Skip to content

Fix losses and aggregators when using CUDA graphs #280

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
ktangsali merged 10 commits into from
Feb 19, 2026
Merged

Fix losses and aggregators when using CUDA graphs #280

Show file tree
Hide file tree
Changes from 5 commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
7f8324c
Pass step as scalar tensor for CUDA graph compatability
jasooney23 Feb 7, 2026
f6e8d25
Update turbulent channel example aggregator for CUDA graphs
jasooney23 Feb 7, 2026
8cbf819
Merge branch 'NVIDIA:main' into main
jasooney23 Feb 7, 2026
5bfb29c
Update aggregator tests
jasooney23 Feb 7, 2026
3a4accd
Merge branch 'main' of github.com:jasooney23/physicsnemo-sym
jasooney23 Feb 7, 2026
eaa6592
Merge branch 'main' into main
ktangsali Feb 19, 2026
aeed9a0
Fix formatting issues
ktangsali Feb 19, 2026
9466760
Fix formatting issues
ktangsali Feb 19, 2026
f938992
Fix formatting issues
ktangsali Feb 19, 2026
5c9efb8
Fix formatting issues
ktangsali Feb 19, 2026
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
7 changes: 3 additions & 4 deletions examples/turbulent_channel/2d/custom_aggregator.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,15 +30,15 @@ class CustomSum(Aggregator):
def __init__(self, params, num_losses, weights=None):
super().__init__(params, num_losses, weights)

def forward(self, losses: Dict[str, torch.Tensor], step: int) -> torch.Tensor:
def forward(self, losses: Dict[str, torch.Tensor], step: torch.Tensor) -> torch.Tensor:
"""
Aggregates the losses by summation

Parameters
----------
losses : Dict[str, torch.Tensor]
A dictionary of losses
step : int
step : torch.Tensor
Optimizer step

Returns
Expand All @@ -54,8 +54,7 @@ def forward(self, losses: Dict[str, torch.Tensor], step: int) -> torch.Tensor:
loss: torch.Tensor = torch.zeros_like(self.init_loss)

smoothness = 0.0005 # use 0.0005 to smoothen the transition over ~10k steps
step_tensor = torch.tensor(step, dtype=torch.float32)
decay_weight = (torch.tanh((20000 - step_tensor) * smoothness) + 1.0) * 0.5
decay_weight = (torch.tanh((20000 - step.float()) * smoothness) + 1.0) * 0.5

# Add losses
for key in losses.keys():
Expand Down
7 changes: 3 additions & 4 deletions examples/turbulent_channel/2d_std_wf/custom_aggregator.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,15 +30,15 @@ class CustomSum(Aggregator):
def __init__(self, params, num_losses, weights=None):
super().__init__(params, num_losses, weights)

def forward(self, losses: Dict[str, torch.Tensor], step: int) -> torch.Tensor:
def forward(self, losses: Dict[str, torch.Tensor], step: torch.Tensor) -> torch.Tensor:
"""
Aggregates the losses by summation

Parameters
----------
losses : Dict[str, torch.Tensor]
A dictionary of losses
step : int
step : torch.Tensor
Optimizer step

Returns
Expand All @@ -54,8 +54,7 @@ def forward(self, losses: Dict[str, torch.Tensor], step: int) -> torch.Tensor:
loss: torch.Tensor = torch.zeros_like(self.init_loss)

smoothness = 0.0005 # use 0.0005 to smoothen the transition over ~10k steps
step_tensor = torch.tensor(step, dtype=torch.float32)
decay_weight = (torch.tanh((20000 - step_tensor) * smoothness) + 1.0) * 0.5
decay_weight = (torch.tanh((20000 - step.float()) * smoothness) + 1.0) * 0.5

# Add losses
for key in losses.keys():
Expand Down
Loading