Skip to content

fix: add PyTorch version check for torch.accelerator in gather_size_by_comm #13093

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Mr-Neutr0n wants to merge 2 commits into
base: main
Choose a base branch
from
Open

fix: add PyTorch version check for torch.accelerator in gather_size_by_comm #13093

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d3493ae
fix: add PyTorch version check for torch.accelerator in gather_size_b…
Mr-Neutr0n Feb 7, 2026
3a12e93
Simplify gather_device selection per @DN6 suggestion
Mr-Neutr0n Jun 1, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 5 additions & 1 deletion src/diffusers/models/_modeling_parallel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,7 @@
import torch.distributed as dist

from ..utils import get_logger
from ..utils.torch_utils import get_device


if TYPE_CHECKING:
Expand Down Expand Up @@ -290,7 +291,10 @@ def gather_size_by_comm(size: int, group: dist.ProcessGroup) -> List[int]:
# HACK: Use Gloo backend for all_gather to avoid H2D and D2H overhead
comm_backends = str(dist.get_backend(group=group))
# NOTE: e.g., dist.init_process_group(backend="cpu:gloo,cuda:nccl")
gather_device = "cpu" if "cpu" in comm_backends else torch.accelerator.current_accelerator()
# get_device() handles accelerator version compatibility internally
# (cuda/npu/xpu/mps/mlu/cpu), so we don't need the hasattr(torch, "accelerator")
# check here.
gather_device = "cpu" if "cpu" in comm_backends else get_device()
gathered_sizes = [torch.empty((1,), device=gather_device, dtype=torch.int64) for _ in range(world_size)]
dist.all_gather(
gathered_sizes,
Expand Down
Loading