mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 27 Packages Projects Releases Wiki Activity

llama: introduce support for model-embedded sampling parameters (#17120)

Browse Source
This commit is contained in:
Aaron Teo
2025-11-25 09:56:07 +08:00
committed by GitHub
parent 3d07caa99b
commit 877566d512
10 changed files with 293 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
common/arg.cpp
+12
View File
@@ -1232,6 +1232,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
[](common_params & params, const std::string & value) {
const auto sampler_names = string_split<std::string>(value, ';');
params.sampling.samplers = common_sampler_types_from_names(sampler_names, true);
params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS;
}
).set_sparam());
add_opt(common_arg(
@@ -1261,6 +1262,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
[](common_params & params, const std::string & value) {
params.sampling.temp = std::stof(value);
params.sampling.temp = std::max(params.sampling.temp, 0.0f);
params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TEMP;
}
).set_sparam());
add_opt(common_arg(
@@ -1268,6 +1270,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
string_format("top-k sampling (default: %d, 0 = disabled)", params.sampling.top_k),
[](common_params & params, int value) {
params.sampling.top_k = value;
params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_K;
}
).set_sparam());
add_opt(common_arg(
@@ -1275,6 +1278,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
string_format("top-p sampling (default: %.1f, 1.0 = disabled)", (double)params.sampling.top_p),
[](common_params & params, const std::string & value) {
params.sampling.top_p = std::stof(value);
params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P;
}
).set_sparam());
add_opt(common_arg(
@@ -1282,6 +1286,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
string_format("min-p sampling (default: %.1f, 0.0 = disabled)", (double)params.sampling.min_p),
[](common_params & params, const std::string & value) {
params.sampling.min_p = std::stof(value);
params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P;
}
).set_sparam());
add_opt(common_arg(
@@ -1296,6 +1301,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sampling.xtc_probability),
[](common_params & params, const std::string & value) {
params.sampling.xtc_probability = std::stof(value);
params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY;
}
).set_sparam());
add_opt(common_arg(
@@ -1303,6 +1309,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
string_format("xtc threshold (default: %.1f, 1.0 = disabled)", (double)params.sampling.xtc_threshold),
[](common_params & params, const std::string & value) {
params.sampling.xtc_threshold = std::stof(value);
params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD;
}
).set_sparam());
add_opt(common_arg(
@@ -1321,6 +1328,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
}
params.sampling.penalty_last_n = value;
params.sampling.n_prev = std::max(params.sampling.n_prev, params.sampling.penalty_last_n);
params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N;
}
).set_sparam());
add_opt(common_arg(
@@ -1328,6 +1336,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
string_format("penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)", (double)params.sampling.penalty_repeat),
[](common_params & params, const std::string & value) {
params.sampling.penalty_repeat = std::stof(value);
params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT;
}
).set_sparam());
add_opt(common_arg(
@@ -1425,6 +1434,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
"(default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)", params.sampling.mirostat),
[](common_params & params, int value) {
params.sampling.mirostat = value;
params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT;
}
).set_sparam());
add_opt(common_arg(
@@ -1432,6 +1442,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
string_format("Mirostat learning rate, parameter eta (default: %.1f)", (double)params.sampling.mirostat_eta),
[](common_params & params, const std::string & value) {
params.sampling.mirostat_eta = std::stof(value);
params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA;
}
).set_sparam());
add_opt(common_arg(
@@ -1439,6 +1450,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
string_format("Mirostat target entropy, parameter tau (default: %.1f)", (double)params.sampling.mirostat_tau),
[](common_params & params, const std::string & value) {
params.sampling.mirostat_tau = std::stof(value);
params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU;
}
).set_sparam());
add_opt(common_arg(
common/common.cpp
+55
View File
@@ -8,6 +8,7 @@
#include "common.h"
#include "log.h"
#include "llama.h"
#include "sampling.h"
#include <algorithm>
#include <cinttypes>
@@ -949,6 +950,58 @@ std::vector<common_file_info> fs_list_files(const std::string & path) {
// Model utils
//
static inline void common_init_sampler_from_model(
const llama_model * model,
common_params_sampling & sparams) {
const uint64_t config = sparams.user_sampling_config;
auto get_int32 = [&](const char * key, int32_t & dst, uint64_t user_config) {
if (config & user_config) return;
char buf[64] = {0};
if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
char * end = nullptr;
int32_t v = strtol(buf, &end, 10);
if (end && end != buf) dst = v;
}
};
auto get_float = [&](const char * key, float & dst, uint64_t user_config) {
if (config & user_config) return;
char buf[128] = {0};
if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
char * end = nullptr;
float v = strtof(buf, &end);
if (end && end != buf) dst = v;
}
};
// Sampling sequence
if (!(config & common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS)) {
char buf[512] = {0};
if (llama_model_meta_val_str(model, llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE), buf, sizeof(buf)) > 0) {
const std::vector<std::string> sampler_names = string_split<std::string>(std::string(buf), ';');
if (!sampler_names.empty()) {
sparams.samplers = common_sampler_types_from_names(sampler_names, true);
}
}
}
get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_K), sparams.top_k, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_K);
get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_P), sparams.top_p, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P);
get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIN_P), sparams.min_p, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P);
get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY), sparams.xtc_probability, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY);
get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD), sparams.xtc_threshold, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD);
get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TEMP), sparams.temp, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TEMP);
get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N), sparams.penalty_last_n, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N);
get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT), sparams.penalty_repeat, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT);
get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT), sparams.mirostat, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT);
get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU), sparams.mirostat_tau, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU);
get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA), sparams.mirostat_eta, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA);
}
struct common_init_result common_init_from_params(common_params & params) {
common_init_result iparams;
auto mparams = common_model_params_to_llama(params);
@@ -960,6 +1013,8 @@ struct common_init_result common_init_from_params(common_params & params) {
return iparams;
}
common_init_sampler_from_model(model, params.sampling);
const llama_vocab * vocab = llama_model_get_vocab(model);
auto cparams = common_context_params_to_llama(params);
common/common.h
+18
View File
@@ -140,6 +140,22 @@ struct common_grammar_trigger {
llama_token token = LLAMA_TOKEN_NULL;
};
enum common_params_sampling_config : uint64_t {
COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS = 1 << 0,
COMMON_PARAMS_SAMPLING_CONFIG_TOP_K = 1 << 1,
COMMON_PARAMS_SAMPLING_CONFIG_TOP_P = 1 << 2,
COMMON_PARAMS_SAMPLING_CONFIG_MIN_P = 1 << 3,
COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY = 1 << 4,
COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD = 1 << 5,
COMMON_PARAMS_SAMPLING_CONFIG_TEMP = 1 << 6,
COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N = 1 << 7,
COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT = 1 << 8,
COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT = 1 << 9,
COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU = 1 << 10,
COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA = 1 << 11,
};
// sampling parameters
struct common_params_sampling {
uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
@@ -172,6 +188,8 @@ struct common_params_sampling {
bool no_perf = false; // disable performance metrics
bool timing_per_token = false;
uint64_t user_sampling_config = 0; // bitfield to track user-specified samplers
std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"}; // default sequence breakers for DRY