Public/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-18 16:55:05 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into pr/18663

Browse Source
This commit is contained in:
Georgi Gerganov
2026-01-09 13:05:16 +02:00
parent 439c3b5021 53eb9435da
commit 91fd50be1b
2 changed files with 25 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
scripts/pr2wt.sh
+20 -8
View File
@@ -4,12 +4,13 @@
#
# - creates a new remote using the fork's clone URL
# - creates a local branch tracking the remote branch
# - creates a new worktree in a parent folder, suffixed with "-pr-${PR}"
# - creates a new worktree in a parent folder, suffixed with "-pr-$PR"
#
# sample usage:
# ./scripts/pr2wt.sh 12345
# ./scripts/pr2wt.sh 12345 opencode
# ./scripts/pr2wt.sh 12345 "cmake -B build && cmake --build build"
# ./scripts/pr2wt.sh 12345 "bash -l"
function usage() {
echo "usage: $0 <pr_number> [cmd]"
@@ -39,7 +40,7 @@ org_repo=${org_repo%.git}
echo "org/repo: $org_repo"
meta=$(curl -sSf -H "Accept: application/vnd.github+json" "https://api.github.com/repos/${org_repo}/pulls/${PR}")
meta=$(curl -sSf -H "Accept: application/vnd.github+json" "https://api.github.com/repos/$org_repo/pulls/$PR")
url_remote=$(echo "$meta" | jq -r '.head.repo.clone_url')
head_ref=$(echo "$meta" | jq -r '.head.ref')
@@ -47,21 +48,32 @@ head_ref=$(echo "$meta" | jq -r '.head.ref')
echo "url: $url_remote"
echo "head_ref: $head_ref"
git remote rm pr/${PR} 2> /dev/null
git remote add pr/${PR} $url_remote
git fetch pr/${PR} $head_ref
url_remote_cur=$(git config --get "remote.pr/$PR.url" 2>/dev/null || true)
if [[ "$url_remote_cur" != "$url_remote" ]]; then
git remote rm pr/$PR 2> /dev/null
git remote add pr/$PR "$url_remote"
fi
git fetch "pr/$PR" "$head_ref"
dir=$(basename $(pwd))
git branch -D pr/$PR 2> /dev/null
git worktree add -b pr/$PR ../$dir-pr-$PR pr/$PR/${head_ref} 2> /dev/null
git worktree add -b pr/$PR ../$dir-pr-$PR pr/$PR/$head_ref 2> /dev/null
wt_path=$(cd ../$dir-pr-$PR && pwd)
echo "git worktree created in $wt_path"
# if a command was provided, execute it
cd $wt_path
git branch --set-upstream-to=pr/$PR/$head_ref
git pull --ff-only || {
echo "error: failed to pull pr/$PR"
exit 1
}
if [[ $# -eq 2 ]]; then
cd ../$dir-pr-$PR
echo "executing: $2"
eval "$2"
fi
tools/server/server-context.cpp
+5 -4
View File
@@ -2639,10 +2639,6 @@ private:
// on successful decode, restore the original batch size
n_batch = llama_n_batch(ctx);
// technically, measuring the time here excludes the sampling time for the last batch
// but on the other hand, we don't want to do too many system calls to measure the time, so it's ok
const int64_t t_current = ggml_time_us();
for (auto & slot : slots) {
// may need to copy state to other slots
if (slot.state == SLOT_STATE_DONE_PROMPT && slot.is_parent()) {
@@ -2711,6 +2707,9 @@ private:
common_sampler_accept(slot.smpl.get(), id, true);
// here we have synchronized the llama_context (due to the sampling above), so we can do time measurement
const int64_t t_current = ggml_time_us();
slot.n_decoded += 1;
if (slot.n_decoded == 1) {
@@ -2754,6 +2753,8 @@ private:
slot.i_batch_dft.clear();
slot.drafted.clear();
const int64_t t_current = ggml_time_us();
slot.n_decoded += ids.size();
slot.t_token_generation = std::max<int64_t>(1, t_current - slot.t_start_generation) / 1e3;