From 1b3b6a506f8538c0192fa659dcb524d394bee7c1 Mon Sep 17 00:00:00 2001
From: "Li, Zonghang" <870644199@qq.com>
Date: Tue, 3 Jun 2025 17:10:09 +0400
Subject: [PATCH] fix: add warm-up in profiling to prevent init delay

---
 common/profiler.cpp |  5 ++++-
 src/llama.cpp       | 10 +++++-----
 2 files changed, 9 insertions(+), 6 deletions(-)

diff --git a/common/profiler.cpp b/common/profiler.cpp
index b54bb0be..a2ac33b5 100644
--- a/common/profiler.cpp
+++ b/common/profiler.cpp
@@ -439,7 +439,7 @@ float device_inp_embd_delay(struct llama_model * model, enum ggml_type src0t, in
     }
 
     // warm-up
-    // ggml_backend_graph_compute(backend, gf);
+    ggml_backend_graph_compute(backend, gf);
 
     const int64_t t_start = ggml_time_us();
     ggml_backend_graph_compute(backend, gf);
@@ -1288,6 +1288,9 @@ static float device_mem_copy(struct llama_model * model, enum profiler_backend_t
         ggml_backend_cpu_set_n_threads(backend, n_threads);
     }
 
+    // warm-up 
+    ggml_backend_graph_compute(backend, gf);
+
     const int64_t t_start = ggml_time_us();
     ggml_backend_graph_compute(backend, gf);
     const int64_t t_end = ggml_time_us();
diff --git a/src/llama.cpp b/src/llama.cpp
index cd5a95b1..2cc8da15 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -18211,7 +18211,7 @@ static int llama_decode_internal(
                     [&]{ llama_kv_cache_clear       (&lctx); },
                     [&]{ llama_send_kv_cache_clear  (&lctx); },
                     is_last_dev)) {
-            LLAMA_LOG_INFO("%s: received signal kv_cache_clear\n", __func__);
+            LLAMA_LOG_DEBUG("%s: received signal kv_cache_clear\n", __func__);
             return -1;
         }
 
@@ -18219,7 +18219,7 @@ static int llama_decode_internal(
                     [&]{ llama_kv_cache_seq_rm      (&lctx, meta.rm_seq_id, meta.rm_p0, meta.rm_p1); },
                     [&]{ llama_send_kv_cache_seq_rm (&lctx, meta.rm_seq_id, meta.rm_p0, meta.rm_p1); },
                     is_last_dev)) {
-            LLAMA_LOG_INFO("%s: received signal kv_cache_seq_rm\n", __func__);
+            LLAMA_LOG_DEBUG("%s: received signal kv_cache_seq_rm\n", __func__);
             return -1;
         }
 
@@ -18227,7 +18227,7 @@ static int llama_decode_internal(
                     [&]{ llama_kv_cache_seq_add     (&lctx, meta.add_seq_id, meta.add_p0, meta.add_p1, meta.add_delta); },
                     [&]{ llama_send_kv_cache_seq_add(&lctx, meta.add_seq_id, meta.add_p0, meta.add_p1, meta.add_delta); },
                     is_last_dev)) {
-            LLAMA_LOG_INFO("%s: received signal kv_cache_seq_add\n", __func__);
+            LLAMA_LOG_DEBUG("%s: received signal kv_cache_seq_add\n", __func__);
             return -1;
         }
 
@@ -18235,7 +18235,7 @@ static int llama_decode_internal(
                     [&]{ llama_kv_cache_seq_cp      (&lctx, meta.cp_src_seq_id, meta.cp_dst_seq_id, meta.cp_p0, meta.cp_p1); },
                     [&]{ llama_send_kv_cache_seq_cp (&lctx, meta.cp_src_seq_id, meta.cp_dst_seq_id, meta.cp_p0, meta.cp_p1); },
                     is_last_dev)) {
-            LLAMA_LOG_INFO("%s: received signal kv_cache_seq_cp\n", __func__);
+            LLAMA_LOG_DEBUG("%s: received signal kv_cache_seq_cp\n", __func__);
             return -1;
         }
 
@@ -18243,7 +18243,7 @@ static int llama_decode_internal(
                     [&]{ llama_kv_cache_seq_div     (&lctx, meta.div_seq_id, meta.div_p0, meta.div_p1, meta.div_factor); },
                     [&]{ llama_send_kv_cache_seq_div(&lctx, meta.div_seq_id, meta.div_p0, meta.div_p1, meta.div_factor); },
                     is_last_dev)) {
-            LLAMA_LOG_INFO("%s: received signal kv_cache_seq_div\n", __func__);
+            LLAMA_LOG_DEBUG("%s: received signal kv_cache_seq_div\n", __func__);
             return -1;
         }
     }