Skip to content

Fix semi-sync training with 1GPU per FT replica #1505

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

Fix semi-sync training with 1GPU per FT replica #1505

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: 25 additions & 0 deletions torchtitan/distributed/parallel_dims.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -105,6 +105,13 @@ def _build_mesh_with_ep(self) -> DeviceMesh:
names.append(name)

logger.info(f"Building {len(dims)}-D device mesh with {names}, {dims}")

# Handle the case where all parallelism dimensions are 1
# We still need to create a mesh with named dimensions for submesh access
if not dims:
dims = [1]
names = ["world"]

mesh = init_device_mesh(device_type, dims, mesh_dim_names=names)

# Create all the submesh here to ensure all required process groups are
Expand Down Expand Up @@ -137,6 +144,11 @@ def _build_mesh_with_ep(self) -> DeviceMesh:
if self.etp == 1 and self.tp_enabled:
ep_mesh_dim_names.append("tp")

# Ensure dp_cp submesh exists even when all parallelism dimensions are 1
# This is required for fault tolerance functionality
if not dp_cp_mesh_dim_names and "world" in names:
dp_cp_mesh_dim_names = ["world"]

mesh[tuple(dp_mesh_dim_names)]._flatten(mesh_dim_name="dp")
mesh[tuple(dp_shard_cp_mesh_dim_names)]._flatten(mesh_dim_name="dp_shard_cp")
mesh[tuple(dp_cp_mesh_dim_names)]._flatten(mesh_dim_name="dp_cp")
Expand All @@ -156,6 +168,13 @@ def _build_mesh_without_ep(self) -> DeviceMesh:
names.append(name)

logger.info(f"Building {len(dims)}-D device mesh with {names}, {dims}")

# Handle the case where all parallelism dimensions are 1
# We still need to create a mesh with named dimensions for submesh access
if not dims:
dims = [1]
names = ["world"]

mesh = init_device_mesh(device_type, dims, mesh_dim_names=names)

# Create all the submesh here to ensure all required process groups are
Expand All @@ -178,6 +197,12 @@ def _build_mesh_without_ep(self) -> DeviceMesh:
dp_shard_cp_mesh_dim_names.append("cp")
dp_cp_mesh_dim_names.append("cp")

# Ensure dp_cp submesh exists even when all parallelism dimensions are 1
# This is required for fault tolerance functionality
if not dp_cp_mesh_dim_names and "world" in names:
dp_cp_mesh_dim_names = ["world"]


if dp_mesh_dim_names != []:
mesh[tuple(dp_mesh_dim_names)]._flatten(mesh_dim_name="dp")
if dp_shard_cp_mesh_dim_names != []:
Expand Down