vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml : ggml_rope now takes a vector with positions instead of n_past

Browse source
This commit is contained in:
Georgi Gerganov 2023-09-17 21:12:51 +03:00
parent 3b4bab6a38
commit 1fb033fd85
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
9 changed files with 270 additions and 131 deletions
Download patch file Download diff file Expand all files Collapse all files

37
examples/baby-llama/baby-llama.cpp
View file

@ -556,6 +556,14 @@ struct ggml_tensor * forward(
struct ggml_tensor * kc = kv_self.k;
struct ggml_tensor * vc = kv_self.v;
struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
{
int * data = (int *) KQ_pos->data;
for (int i = 0; i < N; ++i) {
data[i] = n_past + i;
}
}
// inpL shape [n_embd,N,1,1]
struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
for (int il = 0; il < n_layer; ++il) {
@ -583,8 +591,8 @@ struct ggml_tensor * forward(
// wk shape [n_embd, n_embd, 1, 1]
// Qcur shape [n_embd/n_head, n_head, N, 1]
// Kcur shape [n_embd/n_head, n_head, N, 1]
struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0, 0);
struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0, 0);
// store key and value to memory
{
@ -810,9 +818,18 @@ struct ggml_tensor * forward_batch(
struct ggml_tensor * kc = kv_self.k;
struct ggml_tensor * vc = kv_self.v;
struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
{
int * data = (int *) KQ_pos->data;
for (int i = 0; i < N; ++i) {
data[i] = n_past + i;
}
}
// inpL shape [n_embd,N*n_batch,1]
struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
assert_shape_2d(inpL, n_embd, N*n_batch);
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * inpSA = inpL;
@ -840,8 +857,8 @@ struct ggml_tensor * forward_batch(
// wk shape [n_embd, n_embd, 1, 1]
// Qcur shape [n_embd/n_head, n_head, N, n_batch]
// Kcur shape [n_embd/n_head, n_head, N, n_batch]
struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0, 0);
struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0, 0);
assert_shape_4d(Qcur, n_embd/n_head, n_head, N, n_batch);
assert_shape_4d(Kcur, n_embd/n_head, n_head, N, n_batch);
@ -1100,6 +1117,14 @@ struct ggml_tensor * forward_lora(
struct ggml_tensor * kc = kv_self.k;
struct ggml_tensor * vc = kv_self.v;
struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
{
int * data = (int *) KQ_pos->data;
for (int i = 0; i < N; ++i) {
data[i] = n_past + i;
}
}
// inpL shape [n_embd,N,1,1]
struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
for (int il = 0; il < n_layer; ++il) {
@ -1133,7 +1158,7 @@ struct ggml_tensor * forward_lora(
model->layers[il].wqb,
cur)),
n_embd/n_head, n_head, N),
n_past, n_rot, 0, 0);
KQ_pos, n_rot, 0, 0);
struct ggml_tensor * Kcur = ggml_rope(ctx0,
ggml_reshape_3d(ctx0,
ggml_mul_mat(ctx0,
@ -1142,7 +1167,7 @@ struct ggml_tensor * forward_lora(
model->layers[il].wkb,
cur)),
n_embd/n_head, n_head, N),
n_past, n_rot, 0, 0);
KQ_pos, n_rot, 0, 0);
// store key and value to memory
{