Works with this: https://github.com/kuenishi/chainer/commit/64ab4b6f78d6e92fb3f5ed3d72b3167f6ae3b1af

eval_detection_multi3.py

import argparse
import copy

import chainer
from chainer import iterators
from chainer import function

import chainermn
from chainermn.extensions.multi_node_evaluator import GatherEvaluator

from chainercv.utils import ProgressHook
from chainercv.utils.iterator.unzip import unzip

from eval_detection import models
from eval_detection import setup

chainer.config.cv_resize_backend = "cv2"

def _flatten(iterator):
    return (sample for batch in iterator for sample in batch)


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument('--dataset', choices=('voc', 'coco'))
    parser.add_argument('--model', choices=sorted(models.keys()))
    parser.add_argument('--pretrained-model')
    parser.add_argument('--batchsize', type=int)
    args = parser.parse_args()

    comm = chainermn.create_communicator('pure_nccl')
    device = comm.intra_rank

    dataset, eval_, model, batchsize = setup(
        args.dataset, args.model, args.pretrained_model, args.batchsize)

    chainer.cuda.get_device_from_id(device).use()
    model.to_gpu()

    model.use_preset('evaluate')

    dataset = chainermn.scatter_dataset(dataset, comm, force_equal_length=False)
    hook = ProgressHook(len(dataset))

    iterator = iterators.MultithreadIterator(
        dataset, batchsize, repeat=False, shuffle=False)

    def eval_func(batch):
        in_values = []
        rest_values = []
        for sample in batch:
            in_values.append(sample[0:1])
            rest_values.append(sample[1:])
        in_values = tuple(list(v) for v in zip(*in_values))
        rest_values = tuple(list(v) for v in zip(*rest_values))

        out_values = model.predict(*in_values)

        if comm.rank == 0:
            hook(in_values, out_values, rest_values)

        return (out_values, rest_values)

    def aggregate_func(results):
        out_values, rest_values = unzip(results)
        out_values = tuple(map(_flatten, unzip(out_values)))
        rest_values = tuple(map(_flatten, unzip(rest_values)))
        eval_(out_values, rest_values)        

    evaluator = GatherEvaluator(comm, iterator,
                                model, aggregate_func,
                                eval_func=eval_func)

    evaluator(None)


if __name__ == '__main__':
    main()