Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Fused Rotary Embeddings (fixed) #1108

Merged
merged 29 commits into from
Jan 5, 2024
Merged

Conversation

yang
Copy link
Contributor

@yang yang commented Jan 2, 2024

Orig PR started by Stella:

This is the start of an implementation of fused rotary embeddings for GPT-NeoX #1077. I assumed that it would be pretty straight forward despite my lack of familiarity with CUDA but I've been disabused of that notion.

Currently the issue is that there are errors when trying to compile the cuda extension.

StellaAthena and others added 27 commits November 15, 2023 11:38
Ports the fix from NVIDIA/apex#1750 into this branch.
Just checked and this should work for bf16. Or, at least, the reason I originally thought it wouldn't doesn't apply.
Add `self.rope_fusion = neox_args.rope_fusion` so that `ParallelSelfAttention` knows if we're using rope fusion.
Just needed to bring in the latest headers/sources,
and call into it the right way from transformers.py.
@yang yang requested a review from Quentin-Anthony as a code owner January 2, 2024 19:09
@StellaAthena StellaAthena linked an issue Jan 2, 2024 that may be closed by this pull request
@StellaAthena
Copy link
Member

StellaAthena commented Jan 2, 2024

Supersedes #1103

@Quentin-Anthony Quentin-Anthony merged commit 77605ca into EleutherAI:main Jan 5, 2024
3 of 5 checks passed
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.

Apply new fused rotary embedding
3 participants