nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 4 Packages Projects Releases Wiki Activity

speculative : fix n_outputs_max and remove draft-simple auto-enable (#23988)

Browse Source 
* speculative : add common_speculative_n_max helper function

Extract the speculative max-draft-size logic from server_n_outputs_max
into a reusable common_speculative_n_max() function in common/speculative.

Assisted-by: llama.cpp:local pi

* cont : draft context always has n_parallel outputs

* llama : log n_outputs_max

* speculative : remove draft-simple auto-enable

* ci : enable server tests on PRs
This commit is contained in:
Georgi Gerganov
2026-06-01 22:26:58 +03:00
committed by GitHub
parent 5aa3a64596
commit 5dcb711666
6 changed files with 40 additions and 51 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
common/speculative.cpp
View File

@@ -1317,6 +1317,40 @@ static uint32_t common_get_enabled_speculative_configs(const std::vector<common_
return result;
}
int32_t common_speculative_n_max(const common_params_speculative * spec) {
int32_t n_max = 0;
for (const auto type : spec->types) {
switch (type) {
case COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE:
case COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3:
case COMMON_SPECULATIVE_TYPE_DRAFT_MTP:
n_max = std::max(n_max, std::max(0, spec->draft.n_max));
break;
case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE:
n_max = std::max(n_max, (int32_t) spec->ngram_simple.size_m);
break;
case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K:
n_max = std::max(n_max, (int32_t) spec->ngram_map_k.size_m);
break;
case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V:
n_max = std::max(n_max, (int32_t) spec->ngram_map_k4v.size_m);
break;
case COMMON_SPECULATIVE_TYPE_NGRAM_MOD:
n_max = std::max(n_max, std::max(0, spec->ngram_mod.n_max));
break;
case COMMON_SPECULATIVE_TYPE_NGRAM_CACHE:
n_max = std::max(n_max, (int32_t) 8);
break;
case COMMON_SPECULATIVE_TYPE_NONE:
case COMMON_SPECULATIVE_TYPE_COUNT:
break;
}
}
return n_max;
}
// initialization of the speculative decoding system
//
common_speculative * common_speculative_init(common_params_speculative & params, uint32_t n_seq) {
@@ -1325,8 +1359,6 @@ common_speculative * common_speculative_init(common_params_speculative & params,
{
uint32_t enabled_configs = common_get_enabled_speculative_configs(params.types);
bool has_draft_model_path = !params.draft.mparams.path.empty();
bool has_draft_simple = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE));
bool has_draft_eagle3 = false; // TODO PR-18039: if params.speculative.eagle3
bool has_mtp = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_MTP)) && params.draft.ctx_dft != nullptr;
@@ -1359,16 +1391,6 @@ common_speculative * common_speculative_init(common_params_speculative & params,
if (has_ngram_cache) {
configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, params));
}
if (has_draft_simple) {
if (!has_draft_model_path) {
LOG_WRN("%s: draft model is not specified - cannot use 'draft' type\n", __func__);
has_draft_simple = false;
}
} else if (has_draft_model_path && !has_mtp && !has_draft_eagle3) {
LOG_WRN("%s: draft model is specified but 'draft' speculative type is not explicitly enabled - enabling it\n", __func__);
has_draft_simple = true;
}
if (has_draft_simple) {
configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE, params));
}