Skip to content

Conversation

@bghira
Copy link
Owner

@bghira bghira commented Feb 1, 2026

The Embedding/RMSNorm and others may not be needed on medium-to-large GPUs, and can cause problems.

The sync hooks are added to ensure deterministic operations for RamTorch since it has drift between executions. Disabling them can speed it up a little bit, at the risk of having random differences during training.

note: ramtorch uses stochastic rounding internally anyway during training, accuracy isn't guaranteed.

@bghira bghira merged commit bbc8459 into main Feb 1, 2026
4 checks passed
@bghira bghira deleted the feat/ramtorch-apply-option branch February 1, 2026 03:04
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants