From d3435efc8a2f926f1d7a5f2a11ac161b798c355a Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Fri, 9 Jan 2026 12:16:40 +0200
Subject: [PATCH 1/2] scripts : pr2wt.sh reset to remote head (#18695)

* scripts : pr2wt.sh reset to remote head

* cont : cleaner

* cont : restore --set-upstream-to
---
 scripts/pr2wt.sh | 28 ++++++++++++++++++++--------
 1 file changed, 20 insertions(+), 8 deletions(-)

diff --git a/scripts/pr2wt.sh b/scripts/pr2wt.sh
index 7970bec371..8e5d89462c 100755
--- a/scripts/pr2wt.sh
+++ b/scripts/pr2wt.sh
@@ -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

From 53eb9435da3affa12a38a8b0fb29081698a8d1cc Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Fri, 9 Jan 2026 12:59:50 +0200
Subject: [PATCH 2/2] server : fix timing of prompt/generation (#18713)

---
 tools/server/server-context.cpp | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp
index e1f65dfcce..324c3af30c 100644
--- a/tools/server/server-context.cpp
+++ b/tools/server/server-context.cpp
@@ -2615,10 +2615,6 @@ 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()) {
@@ -2685,6 +2681,9 @@ 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) {
@@ -2728,6 +2727,8 @@ 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;