Release Notes

MindSpore Golden Stick 1.0.0 Release Notes

Major Features and Improvements

  • The Post-Training Quantization algorithm PTQ supports the GPTQ quantization algorithm, which can quantize weight to 8 bit or 4 bit. GPTQ has been added to the precision recovery algorithm set and can be selected through precision_recovery in PTQConfig. Currently, only the GPTQ algorithm is available for precision recovery.

  • The Post-Training Quantization algorithm PTQ supports the AWQ quantization algorithm. It enables AWQ by adding a new outlier suppression method to quantize the weights to 4 bits. The outlier suppression method can be selected through outliers_suppression in PTQConfig, and currently, two methods are available: smooth and awq.

  • The Post-Training Quantization algorithm PTQ supports per-token dynamic quantization for activations, enabling online quantization of activations. This can be configured through PTQConfig with act_quant_granularity=QuantGranularity.PER_TOKEN.

API Change

  • The quantization methods RoundToNearest and SmoothQuant are deprecated. Please replace them with PTQ.

Contributors

Thanks goes to these wonderful people:

huangzhuo, zhangminli, ccsszz, yyyyrf, hangangqiang

Contributions of any kind are welcome!