Skip to content

fix: enable torch.autocast for TP parallelism without FSDP #2213

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
eous wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix: enable torch.autocast for TP parallelism without FSDP #2213

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
25 changes: 10 additions & 15 deletions torchtitan/distributed/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -250,24 +250,19 @@ def maybe_enable_amp(
parallel_dims: ParallelDims, mixed_precision_param: str, device_type: str
) -> contextlib.AbstractContextManager[None]:
if parallel_dims.fsdp_enabled:
# FSDP handles mixed precision internally
# FSDP handles mixed precision internally via MixedPrecisionPolicy
logger.info("Mixed precision training is handled by fully_shard")
return contextlib.nullcontext()
else:
if parallel_dims.tp_enabled or parallel_dims.pp_enabled:
logger.warning(
"Mixed precision training with TP or PP is only supported when FSDP/HSDP/CP is enabled."
)
logger.info("Mixed precision training is disabled")
return contextlib.nullcontext()
else:
# the following code will only be executed for DDP or single-device training
logger.info("Mixed precision training is handled by AMP")
# pyrefly: ignore [bad-return]
return torch.autocast(
device_type,
dtype=TORCH_DTYPE_MAP[mixed_precision_param],
)
# Enable autocast for non-FSDP cases (DDP, TP-only, single-device)
# Note: PP uses its own schedule-based execution and doesn't use this context
# torch.autocast works at the operator level and is orthogonal to parallelism strategy
logger.info("Mixed precision training is handled by AMP")
# pyrefly: ignore [bad-return]
return torch.autocast(
device_type,
dtype=TORCH_DTYPE_MAP[mixed_precision_param],
)


def init_fake_mode(world_size: int, comm_mode: str = "fake_backend"):
Expand Down
Loading