Commit 30b6a755e for llama.cpp

commit 30b6a755e29692e8bc8e072885325716a2fee70f
Author: Georgi Gerganov <ggerganov@gmail.com>
Date:   Wed Sep 9 07:50:38 2026 +0300

    tests : use less threads for data initialization (#28325)

    * tests : use 1 thread for data initialization

    * cont : scale threads with number of elements

    * cont : adjust

diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 01e72041b..8030186fb 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -56,7 +56,7 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m
     std::vector<float> data(nels);
     {
         // parallel initialization
-        static const size_t n_threads = N_THREADS;
+        static const size_t n_threads = std::max<size_t>(1, std::min<size_t>(nels/1024, std::min<size_t>(4, N_THREADS/2)));

         auto init_thread = [&](size_t start, size_t end) {
             thread_local std::default_random_engine gen(std::random_device{}());
@@ -11384,10 +11384,17 @@ static bool op_names_filter_selects(const char * op_names_filter, const char * o
 // Covers padded rows, sinks, kvpad, multi-SIMDgroup reduction, quantized K/V, and MLA views.
 // The override is backend-global, so this runs after all parallel workers have joined.
 static bool run_fa_vec_slice(ggml_backend_t backend, ggml_backend_t backend_cpu, const char * op_names_filter) {
+    const char * LLAMA_TEST_FA_VEC_DISABLE = getenv("LLAMA_TEST_FA_VEC_DISABLE");
+    if (LLAMA_TEST_FA_VEC_DISABLE) {
+        return true;
+    }
+
     if (!op_names_filter_selects(op_names_filter, "FLASH_ATTN_EXT")) {
         return true;
     }

+    printf("Running FA vec slice tests (env LLAMA_TEST_FA_VEC_DISABLE=1 to skip)\n");
+
     auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend));

     auto set_ov   = (set_fa_vec_override_t)   ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_set_fa_vec_override");