fix(wip): use MPS when available except for sparse computations #5
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
This allows the code to run (albeit slowly) on MacOS without a CUDA device. I also experimented with using MPS, and got further than indicated here. As noted in safety-research#1, there are issues with the SparseMPS backend, which is mostly not implemented. Sparse tensors are used for e.g. transcoder activations. I think it might not be terrible to do a workaround where MPS devices are used except for the sparse computations, as unlike with CUDA, the cost of shuttling tensors "between" devices on Apple Silicon is minimal (it's all unified memory). However I haven't actually got this working to be able to test it. In the meantime, the entire codebase now has consistent device selection defaults and the notebooks will run on a Mac.
Selecting MPS as a backend when CUDA is unavailable does not initially work properly, as the code uses sparse tensors and the SparseMPS backend is not implemented. This commit changes the code so that when MPS is in use, sparse tensors are created on the CPU instead, and then moved back to the same device as the rest of the computation. Since Apple Silicon has unified memory, this "movement" should be an accounting change only and a no-op in terms of actual data movement. I haven't fully tested this yet, but it provides a significant performance improvement on an M3 MBA, taking the time to execute all cells in the circuit_tracing_tutorial.ipynb from ~19min to <1min.
Contributor
|
Thank you for the contribution! Eager to get this working on MPS. I can't test this myself at the moment. Could you confirm that:
I'd prefer to handle this as a single change, so closing #4 in favor of this. |
|
Running results in a SIGTRAP for me: full stacktrace |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
Selecting MPS as a backend when CUDA is unavailable does not initially work properly, as the code uses sparse tensors and the SparseMPS backend is not implemented. This commit changes the code so that when MPS is in use, sparse tensors are created on the CPU instead, and then moved back to the same device as the rest of the computation. Since Apple Silicon has unified memory, this "movement" should be an accounting change only and a no-op in terms of actual data movement.
I haven't fully tested this yet, but it provides a significant performance improvement on an M3 MBA, taking the time to execute all cells in the circuit_tracing_tutorial.ipynb from ~19min to <1min.
Note: this stacks on top of #4, cf #1