happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

add args, rename `queue_size` -> `window_size`

This commit is contained in:
ddh0 2025-12-11 15:16:11 -06:00
parent 374bfd4363
commit ffe163911b
4 changed files with 33 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
common/arg.cpp
View File

@ -1501,6 +1501,29 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
}
}
).set_sparam());
add_opt(common_arg(
{"--power-law-target"}, "N",
string_format("Power Law sampler target probability (default: %.2f, 0.0 to 1.0)\n"
"[(more info)](https://github.com/ggml-org/llama.cpp/pull/17927)",
(double)params.sampling.power_law_target),
[](common_params & params, const std::string & value) {
params.sampling.power_law_target = std::stof(value);
}
).set_sparam());
add_opt(common_arg(
{"--power-law-target-range"}, "N",
string_format("Power Law sampler adaptive range +/- from target (default: %.2f, 0.0 = no adaptation)", (double)params.sampling.power_law_target_range),
[](common_params & params, const std::string & value) {
params.sampling.power_law_target_range = std::stof(value);
}
).set_sparam());
add_opt(common_arg(
{"--power-law-window-size"}, "N",
string_format("Power Law sampler rolling window size, in tokens (default: %d)", params.sampling.power_law_window_size),
[](common_params & params, int value) {
params.sampling.power_law_window_size = value;
}
).set_sparam());
add_opt(common_arg(
{"--dynatemp-range"}, "N",
string_format("dynamic temperature range (default: %.1f, 0.0 = disabled)", (double)params.sampling.dynatemp_range),

2
common/common.h
View File

@ -186,7 +186,7 @@ struct common_params_sampling {
int32_t dry_penalty_last_n = -1; // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
float power_law_target = 0.5; // target probability (0.0 to 1.0)
float power_law_target_range = 0.5; // adapt the target within this range (target +/- range)
int32_t power_law_queue_size = 10; // rolling history window size for adaptation
int32_t power_law_window_size = 10; // rolling history window size for target adaptation
int32_t mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
float top_n_sigma = -1.00f; // -1.0 = disabled
float mirostat_tau = 5.00f; // target entropy

4
include/llama.h
View File

@ -1294,8 +1294,8 @@ extern "C" {
/// ref: https://github.com/ggml-org/llama.cpp/pull/17927
LLAMA_API struct llama_sampler * llama_sampler_init_power_law(
float target, // target probability (0.0 to 1.0)
float target_range, // adaptive target range (±range from target)
int32_t queue_size, // rolling history window size for adaptation
float target_range, // adaptive target range (+/- range from target)
int32_t window_size, // rolling history window size for target adaptation
uint32_t seed); // RNG seed
LLAMA_API struct llama_sampler * llama_sampler_init_logit_bias(

14
src/llama-sampling.cpp
View File

@ -2320,7 +2320,7 @@ struct llama_sampler * llama_sampler_init_dry_testing(int32_t context_size, floa
struct llama_sampler_power_law {
const float target;
const float target_range;
const int32_t queue_size;
const int32_t window_size;
const uint32_t seed;
std::mt19937 rng;
@ -2359,7 +2359,7 @@ static void llama_sampler_power_law_apply(struct llama_sampler * smpl, llama_tok
sum_excluding_oldest += ctx->history.rat(i);
}
float next_value = (ctx->target * ctx->queue_size) - sum_excluding_oldest;
float next_value = (ctx->target * ctx->window_size) - sum_excluding_oldest;
computed_target = std::max(min_target, std::min(next_value, max_target));
}
@ -2397,12 +2397,12 @@ static void llama_sampler_power_law_apply(struct llama_sampler * smpl, llama_tok
static void llama_sampler_power_law_reset(struct llama_sampler * smpl) {
auto * ctx = (llama_sampler_power_law *) smpl->ctx;
ctx->history = ring_buffer<float>(ctx->queue_size);
ctx->history = ring_buffer<float>(ctx->window_size);
}
static struct llama_sampler * llama_sampler_power_law_clone(const struct llama_sampler * smpl) {
const auto * ctx = (const llama_sampler_power_law *) smpl->ctx;
auto * result = llama_sampler_init_power_law(ctx->target, ctx->target_range, ctx->queue_size, ctx->seed);
auto * result = llama_sampler_init_power_law(ctx->target, ctx->target_range, ctx->window_size, ctx->seed);
auto * result_ctx = (llama_sampler_power_law *) result->ctx;
result_ctx->rng = ctx->rng;
@ -2427,7 +2427,7 @@ static struct llama_sampler_i llama_sampler_power_law_i = {
struct llama_sampler * llama_sampler_init_power_law(
float target,
float target_range,
int32_t queue_size,
int32_t window_size,
uint32_t seed
) {
auto seed_cur = get_rng_seed(seed);
@ -2436,10 +2436,10 @@ struct llama_sampler * llama_sampler_init_power_law(
/* .ctx = */ new llama_sampler_power_law {
/* .target = */ target,
/* .target_range = */ target_range,
/* .queue_size = */ queue_size,
/* .window_size = */ window_size,
/* .seed = */ seed_cur,
/* .rng = */ std::mt19937(seed_cur),
/* .history = */ ring_buffer<float>(queue_size),
/* .history = */ ring_buffer<float>(window_size),
}
);
}