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

llama : default sampling changes + greedy update (#9897)

Browse source 
* llama : deprecate softmax sampler + fix dist sampler

ggml-ci

* tests : replace macros with functions

ggml-ci

* sampling : change temperature sampler logic

For t <= 0.0f, keep the max logit intact and set the rest to -inf

* cont : no need for special "greedy" logic

top-k == 1 is the same

* tests : init prob correctly

* llama : handle temp <= 0.0 in the temp_ext sampler too

ggml-ci

* cont : avoid extra loop in temperature sampler for sub-zero temp

ggml-ci
This commit is contained in:
Georgi Gerganov 2024-10-21 09:46:40 +03:00 • committed by GitHub
parent bc21975084
commit 55e47786e3
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
7 changed files with 202 additions and 218 deletions
Download patch file Download diff file Expand all files Collapse all files

10
include/llama.h
View file

@ -217,6 +217,7 @@ extern "C" {
typedef struct llama_token_data_array {
// TODO: consider SoA
// NOTE: this pointer can be modified by the samplers
llama_token_data * data;
size_t size;
int64_t selected; // this is the index in the data array (i.e. not the token id)
@ -1069,12 +1070,13 @@ extern "C" {
// available samplers:
LLAMA_API struct llama_sampler * llama_sampler_init_greedy (void);
LLAMA_API struct llama_sampler * llama_sampler_init_dist (uint32_t seed);
LLAMA_API struct llama_sampler * llama_sampler_init_greedy(void);
LLAMA_API struct llama_sampler * llama_sampler_init_dist (uint32_t seed);
/// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
/// NOTE: Avoid using on the full vocabulary as the sorting can become slow. For example, apply top-k or top-p sampling first.
LLAMA_API struct llama_sampler * llama_sampler_init_softmax (void);
DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_softmax (void),
"will be removed in the future (see https://github.com/ggerganov/llama.cpp/pull/9896#discussion_r1800920915)");
/// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
LLAMA_API struct llama_sampler * llama_sampler_init_top_k (int32_t k);
@ -1090,6 +1092,8 @@ extern "C" {
/// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
LLAMA_API struct llama_sampler * llama_sampler_init_typical (float p, size_t min_keep);
/// #details Updates the logits l_i` = l_i/t. When t <= 0.0f, the maximum logit is kept at it's original value, the rest are set to -inf
LLAMA_API struct llama_sampler * llama_sampler_init_temp (float t);
/// @details Dynamic temperature implementation (a.k.a. entropy) described in the paper https://arxiv.org/abs/2309.02772.