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

spec: remove parameter spec-ngram-check-rate

This commit is contained in:
Sascha Rogmann 2026-02-05 22:27:52 +01:00
parent 22cae83218
commit 32d72eee29
7 changed files with 8 additions and 42 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
common/arg.cpp
View File

@ -3437,16 +3437,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.speculative.ngram_size_m = value;
}
).set_examples({LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"--spec-ngram-check-rate"}, "N",
string_format("ngram check rate for ngram-simple/ngram-map speculative decoding (default: %d)", params.speculative.ngram_check_rate),
[](common_params & params, int value) {
if (value < 1) {
throw std::invalid_argument("ngram check rate must be at least 1");
}
params.speculative.ngram_check_rate = value;
}
).set_examples({LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"--spec-ngram-min-hits"}, "N",
string_format("minimum hits for ngram-map speculative decoding (default: %d)", params.speculative.ngram_min_hits),

1
common/common.h
View File

@ -269,7 +269,6 @@ struct common_params_speculative {
uint16_t ngram_size_n = 12; // ngram size for lookup
uint16_t ngram_size_m = 48; // mgram size for speculative tokens
uint16_t ngram_check_rate = 1; // check rate for ngram lookup
uint16_t ngram_min_hits = 1; // minimum hits at ngram/mgram lookup for mgram to be proposed
std::shared_ptr<common_ngram_mod> ngram_mod;

7
common/ngram-map.cpp
View File

@ -231,10 +231,9 @@ void common_ngram_map_draft(common_ngram_map & map,
GGML_ABORT("%s: cur_len exceeds UINT32_MAX: %zu", __func__, cur_len);
}
// Only check every check_rate tokens to save compute
// i.e., perform check if (cur_len - idx_last_check) >= check_rate
if (map.idx_last_check + map.check_rate > cur_len) {
return;
if (map.idx_last_check > cur_len) {
// Should not happen because of common_ngram_map_begin().
GGML_ABORT("%s: map.idx_last_check > cur_len: %zu > %zu", __func__, map.idx_last_check, cur_len);
}
map.idx_last_check = cur_len;

8
common/ngram-map.h
View File

@ -24,7 +24,6 @@
struct common_ngram_simple_config {
uint16_t size_ngram; // size of n-grams to lookup in self-mode
uint16_t size_mgram; // size of m-grams to draft in self-mode
uint16_t check_rate; // check for speculative decoding without draft model for each check_rate token
};
// Searches for a n-gram in the history and checks whether a draft sequence should be generated.
@ -66,15 +65,14 @@ struct common_ngram_map {
bool key_only; // true if only key n-grams are used, no values.
std::vector<common_ngram_map_key> keys; // key n-grams which occur several times in token-history
uint16_t check_rate; // check for speculative decoding without draft model for each check_rate token
uint16_t min_hits; // minimum number of key hits to consider a draft
bool show_key_map_stats = false; // true, if statitics of the key_map should be printed.
bool show_key_map_stats = true; // true, if statistics of the key_map should be printed.
common_ngram_map(uint16_t sz_key, uint16_t sz_value, bool only_keys,
uint16_t check_rate, uint16_t min_hits)
uint16_t min_hits)
: size_key(sz_key), size_value(sz_value), key_only(only_keys),
check_rate(check_rate), min_hits(min_hits) {
min_hits(min_hits) {
key_map.resize(COMMON_NGRAM_HASH_MAP_SIZE); // 2^18 hash entries, 0 entries if key_map shouldn't be used
}

14
common/speculative.cpp
View File

@ -465,8 +465,6 @@ struct common_speculative_state_eagle3 : public common_speculative_state {
struct common_speculative_state_ngram_simple : public common_speculative_state {
common_ngram_simple_config config;
uint16_t check_id = 0; // used to control the frequency of generating drafts
common_speculative_state_ngram_simple(
enum common_speculative_type type,
common_ngram_simple_config config)
@ -481,11 +479,6 @@ struct common_speculative_state_ngram_simple : public common_speculative_state {
const llama_tokens & prompt_tgt,
llama_token id_last,
llama_tokens & result) override {
++check_id;
if (check_id < config.check_rate) {
return;
}
check_id = 0;
result = common_ngram_simple_draft(config, prompt_tgt, id_last);
GGML_UNUSED(params);
@ -752,10 +745,9 @@ static common_ngram_map get_common_ngram_map(const common_speculative_config & c
uint16_t size_key = config.params.ngram_size_n;
uint16_t size_value = config.params.ngram_size_m;
bool key_only = (config.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K);
uint16_t check_rate = config.params.ngram_check_rate;
uint16_t min_hits = config.params.ngram_min_hits;
return common_ngram_map(size_key, size_value, key_only, check_rate, min_hits);
return common_ngram_map(size_key, size_value, key_only, min_hits);
}
static common_speculative_state_ngram_cache create_state_ngram_cache(
@ -895,12 +887,10 @@ common_speculative * common_speculative_init(
uint16_t ngram_size_key = ngram_map.size_key;
uint16_t mgram_size_value = ngram_map.size_value;
uint16_t check_rate = ngram_map.check_rate;
auto config_simple = common_ngram_simple_config {
/* .size_ngram = */ ngram_size_key,
/* .size_mgram = */ mgram_size_value,
/* .check_rate = */ check_rate
/* .size_mgram = */ mgram_size_value
};
auto state = std::make_unique<common_speculative_state_ngram_simple>(
/* .type = */ config.type,

6
docs/speculative.md
View File

@ -119,8 +119,6 @@ If a draft model is combined with a draftless decoding the draftless decoding ha
of lookup n-gram (default: 12)
--spec-ngram-size-m N ngram size M for ngram-simple/ngram-map speculative decoding, length
of draft m-gram (default: 48)
--spec-ngram-check-rate N ngram check rate for ngram-simple/ngram-map speculative decoding
(default: 1)
--spec-ngram-min-hits N minimum hits for ngram-map speculative decoding (default: 1)
```
@ -153,10 +151,6 @@ Sets the size M of the draft m-gram for n-gram map based speculative decoding.
The m-gram size determines how many tokens to draft when a match is found.
Larger values can provide more speedup but may reduce acceptance rate.
### `--spec-ngram-check-rate R`
This option aims at performance if the n-gram lookup in history is to costly. A lookup will be executed at every R tokens (default is 1, every token).
### `--spec-ngram-min-hits H`
This option defines how often a key has to appear in the token history to be used as a draft (default is 1).

4
tools/server/server-task.cpp
View File

@ -80,7 +80,6 @@ json task_params::to_json(bool only_metrics) const {
{"speculative.type", common_speculative_type_to_str(speculative.type)},
{"speculative.ngram_size_n", speculative.ngram_size_n},
{"speculative.ngram_size_m", speculative.ngram_size_m},
{"speculative.ngram_c_rate", speculative.ngram_check_rate},
{"speculative.ngram_m_hits", speculative.ngram_min_hits},
{"timings_per_token", timings_per_token},
{"post_sampling_probs", post_sampling_probs},
@ -144,7 +143,6 @@ json task_params::to_json(bool only_metrics) const {
{"speculative.type", common_speculative_type_to_str(speculative.type)},
{"speculative.ngram_size_n", speculative.ngram_size_n},
{"speculative.ngram_size_m", speculative.ngram_size_m},
{"speculative.ngram_c_rate", speculative.ngram_check_rate},
{"speculative.ngram_m_hits", speculative.ngram_min_hits},
{"timings_per_token", timings_per_token},
{"post_sampling_probs", post_sampling_probs},
@ -257,12 +255,10 @@ task_params server_task::params_from_json_cmpl(
params.speculative.ngram_size_n = json_value(data, "speculative.ngram_size_n", defaults.speculative.ngram_size_n);
params.speculative.ngram_size_m = json_value(data, "speculative.ngram_size_m", defaults.speculative.ngram_size_m);
params.speculative.ngram_check_rate = json_value(data, "speculative.ngram_c_rate", defaults.speculative.ngram_check_rate);
params.speculative.ngram_min_hits = json_value(data, "speculative.ngram_m_hits", defaults.speculative.ngram_min_hits);
params.speculative.ngram_size_n = std::max(std::min(1, (int) params.speculative.ngram_size_n), 1024);
params.speculative.ngram_size_m = std::max(std::min(1, (int) params.speculative.ngram_size_m), 1024);
params.speculative.ngram_check_rate = std::max(std::min(1, (int) params.speculative.ngram_check_rate), 1024);
params.speculative.ngram_min_hits = std::max(std::min(1, (int) params.speculative.ngram_min_hits), 1024);
// Use OpenAI API logprobs only if n_probs wasn't provided