auto range = rewriter.create<mhlo::SubtractOp>(op.getLoc(), limit, start);
    auto abs = rewriter.create<mhlo::AbsOp>(op.getLoc(), range);

    // Delta is not necessarily the same type as start and limit.
    auto abs_cast =
        rewriter.create<mhlo::ConvertOp>(op.getLoc(), compute_type, abs);
    auto delta_cast =
        rewriter.create<mhlo::ConvertOp>(op.getLoc(), compute_type, delta);

      auto tf_reduce_op = rewriter.create<BooleanReduce>(
          reduce_op.getLoc(), reduce_op->getResult(0).getType(), operand,
          reduction_indices,
          /*keep_dim=*/rewriter.getBoolAttr(false));
      auto tf_argreduce_op = rewriter.create<ArgReduce>(
          reduce_op.getLoc(), reduce_op->getResult(1).getType(), operand,
          reduction_indices);

        filter_constant_op->getLoc(), *rewriter.getContext(),
        filter_quantized_type.getScale(), filter_quantized_type.getZeroPoint(),
        /*narrow_range=*/true);
  }

  // Required because the quantized dimension is changed from 3 -> 0.
  auto new_filter_result_type = RankedTensorType::getChecked(
      filter_constant_op.getLoc(), /*shape=*/new_filter_shape,

      OpBuilder builder(op);
      Operation* load = builder.create<mlir::ml_program::GlobalLoadOp>(
          op.getLoc(), globalOp.getType(),
          SymbolRefAttr::get(op->getContext(), globalOp.getSymName()));
      if (globalOp.getType() != op.getValue().getType()) {
        load = builder.create<TF::CastOp>(op.getLoc(), op.getValue().getType(),
                                          load->getResult(0));
      }

// jobControlInterface is an interface that knows how to add or delete jobs
// created as an interface to allow testing.
type jobControlInterface interface {
	// GetJob retrieves a Job.
	GetJob(namespace, name string) (*batchv1.Job, error)
	// CreateJob creates new Jobs according to the spec.
	CreateJob(namespace string, job *batchv1.Job) (*batchv1.Job, error)
	// DeleteJob deletes the Job identified by name.

              builder.create<ConstOp>(op->getLoc(), output_type, attribute);
        } else {
          attr_cst = builder.create<mlir::arith::ConstantOp>(
              op->getLoc(), cast<TypedAttr>(attribute));
        }
        new_operands.push_back(attr_cst);
      }
    }

    // Create the TFR call op
    auto new_op = builder.create<CallOp>(
        op->getLoc(), compose_func_type.getResults(),

  builder.insert(var_handle_op_in_init);
  auto const_op = builder.create<mlir::arith::ConstantOp>(
      session_init_func.getLoc(), tensor_attr.getType(), tensor_attr);

  builder.create<TF::AssignVariableOp>(
      session_init_func.getLoc(), llvm::ArrayRef<mlir::Type>{},
      llvm::ArrayRef<mlir::Value>{var_handle_op_in_init->getResult(0),
                                  const_op.getResult()});
}

    ArrayRef<Value> args;
    builder.create<func::CallOp>(session_initializer.getLoc(),
                                 func.getFunctionType().getResults(),
                                 func.getSymName(), args);
  }
  builder.create<func::ReturnOp>(session_initializer.getLoc());

  session_initializer.erase();
}

}  // namespace

std::unique_ptr<OperationPass<ModuleOp>>

    Type output_type = bmm_op.getResult().getType();
    auto no_input = rewriter.create<TFL::NoValueOp>(
        bmm_op->getLoc(), rewriter.getNoneType(), rewriter.getUnitAttr());
    auto fc_op = rewriter.create<TFL::FullyConnectedOp>(
        bmm_op->getLoc(), ArrayRef<Type>{output_type},
        /*input=*/output_lhs, /*filter=*/output_rhs, /*bias=*/no_input,

  builder.create<mlir::TF::YieldOp>(if_region.getLoc(),
                                    /*operands=*/ArrayRef<Value>{});

  // Create empty else branch region.
  auto& else_branch = host_side_if.getElseBranch();
  else_branch.push_back(new Block);
  builder.setInsertionPointToEnd(&else_branch.front());
  builder.create<mlir::TF::YieldOp>(if_region.getLoc(),