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

llama : add support for control vectors (#5970)

Browse source 
* control vector api and implementation

* control-vectors : minor code style updates

* disable control vector when data == nullptr

use -1 for disabled range (also on init) in case we ever support controlling layer 0 (embeddings)

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Theia Vogel 2024-03-15 13:43:02 -07:00 • committed by GitHub
parent 12247f4c69
commit 877b4d0c62
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 392 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

31
common/common.h
View file

@ -37,10 +37,13 @@ extern char const *LLAMA_COMMIT;
extern char const *LLAMA_COMPILER;
extern char const *LLAMA_BUILD_TARGET;
struct llama_control_vector_load_info;
int32_t get_num_physical_cores();
//
// CLI argument parsing
//
int32_t get_num_physical_cores();
struct gpt_params {
uint32_t seed = LLAMA_DEFAULT_SEED; // RNG seed
@ -103,6 +106,11 @@ struct gpt_params {
std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
std::string lora_base = ""; // base model path for the lora adapter
std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
int32_t control_vector_layer_start = -1; // layer range for control vector
int32_t control_vector_layer_end = -1; // layer range for control vector
int ppl_stride = 0; // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
int ppl_output_type = 0; // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
// (which is more convenient to use for plotting)
@ -269,3 +277,24 @@ void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40
void llama_embd_normalize(const float * inp, float * out, int n);
float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n);
//
// Control vector utils
//
struct llama_control_vector_data {
int n_embd;
// stores data for layers [1, n_layer] where n_layer = data.size() / n_embd
std::vector<float> data;
};
struct llama_control_vector_load_info {
float strength;
std::string fname;
};
// Load control vectors, scale each by strength, and add them together.
// On error, returns {-1, empty}
llama_control_vector_data llama_control_vector_load(const std::vector<llama_control_vector_load_info> & load_infos);