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

gemma : allow offloading the output tensor (#5646)

Browse source
This commit is contained in:
slaren 2024-02-21 22:18:23 +01:00 committed by GitHub
parent 89febfed93
commit ba2135ccae
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 3 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
llama.cpp
View file

@ -4394,6 +4394,8 @@ static bool llm_load_tensors(
// output
model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // same as tok_embd, duplicated to allow offloading
ml.n_created--; // artificial tensor
const int64_t n_ff = hparams.n_ff;
const int64_t n_embd_head_k = hparams.n_embd_head_k;
@ -7525,7 +7527,7 @@ struct llm_build_context {
cb(cur, "result_norm", -1);
// lm_head
cur = ggml_mul_mat(ctx0, model.tok_embd, cur);
cur = ggml_mul_mat(ctx0, model.output, cur);
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);