root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-26 11:24:35 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

add TBD

Browse Source
This commit is contained in:
Meng Zhang 2023-09-15 15:23:50 +08:00
parent 8bc76a225d
commit 101c578715
1 changed files with 5 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
llama.cpp
View File

@ -3570,24 +3570,16 @@ static struct ggml_cgraph * llm_build_starcoder(
offload_func(attn_norm->src[0]);
offload_func(attn_norm);
if (model.layers[il].attn_norm_2) { // Falcon-40B
cur = ggml_norm(ctx0, inpL, norm_eps);
offload_func(cur);
cur = ggml_add(ctx0,
ggml_mul(ctx0, cur, model.layers[il].attn_norm_2),
model.layers[il].attn_norm_2_b);
offload_func(cur->src[0]);
offload_func(cur);
} else { // Falcon 7B
cur = attn_norm;
}
cur = attn_norm;
// compute QKV
cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
offload_func_kq(cur);
// ===== TBD (QKV Split + FF) ====
#define PRINT_SHAPE(x) fprintf(stderr, "%d %s: (%s)\n", __LINE__, #x, llama_format_tensor_shape(x).c_str())
GGML_ASSERT(false);
// Note that the strides for Kcur, Vcur are set up so that the
// resulting views are misaligned with the tensor's storage
// (by applying the K/V offset we shift the tensor's original