root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 10:54:36 +00:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

llama : add flash attention (demo)

Browse Source
This commit is contained in:
Georgi Gerganov 2023-04-05 18:28:01 +03:00
parent 986b6ce9f9
commit 36ddd12924
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
1 changed files with 27 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
llama.cpp
View File

@ -28,6 +28,8 @@
#define LLAMA_USE_SCRATCH
#define LLAMA_MAX_SCRATCH_BUFFERS 16
#define LLAMA_USE_FLASH_ATTN
#define LLAMA_ASSERT(x) \
do { \
if (!(x)) { \
@ -829,6 +831,30 @@ static bool llama_eval_internal(
ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
}
#ifdef LLAMA_USE_FLASH_ATTN
struct ggml_tensor * Q =
ggml_permute(ctx0,
ggml_cpy(ctx0,
Qcur,
ggml_new_tensor_3d(ctx0, GGML_TYPE_F16, n_embd/n_head, n_head, N)),
0, 2, 1, 3);
struct ggml_tensor * K =
ggml_permute(ctx0,
ggml_reshape_3d(ctx0,
ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.k)*n_embd),
n_embd/n_head, n_head, n_past + N),
0, 2, 1, 3);
struct ggml_tensor * V =
ggml_view_3d(ctx0, kv_self.v,
n_past + N, n_embd/n_head, n_head,
n_ctx*ggml_element_size(kv_self.v),
n_ctx*ggml_element_size(kv_self.v)*n_embd/n_head,
il*n_ctx*ggml_element_size(kv_self.v)*n_embd);
struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, true);
#else
struct ggml_tensor * Q =
ggml_permute(ctx0,
Qcur,
@ -872,6 +898,7 @@ static bool llama_eval_internal(
// is there a better way?
struct ggml_tensor * V_cont = ggml_cpy(ctx0, V, ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_embd/n_head, n_head));
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_cont, KQ_soft_max);
#endif
#endif
// KQV_merged = KQV.permute(0, 2, 1, 3)