评估集上loss、学习率为0

[sly123197811]

Reminder

• I have read the README and searched the existing issues.

Reproduction

CUDA_VISIBLE_DEVICES=1 python src/train_bash.py
--stage sft
--do_train True
--model_name_or_path /home/wzb/yan/Meta-Llama-3-8B-Instruct
--finetuning_type lora
--template default
--flash_attn auto
--dataset_dir data
--dataset alpaca_gpt4_zh
--cutoff_len 1024
--learning_rate 5e-05
--num_train_epochs 3.0
--max_samples 100000
--per_device_train_batch_size 2
--gradient_accumulation_steps 8
--lr_scheduler_type cosine
--max_grad_norm 1.0
--logging_steps 5
--save_steps 100
--warmup_steps 0
--optim adamw_torch
--report_to none
--output_dir saves/LLaMA3-8B/lora/train_2024-04-25-15-45-51
--fp16 True
--lora_rank 8
--lora_alpha 16
--lora_dropout 0
--lora_target q_proj,v_proj
--val_size 0.15
--evaluation_strategy steps
--eval_steps 100
--per_device_eval_batch_size 2
--load_best_model_at_end True
--plot_loss True
训练过程曲线正常，评估时，loss和学习率均为0
日志如下：

[INFO|trainer.py:2057] 2024-04-25 17:21:02,619 >> Number of trainable parameters = 3,407,872

[INFO|callbacks.py:145] 2024-04-25 17:26:24,858 >> {'loss': 1.6616, 'learning_rate': 5.0000e-05, 'epoch': 0.00}

[INFO|callbacks.py:145] 2024-04-25 17:31:29,975 >> {'loss': 1.6778, 'learning_rate': 5.0000e-05, 'epoch': 0.00}

[INFO|callbacks.py:145] 2024-04-25 17:36:14,088 >> {'loss': 1.6495, 'learning_rate': 5.0000e-05, 'epoch': 0.01}

[INFO|callbacks.py:145] 2024-04-25 17:41:43,950 >> {'loss': 1.6450, 'learning_rate': 4.9999e-05, 'epoch': 0.01}

[INFO|callbacks.py:145] 2024-04-25 17:46:40,556 >> {'loss': 1.6643, 'learning_rate': 4.9999e-05, 'epoch': 0.01}

[INFO|callbacks.py:145] 2024-04-25 17:52:00,261 >> {'loss': 1.5792, 'learning_rate': 4.9998e-05, 'epoch': 0.01}

[INFO|callbacks.py:145] 2024-04-25 17:57:28,697 >> {'loss': 1.6068, 'learning_rate': 4.9998e-05, 'epoch': 0.01}

[INFO|callbacks.py:145] 2024-04-25 18:03:06,412 >> {'loss': 1.5097, 'learning_rate': 4.9997e-05, 'epoch': 0.02}

[INFO|callbacks.py:145] 2024-04-25 18:08:30,067 >> {'loss': 1.5304, 'learning_rate': 4.9996e-05, 'epoch': 0.02}

[INFO|callbacks.py:145] 2024-04-25 18:14:13,946 >> {'loss': 1.4504, 'learning_rate': 4.9995e-05, 'epoch': 0.02}

[INFO|callbacks.py:145] 2024-04-25 18:19:42,074 >> {'loss': 1.5191, 'learning_rate': 4.9994e-05, 'epoch': 0.02}

[INFO|callbacks.py:145] 2024-04-25 18:25:20,252 >> {'loss': 1.3741, 'learning_rate': 4.9993e-05, 'epoch': 0.02}

[INFO|callbacks.py:145] 2024-04-25 18:31:06,083 >> {'loss': 1.3848, 'learning_rate': 4.9991e-05, 'epoch': 0.03}

[INFO|callbacks.py:145] 2024-04-25 18:36:26,379 >> {'loss': 1.4360, 'learning_rate': 4.9990e-05, 'epoch': 0.03}

[INFO|callbacks.py:145] 2024-04-25 18:42:14,617 >> {'loss': 1.4140, 'learning_rate': 4.9989e-05, 'epoch': 0.03}

[INFO|callbacks.py:145] 2024-04-25 18:46:53,814 >> {'loss': 1.4261, 'learning_rate': 4.9987e-05, 'epoch': 0.03}

[INFO|callbacks.py:145] 2024-04-25 18:51:51,722 >> {'loss': 1.4251, 'learning_rate': 4.9985e-05, 'epoch': 0.03}

[INFO|callbacks.py:145] 2024-04-25 18:56:52,623 >> {'loss': 1.4112, 'learning_rate': 4.9983e-05, 'epoch': 0.03}

[INFO|callbacks.py:145] 2024-04-25 19:01:55,584 >> {'loss': 1.3598, 'learning_rate': 4.9982e-05, 'epoch': 0.04}

[INFO|callbacks.py:145] 2024-04-25 19:07:13,635 >> {'loss': 1.4213, 'learning_rate': 4.9980e-05, 'epoch': 0.04}

[INFO|trainer.py:3614] 2024-04-25 19:07:13,638 >> ***** Running Evaluation *****

[INFO|trainer.py:3616] 2024-04-25 19:07:13,638 >> Num examples = 7323

[INFO|trainer.py:3619] 2024-04-25 19:07:13,639 >> Batch size = 2

[INFO|callbacks.py:145] 2024-04-25 21:25:17,654 >> {'loss': 0.0000, 'learning_rate': 0.0000e+00, 'epoch': 0.04}

[INFO|trainer.py:3305] 2024-04-25 21:25:17,655 >> Saving model checkpoint to saves/LLaMA3-8B/lora/train_2024-04-25-17-18-36/checkpoint-100

Expected behavior

No response

System Info

• transformers version: 4.29.1
• Platform: Linux-5.15.0-82-generic-x86_64-with-glibc2.17
• Python version: 3.8.18
• Huggingface_hub version: 0.22.2
• Safetensors version: not installed
• PyTorch version (GPU?): 2.0.1+cu117 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using GPU in script?:
• Using distributed or parallel set-up in script?:

Others

No response

[hiyouga]

是 log 显示的问题，不影响训练