Adding CUDA graph capture feature for supervised learning. #98
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
Signed-off-by: Josh Romero <joshr@nvidia.com>
Signed-off-by: Josh Romero <joshr@nvidia.com>
Signed-off-by: Josh Romero <joshr@nvidia.com>
Signed-off-by: Josh Romero <joshr@nvidia.com>
romerojosh
force-pushed
the
cuda_graphs_2
branch
from
434417a to
76b058e
Compare
Signed-off-by: Josh Romero <joshr@nvidia.com>
Collaborator
Author
|
/build_and_test |
|
🚀 Build workflow triggered! View run |
|
✅ Build workflow passed! View run |
Signed-off-by: Josh Romero <joshr@nvidia.com>
azrael417
reviewed
Dec 5, 2025
Collaborator
azrael417
left a comment
azrael417
left a comment
There was a problem hiding this comment.
looks good, thanks a lot. much cleaner but I have a few comments still.
|
|
||
| private: | ||
| // Input signature for validating consistent inputs | ||
| struct InputSignature { |
Collaborator
There was a problem hiding this comment.
This structure is the same for all the states, wouldn't it be better to move it outside?
| namespace torchfort { | ||
|
|
||
| // Action to take for current iteration | ||
| enum class GraphAction { |
Collaborator
There was a problem hiding this comment.
do we need this outside the ENABLE_GPU context?
| void launch(cudaStream_t stream) { CHECK_CUDA(cudaGraphLaunch(graph_exec_.get(), stream)); } | ||
|
|
||
| // Get static loss (valid after CAPTURE or REPLAY) | ||
| const torch::Tensor& get_loss() const { return static_loss_; } |
Collaborator
There was a problem hiding this comment.
shall we add asserts to return error when not fully initialized?
| } | ||
|
|
||
| // Extract loss value | ||
| *loss_val = loss.item<float>(); |
Collaborator
There was a problem hiding this comment.
does this work? .item copies the loss back to the CPU, and then all reduce needs a tensor, right? We can just clone the loss tensor, call all reduce on it and then extract the scalar with .item?
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 adds CUDA graphs support for supervised learning problems. The feature is enabled via a new general configuration entry:
enable_cuda_graphs, see updated documentation.Since we are targeting high-performance use cases, this functionality is made to be fairly minimal in terms of features. In particular, we do not maintain internal static entry points to the captured graphs, allow graph recapture for dynamic shapes, etc. Instead, we expect users to provide consistent input data (memory locations, shapes) to be compatible with the CUDA graphs operating model.
Marking this as a draft for now as I still need to implement some tests.