Installation¶

Python¶

Requires Python 3.10 or later. Install the package with the extra matching your preferred backend:

PyTorchONNX (torch-free)Both

pip install "mlconfgen[torch]"

Enables MLConformerGenerator, Inertial Fragment Matching, RL fine-tuning and ONNX export.

pip install "mlconfgen[onnx]"

Enables MLConformerGeneratorONNX only. For GPU inference install the GPU runtime in addition:

pip install onnxruntime-gpu

pip install "mlconfgen[full]"

The core dependencies (rdkit, numpy) are always installed. huggingface_hub is installed on demand the first time weights need to be downloaded.

Verify¶

from mlconfgen import MLConformerGenerator

model = MLConformerGenerator(diffusion_steps=10)   # downloads default weights on first run
print(model.list_weights())

Devices¶

The PyTorch generator runs on CPU by default. Pass a device explicitly for accelerators:

import torch
from mlconfgen import MLConformerGenerator

device = (
    torch.device("cuda:0") if torch.cuda.is_available()
    else torch.device("mps:0") if torch.backends.mps.is_available()
    else torch.device("cpu")
)

model = MLConformerGenerator(device=device)

JavaScript¶

Requires Node 18 or later. The package bundles @rdkit/rdkit; you bring your own ONNX Runtime:

NodeBrowser / WebAssembly

npm install mlconfgen onnxruntime-node

npm install mlconfgen onnxruntime-web

onnxruntime-node is an optional peer dependency — the package core is runtime-neutral and you pass whichever runtime you installed to createGenerator as ort.

Local development¶

git clone https://github.com/Membrizard/ml_conformer_generator.git
cd ml_conformer_generator/js
npm install
npm test

Next step¶

Both libraries need the trained weights. See Model Weights.