From ae547b85023ff75cb0900b04d2998546c1db89a9 Mon Sep 17 00:00:00 2001
From: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Date: Fri, 9 Jan 2026 11:15:05 +0000
Subject: [PATCH] Fix assert when --max-gpu is less than available GPUs

---
 src/llama-load-tensors.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/llama-load-tensors.cpp b/src/llama-load-tensors.cpp
index bf572483..8acc6dc5 100644
--- a/src/llama-load-tensors.cpp
+++ b/src/llama-load-tensors.cpp
@@ -274,7 +274,7 @@ static std::vector<int> create_split(int nr, int granularity, const std::vector<
     }
     while (sum > nchunk) {
         last_split = 0;
-        float best_err = 0;
+        float best_err = -INFINITY;
         int ibest = -1;
         for (int i = 0; i < (int)splits.size(); ++i) {
             if (result[i] > 0) {
@@ -294,7 +294,7 @@ static std::vector<int> create_split(int nr, int granularity, const std::vector<
     }
     while (sum < nchunk) {
         last_split = 0;
-        float best_err = 0;
+        float best_err = -INFINITY;
         int ibest = -1;
         for (int i = 0; i < (int)splits.size(); ++i) {
             float p = splits[i] - last_split;