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

re-enable interleaved 1f1b test #1079

Open
wants to merge 1 commit into
base: gh/H-Huang/3/base
Choose a base branch
from
Open

re-enable interleaved 1f1b test #1079

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
24 changes: 8 additions & 16 deletions test/test_pipeline_schedule.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -299,9 +299,6 @@ def test_1f1b(self):

@skip_if_lt_x_gpu(4)
def test_interleaved_1f1b(self):
# TODO: not working
return

device = torch.device(f"cuda:{self.rank}")
dist.init_process_group(
init_method=self.init_method,
Expand All @@ -320,38 +317,33 @@ def test_interleaved_1f1b(self):
microbatches = [
(torch.randn_like(microbatch),) for _ in range(num_microbatches)
]
target_mbs = [
torch.randn_like(microbatch) for _ in range(num_microbatches)
]

loss_fn = torch.nn.MSELoss()
schedule = ScheduleInterleaved1F1B(
stages,
num_microbatches,
loss_fn=loss_fn,
)
schedule.step_microbatches(microbatches)
schedule.step_microbatches(microbatches, target_mbs=target_mbs)
Copy link
Contributor

Choose a reason for hiding this comment

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

Should we feed target to every rank?

Copy link
Member Author

Choose a reason for hiding this comment

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

we don't have to, but if we do that is also allright too since if it is not the last stage then it wont calculate the loss using the target_mbs.

Copy link
Contributor

Choose a reason for hiding this comment

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

I'd rather not rely on assumptions :)
We should test the real case (strictly speaking even microbatches should not be fed to every rank).


# num local pipeline stages == world_size
num_microbatches = 8
stages = self._create_virtual_pipeline_stages(
model,
microbatch,
device,
self.world_size,
num_microbatches=num_microbatches,
)
microbatches = [
torch.randn_like(microbatch) for _ in range(num_microbatches)
]

schedule = ScheduleInterleaved1F1B(
stages,
num_microbatches,
loss_fn=loss_fn,
)
schedule.step_microbatches(microbatches)

# differing microbatch size
num_microbatches = 64
microbatches = [
torch.randn_like(microbatch) for _ in range(num_microbatches)
]
schedule.step_microbatches(microbatches)
schedule.step_microbatches(microbatches, target_mbs=target_mbs)
Copy link
Contributor

Choose a reason for hiding this comment

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

Same here.


def test_interleaved_1f1b_negative(self):
device = torch.device("cpu")
Expand Down
Loading