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

llama : add infill sampler (#9896)

Browse source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2024-10-15 16:35:33 +03:00 • committed by GitHub
parent 223c25a72f
commit 755a9b2bf0
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
9 changed files with 300 additions and 29 deletions
Download patch file Download diff file Expand all files Collapse all files

28
include/llama.h
View file

@ -953,6 +953,12 @@ extern "C" {
int32_t lstrip,
bool special);
// check if token0 is contained as a prefix in token1
LLAMA_API bool llama_token_is_prefix(
const struct llama_model * model,
llama_token token0,
llama_token token1);
/// @details Convert the provided tokens into text (inverse of llama_tokenize()).
/// @param text The char pointer must be large enough to hold the resulting text.
/// @return Returns the number of chars/bytes on success, no more than text_len_max.
@ -1148,6 +1154,28 @@ extern "C" {
int32_t n_logit_bias,
const llama_logit_bias * logit_bias);
// this sampler is meant to be used for fill-in-the-middle infilling
// it's supposed to be used after top_k + top_p sampling
//
// 1. if the sum of the EOG probs times the number of candidates is higher than the sum of the other probs -> pick EOG
// 2. combine probs of tokens that have the same prefix
//
// example:
//
// - before:
// "hel": 0.5
// "hell": 0.2
// "hello": 0.1
// "dummy": 0.1
//
// - after:
// "hel": 0.8
// "dummy": 0.1
//
// 3. discard non-EOG tokens with low prob
// 4. if no tokens are left -> pick EOT
//
LLAMA_API struct llama_sampler * llama_sampler_init_infill(const struct llama_model * model);
// Returns the seed used by the sampler if applicable, LLAMA_DEFAULT_SEED otherwise
LLAMA_API uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl);