LLaMA¶

Chinese LLaMA¶

环境配置¶

pip install sentencepiece
pip install scikit-learn
pip install --upgrade pytest
pip install deepspeed

模型转换与合并¶

$ python merge_llama_with_chinese_lora.py \
    --base_model /root/llama-7b-hf \
    --lora_model /root/chinese_llama_plus_lora_7b \
    --output_dir /root/chinese-llama-7b \
    --output_type pth

# check sha256sum of file
$ sha256sum /root/chinese-llama-7b/consolidated.00.pth
# f8d380d63f77a08b7f447f5ec63f0bb1cde9ddeae2207e9f86e6b5f0f95a7955

$ python merge_llama_with_chinese_lora.py \
    --base_model /root/llama-7b-hf \
    --lora_model /root/chinese_llama_plus_lora_7b \
    --output_dir /root/chinese-llama-7b \
    --output_type huggingface

Base model: /root/llama-7b-hf
LoRA model(s) ['/root/chinese_llama_plus_lora_7b']:
Loading checkpoint shards: 100%|█████████████████████████████████████| 2/2 [00:10<00:00,  5.33s/it]
Peft version: 0.3.0
Loading LoRA for 7B model
Loading LoRA /root/chinese_llama_plus_lora_7b...
base_model vocab size: 32000
tokenizer vocab size: 49953
Extended vocabulary size to 49953
Loading LoRA weights
Merging with merge_and_unload...
Saving to Hugging Face format...

$ python merge_llama_with_chinese_lora.py \
    --base_model /root/llama-7b-hf \
    --lora_model /root/chinese_llama_plus_lora_7b,/root/chinese_alpaca_plus_lora_7b \
    --output_dir /root/chinese-llama-alpaca-7b \
    --output_type huggingface
Base model: /root/llama-7b-hf
LoRA model(s) ['/root/chinese_llama_plus_lora_7', '/root/chinese_alpaca_plus_lora_7']:
Loading checkpoint shards: 100%|█████████████████████████████████████| 2/2 [00:10<00:00,  5.08s/it]
Peft version: 0.3.0
Loading LoRA for 7B model
Loading LoRA /root/chinese_llama_plus_lora_7...
base_model vocab size: 32000
tokenizer vocab size: 49953
Extended vocabulary size to 49953
Loading LoRA weights
Merging with merge_and_unload...
Loading LoRA /root/chinese_alpaca_plus_lora_7...
base_model vocab size: 49953
tokenizer vocab size: 49954
Extended vocabulary size to 49954
Loading LoRA weights
Merging with merge_and_unload...
Saving to Hugging Face format...

模型推理¶

python inference/inference_hf.py \
    --base_model /root/chinese-llama-7b \
    --lora_model /root/chinese-llama-7b_lora \
    --interactive

python inference/inference_hf.py \
    --base_model /root/chinese-llama-alpaca-7b \
    --with_prompt \
    --interactive

模型训练¶

lr=2e-4
lora_rank=8
lora_alpha=32
lora_trainable="q_proj,v_proj,k_proj,o_proj,gate_proj,down_proj,up_proj"
modules_to_save="embed_tokens,lm_head"
lora_dropout=0.05

pretrained_model=/model/chinese-llama-7b
chinese_tokenizer_path=/model/chinese-llama-7b
dataset_dir=/root/datasets
data_cache=/root/cache
per_device_train_batch_size=1
per_device_eval_batch_size=1
gradient_accumulation_steps=8
output_dir=/model/pstory78

deepspeed_config_file=ds_zero2_no_offload.json

export TF_CPP_MIN_LOG_LEVEL=3
export HF_HUB_OFFLINE=1

torchrun --nnodes 1 --nproc_per_node 1 run_clm_pt_with_peft.py \
    --deepspeed ${deepspeed_config_file} \
    --model_name_or_path ${pretrained_model} \
    --tokenizer_name_or_path ${chinese_tokenizer_path} \
    --dataset_dir ${dataset_dir} \
    --data_cache_dir ${data_cache} \
    --validation_split_percentage 0.001 \
    --per_device_train_batch_size ${per_device_train_batch_size} \
    --per_device_eval_batch_size ${per_device_eval_batch_size} \
    --do_train \
    --do_eval \
    --seed $RANDOM \
    --fp16 \
    --num_train_epochs 10 \
    --report_to tensorboard \
    --lr_scheduler_type cosine \
    --learning_rate ${lr} \
    --warmup_ratio 0.05 \
    --weight_decay 0.01 \
    --logging_strategy steps \
    --logging_steps 50 \
    --eval_steps 500 \
    --evaluation_strategy steps \
    --save_strategy steps \
    --save_total_limit 10 \
    --save_steps 500 \
    --gradient_accumulation_steps ${gradient_accumulation_steps} \
    --preprocessing_num_workers 8 \
    --block_size 512 \
    --output_dir ${output_dir} \
    --overwrite_output_dir \
    --ddp_timeout 30000 \
    --logging_first_step True \
    --lora_rank ${lora_rank} \
    --lora_alpha ${lora_alpha} \
    --trainable ${lora_trainable} \
    --modules_to_save ${modules_to_save} \
    --lora_dropout ${lora_dropout} \
    --torch_dtype float16 \
    --gradient_checkpointing \
    --ddp_find_unused_parameters False

问题报错解决¶

提示：Setting pad_token_id to eos_token_id:2 for open-end generation. Suppress HuggingFace logging warning: “Setting pad_token_id to eos_token_id:{eos_token_id} for open-end generation.” - Stack Overflow

generation_output = model.generate(
    input_ids = inputs["input_ids"].to(device),
    attention_mask = inputs['attention_mask'].to(device),
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.eos_token_id,
    # pad_token_id=tokenizer.pad_token_id,
    **generation_config
)

RuntimeError: Failed to import transformers.testing_utils because of the following error (look up to see its traceback): No module named ‘_pytest’ 没有安装pytest或者版本太老

pip install --upgrade pytest