ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 03:11:51 +00:00
Code Issues Packages Projects Releases Wiki Activity

Implement Adaptive-P Sampler (#1100)

Browse Source 
* initial implementation of adaptive-p sampler

* explicitly mark candidates unsorted + cleanup qualifiers

* cosmetic update

* reorg prototypes

* lockstep with mainline

* add _impl for _init + reorg

* add LLAMA_API to prototypes

* update sharpness to 10

* lockstep: rng seed

* delete llama_sampling member in llama_sampler_adaptive_p

* fix LLAMA_API return type

* lockstep: rng seed cont

* actually correct implementation

* lockstep: sorting behavior

* const -> constexpr for known constants

* add missing space

* fix softmax usage in adaptive p sampler

* cosmetic changes

* implement do-not-sort version of softmax

* simpify rng seed, add static to constexpr

* refactor: remove iface + use shared rng + use actually original probabilities

* adaptive-p: add dedicated rng back in

* fix initial max_logit + add float vector to adaptive p sampler context + stochastic sampling

* adaptive-p: fuse first softmax with transformation

* adaptive-p: implement binary search selection

* adaptive-p: update comment
This commit is contained in:
dungquixote42
2026-01-10 00:58:53 -05:00
committed by GitHub
parent dd3c3f72f2
commit 52ad1c6421
8 changed files with 226 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
src/llama-sampling.h
View File

@@ -61,6 +61,24 @@ struct llama_sampler_dry * llama_sampler_init_dry_impl(
void llama_sampler_dry_apply(struct llama_sampler_dry* smpl, llama_token_data_array* cur_p);
// maintains an exponential moving average of the *ORIGINAL* probabilities of selected tokens
// used to compute an adapted target at each sampling step.
// see llama.h for a full description of the sampler
struct llama_sampler_adaptive_p {
const float target; // target probability (0.0 - 1.0; negative = disabled)
const float decay; // EMA decay; history ≈ 1/(1-decay) tokens (0.0 - 0.99)
std::mt19937 rng; // RNG
float weighted_sum; // sum(p_n * decay^N)
float total_weight; // sum(decay^i), converges to 1/(1-decay)
float max_logit; // maximum logit found during transform
std::vector<float> probs; // cumulative probabilities
};
void llama_sampler_adaptive_p_apply(
struct llama_sampler_adaptive_p * adapt_p_ctx,
llama_token_data_array * candidates);
struct llama_sampler_adaptive_p * llama_sampler_init_adaptive_p_impl(const float target, const float decay, const uint32_t seed);
void llama_sample_repetition_penalties_impl(
@@ -83,6 +101,6 @@ llama_token llama_sample_token_mirostat_v2_impl(struct llama_sampling * smpl, ll
llama_token llama_sample_token_greedy_impl (struct llama_sampling * smpl, llama_token_data_array * candidates);
llama_token llama_sample_token_with_rng_impl (struct llama_sampling * smpl, llama_token_data_array * candidates, std::mt19937 & rng);
llama_token llama_sample_token_impl (struct llama_sampling * smpl, llama_token_data_array * candidates);
llama_token llama_sample_token_adaptive_p_impl (struct llama_sampling * smpl, llama_token_data_array * candidates, struct llama_sampler_adaptive_p * adapt_p_ctx, float * orig_probs);