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

Rebase

Browse source
This commit is contained in:
Nam Nguyen 2023-12-29 08:48:02 -08:00
parent e96fad12c5
commit 6200da58fc
1 changed files with 9 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

10
llama.cpp
View file

@ -3812,7 +3812,10 @@ static bool llm_load_tensors(
model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output);
} }
const uint32_t n_ff = hparams.n_ff; const uint32_t n_ff = hparams.n_ff;
const int64_t n_embd_gqa = n_value_gqa;
GGML_ASSERT(n_embd_gqa == n_embd);
GGML_ASSERT(n_embd_gqa == n_key_gqa);
const int i_gpu_start = n_layer - n_gpu_layers; const int i_gpu_start = n_layer - n_gpu_layers;
@ -5946,6 +5949,11 @@ struct llm_build_context {
struct ggml_cgraph * build_gpt2() { struct ggml_cgraph * build_gpt2() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
const int64_t n_embd_head = hparams.n_value_dim;
const int64_t n_embd_gqa = hparams.n_value_gqa();
GGML_ASSERT(n_embd_head == hparams.n_key_dim);
GGML_ASSERT(n_embd_gqa == n_embd);
struct ggml_tensor * cur; struct ggml_tensor * cur;
struct ggml_tensor * pos; struct ggml_tensor * pos;
struct ggml_tensor * inpL; struct ggml_tensor * inpL;