Skip to content

LLM inference in C/C++

C++ ggml

server : avoid repeated mmproj memory margin addition on sleep/resume (#28233)

TANGBUDU5a9a34d88847
diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp
index b2116b1740..74812189cb 100644
--- a/tools/server/server-context.cpp
+++ b/tools/server/server-context.cpp
@@ -1180,7 +1180,8 @@ private:
}
// optionally fit mmproj memory usage
- if (has_mmproj && params_base.fit_params) {
+ // on resume params_base already holds the margin from the first load
+ if (has_mmproj && params_base.fit_params && !is_resume) {
const auto & mmproj_mem = mmproj_usage.backend_mem_usage;
const int64_t t_elapsed = mmproj_usage_t_us;
if (!mmproj_mem.empty()) {