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

Jump to bottom

Fix param input order for cudagraph #1138

Open
wants to merge 4 commits into
base: main
Choose a base branch
from
Open

Fix param input order for cudagraph #1138

Changes from 1 commit
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
4235bfe
Fix param input order for cudagraph
yifeis-nv Aug 28, 2024
f876a23
Addtional Support for Muti-Chunks
yifeis-nv Sep 6, 2024
2ba9465
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Sep 6, 2024
2a4a478
Merge branch 'main' into yifeis/cudagraph
timmoon10 Oct 3, 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
4 changes: 2 additions & 2 deletions transformer_engine/pytorch/graph.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -171,8 +171,8 @@ def _make_graphed_callables(
]
else:
per_callable_module_params = []
for c in callables:
for i in range(num_microbatches):
for i in range(num_microbatches):
Copy link
Collaborator

@vasunvidia vasunvidia Sep 5, 2024 •
edited
Loading

Choose a reason for hiding this comment

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

The change doesn't appear to fully solve the bug. For example this fix will work only when the number of model chunks (num_model_chunks) is 1. The correct solution will be

for m_chunk in range(num_model_chunks):
    for idx in range(num_microbatches):
        for l_no in range(num_layers):
            per_callable_module_params.append(
                    tuple(callables[m_chunk*num_layers + l_no].parameters()) if isinstance(c, torch.nn.Module) else ()
                )

Can you test if this fix works?

Copy link
Author

Choose a reason for hiding this comment

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

Thanks for your input! This works for my situation.

for c in callables:
per_callable_module_params.append(
tuple(c.parameters()) if isinstance(c, torch.nn.Module) else ()
)
Expand Down