Swap in PyTorch kernels for MHA#13
Closed
wadkisson wants to merge 3 commits into
Closed
Conversation
Member
|
Merged into main as part of the consolidated backend work. Thanks, Will! Closing this PR. |
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
In the current cuda training process, the multi headed attention mechanism is the biggest bottleneck to training transformer models. Currently, training a ~500M parameter GPT model is ~74x as slow as PyTorch.
In this PR, I swap the actual computation of the MHA Q,K,V scores to be done entirely with PyTorch's kernels, Libtorch. This implantation more than doubles the speed of training large GPT models in TorchLean — the same ~500M parameter model ran ended up being ~31x as slow as PyTorch after this fix.
The diff is quite scoped and does three things: Adds Libtorch as a requirement for training TorchLean with cuda, and wires in libtorch for computation of the forward attention mechanism as well as the backward attention mechanism.
Let me know if any changes are needed. Thanks!