sdgoij/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-05-14 21:14:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

spec : disacard last drafted token with low prob (#22506)

Browse Source
This commit is contained in:
Georgi Gerganov
2026-04-29 17:00:00 +03:00
committed by GitHub
parent b1d5f5b449
commit 683c5acb90
2 changed files with 7 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tools/server/server-context.cpp
View File

@@ -354,6 +354,7 @@ struct server_slot {
// generate a new draft
spec_draft = common_speculative_draft(spec.get(), params_spec, tokens, sampled);
n_draft_total += spec_draft.size();
if (spec_draft.size() > (size_t) n_draft_max) {
SLT_WRN(*this, "draft size %d exceeds max %d, truncating\n", (int) spec_draft.size(), n_draft_max);
@@ -3019,7 +3020,6 @@ private:
// update how many tokens out of those tested were accepted
slot.n_draft_accepted += ids.size() - 1;
slot.n_draft_total += n_draft;
// add accepted tokens to the prompt
slot.prompt.tokens.keep_first(slot.prompt.n_tokens() - n_draft);