpCA here
// cur = ln_0_w*cur + ln_0_b
cur = ggml_add(ctx0,
@ -2229,7 +2227,7 @@ static bool whisper_decode_internal(
{
wstate.use_buf(ctx0, 0);
cur = ggml_norm(ctx0, inpFF);
cur = ggml_norm(ctx0, inpFF, hparams.eps);
wstate.use_buf(ctx0, 1);
@ -2284,7 +2282,7 @@ static bool whisper_decode_internal(
{
wstate.use_buf(ctx0, 0);
cur = ggml_norm(ctx0, cur);
cur = ggml_norm(ctx0, cur, hparams.eps);
wstate.use_buf(ctx0, 1);
@ -2308,8 +2306,8 @@ static bool whisper_decode_internal(
// run the computation
{
ggml_build_forward_expand(&gf, logits);
ggml_graph_compute (ctx0, &gf);
ggml_build_forward_expand (&gf, logits);
ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
}
// extract logits for all N tokens
@ -2358,7 +2356,7 @@ static std::string to_timestamp(int64_t t, bool comma = false) {
static float sin_vals[SIN_COS_N_COUNT];
static float cos_vals[SIN_COS_N_COUNT];
// In FFT, we frequently use sine and cosine operations with the same values.
// In FFT, we frequently use sine and cosine operations with the same values.
// We can use precalculated values to speed up the process.
static void fill_sin_cos_table() {
static bool is_filled = false;
@ -5165,17 +5163,15 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) {
struct ggml_cgraph gf = ggml_build_forward(c);
gf.n_threads = n_threads;
double tsum = 0.0;
// heat-up
ggml_graph_compute(ctx0, &gf);
ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
for (int i = 0; i < n_max; ++i) {
const int64_t t0 = ggml_time_us();
ggml_graph_compute(ctx0, &gf);
ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
const int64_t t1 = ggml_time_us();