From d1998bf20e55ad90054dacfa2de514533045f594 Mon Sep 17 00:00:00 2001 From: Andrew Godfrey Date: Fri, 17 Nov 2023 02:23:11 -0800 Subject: [PATCH] finetune : zero the loraB initial vectors (#4082) * finetune : zero the loraB initial vectors Without this, the first iteration is starting out far from the base model, instead of exactly on it. Zeroing loraB is what the paper recommends. loralib also zeroes at least one of the init vector pairs (though it departs from the paper in using a different distribution for the other vector, in some cases). * tabs to spaces * Use ggml_set_zero instead of adding a new function --- examples/finetune/finetune.cpp | 24 ++++++++++++------------ 1 file changed, 12 insertions(+), 12 deletions(-) diff --git a/examples/finetune/finetune.cpp b/examples/finetune/finetune.cpp index 5a6cf22c..7fecce25 100644 --- a/examples/finetune/finetune.cpp +++ b/examples/finetune/finetune.cpp @@ -548,35 +548,35 @@ static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, fl struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max); randomize_tensor_normal(lora->tok_embeddings_a, rnd); - randomize_tensor_normal(lora->tok_embeddings_b, rnd); + ggml_set_zero(lora->tok_embeddings_b); randomize_tensor_normal(lora->norm_a, rnd); - randomize_tensor_normal(lora->norm_b, rnd); + ggml_set_zero(lora->norm_b); randomize_tensor_normal(lora->output_a, rnd); - randomize_tensor_normal(lora->output_b, rnd); + ggml_set_zero(lora->output_b); for (uint32_t i = 0; i < n_layer; ++i) { auto & layer = lora->layers[i]; randomize_tensor_normal(layer.attention_norm_a, rnd); - randomize_tensor_normal(layer.attention_norm_b, rnd); + ggml_set_zero(layer.attention_norm_b); randomize_tensor_normal(layer.wq_a, rnd); - randomize_tensor_normal(layer.wq_b, rnd); + ggml_set_zero(layer.wq_b); randomize_tensor_normal(layer.wk_a, rnd); - randomize_tensor_normal(layer.wk_b, rnd); + ggml_set_zero(layer.wk_b); randomize_tensor_normal(layer.wv_a, rnd); - randomize_tensor_normal(layer.wv_b, rnd); + ggml_set_zero(layer.wv_b); randomize_tensor_normal(layer.wo_a, rnd); - randomize_tensor_normal(layer.wo_b, rnd); + ggml_set_zero(layer.wo_b); randomize_tensor_normal(layer.ffn_norm_a, rnd); - randomize_tensor_normal(layer.ffn_norm_b, rnd); + ggml_set_zero(layer.ffn_norm_b); randomize_tensor_normal(layer.w1_a, rnd); - randomize_tensor_normal(layer.w1_b, rnd); + ggml_set_zero(layer.w1_b); randomize_tensor_normal(layer.w2_a, rnd); - randomize_tensor_normal(layer.w2_b, rnd); + ggml_set_zero(layer.w2_b); randomize_tensor_normal(layer.w3_a, rnd); - randomize_tensor_normal(layer.w3_b, rnd); + ggml_set_zero(layer.w3_b); } free_random_normal_distribution(rnd);