From e19ead6e3f30a8c0c944e238d90caf5902c92415 Mon Sep 17 00:00:00 2001 From: xaedes Date: Fri, 19 May 2023 18:40:20 +0200 Subject: [PATCH] print used memory before and after optimization --- examples/baby-llama/baby-llama-text.cpp | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/examples/baby-llama/baby-llama-text.cpp b/examples/baby-llama/baby-llama-text.cpp index e65d2d186..099863bb8 100644 --- a/examples/baby-llama/baby-llama-text.cpp +++ b/examples/baby-llama/baby-llama-text.cpp @@ -1637,6 +1637,8 @@ int main(int argc, char ** argv) { ggml_build_forward_expand(&gf, e); ggml_graph_compute(ctx0, &gf); + size_t used_mem_before_opt = ggml_used_mem(ctx0); + float error_before_opt = ggml_get_f32_1d(e, 0); struct ggml_opt_params opt_params_adam = ggml_opt_default_params(GGML_OPT_ADAM); @@ -1645,6 +1647,7 @@ int main(int argc, char ** argv) { opt_params_adam.print_backward_graph = false; opt_params_adam.n_threads = gf.n_threads; opt_params_adam.adam.n_iter = 16; + opt_params_adam.adam.alpha = 1e-4; opt_params_lbfgs.print_forward_graph = false; opt_params_lbfgs.print_backward_graph = false; @@ -1658,6 +1661,8 @@ int main(int argc, char ** argv) { ggml_opt(ctx0, opt_params_lbfgs, e); } + size_t used_mem_after_opt = ggml_used_mem(ctx0); + model.train_its += use_adam ? opt_params_adam.adam.n_iter : opt_params_lbfgs.lbfgs.n_iter; model.train_samples += n_batch; @@ -1666,6 +1671,9 @@ int main(int argc, char ** argv) { float error_after_opt = ggml_get_f32_1d(e, 0); + printf("used_mem_before_opt: %zu bytes\n", used_mem_before_opt); + printf("used_mem_after_opt: %zu bytes\n", used_mem_after_opt); + if (ex % 1 == 0) { printf("Example %d\n", ex); printf("error_before_opt: %.6f\n", error_before_opt);