From 9eaf105ae8c9ded91df8a20dff96fe23897db95f Mon Sep 17 00:00:00 2001
From: Kawrakow <iwankawrakow@gmail.com>
Date: Thu, 26 Mar 2026 10:42:35 +0100
Subject: [PATCH] Print info when allocating large amounts of pinned host
 memory (#1517)

---
 ggml/src/ggml-cuda.cu | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu
index 205747e6..a3b0a418 100644
--- a/ggml/src/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda.cu
@@ -1261,9 +1261,19 @@ static void * ggml_cuda_host_malloc(size_t size) {
     if (getenv("GGML_CUDA_NO_PINNED") != nullptr) {
         return nullptr;
     }
+    constexpr double k_warn_limit = 8.0;
 
     void * ptr = nullptr;
+    double size_GiB = size/(1024.*1024.*1024.);
+    auto tim1 = ggml_time_us();
+    if (size_GiB > k_warn_limit) {
+        GGML_CUDA_LOG_INFO("Allocating %.2f GiB of pinned host memory, this can take a while...\n", size_GiB);
+    }
     cudaError_t err = cudaMallocHost((void **) &ptr, size);
+    if (size_GiB > k_warn_limit) {
+        auto tim2 = ggml_time_us();
+        GGML_CUDA_LOG_INFO("    done allocating %.2f GiB in %.1f ms\n", size_GiB, 1e-3*(tim2-tim1));
+    }
     if (err != cudaSuccess) {
         // clear the error
         cudaGetLastError();