Skip to main content

​
Installation

ChemLactica requires Python 3.11 and conda for managing dependencies. Follow these steps to set up your environment.

​
Prerequisites

Before installing ChemLactica, ensure you have:
  • Python 3.11
  • conda (Anaconda or Miniconda)
  • CUDA 11.8 (for GPU acceleration)
While ChemLactica can run on CPU, GPU acceleration is strongly recommended for molecular optimization tasks.

​
Quick installation

1

Clone the repository

Clone the ChemLactica repository from GitHub:
git clone https://github.com/YerevaNN/ChemLactica.git
cd ChemLactica
2

Create conda environment

Create a new conda environment with Python 3.11 using the provided environment file:
conda create -n chemlactica python=3.11 -y -f environment.yml
This will install all required dependencies including:
  • PyTorch 2.1.2 with CUDA 11.8 support
  • Transformers 4.39.0
  • RDKit 2023.9.5 for molecular manipulation
  • Flash Attention 2.5.6 for efficient inference
  • TRL 0.8.6 for fine-tuning
  • And many other dependencies
3

Activate the environment

Activate the newly created conda environment:
conda activate chemlactica

​
Key dependencies

The installation includes these important packages:

​
Core ML frameworks

  • PyTorch 2.1.2: Deep learning framework with CUDA 11.8 support
  • Transformers 4.39.0: Hugging Face library for loading pre-trained models
  • Accelerate 0.28.0: Distributed training and inference
  • TRL 0.8.6: Transformer Reinforcement Learning for fine-tuning

​
Chemistry libraries

  • RDKit 2023.9.5: Comprehensive cheminformatics toolkit
  • ChemLactica 0.0.1: Core ChemLactica package (installed via pip)

​
Performance optimization

  • Flash Attention 2.5.6: Efficient attention mechanism implementation
  • BitsAndBytes 0.43.0: 8-bit optimizer for memory-efficient training
  • Einops 0.7.0: Tensor operations library

​
Data and utilities

  • Datasets 2.18.0: Hugging Face datasets library
  • Pandas 2.2.1: Data manipulation
  • NumPy 1.26.4: Numerical computing
  • Scikit-learn 1.4.1: Machine learning utilities

​
Verify installation

After installation, verify that everything is set up correctly: 
python -c "import torch; print(f'PyTorch: {torch.__version__}')"
python -c "import transformers; print(f'Transformers: {transformers.__version__}')"
python -c "import rdkit; print(f'RDKit: {rdkit.__version__}')"
python -c "import chemlactica; print('ChemLactica installed successfully')"

​
Check CUDA availability

To verify GPU support: 
python -c "import torch; print(f'CUDA available: {torch.cuda.is_available()}'); print(f'CUDA version: {torch.version.cuda}')"
If CUDA is not available, the models will run on CPU, which will be significantly slower for molecular optimization tasks.

​
Download pre-trained models

ChemLactica models are hosted on Hugging Face and will be automatically downloaded when you first use them. You can also download them manually: 
from transformers import AutoModelForCausalLM, AutoTokenizer

# Download Chemlactica-125M
model = AutoModelForCausalLM.from_pretrained("yerevann/chemlactica-125m")
tokenizer = AutoTokenizer.from_pretrained("yerevann/chemlactica-125m")

# Download Chemlactica-1.3B
model = AutoModelForCausalLM.from_pretrained("yerevann/chemlactica-1.3b")
tokenizer = AutoTokenizer.from_pretrained("yerevann/chemlactica-1.3b")

# Download Chemma-2B
model = AutoModelForCausalLM.from_pretrained("yerevann/chemma-2b")
tokenizer = AutoTokenizer.from_pretrained("yerevann/chemma-2b")

​
Troubleshooting

​
CUDA version mismatch

If you encounter CUDA version issues, ensure your NVIDIA drivers are compatible with CUDA 11.8: 
nvidia-smi

​
Memory issues

For the larger models (1.3B and 2B parameters), ensure you have sufficient GPU memory:
  • Chemlactica-125M: ~1GB GPU memory
  • Chemlactica-1.3B: ~5GB GPU memory
  • Chemma-2B: ~8GB GPU memory
Consider using lower precision (bfloat16 or float16) to reduce memory usage: 
import torch

model = AutoModelForCausalLM.from_pretrained(
    "yerevann/chemlactica-125m",
    torch_dtype=torch.bfloat16
)

​
Next steps

Now that you have ChemLactica installed, proceed to the Quick start guide to generate your first molecules!

Build docs developers (and LLMs) love

Get started for freeTalk to us