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

llama : move random seed generation to the samplers (#9398)

Browse source 
* llama_sampler_penalties : clamp penalty_last_n to zero
This commit is contained in:
slaren 2024-09-10 18:04:25 +02:00 • committed by GitHub
parent 00ba2ff781
commit 49006c67b4
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
10 changed files with 92 additions and 34 deletions
Download patch file Download diff file Expand all files Collapse all files

7
common/arg.cpp
View file

@ -173,7 +173,6 @@ static bool gpt_params_parse_ex(int argc, char ** argv, gpt_params_context & ctx
std::string arg;
const std::string arg_prefix = "--";
gpt_params & params = ctx_arg.params;
gpt_sampler_params & sparams = params.sparams;
std::unordered_map<std::string, llama_arg *> arg_to_options;
for (auto & opt : ctx_arg.options) {
@ -283,10 +282,6 @@ static bool gpt_params_parse_ex(int argc, char ** argv, gpt_params_context & ctx
params.kv_overrides.back().key[0] = 0;
}
if (sparams.seed == LLAMA_DEFAULT_SEED) {
sparams.seed = time(NULL);
}
return true;
}
@ -909,7 +904,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
).set_sparam());
add_opt(llama_arg(
{"-s", "--seed"}, "SEED",
format("RNG seed (default: %d, use random seed for < 0)", params.sparams.seed),
format("RNG seed (default: %u, use random seed for %u)", params.sparams.seed, LLAMA_DEFAULT_SEED),
[](gpt_params & params, const std::string & value) {
params.sparams.seed = std::stoul(value);
}

4
common/sampling.cpp
View file

@ -310,6 +310,10 @@ llama_token gpt_sampler_sample(struct gpt_sampler * gsmpl, struct llama_context
return cur_p.data[cur_p.selected].id;
}
uint32_t gpt_sampler_get_seed(const struct gpt_sampler * gsmpl) {
return llama_sampler_get_seed(gsmpl->chain);
}
// helpers
llama_token_data_array * gpt_sampler_get_candidates(struct gpt_sampler * gsmpl) {

2
common/sampling.h
View file

@ -60,6 +60,8 @@ void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler *
//
llama_token gpt_sampler_sample(struct gpt_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first = false);
uint32_t gpt_sampler_get_seed(const struct gpt_sampler * gsmpl);
// helpers
// access the internal list of current candidate tokens