这是一个学习项目

将ray的04_finetuning_llms_with_deepspeed 重新改造会原生accelerate + deepspeed 使用 llama 3.2 1B模型，在本地8G显卡+offload cpu上完成训练
参考的另一个项目为https://github.com/huggingface/accelerate/blob/main/examples/nlp_example.py
其次为https://huggingface.co/learn/nlp-course

一些学习心得

04_finetuning_llms_with_deepspeed从accelerate examples改造来的时候，在数据集处理方面有改动，特别是最终的datasets需要包含labels列
在小容量显卡运行时，需要及时开启 accelerate + deepspeed zero3 offload cpu能力

Name		Name	Last commit message	Last commit date
Latest commit History 16 Commits
data		data
dataset		dataset
deepspeed_configs		deepspeed_configs
lora_configs		lora_configs
.gitignore		.gitignore
01-datasets-tokenizer.ipynb		01-datasets-tokenizer.ipynb
02-transformers-huggingface-learn.ipynb		02-transformers-huggingface-learn.ipynb
03-llama-3.1-3b-accelerate-deepspeed.ipynb		03-llama-3.1-3b-accelerate-deepspeed.ipynb
04-LLaMA3_1-8B-Instruct_Lora.ipynb		04-LLaMA3_1-8B-Instruct_Lora.ipynb
04_finetuning_llms_with_deepspeed.ipynb		04_finetuning_llms_with_deepspeed.ipynb
05-accelerate-deepspeed-train.ipynb		05-accelerate-deepspeed-train.ipynb
README.md		README.md
accelerate_train.py		accelerate_train.py
alpaca_data.json		alpaca_data.json
alpaca_train.py		alpaca_train.py
create_dataset.py		create_dataset.py
finetune_hf_llm.py		finetune_hf_llm.py
llm_accelerator_deepspeed.py		llm_accelerator_deepspeed.py
lora.py		lora.py
nlp_example.py		nlp_example.py
pytorch.ipynb		pytorch.ipynb
run_llama_ft.sh		run_llama_ft.sh