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

ggml : remove ggml_flash_attn and ggml_flash_ff (#7463)

Browse source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2024-05-23 10:00:44 +03:00 • committed by GitHub
parent e84b71c2c6
commit d48c88cbd5
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
5 changed files with 47 additions and 769 deletions
Download patch file Download diff file Expand all files Collapse all files

3
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -341,7 +341,8 @@ static struct ggml_tensor * llama_build_train_graphs(
struct ggml_tensor * t15 = ggml_permute (ctx, t12, 0, 3, 1, 2); set_name(t15, "t15"); assert_shape_4d(t15, N, n_embd/n_head, n_head, n_batch);
struct ggml_tensor * t16;
if (enable_flash_attn) {
t16 = ggml_flash_attn(ctx, t13, t14, t15, true); set_name(t16, "t16"); assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
GGML_ASSERT(false && "TODO: ggml_flash_attn_ext() not yet supported");
//t16 = ggml_flash_attn(ctx, t13, t14, t15, true); set_name(t16, "t16"); assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
} else {
struct ggml_tensor * t16_0 = ggml_mul_mat (ctx, t14, t13); set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch);
struct ggml_tensor * t16_1 = ggml_scale_inplace (ctx, t16_0, kv_scale); set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch);