diff --git a/ggml/src/ggml-metal.metal b/ggml/src/ggml-metal.metal
index 05a4f759b..971257db4 100644
--- a/ggml/src/ggml-metal.metal
+++ b/ggml/src/ggml-metal.metal
@@ -1985,6 +1985,10 @@ kernel void kernel_im2col_ext(
     const int32_t HW = tgpig[0] % KHW;
     
     const int32_t tpitg_0 = (d * ntg[0]) + tpitg[0];
+    if (tpitg_0 >= N) {
+        return;
+    }
+
     const int32_t tpitg_1 = HW / KW;
     const int32_t tpitg_2 = HW % KW;
 
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index a9bcb1c1c..d5ebd5cf6 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -3323,6 +3323,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 1024}, {3, 3, 2, 1024}, 1, 1, 1, 1, 1, 1, true));
     test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 2048}, {3, 3, 1, 2048}, 1, 1, 1, 1, 1, 1, true));
     test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 2048}, {3, 3, 2, 2048}, 1, 1, 1, 1, 1, 1, true));
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 2560}, {3, 3, 1, 2560}, 1, 1, 1, 1, 1, 1, true));
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 2560}, {3, 3, 2, 2560}, 1, 1, 1, 1, 1, 1, true));
     
     // sycl backend will limit task global_range < MAX_INT
     // test cases for 2D im2col with large input W and H (occurs in stable-diffusion)