github/gpt4all
1
0
Fork 0
mirror of https://github.com/nomic-ai/gpt4all.git synced 2025-09-05 10:30:29 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix VRAM leak when model loading fails (#1901)

Browse Source 
Signed-off-by: Jared Van Bortel <jared@nomic.ai>
This commit is contained in:
Jared Van Bortel
2024-02-01 15:45:45 -05:00
committed by GitHub
parent e1eac00ee0
commit 10e3f7bbf5
3 changed files with 16 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
gpt4all-backend/llama.cpp-mainline

Submodule gpt4all-backend/llama.cpp-mainline updated: 47aec1bcc0...cd1b5a104b

16
gpt4all-backend/llamamodel.cpp
View File

@@ -150,7 +150,15 @@ size_t LLamaModel::requiredMem(const std::string &modelPath, int n_ctx, int ngl)
bool LLamaModel::loadModel(const std::string &modelPath, int n_ctx, int ngl)
{
gpt_params params;
// clean up after previous loadModel()
if (d_ptr->model) {
llama_free_model(d_ptr->model);
d_ptr->model = nullptr;
}
if (d_ptr->ctx) {
llama_free(d_ptr->ctx);
d_ptr->ctx = nullptr;
}
if (n_ctx < 8) {
std::cerr << "warning: minimum context size is 8, using minimum size.\n";
@@ -159,6 +167,8 @@ bool LLamaModel::loadModel(const std::string &modelPath, int n_ctx, int ngl)
// -- load the model --
gpt_params params;
d_ptr->model_params = llama_model_default_params();
d_ptr->model_params.use_mmap = params.use_mmap;
@@ -215,8 +225,10 @@ bool LLamaModel::loadModel(const std::string &modelPath, int n_ctx, int ngl)
d_ptr->ctx = llama_new_context_with_model(d_ptr->model, d_ptr->ctx_params);
if (!d_ptr->ctx) {
d_ptr->device = -1;
std::cerr << "LLAMA ERROR: failed to init context for model " << modelPath << std::endl;
llama_free_model(d_ptr->model);
d_ptr->model = nullptr;
d_ptr->device = -1;
return false;
}