Skip to content

Conversation

huangzhuxing
Copy link

Checkpoint export (export_hf_checkpoint.py)

These files contain scripts that merge the QLoRA weights back into the base model
for export to Hugging Face format .

  • Example:
$cd /mnt/e/PycharmProjects/qlora
$export BASE_MODEL=huggyllama/llama-30b 
$export LORA_MODEL=/mnt/e/PycharmProjects/qlora/output/guanaco-33b/checkpoint-1500/adapter_model
$export HF_CHECKPOINT=/mnt/e/PycharmProjects/qlora/output/guanaco-33b/hf

$python export_hf_checkpoint.py
CUDA SETUP: CUDA runtime path found: /home/hzx/.conda/envs/qlora/lib/libcudart.so.11.0
CUDA SETUP: Highest compute capability among GPUs detected: 8.9
CUDA SETUP: Detected CUDA version 118
CUDA SETUP: Loading binary /home/hzx/.local/lib/python3.10/site-packages/bitsandbytes/libbitsandbytes_cuda118.so...
Loading checkpoint shards: 100%|█████████████████████████████████████| 7/7 [00:02<00:00,  3.28it/s]total 63533373

  • If you need to run it directly, you need to copy the three files special_tokens_map.json, tokenizer.model and tokenizer_config.json from the checkpoint-xxxx directory to the /mnt/e/PycharmProjects/qlora/output/guanaco-33b/hf directory to run

  • Final Results

$ ls -lrt /mnt/e/PycharmProjects/qlora/output/guanaco-33b/hf
total 63533869
-rwxrwxrwx 1 hzx hzx        727 Jun  4 17:26 tokenizer_config.json
-rwxrwxrwx 1 hzx hzx         96 Jun  4 17:26 special_tokens_map.json
-rwxrwxrwx 1 hzx hzx     499723 Jun  4 17:26 tokenizer.model
-rwxrwxrwx 1 hzx hzx        607 Jun  5 22:45 config.json
-rwxrwxrwx 1 hzx hzx        137 Jun  5 22:45 generation_config.json
-rwxrwxrwx 1 hzx hzx 9818324627 Jun  5 22:45 pytorch_model-00001-of-00007.bin
-rwxrwxrwx 1 hzx hzx 9869497721 Jun  5 22:46 pytorch_model-00002-of-00007.bin
-rwxrwxrwx 1 hzx hzx 9896734097 Jun  5 22:46 pytorch_model-00003-of-00007.bin
-rwxrwxrwx 1 hzx hzx 9719524707 Jun  5 22:46 pytorch_model-00004-of-00007.bin
-rwxrwxrwx 1 hzx hzx 9869470481 Jun  5 22:46 pytorch_model-00005-of-00007.bin
-rwxrwxrwx 1 hzx hzx 9869470445 Jun  5 22:47 pytorch_model-00006-of-00007.bin
-rwxrwxrwx 1 hzx hzx 6015086981 Jun  5 22:47 pytorch_model-00007-of-00007.bin
-rwxrwxrwx 1 hzx hzx      50084 Jun  5 22:47 pytorch_model.bin.index.json

… base model

for export to Hugging Face format .
@bqcao
Copy link

bqcao commented Jun 7, 2023

I got the following erros:

lora_model = PeftModel.from_pretrained(base_model,LORA_MODEL,device_map={"": DEVICE},torch_dtype=torch.float16)
*** TypeError: init() got an unexpected keyword argument 'device_map'

lora_model = PeftModel.from_pretrained(base_model,LORA_MODEL,torch_dtype=torch.float16)
*** TypeError: init() got an unexpected keyword argument 'torch_dtype'

My peft version is 0.4.0.dev0

Any help to fix the error here?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants