Lightning-AI / Lightning-AI/lightning-thunder
TE: Redundant backward computation in PEFT setting.
Open
@kshitij12345 is already working on this.
Since May 19, 2025.
TransformerEngine
- Dominant language
- Python
- Stars
- 1.5k
- Forks
- 121
- PR merge metrics
- No merged PRs in 30d
Description
Problem - Currently, TE executor just assumes that we will be computing gradients for both input and weight. This is true for full finetuning, but in PEFT scenario the weights are frozen i.e. requires_grad=False. This leads to poor performance as the TE executor ends up doing unnecessary work of computing gradients for weights.
NOTE: We tried fixing in #1838 but it had to be reverted due to #1966
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Assessment
This issue has not been assessed yet.