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

Server: fix seed for multiple slots (#6835)

Browse source 
* Server: add tests for consistent results

* sampling: separate rng per sampling context
This commit is contained in:
Johannes Gäßler 2024-04-24 11:08:36 +02:00 • committed by GitHub
parent c0d1b3e03e
commit 28103f4832
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
11 changed files with 145 additions and 30 deletions
Download patch file Download diff file Expand all files Collapse all files

13
common/sampling.cpp
View file

@ -1,4 +1,6 @@
#define LLAMA_API_INTERNAL
#include "sampling.h"
#include <random>
struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params) {
struct llama_sampling_context * result = new llama_sampling_context();
@ -33,6 +35,8 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
result->prev.resize(params.n_prev);
llama_sampling_set_rng_seed(result, params.seed);
return result;
}
@ -62,6 +66,13 @@ void llama_sampling_reset(llama_sampling_context * ctx) {
ctx->cur.clear();
}
void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed) {
if (seed == LLAMA_DEFAULT_SEED) {
seed = time(NULL);
}
ctx->rng.seed(seed);
}
void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst) {
if (dst->grammar) {
llama_grammar_free(dst->grammar);
@ -203,7 +214,7 @@ static llama_token llama_sampling_sample_impl(
sampler_queue(ctx_main, params, cur_p, min_keep);
id = llama_sample_token(ctx_main, &cur_p);
id = llama_sample_token_with_rng(ctx_main, &cur_p, ctx_sampling->rng);
//{
// const int n_top = 10;