root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-31 22:04:35 +00:00
Code Issues Actions 7 Packages Projects Releases Wiki Activity

llama : read new cls tensors [no ci]

Browse Source
This commit is contained in:
Georgi Gerganov 2024-09-17 16:38:38 +03:00
parent 49f90de363
commit dc0cdd8760
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
1 changed files with 16 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
src/llama.cpp
View File

@ -602,6 +602,8 @@ enum llm_tensor {
LLM_TENSOR_ENC_FFN_DOWN,
LLM_TENSOR_ENC_FFN_UP,
LLM_TENSOR_ENC_OUTPUT_NORM,
LLM_TENSOR_CLS,
LLM_TENSOR_CLS_OUT,
};
static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES = {
@ -789,6 +791,8 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
{ LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
{ LLM_TENSOR_CLS, "cls" },
{ LLM_TENSOR_CLS_OUT, "cls.output" },
},
},
{
@ -2882,6 +2886,12 @@ struct llama_model {
struct ggml_tensor * output_b;
struct ggml_tensor * output_norm_enc;
// classifier
struct ggml_tensor * cls;
struct ggml_tensor * cls_b;
struct ggml_tensor * cls_out;
struct ggml_tensor * cls_out_b;
std::vector<llama_layer> layers;
llama_split_mode split_mode;
@ -7351,6 +7361,12 @@ static bool llm_load_tensors(
if (model.arch == LLM_ARCH_BERT) {
model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, n_ctx_train});
model.cls = ml.create_tensor(ctx_output, tn(LLM_TENSOR_CLS, "weight"), {n_embd, n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
model.cls_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_CLS, "bias"), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
model.cls_out = ml.create_tensor(ctx_output, tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
model.cls_out_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_CLS_OUT, "bias"), {1}, llama_model_loader::TENSOR_NOT_REQUIRED);
}
model.tok_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd});