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

Went back to the original size calculation for now.

Browse source 
Though it appears not to matter.
This commit is contained in:
John 2023-06-16 20:20:30 +02:00
parent 3778836046
commit f143d0e6bf
1 changed files with 10 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

20
libfalcon.cpp
View file

@ -1402,9 +1402,9 @@ static bool falcon_eval_internal(
struct ggml_tensor * layernorm_output; struct ggml_tensor * layernorm_output;
// ggml_type wtype = GGML_TYPE_F32; ggml_type wtype = GGML_TYPE_F32;
// ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype)); // ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
// const int sizeof_wtype = ggml_type_sizef(wtype); const int sizeof_wtype = ggml_type_sizef(wtype);
const int i_gpu_start = n_layer - n_gpu_layers; const int i_gpu_start = n_layer - n_gpu_layers;
(void) i_gpu_start; (void) i_gpu_start;
@ -1462,23 +1462,23 @@ static bool falcon_eval_internal(
struct ggml_tensor * Qcur = ggml_view_3d( struct ggml_tensor * Qcur = ggml_view_3d(
ctx0, cur, head_dim, n_head, N, ctx0, cur, head_dim, n_head, N,
head_dim * ggml_element_size(cur), head_dim * sizeof_wtype,
head_dim * (n_head + 2 * n_head_kv) * ggml_element_size(cur), head_dim * (n_head + 2 * n_head_kv) * sizeof_wtype,
0); 0);
ggml_set_name(Qcur, "Qcur"); ggml_set_name(Qcur, "Qcur");
struct ggml_tensor * Kcur = ggml_view_3d( struct ggml_tensor * Kcur = ggml_view_3d(
ctx0, cur, head_dim, n_head_kv, N, ctx0, cur, head_dim, n_head_kv, N,
head_dim * ggml_element_size(cur), head_dim * sizeof_wtype,
head_dim * (n_head + 2 * n_head_kv) * ggml_element_size(cur), head_dim * (n_head + 2 * n_head_kv) * sizeof_wtype,
head_dim * n_head * ggml_element_size(cur)); head_dim * n_head * sizeof_wtype);
ggml_set_name(Kcur, "Kcur"); ggml_set_name(Kcur, "Kcur");
struct ggml_tensor * Vcur = ggml_view_3d( struct ggml_tensor * Vcur = ggml_view_3d(
ctx0, cur, head_dim, n_head_kv, N, ctx0, cur, head_dim, n_head_kv, N,
head_dim * ggml_element_size(cur), head_dim * sizeof_wtype,
head_dim * (n_head + 2 * n_head_kv) * ggml_element_size(cur), head_dim * (n_head + 2 * n_head_kv) * sizeof_wtype,
head_dim * (n_head + n_head_kv) * ggml_element_size(cur)); head_dim * (n_head + n_head_kv) * sizeof_wtype);
ggml_set_name(Vcur, "Vcur"); ggml_set_name(Vcur, "Vcur");
// using mode = 2 for neox mode // using mode = 2 for neox mode