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

hopefully this fixes the dotnet nonsense

Browse source
This commit is contained in:
Concedo 2023-11-03 11:23:56 +08:00
parent bc2027b008
commit 8c14c81b33
3 changed files with 6 additions and 6 deletions
Download patch file Download diff file Expand all files Collapse all files

4
otherarch/gptj_v3.cpp
View file

@ -494,8 +494,8 @@ bool gptj_eval(
} }
} }
struct ggml_tensor *Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur), n_embd / n_head, n_head, N), KQ_pos, n_rot, 0, n_ctx, 0, freq_base, freq_scale, NAN, 1, 32, 1); struct ggml_tensor *Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur), n_embd / n_head, n_head, N), KQ_pos, n_rot, 0, n_ctx, 0, freq_base, freq_scale, 0, 1, 32, 1);
struct ggml_tensor *Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur), n_embd / n_head, n_head, N), KQ_pos, n_rot, 0, n_ctx, 0, freq_base, freq_scale, NAN, 1, 32, 1); struct ggml_tensor *Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur), n_embd / n_head, n_head, N), KQ_pos, n_rot, 0, n_ctx, 0, freq_base, freq_scale, 0, 1, 32, 1);
// store key and value to memory // store key and value to memory
{ {

4
otherarch/llama_v3.cpp
View file

@ -1614,11 +1614,11 @@ static struct ggml_cgraph * llama_v3_build_graph(
} }
#endif #endif
struct ggml_tensor *Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, N), KQ_pos, n_embd_head, 0, 0, 0, freq_base, freq_scale, NAN, 1, 32, 1); struct ggml_tensor *Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, N), KQ_pos, n_embd_head, 0, 0, 0, freq_base, freq_scale, 0, 1, 32, 1);
offload_func_kq(Kcur); offload_func_kq(Kcur);
ggml_set_name(Kcur, "Kcur"); ggml_set_name(Kcur, "Kcur");
struct ggml_tensor *Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, N), KQ_pos, n_embd_head, 0, 0, 0, freq_base, freq_scale, NAN, 1, 32, 1); struct ggml_tensor *Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, N), KQ_pos, n_embd_head, 0, 0, 0, freq_base, freq_scale, 0, 1, 32, 1);
offload_func_kq(Qcur); offload_func_kq(Qcur);
ggml_set_name(Qcur, "Qcur"); ggml_set_name(Qcur, "Qcur");

4
otherarch/neox_v3.cpp
View file

@ -522,8 +522,8 @@ bool gpt_neox_eval(
} }
// using mode = 2 for GPT-NeoX mode // using mode = 2 for GPT-NeoX mode
Qcur = ggml_rope_custom_inplace(ctx0, Qcur, KQ_pos, n_rot, 2, n_ctx, 0, freq_base, freq_scale, NAN, 1, 32, 1); Qcur = ggml_rope_custom_inplace(ctx0, Qcur, KQ_pos, n_rot, 2, n_ctx, 0, freq_base, freq_scale, 0, 1, 32, 1);
Kcur = ggml_rope_custom_inplace(ctx0, Kcur, KQ_pos, n_rot, 2, n_ctx, 0, freq_base, freq_scale, NAN, 1, 32, 1); Kcur = ggml_rope_custom_inplace(ctx0, Kcur, KQ_pos, n_rot, 2, n_ctx, 0, freq_base, freq_scale, 0, 1, 32, 1);
// store key and value to memory // store key and value to memory
{ {