Release Notes

MindSpore Golden Stick 1.2.0 Release Notes

Major Features and Improvements

• The Post-Training Quantization algorithm PTQ supports the MOE structure with SmoothQuant-A8W8 quantization algorithm and GPTQ-A16W4 low-bit quantization algorithm. These have been adapted for the DeepSeekV3/R1 model.

• Added OutlierSuppression-Lite(OSL), an outlier suppression technique. OSL is an extension of SmoothQuant that tunes the migration strength alpha through hyperparameter optimization at the matrix granularity to achieve more fine-grained network self-adaptive calibration. In the DeepSeek V3-0324 network function call scenario, OSL preserves higher accuracy and achieves BFCL scores on par with the official FP8 baselines.

• [Demo] The Post-Training Quantization algorithm PTQ supports the A8W4 quantization algorithm. These have been adapted for the DeepSeekV3/R1 model.

• Added the loading and evaluation of the datasets wikitext, boolq, ceval, and gsm8k.

• Accuracy of DeepSeekR1:

  method	ceval	gsm8k
  BF16	89.67	91.74
  SmoothQuant-A8W8	89.45	92.42
  OSL-A8W8	89.9	91.81
  GPTQ-A16W4	89.52	91.12
  A8W4	89.0	91.51

Contributors

Thanks goes to these wonderful people:

tongl, zhuxiaochen, guoguopot, huangzhuo, ccsszz, yyyyrf, hangangqiang, HeadSnake

Contributions of any kind are welcome!