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

Maybe adding a memory leak? But it werks now.

Browse source
This commit is contained in:
trollkotze 2024-03-25 04:33:44 +01:00
parent 7dbed974dc
commit 92070cab2a
1 changed files with 7 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

10
llama.cpp
View file

@ -1950,6 +1950,7 @@ struct llama_control_vector {
}
~llama_control_vector() {
LLAMA_LOG_ERROR("Kill the control vector\n");
for (struct ggml_context * ctx : ctxs) {
ggml_free(ctx);
}
@ -13994,9 +13995,9 @@ int32_t llama_model_apply_lora_from_file(const struct llama_model * model, const
}
static bool llama_control_vector_init(struct llama_control_vector & cvec, const llama_model & model) {
GGML_ASSERT(cvec.tensors.empty());
GGML_ASSERT(cvec.ctxs.empty());
GGML_ASSERT(cvec.bufs.empty());
cvec.tensors.clear();
cvec.ctxs.clear();
cvec.bufs.clear();
// count layer buffer types
std::map<ggml_backend_buffer_type_t, int> buft_layer_count;
@ -14062,11 +14063,10 @@ int32_t llama_control_vector_apply(struct llama_context * lctx, const float * da
return 1;
}
if (cvec.tensors.empty()) {
if (!llama_control_vector_init(cvec, model)) {
LLAMA_LOG_ERROR("%s: FUCKING BITCH\n", __func__);
return 1;
}
}
cvec.layer_start = il_start;
cvec.layer_end = il_end;