GPTQ model merged with LoRA become big size after merging

[dahara111]

Please check that this issue hasn't been reported before.

• I searched previous Bug Reports didn't find any similar reports.

Expected Behavior

If I create a LORA for a GPTQ model and merge it with the original GPTQ model, I should be able to create a model with approximately the same size.

Current behaviour

I used examples/llama-2/gptq-lora.yml.

After creating lora with reference to , I merged it with the following command.

CUDA_VISIBLE_DEVICES="" python3 -m axolotl.cli.merge_lora ../../gptq_qlora/axolotl/gptq-lora.yml --lora_model_dir="../../gptq_qlora/axolotl/checkpoint-800" --load_in_8bit=False --load_in_4bit=False

The original GPTQ version model is 3GB.
The created LoRA is 100MB (optimizer.pt + adapter_model.bin).
However, after merging LoRA, the size is 12.6G (pytorch_model-00001-of-00002.bin + pytorch_model-00002-of-00002.bin).
--load_in_4bit=True made no difference

Steps to reproduce

create lora from gptq model and merge with gptq model.

Config yaml

base_model: webbigdata/ALMA-7B-Ja-GPTQ-Ja-En
is_llama_derived_model: false
gptq: true
gptq_disable_exllama: true
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
tokenizer_use_fast: true
tokenizer_legacy: true
trust_remote_code: true
load_in_8bit: false
load_in_4bit: false
strict: false
push_dataset_to_hub:
hf_use_auth_token: true
datasets:
  - path: alma-ja-v2_gptq/converted_file.jsonl
    type: alpaca
dataset_prepared_path:
val_set_size: 0.01
adapter: lora
lora_model_dir:
sequence_len: 4096
sample_packing:
lora_r: 8
lora_alpha: 32
lora_dropout: 0.05
lora_target_modules:
  - k_proj
  - o_proj
  - q_proj
  - v_proj
lora_target_linear:
lora_fan_in_fan_out:
wandb_project: test
wandb_watch:
wandb_run_id:
wandb_log_model:
output_dir: ./model-out
gradient_accumulation_steps: 2
micro_batch_size: 2
num_epochs: 4
optimizer: adamw_torch
adam_beta2: 0.95
adam_eps: 0.00001
max_grad_norm: 1.0
torchdistx_path:
lr_scheduler: cosine
lr_quadratic_warmup: true
learning_rate: 0.000017
train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
float16: true
tf32: true
gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention:
sdp_attention:
flash_optimum:
warmup_steps: 100
eval_steps: 50
save_steps: 200
debug:
deepspeed:
weight_decay: 0.1
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"

Possible solution

No response

Which Operating Systems are you using?

• Linux
• macOS
• Windows

Python Version

3.10

axolotl branch-commit

axolotl 0.3.0

Acknowledgements

• My issue title is concise, descriptive, and in title casing.
• I have searched the existing issues to make sure this bug has not been reported yet.
• I am using the latest version of axolotl.
• I have provided enough information for the maintainers to reproduce and diagnose the issue.

[Praful932]

Hi, afaik, merging isn't supported for gptq models - link

[webbigdata-jp]

@Praful932
Hi, Thank you very much.
Okay, So, if I want a single model that has been fine-tuned and GPTQ quantized,

Lora fine-tuning -> merge into one file -> GPTQ -> OK
GPTQ -> fine-tuning -> I can, but merging into one file is not supported.

[Foreist]

@Praful932 안녕하세요. 정말 감사합니다. 좋습니다. 미세 조정되고 GPTQ 양자화된 단일 모델을 원한다면

Lora 미세 조정 -> 하나의 파일로 병합 -> GPTQ -> OK GPTQ -> 미세 조정 -> 가능하지만 하나의 파일로 병합은 지원되지 않습니다.

Can anyone tell me how to apply GPTQ with axolotl after finetuning with LoRA and changing to a single file?