From 00565e85a02d805e3155336e757b87cd14efa6b4 Mon Sep 17 00:00:00 2001 From: Galaxy1458 <55453380+Galaxy1458@users.noreply.github.com> Date: Tue, 14 May 2024 21:49:12 +0800 Subject: [PATCH] Update llama_npu_opt_lora.sh --- llm/llama/npu/llama_npu_opt_lora.sh | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/llm/llama/npu/llama_npu_opt_lora.sh b/llm/llama/npu/llama_npu_opt_lora.sh index fd1b004a8af5..fa596b2ba648 100644 --- a/llm/llama/npu/llama_npu_opt_lora.sh +++ b/llm/llama/npu/llama_npu_opt_lora.sh @@ -38,7 +38,7 @@ python -u -m paddle.distributed.launch \ --dataset_name_or_path "data/" \ --output_dir "./output/lora_bf16_llama_N1C8" \ --per_device_train_batch_size 2 \ - --gradient_accumulation_steps 16 \ + --gradient_accumulation_steps 32 \ --per_device_eval_batch_size 1 \ --eval_accumulation_steps 1 \ --max_steps ${max_steps} \ @@ -57,7 +57,7 @@ python -u -m paddle.distributed.launch \ --eval_with_do_generation false \ --metric_for_best_model "accuracy" \ --recompute false \ - --tensor_parallel_degree 8 \ + --tensor_parallel_degree 4 \ --pipeline_parallel_degree 1 \ --zero_padding 0 \ --sequence_parallel 1 \