Fix scale bug in quantized batch_dot (#20735)

apache · Nov 17, 2021 · 52bc1bf · 52bc1bf
1 parent 36ed5e0
commit 52bc1bf
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 2 deletions.
diff --git a/src/operator/nn/dnnl/dnnl_batch_dot-inl.h b/src/operator/nn/dnnl/dnnl_batch_dot-inl.h
@@ -48,7 +48,7 @@ struct DNNLDotParam : public dmlc::Parameter<DNNLDotParam> {
 
   dmlc::optional<float> min_calib_range;  // min float value calculated from calibration dataset
   dmlc::optional<float> max_calib_range;  // max float value calculated from calibration dataset
-  bool enable_float_output;               // min float value calculated from calibration dataset
+  bool enable_float_output;
   DMLC_DECLARE_PARAMETER(DNNLDotParam) {
     DMLC_DECLARE_FIELD(transpose_a)
         .describe("If true then transpose the first input before dot.")

diff --git a/src/operator/nn/dnnl/dnnl_batch_dot.cc b/src/operator/nn/dnnl/dnnl_batch_dot.cc
@@ -77,7 +77,7 @@ dnnl::primitive_attr GetQuantizationAttributes(const DNNLDotParam& param,
     // fused requantize => output is int
     out_scale_ = GetQuantizeScale(outputs[DotOut::out].dtype(),
                                   param.min_calib_range.value(),
-                                  param.min_calib_range.value()) /
+                                  param.max_calib_range.value()) /
                  lhs_scale_ / rhs_scale_;
     attr.set_output_scales(0, {out_scale_});
   } else if (param.enable_float_output) {