common, server : surface min_keep as its own parameter (#5567)

* Feature - surface min_keep as its own parameter

* Updated README with min_keep param
This commit is contained in:
Robey Holderith 2024-02-18 11:11:16 -08:00 committed by GitHub
parent c145f8a132
commit 5ee99c32f5
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 14 additions and 1 deletions

View file

@ -1704,6 +1704,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
}
fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
fprintf(stream, "min_keep: %d # default: 0 (disabled)\n", sparams.min_keep);
fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
fprintf(stream, "mirostat_ent: %f # default: 5.0\n", sparams.mirostat_tau);
fprintf(stream, "mirostat_lr: %f # default: 0.1\n", sparams.mirostat_eta);

View file

@ -248,7 +248,10 @@ static llama_token llama_sampling_sample_impl(
llama_sample_temp(ctx_main, &cur_p, temp);
id = llama_sample_token_mirostat_v2(ctx_main, &cur_p, mirostat_tau, mirostat_eta, &ctx_sampling->mirostat_mu);
} else {
sampler_queue(ctx_main, params, cur_p, 1);
// temperature sampling
size_t min_keep = std::max(1, params.min_keep);
sampler_queue(ctx_main, params, cur_p, min_keep);
id = llama_sample_token(ctx_main, &cur_p);

View file

@ -22,6 +22,7 @@ enum class llama_sampler_type : char {
typedef struct llama_sampling_params {
int32_t n_prev = 64; // number of previous tokens to remember
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
int32_t min_keep = 0; // 0 = disabled, otherwise samplers should return at least min_keep tokens
int32_t top_k = 40; // <= 0 to use vocab size
float top_p = 0.95f; // 1.0 = disabled
float min_p = 0.05f; // 0.0 = disabled