if (IsCompatibleTypeWithTFLCastOp(value.getType()) &&
            IsCompatibleTypeWithTFLCastOp(type)) {
          auto cast = b.create<CastOp>(yield_op->getLoc(), type, value);
          args.push_back(cast);
        } else {
          auto cast = b.create<TF::CastOp>(yield_op->getLoc(), type, value);
          args.push_back(cast);
        }
      }
    }
    args.append(new_args.begin(), new_args.end());
  } else {

    auto var_handle_op =
        builder.create<TF::VarHandleOp>(const_op.getLoc(),
                                        /*resource=*/resource_type,
                                        /*container=*/"", shared_name);

    auto read_variable_op = builder.create<TF::ReadVariableOp>(
        const_op.getLoc(), const_op.getType(), var_handle_op);

        mlir::cast<TensorType>(input.getType()), kNchwToNhwcPermutation);

    auto input_transpose_op = rewriter.create<mlir::stablehlo::TransposeOp>(
        op.getLoc(), /*resultType0=*/new_input_tensor_type, /*operand=*/input,
        rewriter.getDenseI64ArrayAttr(kNchwToNhwcPermutation));

    // Transpose the filter tensor: [o, i, 0, 1] => [0, 1, i, o]
    Value filter = op->getOperand(1);

      return InlineRegion(if_op.getLoc(), rewriter, if_op,
                          &if_op.getThenRegion());
    }

    // Else branch
    if (matchPattern(if_op.getCondition(), m_Zero())) {
      if (if_op.getElseRegion().empty()) {
        // Remove the op
        rewriter.eraseOp(if_op);
        return success();
      } else {
        return InlineRegion(if_op.getLoc(), rewriter, if_op,

  }

  auto compile_op = builder->create<TF::_TPUCompileMlirOp>(
      cluster_func.getLoc(),
      /*compilation_status=*/compilation_status_type, /*program=*/
      llvm::SmallVector<Type, 8>(num_cores_per_replica, program_type),
      compile_op_operands, txt_module, txt_metadata);

  return tensorflow::WrapOpInLaunch(builder, compile_op.getLoc(), compile_op,
                                    compilation_device);
}

  // Unformat op.
  auto unformat_op = builder.create<TF::TPUReshardVariablesOp>(
      while_op.getLoc(), llvm::ArrayRef<Type>{}, unformat_operands);
  WrapOpInLaunch(&builder, execute_launch.getLoc(), unformat_op,
                 execute_launch.getDevice());
  builder.create<tf_device::ReturnOp>(while_op.getLoc(), ArrayRef<Value>{});

  return true;
}

void TPUVariableRuntimeReformattingPass::runOnOperation() {

      block_arg.getLoc(),
      DenseIntElementsAttr::get(
          RankedTensorType::get({1, num_replicas}, builder.getIntegerType(32)),
          group_assignment_val));

  StringAttr reduce_op = builder.getStringAttr("Add");
  StringAttr mode = builder.getStringAttr("CrossReplica");
  return builder.create<XlaAllReduceOp>(block_arg.getLoc(), block_arg.getType(),

    auto yield = builder->create<tf_executor::YieldOp>(
        terminator->getLoc(), terminator->getOperands());
    terminator->erase();

    // Create new island for each region.
    builder->setInsertionPoint(island_op);
    auto execute_island = builder->create<tf_executor::IslandOp>(
        island_op.getLoc(), yield.getOperandTypes(),
        island_op.getControl().getType(), island_op.getControlInputs());

    auto *new_ewise_op =
        rewriter.create(op->getLoc(), op->getName().getIdentifier(),
                        {tpose_arg1.getOperand(0), tpose_arg2.getOperand(0)},
                        new_out_type, op->getAttrs());

    // Apply original tranpose to output of ewise op.
    auto out_tpose_op = rewriter.create<TFL::TransposeOp>(
        new_ewise_op->getLoc(), op->getResult(0).getType(),

    }
    Value broadcast_multiplier = rewriter.create<mhlo::BroadcastInDimOp>(
        mul_op.getLoc(), filter.getType(), multiplier, broadcast_dims);
    Value new_filter = rewriter.create<mhlo::MulOp>(
        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,