readme.md

Simple QLoRA Model Inference

This guide demonstrates how to perform inference using a QLoRA (Quantized Low-Rank Adaptation) fine-tuned model with a single code cell.

Requirements

• Python 3.7+
• PyTorch
• Transformers
• PEFT (Parameter-Efficient Fine-Tuning)
• bitsandbytes

Install the required packages:

pip install torch transformers peft bitsandbytes

Inference Code

Copy and paste the following code into a Python script or Jupyter notebook cell:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

# Set up model paths
BASE_MODEL_PATH = "path/to/your/base/model"
ADAPTER_PATH = "path/to/your/qlora/adapter"

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_PATH)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"

# Load quantized model with adapter
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
)
model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL_PATH,
    quantization_config=bnb_config,
    device_map="auto"
)
model = PeftModel.from_pretrained(model, ADAPTER_PATH)

# Generate text
prompt = "Explain quantum computing in simple terms:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100, do_sample=True, temperature=0.7)
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)

print(generated_text)

Usage

1. Replace BASE_MODEL_PATH with the path to your base model.
2. Replace ADAPTER_PATH with the path to your QLoRA adapter.
3. Modify the prompt variable to use your desired input text.
4. Run the code cell.

Customization

• Adjust max_new_tokens, temperature, and other generation parameters in the model.generate() function call to control the output.

Troubleshooting

• If you encounter CUDA out-of-memory errors, try reducing max_new_tokens or using a smaller model.
• Ensure your GPU drivers and CUDA toolkit are up-to-date.

For more advanced usage or optimizations, refer to the Hugging Face documentation for Transformers and PEFT.