ref url:

https://modelscope.cn/models/ZJUNLP/DeepKE-LLM/summary

Linux增大Swap分区，增加虚拟内存，以解决内存不足等问题_增大swapfile能缓解内存不足吗-CSDN博客

Add Local visual RAM 100G:

##增加虚拟内存
mkdir /data/VisualRAM
cd VisualRAMdd if=/dev/zero of=swapfile bs=1G count=100
sudo chmod 600 swapfile
mkswap swapfile
sudo swapon swapfile

 add to linux restart config for auto upload device:

sudo gedit /etc/fstab
/data/VisualRAM$/swapfile none swap defaults 0 0
sudo swapon --show

Download llama13b huggingface model:

#模型下载
from modelscope import snapshot_download
model_dir = snapshot_download('skyline2006/llama-13b',cache_dir='/YourModelCatchDir')

Download zhixi diff 16pf:

python tools/download.py --download_path ./zhixi-diff-fp16 --only_base --fp16

merge llama 13b base to zhixi diff 16pf:

first modify tools/weight_diff.py , goto line 107 nearly,  comment a few lines,like this:

    if check_integrity_naively:# This is not a rigorous, cryptographically strong integrity check :)allsum = sum(state_dict_recovered[key].sum() for key in state_dict_recovered)print("allsum={0}, code check number is 94052.2891 or 94046.1875".format(allsum))# comment below codes::# assert torch.allclose(#     allsum, torch.full_like(allsum, fill_value=94052.2891 if not is_fp16 else 94046.1875), atol=1e-2, rtol=0# ), "Naive integrity check failed. This could imply that some of the checkpoint files are corrupted."

second: run code below

python tools/weight_diff.py recover --path_raw ./skyline2006/llama-13b --path_diff ./zhixi-diff-fp16 --path_tuned ./zhixi  --is_fp16 True

Now , all works is ready ,  begin run example:

fix bug:

python examples/generate_finetune.py --base_model ./zhixi

 error:

│                                                                                                  │
│ /home/gw00241401/anaconda3/envs/zhixi/lib/python3.9/site-packages/torch/nn/modules/linear.py:114 │
│ in forward                                                                                       │
│                                                                                                  │
│   111 │   │   │   init.uniform_(self.bias, -bound, bound)                                        │
│   112 │                                                                                          │
│   113 │   def forward(self, input: Tensor) -> Tensor:                                            │
│ ❱ 114 │   │   return F.linear(input, self.weight, self.bias)                                     │
│   115 │                                                                                          │
│   116 │   def extra_repr(self) -> str:                                                           │
│   117 │   │   return f'in_features={self.in_features}, out_features={self.out_features}, bias=   │
╰──────────────────────────────────────────────────────────────────────────────────────────────────╯
RuntimeError: "addmm_impl_cpu_" not implemented for 'Half'

fix method:

1.modify  examples/generate_finetune.py

goto about 90 line,like this is ok

if not load_8bit:if device == "cpu":model.float()else:model.half()  

Error2:

OutOfMemoryError: CUDA out of memory. Tried to allocate 314.00 MiB. GPU 0 has a total capacty of 7.78 GiB of which 260.06 MiB is free. Including non-PyTorch memory, this process has 
6.96 GiB memory in use. Of the allocated memory 6.49 GiB is allocated by PyTorch, and 360.60 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large 
try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

because your GPU is not enough!  go to set CPU run model

for me ,like to du  this:

modify examples/generate_fune.py, goto line 10 about, simply added code at last line

if torch.cuda.is_available():device = "cuda"
else:device = "cpu"
##added
device = "cpu"