Skip to content

Fix double sequence partition during training with context-parallel #3498

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

Merged
NanoCode012 merged 1 commit into from
Mar 20, 2026
Merged

Fix double sequence partition during training with context-parallel #3498

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
23 changes: 7 additions & 16 deletions src/axolotl/monkeypatch/accelerate/parallelism_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -78,30 +78,21 @@ def patch_parallelism_config():


def patch_prepare_cp():
import functools
import contextlib

import torch
from accelerate import Accelerator

def patched_prepare_cp(self, *args):
if self.parallelism_config.cp_backend == "deepspeed":
return args

from accelerate.big_modeling import _attach_context_parallel_hooks
from torch.distributed.tensor.experimental import context_parallel
from torch.distributed.tensor.experimental._attention import set_rotate_method

cp_comm_strategy = self.parallelism_config.cp_handler.cp_comm_strategy
set_rotate_method(cp_comm_strategy)

self._cp_context = functools.partial(
context_parallel, mesh=self.torch_device_mesh["cp"]
)

for arg in args:
if isinstance(arg, torch.nn.Module):
_attach_context_parallel_hooks(arg)
@contextlib.contextmanager
def _noop_cp_context(
buffers=None, buffer_seq_dims=None, no_restore_buffers=None
):
yield

self._cp_context = _noop_cp_context
return args

Accelerator._prepare_cp = patched_prepare_cp
Loading