CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-30 08:42:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama : MiniCPM support tied embeddings (#7664)

Browse Source 
* support lm_head

* remove the code block

---------

Co-authored-by: zhangkaihuo <zhangkaihuo@modelbest.cn>
This commit is contained in:
zhangkaihuo
2024-06-03 15:49:30 +08:00
committed by GitHub
parent 549279d804
commit 6f28a333c1
2 changed files with 6 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
gguf-py/gguf/constants.py
View File

@@ -645,6 +645,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
], ],
MODEL_ARCH.MINICPM: [ MODEL_ARCH.MINICPM: [
MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.OUTPUT_NORM, MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_NORM, MODEL_TENSOR.ATTN_NORM,

4
llama.cpp
View File

@@ -5124,14 +5124,12 @@ static bool llm_load_tensors(
// output // output
{ {
model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
if (model.arch != LLM_ARCH_MINICPM){
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED); model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
// if output is NULL, init from the input tok embed // if output is NULL, init from the input tok embed
if (model.output == NULL) { if (model.output == NULL) {
model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
} }
} }
}
for (int i = 0; i < n_layer; ++i) { for (int i = 0; i < n_layer; ++i) {
ggml_context * ctx_layer = ctx_for_layer(i); ggml_context * ctx_layer = ctx_for_layer(i);
@@ -10212,7 +10210,7 @@ struct llm_build_context {
cb(cur, "lmhead_scaling", -1); cb(cur, "lmhead_scaling", -1);
// lm_head // lm_head
cur = ggml_mul_mat(ctx0, model.tok_embd, cur); cur = ggml_mul_mat(ctx0, model.output, cur);
cb(cur, "result_output", -1); cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur); ggml_build_forward_expand(gf, cur);