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

perplexity anti-mode improvements

Browse source
This commit is contained in:
KerfuffleV2 2023-10-21 04:26:47 -06:00
parent d6b44fb3ae
commit 8a569cfee5
1 changed files with 17 additions and 15 deletions
Download patch file Download diff file Expand all files Collapse all files

32
examples/perplexity/perplexity.cpp
View file

@ -352,13 +352,13 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
std::vector<int32_t> extremes; std::vector<int32_t> extremes;
extremes.resize(n_layers); extremes.resize(n_layers);
std::fill(extremes.begin(), extremes.end(), 0); std::fill(extremes.begin(), extremes.end(), 0);
if (anti_mode) { // if (anti_mode) {
// No pointing in starting with first/last layer disabled. // // No point in starting with first/last layer disabled.
skip_types[0] = 15; // skip_types[0] = 15;
skip_types[n_layers - 1] = 15; // skip_types[n_layers - 1] = 15;
skips.push_back(0); skips.push_back(0 + n_layers); // skips.push_back(0); skips.push_back(0 + n_layers);
skips.push_back(n_layers - 1); skips.push_back(n_layers - 1 + n_layers); // skips.push_back(n_layers - 1); skips.push_back(n_layers - 1 + n_layers);
} // }
int32_t curr_best_layer = -1, curr_best_type = 0; int32_t curr_best_layer = -1, curr_best_type = 0;
double curr_best_ppl = -1, ref_ppl = -1; double curr_best_ppl = -1, ref_ppl = -1;
const int32_t mask = anti_mode ? 3 : 0; const int32_t mask = anti_mode ? 3 : 0;
@ -389,7 +389,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
} }
if (skip_layer >= n_layers) { if (skip_layer >= n_layers) {
if (curr_best_layer == -1) break; if (curr_best_layer == -1) break;
if (prune_target > 0 && pass_results.size() >= prune_target * 2) { if (anti_mode || (prune_target > 0 && pass_results.size() >= prune_target * 2)) {
std::sort(pass_results.begin(), pass_results.end(), std::sort(pass_results.begin(), pass_results.end(),
[](const std::tuple<int32_t, int32_t, double> & a, const std::tuple<int32_t, int32_t, double> & b) { [](const std::tuple<int32_t, int32_t, double> & a, const std::tuple<int32_t, int32_t, double> & b) {
if (anti_mode) return std::get<2>(b) > std::get<2>(a); if (anti_mode) return std::get<2>(b) > std::get<2>(a);
@ -399,14 +399,14 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
const size_t num_prune = std::min(pass_results.size(), prune_target); const size_t num_prune = std::min(pass_results.size(), prune_target);
for (size_t temp = 0, pruned = 0; temp < pass_results.size(); temp++) { for (size_t temp = 0, pruned = 0; temp < pass_results.size(); temp++) {
int32_t lidx = std::get<0>(pass_results[temp]); int32_t lidx = std::get<0>(pass_results[temp]);
if (lidx == curr_best_layer && std::get<1>(pass_results[temp]) == curr_best_type) continue;
extremes[lidx] |= std::get<1>(pass_results[temp]);
printf("\nPrune[%zu]: %d (%d) - %.2f\n", pruned + 1, lidx,
std::get<1>(pass_results[temp]), std::get<2>(pass_results[temp]));
if (anti_mode) { if (anti_mode) {
skip_types[lidx] |= std::get<1>(pass_results[temp]); skip_types[lidx] |= std::get<1>(pass_results[temp]);
skips.push_back(std::get<1>(pass_results[temp]) == 1 ? lidx : lidx + n_layers); skips.push_back(std::get<1>(pass_results[temp]) == 1 ? lidx : lidx + n_layers);
} }
if (lidx == curr_best_layer && std::get<1>(pass_results[temp]) == curr_best_type) continue;
extremes[lidx] |= std::get<1>(pass_results[temp]);
printf("\nPrune[%zu]: %d (%d) - %.2f\n", pruned + 1, lidx,
std::get<1>(pass_results[temp]), std::get<2>(pass_results[temp]));
if (++pruned >= num_prune) break; if (++pruned >= num_prune) break;
} }
} }
@ -414,9 +414,11 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
printf("\n\nADD %c%3d - ppl vs ref %.4f", printf("\n\nADD %c%3d - ppl vs ref %.4f",
int(label[curr_best_type]), curr_best_layer, int(label[curr_best_type]), curr_best_layer,
curr_best_ppl - ref_ppl); curr_best_ppl - ref_ppl);
if (!anti_mode && curr_best_ppl > ref_ppl * 1.75) break; if (!anti_mode) {
skip_types[curr_best_layer] += curr_best_type; if (curr_best_ppl > ref_ppl * 1.75) break;
skips.push_back(curr_best_type == 1 ? curr_best_layer : curr_best_layer + n_layers); skip_types[curr_best_layer] += curr_best_type;
skips.push_back(curr_best_type == 1 ? curr_best_layer : curr_best_layer + n_layers);
}
curr_best_layer = -1; curr_best_layer = -1;
curr_best_ppl = -1; curr_best_ppl = -1;
curr_best_type = 0; curr_best_type = 0;