root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-27 03:44:35 +00:00
Code Issues Actions 4 Packages Projects Releases Wiki Activity

llama : comments

Browse Source
This commit is contained in:
Georgi Gerganov 2023-10-28 22:50:08 +03:00
parent 3af8771389
commit 51c4f9ee9f
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
1 changed files with 4 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
llama.cpp
View File

@ -5452,15 +5452,12 @@ static struct ggml_cgraph * llama_build_graph(
} while (0); } while (0);
// offload layers // offload layers
{ {
const int n_layer = model.hparams.n_layer; const int n_layer = model.hparams.n_layer;
const int n_gpu_layers = model.n_gpu_layers; const int n_gpu_layers = model.n_gpu_layers;
const int i_gpu_start = n_layer - n_gpu_layers; const int i_gpu_start = n_layer - n_gpu_layers;
GGML_UNUSED(i_gpu_start);
// offload functions set the tensor output backend to GPU // offload functions set the tensor output backend to GPU
// tensors are GPU-accelerated if any input or the output has been offloaded // tensors are GPU-accelerated if any input or the output has been offloaded
offload_func_t offload_func_nr = ggml_offload_nop; // nr = non-repeating offload_func_t offload_func_nr = ggml_offload_nop; // nr = non-repeating
@ -5588,13 +5585,16 @@ static struct ggml_cgraph * llama_build_graph(
const std::string name = cur->name; const std::string name = cur->name;
if (k_offload_func.find(name) == k_offload_func.end()) { if (k_offload_func.find(name) == k_offload_func.end()) {
// if a tensor that is not view hasn't been offloaded, we warn the user
if (worst_case && cur->view_src == nullptr) { if (worst_case && cur->view_src == nullptr) {
LLAMA_LOG_WARN("%s: node %4d %32s: not offloaded (ref: %s)\n", __func__, LLAMA_LOG_WARN("%s: node %4d %32s: not offloaded (ref: %s)\n", __func__,
i, name.c_str(), "https://github.com/ggerganov/llama.cpp/pull/3837"); i, name.c_str(), "https://github.com/ggerganov/llama.cpp/pull/3837");
} }
continue; continue;
} }
// count the number of layers and respect the provided n_gpu_layers
offload_func_t f = k_offload_func.at(name); offload_func_t f = k_offload_func.at(name);
if (f == offload_func) { if (f == offload_func) {
if (ofn[name]++ < i_gpu_start) { if (ofn[name]++ < i_gpu_start) {
@ -5602,6 +5602,7 @@ static struct ggml_cgraph * llama_build_graph(
} }
} }
// apply offload function to the tensor
f(cur); f(cur);
if (worst_case && cur->view_src == nullptr) { if (worst_case && cur->view_src == nullptr) {