whisper : switch back to F32 mask (#0)
This commit is contained in:
parent
e93081f83f
commit
2b434c449e
@ -2294,8 +2294,6 @@ static struct ggml_cgraph * whisper_build_graph_decoder(
|
||||
ggml_set_name(KQ_mask, "KQ_mask");
|
||||
ggml_set_input(KQ_mask);
|
||||
|
||||
struct ggml_tensor * KQ_mask_f16 = ggml_cast(ctx0, KQ_mask, GGML_TYPE_F16);
|
||||
|
||||
// token encoding + position encoding
|
||||
struct ggml_tensor * cur =
|
||||
ggml_add(ctx0,
|
||||
@ -2379,7 +2377,7 @@ static struct ggml_cgraph * whisper_build_graph_decoder(
|
||||
// K * Q
|
||||
struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
|
||||
|
||||
struct ggml_tensor * KQ_soft_max = ggml_soft_max_ext(ctx0, KQ, KQ_mask_f16, 1.0f, 0.0f);
|
||||
struct ggml_tensor * KQ_soft_max = ggml_soft_max_ext(ctx0, KQ, KQ_mask, 1.0f, 0.0f);
|
||||
|
||||
struct ggml_tensor * V =
|
||||
ggml_view_3d(ctx0, kv_self.v,
|
||||
|
Loading…
Reference in New Issue
Block a user