Skip to content

Support tuple of tensors in estimate_strategy_runtime_cost #102

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

Merged
merged 2 commits into from
Aug 21, 2025
Merged

Support tuple of tensors in estimate_strategy_runtime_cost #102

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6ca8611
Support tuple of tensors in estimate_strategy_runtime_cost
fmassa Aug 21, 2025
0a4ba18
Fix bad copy-paste
fmassa Aug 21, 2025
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
18 changes: 8 additions & 10 deletions autoparallel/compute_estimation.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -213,25 +213,23 @@ def estimate_strategy_runtime_cost(node, strategy):
args = tree_map_only(torch.fx.Node, lambda x: x.meta["val"], node.args)
kwargs = tree_map_only(torch.fx.Node, lambda x: x.meta["val"], node.kwargs)

fake_mode = torch._guards.detect_fake_mode(args)
Copy link
Contributor

@wconstab wconstab Aug 21, 2025

Choose a reason for hiding this comment

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

is this bc we are already under a fake mode now, but we weren't in the initial autop?

Copy link
Contributor Author

@fmassa fmassa Aug 21, 2025

Choose a reason for hiding this comment

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

Yes, that's right, now the whole AutoParallel is running under fake mode, so we can remove it

Copy link
Member

@xmfan xmfan Aug 21, 2025

Choose a reason for hiding this comment

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

this is removed because we're already running this in a fake mode?


if len(kwargs) > 0:
for k, v in kwargs.items():
assert not isinstance(v, torch.Tensor), f"{node} {v}"
args_sizes_strides = tuple(
_get_sharded_shape_stride(spec) for spec in strategy.input_specs
)

flat_args, treespec = tree_flatten(args)
Copy link
Contributor

@wconstab wconstab Aug 21, 2025

Choose a reason for hiding this comment

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

shouldn't we just call tree_map_only(Tensor, torch.empty) here instead of doing the for loop?

Copy link
Contributor Author

@fmassa fmassa Aug 21, 2025

Choose a reason for hiding this comment

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

We need to get the size from the args_sizes_strides (which comes from the spec), so I think we might need this indirection.

But if there is a cleaner way of doing this I'm happy to change the code!

new_flat_args = []
counter = 0
args = list(args)
for i, arg in enumerate(args):
if isinstance(arg, torch.Tensor):
with fake_mode:
sizes, strides = args_sizes_strides[counter]
args[i] = torch.empty_strided(
sizes, strides, device=arg.device, dtype=arg.dtype
)
for x in flat_args:
if isinstance(x, torch.Tensor):
sizes, strides = args_sizes_strides[counter]
x = torch.empty_strided(sizes, strides, device=x.device, dtype=x.dtype)
counter += 1
new_flat_args.append(x)
args = treespec.unflatten(new_flat_args)

# TODO: maybe cache the flop_counter to avoid recreating it
# all the time
Expand Down