3090 cudamemory 不够用25GB显存

illumionous commented 1 year ago

提交前必须检查以下项目

[X] 请确保使用的是仓库最新代码（git pull），一些问题已被解决和修复。
[X] 由于相关依赖频繁更新，请确保按照Wiki中的相关步骤执行
[X] 我已阅读FAQ章节并且已在Issue中对问题进行了搜索，没有找到相似问题和解决方案
[X] 第三方插件问题：例如llama.cpp、text-generation-webui、LlamaChat等，同时建议到对应的项目中查找解决方案
[X] 模型正确性检查：务必检查模型的SHA256.md，模型不对的情况下无法保证效果和正常运行

问题类型

模型训练与精调

基础模型

LLaMA-13B

操作系统

Linux

详细描述问题

r=3e-4
lora_rank=8
lora_alpha=16
lora_trainable="q_proj,v_proj,k_proj,o_proj,gate_proj,down_proj,up_proj"
#lora_trainable="q_proj,v_proj"
#modules_to_save="embed_tokens,lm_head"
modules_to_save=""
lora_dropout=0.05

pretrained_model=/home/wilsontam/pyllama_data/13B_hf
tokenizer_path=/home/wilsontam/pyllama_data/13B_hf

dataset_dir=data.instruct2/train
per_device_train_batch_size=1
per_device_eval_batch_size=1
training_steps=300
gradient_accumulation_steps=16 # 1 GPU, 32 batch per GPU = 128 batch for 1step update
output_dir=outs/13B_instruct_8-1
#peft_model=path/to/peft/model/dir
validation_file=data.instruct2/val/8-1_val.json
force_resize_embeddings=False

当我用这个脚本去精调指令时，发生memoryout，首先是词袋是The vocab size of the tokenizer must be 49954, but found49953，我修改代码定义为49953，然后又发生memory不够

依赖情况（代码类问题务必提供）

# 请在此处粘贴依赖情况

依赖完全依照教程

运行日志或截图

rainable params: 31293440 || all params: 13231006720 || trainable%: 0.23651594064038084
08/02/2023 12:34:08 - INFO - __main__ - model.modules_to_save: None
Traceback (most recent call last):
  File "/home/nlp/lama/Chinese-LLaMA-Alpaca/./scripts/training/run_clm_sft_with_peft.py", line 441, in <module>
    main()
  File "/home/nlp/lama/Chinese-LLaMA-Alpaca/./scripts/training/run_clm_sft_with_peft.py", line 397, in main
    trainer = Trainer(
  File "/home/nlp/.conda/envs/finetune/lib/python3.10/site-packages/transformers/trainer.py", line 499, in __init__
    self._move_model_to_device(model, args.device)
  File "/home/nlp/.conda/envs/finetune/lib/python3.10/site-packages/transformers/trainer.py", line 741, in _move_model_to_device
    model = model.to(device)
  File "/home/nlp/.conda/envs/finetune/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1145, in to
    return self._apply(convert)
  File "/home/nlp/.conda/envs/finetune/lib/python3.10/site-packages/torch/nn/modules/module.py", line 797, in _apply
    module._apply(fn)
  File "/home/nlp/.conda/envs/finetune/lib/python3.10/site-packages/torch/nn/modules/module.py", line 797, in _apply
    module._apply(fn)
  File "/home/nlp/.conda/envs/finetune/lib/python3.10/site-packages/torch/nn/modules/module.py", line 797, in _apply
    module._apply(fn)
  [Previous line repeated 4 more times]
  File "/home/nlp/.conda/envs/finetune/lib/python3.10/site-packages/torch/nn/modules/module.py", line 820, in _apply
    param_applied = fn(param)
  File "/home/nlp/.conda/envs/finetune/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1143, in convert
    return t.to(device, dtype if t.is_floating_point() or t.is_complex() else None, non_blocking)
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 136.00 MiB (GPU 0; 23.70 GiB total capacity; 22.79 GiB already allocated; 130.00 MiB free; 22.79 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
ERROR:torch.distributed.elastic.multiprocessing.api:failed (exitcode: 1) local_rank: 0 (pid: 3411336) of binary: /home/nlp/.conda/envs/finetune/bin/python
Traceback (most recent call last):
  File "/home/nlp/.conda/envs/finetune/bin/torchrun", line 8, in <module>
    sys.exit(main())
  File "/home/nlp/.conda/envs/finetune/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 346, in wrapper
    return f(*args, **kwargs)
  File "/home/nlp/.conda/envs/finetune/lib/python3.10/site-packages/torch/distributed/run.py", line 794, in main
    run(args)
  File "/home/nlp/.conda/envs/finetune/lib/python3.10/site-packages/torch/distributed/run.py", line 785, in run
    elastic_launch(
  File "/home/nlp/.conda/envs/finetune/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 134, in __call__
    return launch_agent(self._config, self._entrypoint, list(args))
  File "/home/nlp/.conda/envs/finetune/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 250, in launch_agent
    raise ChildFailedError(
torch.distributed.elastic.multiprocessing.errors.ChildFailedError: 
============================================================
./scripts/training/run_clm_sft_with_peft.py FAILED

bigcash commented 1 year ago

13B应该不止需要25G显存吧，用zero3训练吧

github-actions[bot] commented 1 year ago

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your consideration.

github-actions[bot] commented 1 year ago

Closing the issue, since no updates observed. Feel free to re-open if you need any further assistance.

ymcui / Chinese-LLaMA-Alpaca