evilpsycho / play-with-llms Goto Github PK

是否有增加多轮对话计划。另外是否可以支持adalora

peft==0.4.0.dev0
要怎么下载啊，pip下载时候报错，没有这个版本。添加了各个镜像源也不行

from model import load_model 这个load_model要怎么下载这个包呢

train:
/opt/conda/envs/py310/bin/python train.py --model_name baichuan
--model_path /Data/cj/ML/baichuan-inc/baichuan-7B
--quantization float16
--data_name belle_open_source_500k
--data_path ./data/Belle_open_source_0.5M.json
--train_size 20000
--output_dir ./output/baichuan_lorasft
--seed 42
--max_len 1024
--lora_rank 8
--num_train_epochs 1
--learning_rate 3e-4
--per_device_train_batch_size 4
--per_device_eval_batch_size 8
--gradient_accumulation_steps 1
--logging_steps 10
--evaluation_strategy steps
--save_strategy steps
--eval_steps 100
--save_steps 100
--report_to tensorboard
--save_total_limit 3
--load_best_model_at_end true
--optim adamw_torch

error:
/Data/cj/ML/train_custom_LLM-main/train.py:105 in │
│ │
│ 102 │ model.save_pretrained(training_args.output_dir) │
│ 103 │
│ 104 if name == "main": │
│ ❱ 105 │ main() │
│ │
│ /Data/cj/ML/train_custom_LLM-main/train.py:74 in main │
│ │
│ 71 │ ) │
│ 72 │ │
│ 73 │ print(config) │
│ ❱ 74 │ model = get_peft_model(model, config) │
│ 75 │ │
│ 76 │ ############# prepare data ########### │
│ 77 │ data = eval(args.data_name)(args.data_path, tokenizer, args.max_le │
│
ValueError: Target modules [] not found in the base model. Please check the target modules and try again.

后面p-tuning能讲讲吗

哈喽，当前拉的最新peft 没有0.4.0dev

请问多轮对话微调和推理具体如何实现呀，发现训练的时候dataset.py实际没有用到

This repo with videos is helpful to me.
Nevetheless, I still met some issues.
May the following tips save your time.

1. [Linux:Yes, Windows:No] Bitsandbytes is officially maintained on Linux. Hence, you may not use the 4-bit, 8-bit model on a Windows platform.
2. [Need fine-tune:Yes, Directly use:No] Firstly, you should train the 4-bit model from scratch. Secondly, you can get the fine-tuned model in the output folder. Finally, you can check the results in the notebook and compare the performance with/without fine-tuning.
3. [P40x1-4-bit-model:Yes, V100x1-bfloat16-model:No] Currently, you can fine-tune the 4-bit-model with a P40-GPU (24G) on Linux, but you can not fine-tune the bfloat16 model with a V100-GPU(32G) on Windows.
4. [Training time based on the default settings]: 4950 iteration in total, 300 iterations for the first 55 min, hence, the whole fine-tunning would take about 15 hours.

可以在windows10上跑吗？

对于CUDA大于等于12.0的，需要直接安装bitsandbytes最新版本（我这边显示不兼容）：pip install -U bitsandbytes
除了requirements.txt里的，还要安装其他包：datasets、sentencepiece、scipy

我有一点疑惑，请问为什么没有mask将instruction+input 的损失mask掉。看代码的意思是输入为instruction+input+output,然后预测目标也是instruction+input+output。

ValueError: Can't find 'adapter_config.json' at './output/baichuan_lorasft/'

when i try to train the model，It was all normal at first，but some times later ,it get a strange mistake，how can i solve it

在微调的时候，200个迭代之后，开始eval，显存会直接飙升一截，并且不再下降。不知道这个泄露是什么引起的

result = tokenizer(
prompt,
truncation=True,
max_length=max_len,
padding=False,
return_tensors=None,
)
为什么result的input 没有 bos 和 eos 标志
而且 result["attention_mask"].append(1) 报错
提示result["attention_mask"]为numpy类型 [[Fasle,.........True],[False,False,False...False]]