vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 01:24:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Remove Unnecessary Rep Counting (#1394)

Browse source 
* stop counting reps

* fix range-based initializer

* strike that - reverse it
This commit is contained in:
Reithan 2025-02-28 18:10:35 -08:00 • committed by GitHub
parent c088355d01
commit e85c0e6901
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 8 additions and 20 deletions
Download patch file Download diff file Expand all files Collapse all files

26
gpttype_adapter.cpp
View file

@ -11,6 +11,7 @@
#include <time.h> #include <time.h>
#include <mutex> #include <mutex>
#include <unordered_map> #include <unordered_map>
#include <unordered_set>
#include "model_adapter.h" #include "model_adapter.h"
#include "otherarch.h" #include "otherarch.h"
#include "llama.h" #include "llama.h"
@ -1188,18 +1189,8 @@ void sample_rep_pen(int n_ctx, int rep_pen_range, float rep_pen, float rep_pen_s
const int64_t t_start_sample_us = ggml_time_us(); const int64_t t_start_sample_us = ggml_time_us();
// Create a frequency map to count occurrences of each token in last_tokens // Create a frequency map to count occurrences of each token in last_tokens
std::unordered_map<llama_token, int> token_count_near; std::unordered_set<llama_token> tokens_near(last_tokens + last_n_repeat / 2, last_tokens + last_n_repeat);
std::unordered_map<llama_token, int> token_count_far; std::unordered_set<llama_token> tokens_far(last_tokens, last_tokens + last_n_repeat / 2);
for (size_t i = 0; i < last_n_repeat; ++i) {
if((i*2) >= last_n_repeat)
{
token_count_near[last_tokens[i]]++;
}
else
{
token_count_far[last_tokens[i]]++;
}
}
float rep_pen_reduced = rep_pen; float rep_pen_reduced = rep_pen;
if(rep_pen_reduced>1.0f) if(rep_pen_reduced>1.0f)
@ -1207,15 +1198,13 @@ void sample_rep_pen(int n_ctx, int rep_pen_range, float rep_pen, float rep_pen_s
rep_pen_reduced = 1.0f + ((rep_pen-1.0f)*rep_pen_slope); rep_pen_reduced = 1.0f + ((rep_pen-1.0f)*rep_pen_slope);
} }
for (size_t i = 0; i < candidates->size; ++i) { for (size_t i = 0; i < candidates->size; ++i) {
const auto token_in_near = token_count_near.find(candidates->data[i].id); const bool token_in_near = tokens_near.find(candidates->data[i].id) != tokens_near.end();
const auto token_in_far = token_count_far.find(candidates->data[i].id); const bool token_in_far = tokens_far.find(candidates->data[i].id) != tokens_far.end();
bool in_near = (token_in_near != token_count_near.end()); if (!token_in_near && !token_in_far) {
bool in_far = (token_in_far != token_count_far.end());
if (!in_near && !in_far) {
continue; continue;
} }
float penalty = (in_near?rep_pen:rep_pen_reduced); float penalty = (token_in_near?rep_pen:rep_pen_reduced);
// The academic publication that described this technique actually just only divided, but that would cause tokens with negative logits to become more likely, which is obviously wrong. // The academic publication that described this technique actually just only divided, but that would cause tokens with negative logits to become more likely, which is obviously wrong.
// This is common fix for this problem, which is to multiply by the penalty instead of dividing. // This is common fix for this problem, which is to multiply by the penalty instead of dividing.
@ -1229,7 +1218,6 @@ void sample_rep_pen(int n_ctx, int rep_pen_range, float rep_pen, float rep_pen_s
} }
candidates->sorted = false; candidates->sorted = false;
} }
void sample_top_p(llama_token_data_array * cur_p, float p, size_t min_keep) { void sample_top_p(llama_token_data_array * cur_p, float p, size_t min_keep) {