diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp index 9973572f62..01ac24200a 100644 --- a/tools/server/server-context.cpp +++ b/tools/server/server-context.cpp @@ -196,10 +196,14 @@ struct server_slot { } void init_sampler() const { - const int64_t t_start = ggml_time_us(); - common_sampler_reset(smpl.get()); + if (!task->need_sampling()) { + return; + } + + const int64_t t_start = ggml_time_us(); + int n_text = 0; for (int i = 0; i < (int) prompt.tokens.size(); i++) {