Python torch.distributed.all_gather() Examples

def gather_tensors(input_array):
    world_size = dist.get_world_size()
    ## gather shapes first
    myshape = input_array.shape
    mycount = input_array.size
    shape_tensor = torch.Tensor(np.array(myshape)).cuda()
    all_shape = [torch.Tensor(np.array(myshape)).cuda() for i in range(world_size)]
    dist.all_gather(all_shape, shape_tensor)
    ## compute largest shapes
    all_shape = [x.cpu().numpy() for x in all_shape]
    all_count = [int(x.prod()) for x in all_shape]
    all_shape = [list(map(int, x)) for x in all_shape]
    max_count = max(all_count)
    ## padding tensors and gather them
    output_tensors = [torch.Tensor(max_count).cuda() for i in range(world_size)]
    padded_input_array = np.zeros(max_count)
    padded_input_array[:mycount] = input_array.reshape(-1)
    input_tensor = torch.Tensor(padded_input_array).cuda()
    dist.all_gather(output_tensors, input_tensor)
    ## unpadding gathered tensors
    padded_output = [x.cpu().numpy() for x in output_tensors]
    output = [x[:all_count[i]].reshape(all_shape[i]) for i,x in enumerate(padded_output)]
    return output 

def all_gather(tensors):
    """
    All gathers the provided tensors from all processes across machines.
    Args:
        tensors (list): tensors to perform all gather across all processes in
        all machines.
    """

    gather_list = []
    output_tensor = []
    world_size = dist.get_world_size()
    for tensor in tensors:
        tensor_placeholder = [
            torch.ones_like(tensor) for _ in range(world_size)
        ]
        dist.all_gather(tensor_placeholder, tensor, async_op=False)
        gather_list.append(tensor_placeholder)
    for gathered_tensor in gather_list:
        output_tensor.append(torch.cat(gathered_tensor, dim=0))
    return output_tensor 

def _pad_to_largest_tensor(tensor, group):
    """
    Returns:
        list[int]: size of the tensor, on each rank
        Tensor: padded tensor that has the max size
    """
    world_size = dist.get_world_size(group=group)
    assert (
            world_size >= 1
    ), "comm.gather/all_gather must be called from ranks within the given group!"
    local_size = torch.tensor([tensor.numel()], dtype=torch.int64, device=tensor.device)
    size_list = [
        torch.zeros([1], dtype=torch.int64, device=tensor.device) for _ in range(world_size)
    ]
    dist.all_gather(size_list, local_size, group=group)
    size_list = [int(size.item()) for size in size_list]

    max_size = max(size_list)

    # we pad the tensor because torch all_gather does not support
    # gathering tensors of different shapes
    if local_size != max_size:
        padding = torch.zeros((max_size - local_size,), dtype=torch.uint8, device=tensor.device)
        tensor = torch.cat((tensor, padding), dim=0)
    return size_list, tensor 

def _pad_to_largest_tensor(tensor, group):
    """
    Returns:
        list[int]: size of the tensor, on each rank
        Tensor: padded tensor that has the max size
    """
    world_size = dist.get_world_size(group=group)
    assert (
        world_size >= 1
    ), "comm.gather/all_gather must be called from ranks within the given group!"
    local_size = torch.tensor([tensor.numel()], dtype=torch.int64, device=tensor.device)
    size_list = [
        torch.zeros([1], dtype=torch.int64, device=tensor.device) for _ in range(world_size)
    ]
    dist.all_gather(size_list, local_size, group=group)
    size_list = [int(size.item()) for size in size_list]

    max_size = max(size_list)

    # we pad the tensor because torch all_gather does not support
    # gathering tensors of different shapes
    if local_size != max_size:
        padding = torch.zeros((max_size - local_size,), dtype=torch.uint8, device=tensor.device)
        tensor = torch.cat((tensor, padding), dim=0)
    return size_list, tensor 

def _gather_values(*tensors, group, world_size):
        # Start gather operations asynchronously
        gathered, gather_ops = [], []
        for t in tensors:
            t_all = t.new_empty(world_size, *t.shape)
            t_op = distributed.all_gather(list(t_all.unbind(0)), t, group=group, async_op=True)

            gathered.append(t_all)
            gather_ops.append(t_op)

        # Wait
        for op in gather_ops:
            op.wait()

        # Return results
        return tuple(gathered) 

def merge(img_ids, eval_imgs):
    all_img_ids = all_gather(img_ids)
    all_eval_imgs = all_gather(eval_imgs)

    merged_img_ids = []
    for p in all_img_ids:
        merged_img_ids.extend(p)

    merged_eval_imgs = []
    for p in all_eval_imgs:
        merged_eval_imgs.append(p)

    merged_img_ids = np.array(merged_img_ids)
    merged_eval_imgs = np.concatenate(merged_eval_imgs, 2)

    # keep only unique (and in sorted order) images
    merged_img_ids, idx = np.unique(merged_img_ids, return_index=True)
    merged_eval_imgs = merged_eval_imgs[..., idx]

    return merged_img_ids, merged_eval_imgs 

def _pad_to_largest_tensor(tensor, group):
    """
    Returns:
        list[int]: size of the tensor, on each rank
        Tensor: padded tensor that has the max size
    """
    world_size = dist.get_world_size(group=group)
    assert (
        world_size >= 1
    ), "comm.gather/all_gather must be called from ranks within the given group!"
    local_size = torch.tensor([tensor.numel()], dtype=torch.int64, device=tensor.device)
    size_list = [
        torch.zeros([1], dtype=torch.int64, device=tensor.device) for _ in range(world_size)
    ]
    dist.all_gather(size_list, local_size, group=group)
    size_list = [int(size.item()) for size in size_list]

    max_size = max(size_list)

    # we pad the tensor because torch all_gather does not support
    # gathering tensors of different shapes
    if local_size != max_size:
        padding = torch.zeros((max_size - local_size,), dtype=torch.uint8, device=tensor.device)
        tensor = torch.cat((tensor, padding), dim=0)
    return size_list, tensor 

def _pad_to_largest_tensor(tensor, group):
    """
    Returns:
        list[int]: size of the tensor, on each rank
        Tensor: padded tensor that has the max size
    """
    world_size = dist.get_world_size(group=group)
    assert (
        world_size >= 1
    ), "comm.gather/all_gather must be called from ranks within the given group!"
    local_size = torch.tensor([tensor.numel()],
                              dtype=torch.int64,
                              device=tensor.device)
    size_list = [
        torch.zeros([1], dtype=torch.int64, device=tensor.device)
        for _ in range(world_size)
    ]
    dist.all_gather(size_list, local_size, group=group)
    size_list = [int(size.item()) for size in size_list]

    max_size = max(size_list)

    # we pad the tensor because torch all_gather does not support
    # gathering tensors of different shapes
    if local_size != max_size:
        padding = torch.zeros((max_size - local_size, ),
                              dtype=torch.uint8,
                              device=tensor.device)
        tensor = torch.cat((tensor, padding), dim=0)
    return size_list, tensor 

def all_gather_unaligned(data, group=None):
    """
    Run all_gather on arbitrary picklable data (not necessarily tensors).

    Args:
        data: any picklable object
        group: a torch process group. By default, will use a group which
            contains all ranks on gloo backend.

    Returns:
        list[data]: list of data gathered from each rank
    """
    if get_world_size() == 1:
        return [data]
    if group is None:
        group = _get_global_gloo_group()
    if dist.get_world_size(group) == 1:
        return [data]

    tensor = _serialize_to_tensor(data, group)

    size_list, tensor = _pad_to_largest_tensor(tensor, group)
    max_size = max(size_list)

    # receiving Tensor from all ranks
    tensor_list = [
        torch.empty((max_size,), dtype=torch.uint8, device=tensor.device)
        for _ in size_list
    ]
    dist.all_gather(tensor_list, tensor, group=group)

    data_list = []
    for size, tensor in zip(size_list, tensor_list):
        buffer = tensor.cpu().numpy().tobytes()[:size]
        data_list.append(pickle.loads(buffer))

    return data_list 

def forward(ctx, input, num_sync_devices, num_groups):
        """
        Perform forwarding, gathering the stats across different process/ GPU
        group.
        """
        ctx.num_sync_devices = num_sync_devices
        ctx.num_groups = num_groups

        input_list = [
            torch.zeros_like(input) for k in range(du.get_local_size())
        ]
        dist.all_gather(
            input_list, input, async_op=False, group=du._LOCAL_PROCESS_GROUP
        )

        inputs = torch.stack(input_list, dim=0)
        if num_groups > 1:
            rank = du.get_local_rank()
            group_idx = rank // num_sync_devices
            inputs = inputs[
                group_idx
                * num_sync_devices : (group_idx + 1)
                * num_sync_devices
            ]
        inputs = torch.sum(inputs, dim=0)
        return inputs 

def shared_random_seed():
    """
    Returns:
        int: a random number that is the same across all workers.
            If workers need a shared RNG, they can use this shared seed to
            create one.

    All workers must call this function, otherwise it will deadlock.
    """
    ints = np.random.randint(2**31)
    all_ints = all_gather(ints)
    return all_ints[0] 

def all_gather(data):
    """
    Run all_gather on arbitrary picklable data (not necessarily tensors)
    Args:
        data: any picklable object
    Returns:
        list[data]: list of data gathered from each rank
    """
    world_size = get_world_size()
    if world_size == 1:
        return [data]

    # serialized to a Tensor
    buffer = pickle.dumps(data)
    storage = torch.ByteStorage.from_buffer(buffer)
    tensor = torch.ByteTensor(storage).to("cuda")

    # obtain Tensor size of each rank
    local_size = torch.IntTensor([tensor.numel()]).to("cuda")
    size_list = [torch.IntTensor([0]).to("cuda") for _ in range(world_size)]
    dist.all_gather(size_list, local_size)
    size_list = [int(size.item()) for size in size_list]
    max_size = max(size_list)

    # receiving Tensor from all ranks
    # we pad the tensor because torch all_gather does not support
    # gathering tensors of different shapes
    tensor_list = []
    for _ in size_list:
        tensor_list.append(torch.ByteTensor(size=(max_size,)).to("cuda"))
    if local_size != max_size:
        padding = torch.ByteTensor(size=(max_size - local_size,)).to("cuda")
        tensor = torch.cat((tensor, padding), dim=0)
    dist.all_gather(tensor_list, tensor)

    data_list = []
    for size, tensor in zip(size_list, tensor_list):
        buffer = tensor.cpu().numpy().tobytes()[:size]
        data_list.append(pickle.loads(buffer))

    return data_list 

def all_gather(data):
    """
    Run all_gather on arbitrary picklable data (not necessarily tensors)
    Args:
        data: any picklable object
    Returns:
        list[data]: list of data gathered from each rank
    """
    world_size = get_world_size()
    if world_size == 1:
        return [data]

    # serialized to a Tensor
    buffer = pickle.dumps(data)
    storage = torch.ByteStorage.from_buffer(buffer)
    tensor = torch.ByteTensor(storage).to("cuda")

    # obtain Tensor size of each rank
    local_size = torch.IntTensor([tensor.numel()]).to("cuda")
    size_list = [torch.IntTensor([0]).to("cuda") for _ in range(world_size)]
    dist.all_gather(size_list, local_size)
    size_list = [int(size.item()) for size in size_list]
    max_size = max(size_list)

    # receiving Tensor from all ranks
    # we pad the tensor because torch all_gather does not support
    # gathering tensors of different shapes
    tensor_list = []
    for _ in size_list:
        tensor_list.append(torch.ByteTensor(size=(max_size,)).to("cuda"))
    if local_size != max_size:
        padding = torch.ByteTensor(size=(max_size - local_size,)).to("cuda")
        tensor = torch.cat((tensor, padding), dim=0)
    dist.all_gather(tensor_list, tensor)

    data_list = []
    for size, tensor in zip(size_list, tensor_list):
        buffer = tensor.cpu().numpy().tobytes()[:size]
        data_list.append(pickle.loads(buffer))

    return data_list 

def shared_random_seed():
    """
    Returns:
        int: a random number that is the same across all workers.
            If workers need a shared RNG, they can use this shared seed to
            create one.

    All workers must call this function, otherwise it will deadlock.
    """
    ints = np.random.randint(2 ** 31)
    all_ints = all_gather(ints)
    return all_ints[0] 

def all_gather(data, group=None):
    """
    Run all_gather on arbitrary picklable data (not necessarily tensors).

    Args:
        data: any picklable object
        group: a torch process group. By default, will use a group which
            contains all ranks on gloo backend.

    Returns:
        list[data]: list of data gathered from each rank
    """
    if get_world_size() == 1:
        return [data]
    if group is None:
        group = _get_global_gloo_group()
    if dist.get_world_size(group) == 1:
        return [data]

    tensor = _serialize_to_tensor(data, group)

    size_list, tensor = _pad_to_largest_tensor(tensor, group)
    max_size = max(size_list)

    # receiving Tensor from all ranks
    tensor_list = [
        torch.empty((max_size, ), dtype=torch.uint8, device=tensor.device)
        for _ in size_list
    ]
    dist.all_gather(tensor_list, tensor, group=group)

    data_list = []
    for size, tensor in zip(size_list, tensor_list):
        buffer = tensor.cpu().numpy().tobytes()[:size]
        data_list.append(pickle.loads(buffer))

    return data_list 

def all_gather(data):
    """
    Run all_gather on arbitrary picklable data (not necessarily tensors)
    Args:
        data: any picklable object
    Returns:
        list[data]: list of data gathered from each rank
    """
    world_size = get_world_size()
    if world_size == 1:
        return [data]

    # serialized to a Tensor
    buffer = pickle.dumps(data)
    storage = torch.ByteStorage.from_buffer(buffer)
    tensor = torch.ByteTensor(storage).to("cuda")

    # obtain Tensor size of each rank
    local_size = torch.IntTensor([tensor.numel()]).to("cuda")
    size_list = [torch.IntTensor([0]).to("cuda") for _ in range(world_size)]
    dist.all_gather(size_list, local_size)
    size_list = [int(size.item()) for size in size_list]
    max_size = max(size_list)

    # receiving Tensor from all ranks
    # we pad the tensor because torch all_gather does not support
    # gathering tensors of different shapes
    tensor_list = []
    for _ in size_list:
        tensor_list.append(torch.ByteTensor(size=(max_size,)).to("cuda"))
    if local_size != max_size:
        padding = torch.ByteTensor(size=(max_size - local_size,)).to("cuda")
        tensor = torch.cat((tensor, padding), dim=0)
    dist.all_gather(tensor_list, tensor)

    data_list = []
    for size, tensor in zip(size_list, tensor_list):
        buffer = tensor.cpu().numpy().tobytes()[:size]
        data_list.append(pickle.loads(buffer))

    return data_list 

def forward(ctx, input):
        input_list = [torch.zeros_like(input) for k in range(dist.get_world_size())]
        # Use allgather instead of allreduce since I don't trust in-place operations ..
        dist.all_gather(input_list, input, async_op=False)
        inputs = torch.stack(input_list, dim=0)
        return torch.sum(inputs, dim=0) 

def all_gather(data):
    """
    Run all_gather on arbitrary picklable data (not necessarily tensors)
    Args:
        data: any picklable object
    Returns:
        list[data]: list of data gathered from each rank
    """
    world_size = get_world_size()
    if world_size == 1:
        return [data]

    # serialized to a Tensor
    buffer = pickle.dumps(data)
    storage = torch.ByteStorage.from_buffer(buffer)
    tensor = torch.ByteTensor(storage).to("cuda")

    # obtain Tensor size of each rank
    local_size = torch.IntTensor([tensor.numel()]).to("cuda")
    size_list = [torch.IntTensor([0]).to("cuda") for _ in range(world_size)]
    dist.all_gather(size_list, local_size)
    size_list = [int(size.item()) for size in size_list]
    max_size = max(size_list)

    # receiving Tensor from all ranks
    # we pad the tensor because torch all_gather does not support
    # gathering tensors of different shapes
    tensor_list = []
    for _ in size_list:
        tensor_list.append(torch.ByteTensor(size=(max_size,)).to("cuda"))
    if local_size != max_size:
        padding = torch.ByteTensor(size=(max_size - local_size,)).to("cuda")
        tensor = torch.cat((tensor, padding), dim=0)
    dist.all_gather(tensor_list, tensor)

    data_list = []
    for size, tensor in zip(size_list, tensor_list):
        buffer = tensor.cpu().numpy().tobytes()[:size]
        data_list.append(pickle.loads(buffer))

    return data_list 

def all_gather(data):
    """
    Run all_gather on arbitrary picklable data (not necessarily tensors)
    Args:
        data: any picklable object
    Returns:
        list[data]: list of data gathered from each rank
    """
    world_size = get_world_size()
    if world_size == 1:
        return [data]

    # serialized to a Tensor
    buffer = pickle.dumps(data)
    storage = torch.ByteStorage.from_buffer(buffer)
    tensor = torch.ByteTensor(storage).to("cuda")

    # obtain Tensor size of each rank
    local_size = torch.IntTensor([tensor.numel()]).to("cuda")
    size_list = [torch.IntTensor([0]).to("cuda") for _ in range(world_size)]
    dist.all_gather(size_list, local_size)
    size_list = [int(size.item()) for size in size_list]
    max_size = max(size_list)

    # receiving Tensor from all ranks
    # we pad the tensor because torch all_gather does not support
    # gathering tensors of different shapes
    tensor_list = []
    for _ in size_list:
        tensor_list.append(torch.ByteTensor(size=(max_size,)).to("cuda"))
    if local_size != max_size:
        padding = torch.ByteTensor(size=(max_size - local_size,)).to("cuda")
        tensor = torch.cat((tensor, padding), dim=0)
    dist.all_gather(tensor_list, tensor)

    data_list = []
    for size, tensor in zip(size_list, tensor_list):
        buffer = tensor.cpu().numpy().tobytes()[:size]
        data_list.append(pickle.loads(buffer))

    return data_list 

def _pad_to_largest_tensor(tensor, group):
    """
    Padding all the tensors from different GPUs to the largest ones.
    Args:
        tensor (tensor): tensor to pad.
        group (group): pytorch dist group.
    Returns:
        list[int]: size of the tensor, on each rank
        Tensor: padded tensor that has the max size
    """
    world_size = dist.get_world_size(group=group)
    assert (
        world_size >= 1
    ), "comm.gather/all_gather must be called from ranks within the given group!"
    local_size = torch.tensor(
        [tensor.numel()], dtype=torch.int64, device=tensor.device
    )
    size_list = [
        torch.zeros([1], dtype=torch.int64, device=tensor.device)
        for _ in range(world_size)
    ]
    dist.all_gather(size_list, local_size, group=group)
    size_list = [int(size.item()) for size in size_list]

    max_size = max(size_list)

    # we pad the tensor because torch all_gather does not support
    # gathering tensors of different shapes
    if local_size != max_size:
        padding = torch.zeros(
            (max_size - local_size,), dtype=torch.uint8, device=tensor.device
        )
        tensor = torch.cat((tensor, padding), dim=0)
    return size_list, tensor 

def forward(ctx, input):
        input_list = [torch.zeros_like(input) for k in range(dist.get_world_size())]
        # Use allgather instead of allreduce since I don't trust in-place operations ..
        dist.all_gather(input_list, input, async_op=False)
        inputs = torch.stack(input_list, dim=0)
        return torch.sum(inputs, dim=0) 

def forward(ctx, input):
        world_size = distributed.get_world_size()
        # create a destination list for the allgather.  I'm assuming you're gathering from 3 workers.
        allgather_list = [torch.empty_like(input) for _ in range(world_size)]
        #if distributed.get_rank() == 0:
        #    import IPython;IPython.embed()
        distributed.all_gather(allgather_list, input)
        return torch.cat(allgather_list, dim=0) 

def all_gather(data, group=None):
    """
    Run all_gather on arbitrary picklable data (not necessarily tensors).

    Args:
        data: any picklable object
        group: a torch process group. By default, will use a group which
            contains all ranks on gloo backend.

    Returns:
        list[data]: list of data gathered from each rank
    """
    if get_world_size() == 1:
        return [data]
    if group is None:
        group = _get_global_gloo_group()
    if dist.get_world_size(group) == 1:
        return [data]

    tensor = _serialize_to_tensor(data, group)

    size_list, tensor = _pad_to_largest_tensor(tensor, group)
    max_size = max(size_list)

    # receiving Tensor from all ranks
    tensor_list = [
        torch.empty((max_size,), dtype=torch.uint8, device=tensor.device) for _ in size_list
    ]
    dist.all_gather(tensor_list, tensor, group=group)

    data_list = []
    for size, tensor in zip(size_list, tensor_list):
        buffer = tensor.cpu().numpy().tobytes()[:size]
        data_list.append(pickle.loads(buffer))

    return data_list 

def forward(ctx, input):
        input_list = [torch.zeros_like(input) for k in range(dist.get_world_size())]
        # Use allgather instead of allreduce since I don't trust in-place operations ..
        dist.all_gather(input_list, input, async_op=False)
        inputs = torch.stack(input_list, dim=0)
        return torch.sum(inputs, dim=0) 

def shared_random_seed():
    """
    Returns:
        int: a random number that is the same across all workers.
            If workers need a shared RNG, they can use this shared seed to
            create one.

    All workers must call this function, otherwise it will deadlock.
    """
    ints = np.random.randint(2 ** 31)
    all_ints = all_gather(ints)
    return all_ints[0] 

def all_gather(data, group=None):
    """
    Run all_gather on arbitrary picklable data (not necessarily tensors).

    Args:
        data: any picklable object
        group: a torch process group. By default, will use a group which
            contains all ranks on gloo backend.

    Returns:
        list[data]: list of data gathered from each rank
    """
    if get_world_size() == 1:
        return [data]
    if group is None:
        group = _get_global_gloo_group()
    if dist.get_world_size(group) == 1:
        return [data]

    tensor = _serialize_to_tensor(data, group)

    size_list, tensor = _pad_to_largest_tensor(tensor, group)
    max_size = max(size_list)

    # receiving Tensor from all ranks
    tensor_list = [
        torch.empty((max_size,), dtype=torch.uint8, device=tensor.device) for _ in size_list
    ]
    dist.all_gather(tensor_list, tensor, group=group)

    data_list = []
    for size, tensor in zip(size_list, tensor_list):
        buffer = tensor.cpu().numpy().tobytes()[:size]
        data_list.append(pickle.loads(buffer))

    return data_list 

def all_gather(data):
    """
    Run all_gather on arbitrary picklable data (not necessarily tensors)
    Args:
        data: any picklable object
    Returns:
        list[data]: list of data gathered from each rank
    """
    world_size = get_world_size()
    if world_size == 1:
        return [data]

    # serialized to a Tensor
    buffer = pickle.dumps(data)
    storage = torch.ByteStorage.from_buffer(buffer)
    tensor = torch.ByteTensor(storage).to("cuda")

    # obtain Tensor size of each rank
    local_size = torch.IntTensor([tensor.numel()]).to("cuda")
    size_list = [torch.IntTensor([0]).to("cuda") for _ in range(world_size)]
    dist.all_gather(size_list, local_size)
    size_list = [int(size.item()) for size in size_list]
    max_size = max(size_list)

    # receiving Tensor from all ranks
    # we pad the tensor because torch all_gather does not support
    # gathering tensors of different shapes
    tensor_list = []
    for _ in size_list:
        tensor_list.append(torch.ByteTensor(size=(max_size,)).to("cuda"))
    if local_size != max_size:
        padding = torch.ByteTensor(size=(max_size - local_size,)).to("cuda")
        tensor = torch.cat((tensor, padding), dim=0)
    dist.all_gather(tensor_list, tensor)

    data_list = []
    for size, tensor in zip(size_list, tensor_list):
        buffer = tensor.cpu().numpy().tobytes()[:size]
        data_list.append(pickle.loads(buffer))

    return data_list 

def all_gather(data):
    """
    Run all_gather on arbitrary picklable data (not necessarily tensors)
    Args:
        data: any picklable object
    Returns:
        list[data]: list of data gathered from each rank
    """
    world_size = get_world_size()
    if world_size == 1:
        return [data]

    # serialized to a Tensor
    buffer = pickle.dumps(data)
    storage = torch.ByteStorage.from_buffer(buffer)
    tensor = torch.ByteTensor(storage).to("cuda")

    # obtain Tensor size of each rank
    local_size = torch.IntTensor([tensor.numel()]).to("cuda")
    size_list = [torch.IntTensor([0]).to("cuda") for _ in range(world_size)]
    dist.all_gather(size_list, local_size)
    size_list = [int(size.item()) for size in size_list]
    max_size = max(size_list)

    # receiving Tensor from all ranks
    # we pad the tensor because torch all_gather does not support
    # gathering tensors of different shapes
    tensor_list = []
    for _ in size_list:
        tensor_list.append(torch.ByteTensor(size=(max_size,)).to("cuda"))
    if local_size != max_size:
        padding = torch.ByteTensor(size=(max_size - local_size,)).to("cuda")
        tensor = torch.cat((tensor, padding), dim=0)
    dist.all_gather(tensor_list, tensor)

    data_list = []
    for size, tensor in zip(size_list, tensor_list):
        buffer = tensor.cpu().numpy().tobytes()[:size]
        data_list.append(pickle.loads(buffer))

    return data_list 

def all_gather(data):
    """
    Run all_gather on arbitrary picklable data (not necessarily tensors)
    Args:
        data: any picklable object
    Returns:
        list[data]: list of data gathered from each rank
    """
    world_size = get_world_size()
    if world_size == 1:
        return [data]

    # serialized to a Tensor
    buffer = pickle.dumps(data)
    storage = torch.ByteStorage.from_buffer(buffer)
    tensor = torch.ByteTensor(storage).to("cuda")

    # obtain Tensor size of each rank
    local_size = torch.LongTensor([tensor.numel()]).to("cuda")
    size_list = [torch.LongTensor([0]).to("cuda") for _ in range(world_size)]
    dist.all_gather(size_list, local_size)
    size_list = [int(size.item()) for size in size_list]
    max_size = max(size_list)

    # receiving Tensor from all ranks
    # we pad the tensor because torch all_gather does not support
    # gathering tensors of different shapes
    tensor_list = []
    for _ in size_list:
        tensor_list.append(torch.ByteTensor(size=(max_size,)).to("cuda"))
    if local_size != max_size:
        padding = torch.ByteTensor(size=(max_size - local_size,)).to("cuda")
        tensor = torch.cat((tensor, padding), dim=0)
    dist.all_gather(tensor_list, tensor)

    data_list = []
    for size, tensor in zip(size_list, tensor_list):
        buffer = tensor.cpu().numpy().tobytes()[:size]
        data_list.append(pickle.loads(buffer))

    return data_list 

def collect_results_gpu(result_part, size):
    rank, world_size = get_dist_info()
    # dump result part to tensor with pickle
    part_tensor = torch.tensor(
        bytearray(pickle.dumps(result_part)), dtype=torch.uint8, device='cuda')
    # gather all result part tensor shape
    shape_tensor = torch.tensor(part_tensor.shape, device='cuda')
    shape_list = [shape_tensor.clone() for _ in range(world_size)]
    dist.all_gather(shape_list, shape_tensor)
    # padding result part tensor to max length
    shape_max = torch.tensor(shape_list).max()
    part_send = torch.zeros(shape_max, dtype=torch.uint8, device='cuda')
    part_send[:shape_tensor[0]] = part_tensor
    part_recv_list = [
        part_tensor.new_zeros(shape_max) for _ in range(world_size)
    ]
    # gather all result part
    dist.all_gather(part_recv_list, part_send)

    if rank == 0:
        part_list = []
        for recv, shape in zip(part_recv_list, shape_list):
            part_list.append(
                pickle.loads(recv[:shape[0]].cpu().numpy().tobytes()))
        # sort the results
        ordered_results = []
        for res in zip(*part_list):
            ordered_results.extend(list(res))
        # the dataloader may pad some samples
        ordered_results = ordered_results[:size]
        return ordered_results