FilesLast commitLast update
8 days ago
8 days ago
8 days ago
8 days ago
8 days ago
8 days ago
README

AscendWeightQuant

English | 简体中文

If you want to contribute the AscendC implementation of this operator, refer to contribution process.

Description

Weight quantization operator in the quantization pipeline (opFile=Null, no kernel implementation, handled only by fusion passes at graph compilation time).

  • Inputs: x (int8, HWCN, required), offset (int8, HWCN, required)
  • Output: y (int8, HWCN, required)
  • Attribute: dst_type (optional int, supports DT_INT8/DT_INT4, defaults to DT_INT8)
  • Supported SoCs: ascend950 / ascend960dt / ascend960pr
  • aclnn: aclnn_exclude (not exposed as an aclnn interface)