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

Fixes suggested by @mmnga

Browse source
This commit is contained in:
Galunid 2023-10-12 23:56:24 +02:00
parent 80b2e729fa
commit 605e701cb4
1 changed files with 14 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

18
llama.cpp
View file

@ -2145,7 +2145,7 @@ static void llm_load_hparams(
GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS));
switch (hparams.n_layer) {
case 32: model.type = e_model::MODEL_3B; fprintf(stdout, "Selected 3b model!"); break;
case 32: model.type = e_model::MODEL_3B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
@ -2191,7 +2191,6 @@ static void llm_load_vocab(
std::string tokenizer_name;
GGUF_GET_KEY(ctx, tokenizer_name, gguf_get_val_str, GGUF_TYPE_STRING, true, kv(LLM_KV_TOKENIZER_MODEL));
if (tokenizer_name == "llama") {
vocab.type = LLAMA_VOCAB_TYPE_SPM;
@ -5962,6 +5961,11 @@ static struct ggml_cgraph * llm_build_stablelm(
cur = ggml_mul(ctx0, cur, model.layers[il].attn_norm);
offload_func(cur);
ggml_set_name(cur, "attention_norm_0");
// add bias
cur = ggml_add(ctx0, cur, model.layers[il].attn_norm_b);
offload_func(cur);
ggml_set_name(cur, "attention_norm_0");
}
// self-attention
@ -5975,11 +5979,11 @@ static struct ggml_cgraph * llm_build_stablelm(
offload_func_kq(tmpq);
ggml_set_name(tmpq, "tmpq");
struct ggml_tensor * Kcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), KQ_pos, n_embd_head, 0, 0, freq_base, freq_scale);
struct ggml_tensor * Kcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), KQ_pos, n_embd_head, 2, 0, freq_base, freq_scale);
offload_func_kq(Kcur);
ggml_set_name(Kcur, "Kcur");
struct ggml_tensor * Qcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens), KQ_pos, n_embd_head, 0, 0, freq_base, freq_scale);
struct ggml_tensor * Qcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens), KQ_pos, n_embd_head, 2, 0, freq_base, freq_scale);
offload_func_kq(Qcur);
ggml_set_name(Qcur, "Qcur");
@ -6100,6 +6104,12 @@ static struct ggml_cgraph * llm_build_stablelm(
cur = ggml_mul(ctx0, cur, model.layers[il].ffn_norm);
offload_func(cur);
ggml_set_name(cur, "ffn_norm");
// add bias
// cur = cur*ffn_norm(broadcasted)
cur = ggml_add(ctx0, cur, model.layers[il].ffn_norm_b);
offload_func(cur);
ggml_set_name(cur, "ffn_norm");
}
struct ggml_tensor * tmp = ggml_mul_mat(ctx0,