mul_op.getLoc(), filter.getType(), filter, broadcast_multiplier);
    Value new_conv = rewriter.create<mhlo::ConvolutionOp>(
        mul_op.getLoc(), conv_op.getType(), conv_op.getLhs(), new_filter,
        conv_op.getWindowStridesAttr(), conv_op.getPaddingAttr(),
        conv_op.getLhsDilationAttr(), conv_op.getRhsDilationAttr(),
        conv_op.getWindowReversalAttr(), conv_op.getDimensionNumbers(),

        op.getLoc(), /*resultType0=*/new_conv_output_tensor_type,
        /*lhs=*/input_transpose_op,
        /*rhs=*/filter_transpose_op,
        /*window_strides=*/op.getWindowStridesAttr(),
        /*padding=*/op.getPaddingAttr(),
        /*lhs_dilation=*/op.getLhsDilationAttr(),
        /*rhs_dilation=*/op.getRhsDilationAttr(),
        /*window_reversal=*/op.getWindowReversalAttr(),

  } else {
    padding_nums.resize(padding_nums_size);
    for (int i = 0; i < dnums.input_spatial_dimensions_size(); ++i) {
      const int64_t stride =
          mlir::cast<IntegerAttr>(op.getWindowStridesAttr()[i]).getInt();
      const int64_t lhs_size_dilated =
          ::tensorflow::UniformQuantizedConvolutionParams::DilatedSize(
              lhs_shape.getDimSize(dnums.input_spatial_dimensions(i)),

      // Create a Smaller Convolution (Ensure compatibility)
      auto conv_result = rewriter.create<mhlo::ConvolutionOp>(
          conv_op.getLoc(), new_output_type, sliced_input, sliced_kernel,
          conv_op.getWindowStridesAttr(), conv_op.getPaddingAttr(),
          conv_op.getLhsDilationAttr(), conv_op.getRhsDilationAttr(),
          conv_op.getWindowReversalAttr(), conv_op.getDimensionNumbers(), 1, 1,
          conv_op.getPrecisionConfigAttr());