CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-31 08:51:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

samplers : Min-P sampler implementation [alternative to Top P/Top K] (#3841)

Browse Source 
* Introduce the new Min-P sampler by @kalomaze
   The Min-P sampling method was designed as an alternative to Top-P, and aims to ensure a balance of quality and variety. The parameter *p* represents the minimum probability for a token to be considered, relative to the probability of the most likely token.

* Min-P enabled and set to 0.05 default

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: cebtenzzre <cebtenzzre@gmail.com>
This commit is contained in:
kalomaze
2023-10-31 14:44:49 -05:00
committed by GitHub
parent 07178c98e1
commit 238657db23
6 changed files with 54 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
llama.cpp
View File

@@ -7368,6 +7368,32 @@ void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * can
}
}
void llama_sample_min_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep) {
if (p <= 0.0f || !candidates->size) {
return;
}
llama_sample_softmax(ctx, candidates);
const int64_t t_start_sample_us = ggml_time_us();
float scale = candidates->data[0].p; // scale by max prob
size_t i = 1; // first token always matches
for (; i < candidates->size; ++i) {
if (candidates->data[i].p < p * scale && i >= min_keep) {
break; // prob too small
}
}
// Resize the output vector to keep only the matching tokens
candidates->size = i;
if (ctx) {
ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
}
}
void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep) {
if (z >= 1.0f || candidates->size <= 2) {
return;