public_git
/
whisper.cpp
mirror of https://github.com/ggml-org/whisper.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge 640443521f into 1fe009caed

This commit is contained in:
Miguel Cruz 2026-08-14 15:38:04 -04:00 committed by GitHub
parent 1fe009caed 640443521f
commit 6de084d7c5
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 11 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
src/whisper.cpp
View File

@ -2323,7 +2323,11 @@ static struct ggml_cgraph * whisper_build_graph_cross(
struct ggml_tensor * k;
struct ggml_tensor * v;
if (wctx.params.flash_attn) {
// Use non-flash layout for cross-attention KV when DTW is enabled,
// since DTW needs the explicit cross-attention weights (KQ_soft_max)
const bool flash_cross = wctx.params.flash_attn && !wctx.params.dtw_token_timestamps;
if (flash_cross) {
k = ggml_view_1d(ctx0, wstate.kv_cross.k, n_state*n_ctx,
(ggml_element_size(wstate.kv_cross.k)*n_state)*(il*n_ctx_pad));
@ -2683,7 +2687,11 @@ static struct ggml_cgraph * whisper_build_graph_decoder(
ggml_reshape_3d(ctx0, Qcur, n_state_head, n_head, n_tokens),
0, 2, 1, 3);
if (wctx.params.flash_attn) {
// Use non-flash path for cross-attention when DTW is enabled,
// since DTW needs the explicit cross-attention weights (KQ_soft_max)
const bool flash_cross = wctx.params.flash_attn && !wctx.params.dtw_token_timestamps;
if (flash_cross) {
struct ggml_tensor * Kcross =
ggml_view_3d(ctx0, wstate.kv_cross.k,
n_state_head, n_audio_ctx_pad, n_head,
@ -3716,8 +3724,7 @@ struct whisper_context * whisper_init_with_params_no_state(struct whisper_model_
ggml_time_init();
if (params.flash_attn && params.dtw_token_timestamps) {
WHISPER_LOG_WARN("%s: dtw_token_timestamps is not supported with flash_attn - disabling\n", __func__);
params.dtw_token_timestamps = false;
WHISPER_LOG_INFO("%s: flash_attn with dtw - disabling flash attention for cross-attention only\n", __func__);
}
WHISPER_LOG_INFO("%s: use gpu = %d\n", __func__, params.use_gpu);