Atom Quant

Atom Quant AKA: aq is a easy quantization lib supports most decent and fashion quantization method through torch.fx. Unlike original pytorch fx quantization support, we add a fully deploy chain from PTQ and QAT quantization to exporting onnx and then shiping to target inference framework.

atomquant can be easily use to quant any model without a specific dataloader or evaluator, you can even evaluator quantization performance without any GT.

We also support different quantization from vendor package, such as onnxruntime, pytorch_quantization, make it more easy to use and with fully examples.

There are 3 main components in atomquant:

• onnx: directly quantize on onnx model (via onnxruntime);
• atom: Our built-in quantization method;
• tensorrt: Quantization specific for convert to TensorRT engine usage;

Install

atomquant can be installed via:

pip install atomquant

Model Zoo

Here, we provide some models quantized for coco, it devided into CPU use, or TensorRT use. Related training code also available:

Examples

1. Quant Classification

2. Quant GPT3

3. Quant VITS

4. Quant AlphaPose

5. Quant YOLOv7