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

Steer with inpSA instead of with inpL

Browse Source 
Signed-off-by: Henri Vasserman <henv@hot.ee>
This commit is contained in:
Laura 2023-05-18 23:47:10 +02:00 committed by Henri Vasserman
parent 1b0ff2cf6a
commit 7f59af52a9
No known key found for this signature in database
GPG Key ID: 2995FC0F58B1A986
2 changed files with 15 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

25
examples/main/main.cpp
View File

@ -176,28 +176,27 @@ int main(int argc, char ** argv) {
if (!params.steering_add.empty() || !params.steering_sub.empty()) if (!params.steering_add.empty() || !params.steering_sub.empty())
{ {
params.steering_add.insert(0, 1, ' ');
params.steering_sub.insert(0, 1, ' ');
auto add_tokens = ::llama_tokenize(ctx, params.steering_add, true); auto add_tokens = ::llama_tokenize(ctx, params.steering_add, true);
auto sub_tokens = ::llama_tokenize(ctx, params.steering_sub, true); auto sub_tokens = ::llama_tokenize(ctx, params.steering_sub, true);
//if (add_tokens.size() != sub_tokens.size()) {
// while (add_tokens.size() < sub_tokens.size()) { if (add_tokens.size() != sub_tokens.size()) {
// add_tokens.push_back(llama_token_nl()); while (add_tokens.size() < sub_tokens.size()) {
// } add_tokens.push_back(llama_token_nl());
// while (sub_tokens.size() < add_tokens.size()) { }
// sub_tokens.push_back(llama_token_nl()); while (sub_tokens.size() < add_tokens.size()) {
// } sub_tokens.push_back(llama_token_nl());
//} }
//const int N = embd_inp.size(); }
llama_set_steering_write(ctx, params.steering_source, +1.0f); llama_set_steering_write(ctx, params.steering_source, +1.0f);
llama_eval(ctx, add_tokens.data(), std::min((int)add_tokens.size(), n_ctx), 0, params.n_threads); llama_eval(ctx, add_tokens.data(), std::min((int)add_tokens.size(), n_ctx), 0, params.n_threads);
llama_set_steering_write(ctx, params.steering_layer, -1.0f); llama_set_steering_write(ctx, params.steering_source, -1.0f);
llama_eval(ctx, sub_tokens.data(), std::min((int)sub_tokens.size(), n_ctx), 0, params.n_threads); llama_eval(ctx, sub_tokens.data(), std::min((int)sub_tokens.size(), n_ctx), 0, params.n_threads);
llama_set_steering_read(ctx, params.steering_layer, params.steering_mul); llama_set_steering_read(ctx, params.steering_layer, params.steering_mul);
std::cout << "Steering: `" << params.steering_add << "` - `" << params.steering_sub << "` * " << params.steering_mul << "\n";
} }
// debug message about similarity of saved session, if applicable // debug message about similarity of saved session, if applicable

5
llama.cpp
View File

@ -32,6 +32,7 @@
#include <mutex> #include <mutex>
#include <sstream> #include <sstream>
#include <numeric> #include <numeric>
#include <iostream>
#define LLAMA_USE_SCRATCH #define LLAMA_USE_SCRATCH
#define LLAMA_MAX_SCRATCH_BUFFERS 16 #define LLAMA_MAX_SCRATCH_BUFFERS 16
@ -1187,8 +1188,8 @@ static bool llama_eval_internal(
ggml_add(ctx0, ggml_scale(ctx0, inpL, scal), steer), steer)); ggml_add(ctx0, ggml_scale(ctx0, inpL, scal), steer), steer));
break; break;
} }
// std::cout << "\nAdding steering vector to inpL " << il << "\n";
inpL = ggml_add(ctx0, ggml_scale(ctx0, steer, scal), inpL); inpSA = ggml_add(ctx0, ggml_scale(ctx0, steer, scal), inpSA);
} }
// norm // norm