enable offload #612
Open
enable offload #612
+443
−0
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
dhpitt
reviewed
Jun 17, 2025
There was a problem hiding this comment.
This is really cool @wdlctc , thank you for opening! Since the helper function totally reimplements SpectralConv.forward and assumes full access to the method, I think it'll be much more maintainable if it directly interfaces with the forward calls themselves instead of living as helper functions in neuralop.training.
| enable_activation_offload_for_SpectralConv(conv) | ||
|
|
||
|
|
||
| def enable_activation_offload_for_SpectralConv(SpectralConv): |
There was a problem hiding this comment.
I wonder if this might be better as a module in neuralop.layers, or as a param to the original SpectralConv
| def forward( | ||
| self, x: torch.Tensor, output_shape: Optional[Tuple[int]] = None | ||
| ): | ||
| with torch.autograd.graph.save_on_cpu(pin_memory=True): |
There was a problem hiding this comment.
This looks like the only thing that needs to be changed, right?
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR introduces activation offloading support for Fourier Neural Operator (FNO) training in neuralop, targeting reduced GPU memory consumption during forward and backward passes.
Defines enable_activation_offload_for_FNO and supporting functions.
Wraps key forward passes (FNO, FNOBlocks, SpectralConv) with torch.autograd.graph.save_on_cpu(pin_memory=True) for CPU offloading of saved activations during training.
New training script demonstrating end-to-end training with activation offloading.
Compatible with distributed training, WandB logging, and multiresolution datasets (e.g., Darcy Flow).
Preserves model configuration and training loop from standard train.py but optionally enables memory-efficient execution.