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

ggml : change ggml_scale to take a float instead of tensor

Browse source
This commit is contained in:
Georgi Gerganov 2023-12-21 20:50:24 +02:00
parent 8fe03ffdda
commit 199f6bdc46
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
10 changed files with 68 additions and 186 deletions
Download patch file Download diff file Expand all files Collapse all files

8
examples/llava/clip.cpp
View file

@ -330,12 +330,6 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
ggml_repeat(ctx0, model.pre_ln_b, embeddings));
}
struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
ggml_allocr_alloc(ctx->alloc, KQ_scale);
if (!ggml_allocr_is_measure(ctx->alloc)) {
ggml_set_f32(KQ_scale, 1.0f / sqrt((float)d_head));
}
// loop over layers
for (int il = 0; il < n_layer - 1; il++) {
struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states
@ -356,7 +350,7 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
struct ggml_tensor * Q =
ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].q_b, cur), ggml_mul_mat(ctx0, model.layers[il].q_w, cur));
Q = ggml_scale_inplace(ctx0, Q, KQ_scale);
Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_positions, batch_size);
Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
Q = ggml_reshape_3d(ctx0, Q, d_head, num_positions, n_head * batch_size);