send_to_device crashes on defaultdict batches

[aswanth-07]

System Info

• Accelerate version: 1.14.0; also reproduced on 1.15.0.dev0 at 16cb6eb8
• Platform: Windows-10-10.0.26200-SP0
• Python version: 3.11.9
• NumPy version: 2.4.6
• PyTorch version: 2.13.0+cpu
• PyTorch accelerator: N/A
• Accelerate default config: Not found

Information

• The official example scripts
• My own modified scripts

Tasks

• One of the official example tasks
• My own task or dataset

Reproduction

Accelerator.prepare() fails during iteration when a PyTorch DataLoader collator returns a standard-library collections.defaultdict:

from collections import defaultdict

import torch
from torch.utils.data import DataLoader

from accelerate import Accelerator


def collate(examples):
    return defaultdict(list, input_ids=torch.tensor(examples))


accelerator = Accelerator(cpu=True)
dataloader = accelerator.prepare(
    DataLoader([1, 2], batch_size=2, collate_fn=collate)
)
print(next(iter(dataloader)))

Actual result:

TypeError: first argument must be callable or None

This also reproduces directly with the public utility:

from accelerate.utils import send_to_device

send_to_device(
    defaultdict(list, input_ids=torch.tensor([1, 2])),
    "cpu",
)

send_to_device reconstructs every Mapping as type(tensor)(new_mapping). For defaultdict, the first positional constructor argument is the default factory rather than initial data, which causes the exception.

Relevant source:

accelerate/src/accelerate/utils/operations.py

Lines 181 to 190 in 16cb6eb

	skip_keys = [skip_keys]
	elif skip_keys is None:
	skip_keys = []
	return type(tensor)(
	{
	k: t if k in skip_keys else send_to_device(t, device, non_blocking=non_blocking, skip_keys=skip_keys)
	for k, t in tensor.items()
	}
	)
	else:

I have a focused fix and regression test prepared.

Expected behavior

Prepared dataloaders should preserve the mapping type, tensor values, and default_factory, and yield the batch without error. For the reproduction above, the expected batch is a defaultdict containing tensor([1, 2]), with missing keys still producing [].

AI disclosure: I used Codex to help identify and independently reproduce this behavior on the current release and main, search for duplicates, test a focused fix, and draft this report. I reviewed the reproduction, patch, and tests and will personally handle follow-up.

[aswanth-07]

Closing this with #4155 after the same scope review. I do not have enough user demand to justify a shared utility special case.