open-mmlab/mmdetection · warning

group` is deprecated. Currently only supports NCCL backend.

Error message

group` is deprecated. Currently only supports NCCL backend.

What it means

all_reduce_dict in mmdet/utils/dist_utils.py warns that the group argument is deprecated and only the NCCL backend is supported, before performing an all-reduce of a dict of tensors across ranks (used by hooks like before_val_epoch to aggregate validation metrics). On world_size == 1 it returns immediately after the warning, so single-process runs only see noise.

Source

Thrown at mmdet/utils/dist_utils.py:110

    The code is modified from https://github.com/Megvii-
    BaseDetection/YOLOX/blob/main/yolox/utils/allreduce_norm.py.

    NOTE: make sure that py_dict in different ranks has the same keys and
    the values should be in the same shape. Currently only supports
    nccl backend.

    Args:
        py_dict (dict): Dict to be applied all reduce op.
        op (str): Operator, could be 'sum' or 'mean'. Default: 'sum'
        group (:obj:`torch.distributed.group`, optional): Distributed group,
            Default: None.
        to_float (bool): Whether to convert all values of dict to float.
            Default: True.

    Returns:
        OrderedDict: reduced python dict object.
    """
    warnings.warn(
        'group` is deprecated. Currently only supports NCCL backend.')
    _, world_size = get_dist_info()
    if world_size == 1:
        return py_dict

    # all reduce logic across different devices.
    py_key = list(py_dict.keys())
    if not isinstance(py_dict, OrderedDict):
        py_key_tensor = obj2tensor(py_key)
        dist.broadcast(py_key_tensor, src=0)
        py_key = tensor2obj(py_key_tensor)

    tensor_shapes = [py_dict[k].shape for k in py_key]
    tensor_numels = [py_dict[k].numel() for k in py_key]

    if to_float:
        warnings.warn('Note: the "to_float" is True, you need to '
                      'ensure that the behavior is reasonable.')

View on GitHub (pinned to cfd5d3a985)

Solutions

1. If on GPUs, ensure distributed is initialized with NCCL: torch.distributed.init_process_group(backend='nccl') and launch via the provided dist scripts.
2. If you called all_reduce_dict directly, drop the group argument and rely on the default process group.
3. For CPU/multi-machine non-NCCL setups, replace all_reduce_dict with your own reduce (e.g. all_gather + torch.distributed.all_reduce with Gloo) since NCCL is unsupported here.
4. Suppress the notice in logs when behavior is already correct: warnings.filterwarnings('ignore', message=".*group` is deprecated.*").

Example fix

# before
reduced = all_reduce_dict(metrics, group=my_group)
# after
reduced = all_reduce_dict(metrics)  # group deprecated; NCCL default process group used

import torch, torch.distributed as dist

def nccl_ready() -> bool:
    return dist.is_available() and dist.is_initialized() and dist.get_backend() == 'nccl'

def safe_all_reduce_dict(py_dict):
    import warnings
    from mmdet.utils import all_reduce_dict, get_dist_info
    _, world_size = get_dist_info()
    if world_size == 1:
        return py_dict  # skip warn + reduce on single process
    assert dist.get_backend() == dist.Backend.NCCL, 'all_reduce_dict requires NCCL'
    with warnings.catch_warnings():
        warnings.simplefilter('ignore')
        return all_reduce_dict(py_dict)

When it happens

Trigger: Distributed evaluation (mmdet hooks calling all_reduce_dict in before_val_epoch) or any direct call to all_reduce_dict(py_dict). Triggered regardless of backend because the warning fires unconditionally at function entry; actual reduction requires NCCL (GPU) distributed init.

Common situations: Multi-GPU validation with torch.distributed launched via tools/dist_train.sh; running on CPU-only distributed (Gloo) where the NCCL-only assumption breaks; single-GPU runs where the warning is pure noise since world_size==1 short-circuits.

Related errors

• Note: the "to_float" is True, you need to ensure that the be
• The `file_client_args` is deprecated, please use `backend_ar
• device must be 'cpu' or 'gpu', but got {device}
• GPU collecting has not been supported yet
• The `file_client_args` is deprecated, please use `backend_ar

AI-assisted analysis of open-mmlab/mmdetection@cfd5d3a985 (2026-08-27). Data as JSON: /api/errors/d616c8756e086909. Report an issue: GitHub.