auto shape = GetAsVector(quant_tensor->shape());
        if (kUseUpdatedHybridSchemeDefault) {
          EXPECT_EQ(quant_tensor->quantization()->scale()->size(), shape[0]);
        } else {
          EXPECT_EQ(quant_tensor->quantization()->scale()->size(), 1);
        }
      } else {
        EXPECT_EQ(quant_tensor->type(), TensorType_FLOAT32);
      }
    }

		argLen:         2,
		hasSideEffects: true,
		generic:        true,
	},
}

func (o Op) Asm() obj.As          { return opcodeTable[o].asm }
func (o Op) Scale() int16         { return int16(opcodeTable[o].scale) }
func (o Op) String() string       { return opcodeTable[o].name }
func (o Op) SymEffect() SymEffect { return opcodeTable[o].symEffect }

		a.writeNameTo(w, abiDetail)

	case TYPE_MEM:
		a.WriteNameTo(w)
		if a.Index != REG_NONE {
			if a.Scale == 0 {
				// arm64 shifted or extended register offset, scale = 0.
				fmt.Fprintf(w, "(%v)", Rconv(int(a.Index)))
			} else {
				fmt.Fprintf(w, "(%v*%d)", Rconv(int(a.Index)), int(a.Scale))
			}
		}

	case TYPE_CONST:
		io.WriteString(w, "$")
		a.WriteNameTo(w)
		if a.Reg != 0 {

            "type": "string"
          },
          "replicas": {

//              scale * rsqrt(variance + epsilon)
// CHECK:  %[[MUL1:.*]] = "tf.Mul"(%[[ARG1:.*]], %[[RSQRT]])
//              x * scale * rsqrt(variance + epsilon)
// CHECK:  %[[MUL2:.*]] = "tf.Mul"(%[[ARG0:.*]], %[[MUL1]])
//              mean * scale * rsqrt(variance + epsilon)
// CHECK:  %[[MUL3:.*]] = "tf.Mul"(%[[ARG3:.*]], %[[MUL1]])
//              offset - mean * scale * rsqrt(variance + epsilon)

			for _, result := range results {
				totalPut += result.Uploads
				totalGet += result.Downloads
			}

			if totalGet < throughputHighestGet {
				// Following check is for situations
				// when Writes() scale higher than Reads()
				// - practically speaking this never happens
				// and should never happen - however it has
				// been seen recently due to hardware issues
				// causes Reads() to go slower than Writes().

      This transformation pass modifies the input and output types of the function
      to what are specified. The task was not just adding cast operations, but,
      instead, using tfl.quantize and tfl.dequantize ops to scale the tensors.
  }];
  let constructor = "CreateModifyIONodesPass()";
  let dependentDialects = ["TFL::TensorFlowLiteDialect"];
  let options = [
      ListOption<"io_node_types_", "test-io-types", "std::string",

		return endpoints
	}

	// A new array of endpoints to be returned that will have both local and
	// remote gateways (if any)
	filtered := make([]*LocalityEndpoints, 0)

	// Scale all weights by the lcm of gateways per network and gateways per cluster.
	// This will allow us to more easily spread traffic to the endpoint across multiple
	// network gateways, increasing reliability of the endpoint.

}

// Creates a new `tfl.qconst` op for the bias. The bias values are 0s, because
// this bias a dummy bias (note that bias fusion is not considered for this
// transformation). The quantization scale for the bias is input scale *
// filter scale. `filter_const_op` is used to retrieve the filter scales and
// the size of the bias constant.
TFL::QConstOp CreateTflConstOpForDummyBias(

	unsafePoint       bool      // this op is an unsafe point, i.e. not safe for async preemption
	symEffect         SymEffect // effect this op has on symbol in aux
	scale             uint8     // amd64/386 indexed load scale
}

type inputInfo struct {
	idx  int     // index in Args array
	regs regMask // allowed input registers
}

type outputInfo struct {
	idx  int     // index in output tuple