Skip to content
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

fix 2D parallel crash caused by all-reduce on 2D world_mesh #105

Merged
merged 2 commits into from
Mar 2, 2024
Merged

fix 2D parallel crash caused by all-reduce on 2D world_mesh #105

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
497fded
fix 2D parallel crash caused by all-reduce on 2D world_mesh
tianyu-l Mar 2, 2024
b05fad1
Update on "fix 2D parallel crash caused by all-reduce on 2D world_mesh"
tianyu-l Mar 2, 2024
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
2 changes: 1 addition & 1 deletion torchtrain/parallelisms/parallelize_llama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -129,7 +129,7 @@ def parallelize_llama(model, world_mesh, parallel_dims, job_config: JobConfig):
if parallel_dims.sp_enabled:
# First we apply Sequence Parallelism if it's enabled
tp_mesh = world_mesh["sp"] if world_mesh.ndim > 1 else world_mesh
sp_degree = job_config.training.sequence_parallelism_degree
sp_degree = job_config.training.sequence_parallel_degree
# First:
# 1. parallelize the first embedding and the last linear proj layer
# 2. shard the first layer of transformer block
Expand Down
10 changes: 5 additions & 5 deletions train.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,9 +102,9 @@ def main(job_config: JobConfig):

# build dataloader
# need dp world size and rank
# TODO: dp might not always be 0 so we need to handle that more carefully
dp_degree = world_mesh.size(0)
dp_rank = world_mesh.get_local_rank(0)
dp_mesh = world_mesh["dp"]
dp_degree = dp_mesh.size()
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

There is also a config called data_parallel_degree. Should we use that ?

dp_rank = dp_mesh.get_local_rank()
build_dataloader_fn = dataloader_fn[job_config.training.dataset]
data_loader = build_dataloader_fn(
tokenizer,
Expand Down Expand Up @@ -253,8 +253,8 @@ def main(job_config: JobConfig):
np.max(losses_since_last_log),
)
global_avg_loss, global_max_loss = (
dist_mean(avg_loss, world_mesh),
dist_max(max_loss, world_mesh),
dist_mean(avg_loss, dp_mesh),
dist_max(max_loss, dp_mesh),
)

time_delta = timer() - time_last_log
Expand Down
Loading