---
library_name: peft
license: apache-2.0
language:
- en
datasets:
- Abirate/english_quotes
pipeline_tag: text-generation
---

# Quantization 4Bits - 5.02 GB GPU memory usage for inference:

** Vide same fine-tuning for GPT-J-6B: [https://huggingface.co/nlpulse/gpt-j-6b-english_quotes](https://huggingface.co/nlpulse/gpt-j-6b-english_quotes)

```
$ nvidia-smi
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 525.125.06   Driver Version: 525.125.06   CUDA Version: 12.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   1  NVIDIA GeForce ...  Off  | 00000000:04:00.0 Off |                  N/A |
| 65%   74C    P2   169W / 170W |   5028MiB / 12288MiB |     97%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
```

## Fine-tuning
```
3 epochs, all dataset samples (split=train), 939 steps
1 x GPU NVidia RTX 3060 12GB - max. GPU memory: 6.85 GB
Duration: 1h54min

$ nvidia-smi && free -h
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 525.125.06   Driver Version: 525.125.06   CUDA Version: 12.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   1  NVIDIA GeForce ...  Off  | 00000000:04:00.0 Off |                  N/A |
|100%   87C    P2   168W / 170W |   6854MiB / 12288MiB |     98%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
               total        used        free      shared  buff/cache   available
Mem:            77Gi        13Gi       1.1Gi       116Mi        63Gi        63Gi
Swap:           37Gi       3.8Gi        34Gi

```

## Inference 
```
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftConfig, PeftModel

model_path = "nlpulse/llama2-7b-chat-english_quotes"

# tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path, use_auth_token=True)
tokenizer.pad_token = tokenizer.eos_token

# quantization config
quant_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16
)

# model adapter PEFT LoRA
config = PeftConfig.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path,
        quantization_config=quant_config, device_map={"":0}, use_auth_token=True)
model = PeftModel.from_pretrained(model, model_path)

# inference
device = "cuda"
text_list = ["Ask not what your country", "Be the change that", "You only live once, but", "I'm selfish, impatient and"]
for text in text_list:
    inputs = tokenizer(text, return_tensors="pt").to(device)
    outputs = model.generate(**inputs, max_new_tokens=60)
    print('>> ', text, " => ", tokenizer.decode(outputs[0], skip_special_tokens=True))

```

## Requirements
```
pip install -U bitsandbytes
pip install -U git+https://github.com/huggingface/transformers.git 
pip install -U git+https://github.com/huggingface/peft.git
pip install -U accelerate
pip install -U datasets
pip install -U scipy
```

## Scripts
[https://github.com/nlpulse-io/sample_codes/tree/main/fine-tuning/peft_quantization_4bits/llama2-7b-chat](https://github.com/nlpulse-io/sample_codes/tree/main/fine-tuning/peft_quantization_4bits/llama2-7b-chat)


## References
[QLoRa: Fine-Tune a Large Language Model on Your GPU](https://towardsdatascience.com/qlora-fine-tune-a-large-language-model-on-your-gpu-27bed5a03e2b)

[Making LLMs even more accessible with bitsandbytes, 4-bit quantization and QLoRA](https://huggingface.co/blog/4bit-transformers-bitsandbytes)


## Training procedure


The following `bitsandbytes` quantization config was used during training:
- load_in_8bit: False
- load_in_4bit: True
- llm_int8_threshold: 6.0
- llm_int8_skip_modules: None
- llm_int8_enable_fp32_cpu_offload: False
- llm_int8_has_fp16_weight: False
- bnb_4bit_quant_type: nf4
- bnb_4bit_use_double_quant: True
- bnb_4bit_compute_dtype: bfloat16


### Framework versions


- PEFT 0.4.0.dev0